Fri, 16 Mar 2012 16:14:04 +0100
7154517: Build error in hotspot-gc without precompiled headers
Reviewed-by: jcoomes, brutisso
1 /*
2 * Copyright (c) 2005, 2012, 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 */
25 #include "precompiled.hpp"
26 #include "classfile/symbolTable.hpp"
27 #include "classfile/systemDictionary.hpp"
28 #include "code/codeCache.hpp"
29 #include "gc_implementation/parallelScavenge/gcTaskManager.hpp"
30 #include "gc_implementation/parallelScavenge/generationSizer.hpp"
31 #include "gc_implementation/parallelScavenge/parallelScavengeHeap.inline.hpp"
32 #include "gc_implementation/parallelScavenge/pcTasks.hpp"
33 #include "gc_implementation/parallelScavenge/psAdaptiveSizePolicy.hpp"
34 #include "gc_implementation/parallelScavenge/psCompactionManager.inline.hpp"
35 #include "gc_implementation/parallelScavenge/psMarkSweep.hpp"
36 #include "gc_implementation/parallelScavenge/psMarkSweepDecorator.hpp"
37 #include "gc_implementation/parallelScavenge/psOldGen.hpp"
38 #include "gc_implementation/parallelScavenge/psParallelCompact.hpp"
39 #include "gc_implementation/parallelScavenge/psPermGen.hpp"
40 #include "gc_implementation/parallelScavenge/psPromotionManager.inline.hpp"
41 #include "gc_implementation/parallelScavenge/psScavenge.hpp"
42 #include "gc_implementation/parallelScavenge/psYoungGen.hpp"
43 #include "gc_implementation/shared/isGCActiveMark.hpp"
44 #include "gc_interface/gcCause.hpp"
45 #include "memory/gcLocker.inline.hpp"
46 #include "memory/referencePolicy.hpp"
47 #include "memory/referenceProcessor.hpp"
48 #include "oops/methodDataOop.hpp"
49 #include "oops/oop.inline.hpp"
50 #include "oops/oop.pcgc.inline.hpp"
51 #include "runtime/fprofiler.hpp"
52 #include "runtime/safepoint.hpp"
53 #include "runtime/vmThread.hpp"
54 #include "services/management.hpp"
55 #include "services/memoryService.hpp"
56 #include "utilities/events.hpp"
57 #include "utilities/stack.inline.hpp"
59 #include <math.h>
61 // All sizes are in HeapWords.
62 const size_t ParallelCompactData::Log2RegionSize = 9; // 512 words
63 const size_t ParallelCompactData::RegionSize = (size_t)1 << Log2RegionSize;
64 const size_t ParallelCompactData::RegionSizeBytes =
65 RegionSize << LogHeapWordSize;
66 const size_t ParallelCompactData::RegionSizeOffsetMask = RegionSize - 1;
67 const size_t ParallelCompactData::RegionAddrOffsetMask = RegionSizeBytes - 1;
68 const size_t ParallelCompactData::RegionAddrMask = ~RegionAddrOffsetMask;
70 const ParallelCompactData::RegionData::region_sz_t
71 ParallelCompactData::RegionData::dc_shift = 27;
73 const ParallelCompactData::RegionData::region_sz_t
74 ParallelCompactData::RegionData::dc_mask = ~0U << dc_shift;
76 const ParallelCompactData::RegionData::region_sz_t
77 ParallelCompactData::RegionData::dc_one = 0x1U << dc_shift;
79 const ParallelCompactData::RegionData::region_sz_t
80 ParallelCompactData::RegionData::los_mask = ~dc_mask;
82 const ParallelCompactData::RegionData::region_sz_t
83 ParallelCompactData::RegionData::dc_claimed = 0x8U << dc_shift;
85 const ParallelCompactData::RegionData::region_sz_t
86 ParallelCompactData::RegionData::dc_completed = 0xcU << dc_shift;
88 SpaceInfo PSParallelCompact::_space_info[PSParallelCompact::last_space_id];
89 bool PSParallelCompact::_print_phases = false;
91 ReferenceProcessor* PSParallelCompact::_ref_processor = NULL;
92 klassOop PSParallelCompact::_updated_int_array_klass_obj = NULL;
94 double PSParallelCompact::_dwl_mean;
95 double PSParallelCompact::_dwl_std_dev;
96 double PSParallelCompact::_dwl_first_term;
97 double PSParallelCompact::_dwl_adjustment;
98 #ifdef ASSERT
99 bool PSParallelCompact::_dwl_initialized = false;
100 #endif // #ifdef ASSERT
102 #ifdef VALIDATE_MARK_SWEEP
103 GrowableArray<void*>* PSParallelCompact::_root_refs_stack = NULL;
104 GrowableArray<oop> * PSParallelCompact::_live_oops = NULL;
105 GrowableArray<oop> * PSParallelCompact::_live_oops_moved_to = NULL;
106 GrowableArray<size_t>* PSParallelCompact::_live_oops_size = NULL;
107 size_t PSParallelCompact::_live_oops_index = 0;
108 size_t PSParallelCompact::_live_oops_index_at_perm = 0;
109 GrowableArray<void*>* PSParallelCompact::_other_refs_stack = NULL;
110 GrowableArray<void*>* PSParallelCompact::_adjusted_pointers = NULL;
111 bool PSParallelCompact::_pointer_tracking = false;
112 bool PSParallelCompact::_root_tracking = true;
114 GrowableArray<HeapWord*>* PSParallelCompact::_cur_gc_live_oops = NULL;
115 GrowableArray<HeapWord*>* PSParallelCompact::_cur_gc_live_oops_moved_to = NULL;
116 GrowableArray<size_t> * PSParallelCompact::_cur_gc_live_oops_size = NULL;
117 GrowableArray<HeapWord*>* PSParallelCompact::_last_gc_live_oops = NULL;
118 GrowableArray<HeapWord*>* PSParallelCompact::_last_gc_live_oops_moved_to = NULL;
119 GrowableArray<size_t> * PSParallelCompact::_last_gc_live_oops_size = NULL;
120 #endif
122 void SplitInfo::record(size_t src_region_idx, size_t partial_obj_size,
123 HeapWord* destination)
124 {
125 assert(src_region_idx != 0, "invalid src_region_idx");
126 assert(partial_obj_size != 0, "invalid partial_obj_size argument");
127 assert(destination != NULL, "invalid destination argument");
129 _src_region_idx = src_region_idx;
130 _partial_obj_size = partial_obj_size;
131 _destination = destination;
133 // These fields may not be updated below, so make sure they're clear.
134 assert(_dest_region_addr == NULL, "should have been cleared");
135 assert(_first_src_addr == NULL, "should have been cleared");
137 // Determine the number of destination regions for the partial object.
138 HeapWord* const last_word = destination + partial_obj_size - 1;
139 const ParallelCompactData& sd = PSParallelCompact::summary_data();
140 HeapWord* const beg_region_addr = sd.region_align_down(destination);
141 HeapWord* const end_region_addr = sd.region_align_down(last_word);
143 if (beg_region_addr == end_region_addr) {
144 // One destination region.
145 _destination_count = 1;
146 if (end_region_addr == destination) {
147 // The destination falls on a region boundary, thus the first word of the
148 // partial object will be the first word copied to the destination region.
149 _dest_region_addr = end_region_addr;
150 _first_src_addr = sd.region_to_addr(src_region_idx);
151 }
152 } else {
153 // Two destination regions. When copied, the partial object will cross a
154 // destination region boundary, so a word somewhere within the partial
155 // object will be the first word copied to the second destination region.
156 _destination_count = 2;
157 _dest_region_addr = end_region_addr;
158 const size_t ofs = pointer_delta(end_region_addr, destination);
159 assert(ofs < _partial_obj_size, "sanity");
160 _first_src_addr = sd.region_to_addr(src_region_idx) + ofs;
161 }
162 }
164 void SplitInfo::clear()
165 {
166 _src_region_idx = 0;
167 _partial_obj_size = 0;
168 _destination = NULL;
169 _destination_count = 0;
170 _dest_region_addr = NULL;
171 _first_src_addr = NULL;
172 assert(!is_valid(), "sanity");
173 }
175 #ifdef ASSERT
176 void SplitInfo::verify_clear()
177 {
178 assert(_src_region_idx == 0, "not clear");
179 assert(_partial_obj_size == 0, "not clear");
180 assert(_destination == NULL, "not clear");
181 assert(_destination_count == 0, "not clear");
182 assert(_dest_region_addr == NULL, "not clear");
183 assert(_first_src_addr == NULL, "not clear");
184 }
185 #endif // #ifdef ASSERT
188 #ifndef PRODUCT
189 const char* PSParallelCompact::space_names[] = {
190 "perm", "old ", "eden", "from", "to "
191 };
193 void PSParallelCompact::print_region_ranges()
194 {
195 tty->print_cr("space bottom top end new_top");
196 tty->print_cr("------ ---------- ---------- ---------- ----------");
198 for (unsigned int id = 0; id < last_space_id; ++id) {
199 const MutableSpace* space = _space_info[id].space();
200 tty->print_cr("%u %s "
201 SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " "
202 SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " ",
203 id, space_names[id],
204 summary_data().addr_to_region_idx(space->bottom()),
205 summary_data().addr_to_region_idx(space->top()),
206 summary_data().addr_to_region_idx(space->end()),
207 summary_data().addr_to_region_idx(_space_info[id].new_top()));
208 }
209 }
211 void
212 print_generic_summary_region(size_t i, const ParallelCompactData::RegionData* c)
213 {
214 #define REGION_IDX_FORMAT SIZE_FORMAT_W(7)
215 #define REGION_DATA_FORMAT SIZE_FORMAT_W(5)
217 ParallelCompactData& sd = PSParallelCompact::summary_data();
218 size_t dci = c->destination() ? sd.addr_to_region_idx(c->destination()) : 0;
219 tty->print_cr(REGION_IDX_FORMAT " " PTR_FORMAT " "
220 REGION_IDX_FORMAT " " PTR_FORMAT " "
221 REGION_DATA_FORMAT " " REGION_DATA_FORMAT " "
222 REGION_DATA_FORMAT " " REGION_IDX_FORMAT " %d",
223 i, c->data_location(), dci, c->destination(),
224 c->partial_obj_size(), c->live_obj_size(),
225 c->data_size(), c->source_region(), c->destination_count());
227 #undef REGION_IDX_FORMAT
228 #undef REGION_DATA_FORMAT
229 }
231 void
232 print_generic_summary_data(ParallelCompactData& summary_data,
233 HeapWord* const beg_addr,
234 HeapWord* const end_addr)
235 {
236 size_t total_words = 0;
237 size_t i = summary_data.addr_to_region_idx(beg_addr);
238 const size_t last = summary_data.addr_to_region_idx(end_addr);
239 HeapWord* pdest = 0;
241 while (i <= last) {
242 ParallelCompactData::RegionData* c = summary_data.region(i);
243 if (c->data_size() != 0 || c->destination() != pdest) {
244 print_generic_summary_region(i, c);
245 total_words += c->data_size();
246 pdest = c->destination();
247 }
248 ++i;
249 }
251 tty->print_cr("summary_data_bytes=" SIZE_FORMAT, total_words * HeapWordSize);
252 }
254 void
255 print_generic_summary_data(ParallelCompactData& summary_data,
256 SpaceInfo* space_info)
257 {
258 for (unsigned int id = 0; id < PSParallelCompact::last_space_id; ++id) {
259 const MutableSpace* space = space_info[id].space();
260 print_generic_summary_data(summary_data, space->bottom(),
261 MAX2(space->top(), space_info[id].new_top()));
262 }
263 }
265 void
266 print_initial_summary_region(size_t i,
267 const ParallelCompactData::RegionData* c,
268 bool newline = true)
269 {
270 tty->print(SIZE_FORMAT_W(5) " " PTR_FORMAT " "
271 SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " "
272 SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d",
273 i, c->destination(),
274 c->partial_obj_size(), c->live_obj_size(),
275 c->data_size(), c->source_region(), c->destination_count());
276 if (newline) tty->cr();
277 }
279 void
280 print_initial_summary_data(ParallelCompactData& summary_data,
281 const MutableSpace* space) {
282 if (space->top() == space->bottom()) {
283 return;
284 }
286 const size_t region_size = ParallelCompactData::RegionSize;
287 typedef ParallelCompactData::RegionData RegionData;
288 HeapWord* const top_aligned_up = summary_data.region_align_up(space->top());
289 const size_t end_region = summary_data.addr_to_region_idx(top_aligned_up);
290 const RegionData* c = summary_data.region(end_region - 1);
291 HeapWord* end_addr = c->destination() + c->data_size();
292 const size_t live_in_space = pointer_delta(end_addr, space->bottom());
294 // Print (and count) the full regions at the beginning of the space.
295 size_t full_region_count = 0;
296 size_t i = summary_data.addr_to_region_idx(space->bottom());
297 while (i < end_region && summary_data.region(i)->data_size() == region_size) {
298 print_initial_summary_region(i, summary_data.region(i));
299 ++full_region_count;
300 ++i;
301 }
303 size_t live_to_right = live_in_space - full_region_count * region_size;
305 double max_reclaimed_ratio = 0.0;
306 size_t max_reclaimed_ratio_region = 0;
307 size_t max_dead_to_right = 0;
308 size_t max_live_to_right = 0;
310 // Print the 'reclaimed ratio' for regions while there is something live in
311 // the region or to the right of it. The remaining regions are empty (and
312 // uninteresting), and computing the ratio will result in division by 0.
313 while (i < end_region && live_to_right > 0) {
314 c = summary_data.region(i);
315 HeapWord* const region_addr = summary_data.region_to_addr(i);
316 const size_t used_to_right = pointer_delta(space->top(), region_addr);
317 const size_t dead_to_right = used_to_right - live_to_right;
318 const double reclaimed_ratio = double(dead_to_right) / live_to_right;
320 if (reclaimed_ratio > max_reclaimed_ratio) {
321 max_reclaimed_ratio = reclaimed_ratio;
322 max_reclaimed_ratio_region = i;
323 max_dead_to_right = dead_to_right;
324 max_live_to_right = live_to_right;
325 }
327 print_initial_summary_region(i, c, false);
328 tty->print_cr(" %12.10f " SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10),
329 reclaimed_ratio, dead_to_right, live_to_right);
331 live_to_right -= c->data_size();
332 ++i;
333 }
335 // Any remaining regions are empty. Print one more if there is one.
336 if (i < end_region) {
337 print_initial_summary_region(i, summary_data.region(i));
338 }
340 tty->print_cr("max: " SIZE_FORMAT_W(4) " d2r=" SIZE_FORMAT_W(10) " "
341 "l2r=" SIZE_FORMAT_W(10) " max_ratio=%14.12f",
342 max_reclaimed_ratio_region, max_dead_to_right,
343 max_live_to_right, max_reclaimed_ratio);
344 }
346 void
347 print_initial_summary_data(ParallelCompactData& summary_data,
348 SpaceInfo* space_info) {
349 unsigned int id = PSParallelCompact::perm_space_id;
350 const MutableSpace* space;
351 do {
352 space = space_info[id].space();
353 print_initial_summary_data(summary_data, space);
354 } while (++id < PSParallelCompact::eden_space_id);
356 do {
357 space = space_info[id].space();
358 print_generic_summary_data(summary_data, space->bottom(), space->top());
359 } while (++id < PSParallelCompact::last_space_id);
360 }
361 #endif // #ifndef PRODUCT
363 #ifdef ASSERT
364 size_t add_obj_count;
365 size_t add_obj_size;
366 size_t mark_bitmap_count;
367 size_t mark_bitmap_size;
368 #endif // #ifdef ASSERT
370 ParallelCompactData::ParallelCompactData()
371 {
372 _region_start = 0;
374 _region_vspace = 0;
375 _region_data = 0;
376 _region_count = 0;
377 }
379 bool ParallelCompactData::initialize(MemRegion covered_region)
380 {
381 _region_start = covered_region.start();
382 const size_t region_size = covered_region.word_size();
383 DEBUG_ONLY(_region_end = _region_start + region_size;)
385 assert(region_align_down(_region_start) == _region_start,
386 "region start not aligned");
387 assert((region_size & RegionSizeOffsetMask) == 0,
388 "region size not a multiple of RegionSize");
390 bool result = initialize_region_data(region_size);
392 return result;
393 }
395 PSVirtualSpace*
396 ParallelCompactData::create_vspace(size_t count, size_t element_size)
397 {
398 const size_t raw_bytes = count * element_size;
399 const size_t page_sz = os::page_size_for_region(raw_bytes, raw_bytes, 10);
400 const size_t granularity = os::vm_allocation_granularity();
401 const size_t bytes = align_size_up(raw_bytes, MAX2(page_sz, granularity));
403 const size_t rs_align = page_sz == (size_t) os::vm_page_size() ? 0 :
404 MAX2(page_sz, granularity);
405 ReservedSpace rs(bytes, rs_align, rs_align > 0);
406 os::trace_page_sizes("par compact", raw_bytes, raw_bytes, page_sz, rs.base(),
407 rs.size());
408 PSVirtualSpace* vspace = new PSVirtualSpace(rs, page_sz);
409 if (vspace != 0) {
410 if (vspace->expand_by(bytes)) {
411 return vspace;
412 }
413 delete vspace;
414 // Release memory reserved in the space.
415 rs.release();
416 }
418 return 0;
419 }
421 bool ParallelCompactData::initialize_region_data(size_t region_size)
422 {
423 const size_t count = (region_size + RegionSizeOffsetMask) >> Log2RegionSize;
424 _region_vspace = create_vspace(count, sizeof(RegionData));
425 if (_region_vspace != 0) {
426 _region_data = (RegionData*)_region_vspace->reserved_low_addr();
427 _region_count = count;
428 return true;
429 }
430 return false;
431 }
433 void ParallelCompactData::clear()
434 {
435 memset(_region_data, 0, _region_vspace->committed_size());
436 }
438 void ParallelCompactData::clear_range(size_t beg_region, size_t end_region) {
439 assert(beg_region <= _region_count, "beg_region out of range");
440 assert(end_region <= _region_count, "end_region out of range");
442 const size_t region_cnt = end_region - beg_region;
443 memset(_region_data + beg_region, 0, region_cnt * sizeof(RegionData));
444 }
446 HeapWord* ParallelCompactData::partial_obj_end(size_t region_idx) const
447 {
448 const RegionData* cur_cp = region(region_idx);
449 const RegionData* const end_cp = region(region_count() - 1);
451 HeapWord* result = region_to_addr(region_idx);
452 if (cur_cp < end_cp) {
453 do {
454 result += cur_cp->partial_obj_size();
455 } while (cur_cp->partial_obj_size() == RegionSize && ++cur_cp < end_cp);
456 }
457 return result;
458 }
460 void ParallelCompactData::add_obj(HeapWord* addr, size_t len)
461 {
462 const size_t obj_ofs = pointer_delta(addr, _region_start);
463 const size_t beg_region = obj_ofs >> Log2RegionSize;
464 const size_t end_region = (obj_ofs + len - 1) >> Log2RegionSize;
466 DEBUG_ONLY(Atomic::inc_ptr(&add_obj_count);)
467 DEBUG_ONLY(Atomic::add_ptr(len, &add_obj_size);)
469 if (beg_region == end_region) {
470 // All in one region.
471 _region_data[beg_region].add_live_obj(len);
472 return;
473 }
475 // First region.
476 const size_t beg_ofs = region_offset(addr);
477 _region_data[beg_region].add_live_obj(RegionSize - beg_ofs);
479 klassOop klass = ((oop)addr)->klass();
480 // Middle regions--completely spanned by this object.
481 for (size_t region = beg_region + 1; region < end_region; ++region) {
482 _region_data[region].set_partial_obj_size(RegionSize);
483 _region_data[region].set_partial_obj_addr(addr);
484 }
486 // Last region.
487 const size_t end_ofs = region_offset(addr + len - 1);
488 _region_data[end_region].set_partial_obj_size(end_ofs + 1);
489 _region_data[end_region].set_partial_obj_addr(addr);
490 }
492 void
493 ParallelCompactData::summarize_dense_prefix(HeapWord* beg, HeapWord* end)
494 {
495 assert(region_offset(beg) == 0, "not RegionSize aligned");
496 assert(region_offset(end) == 0, "not RegionSize aligned");
498 size_t cur_region = addr_to_region_idx(beg);
499 const size_t end_region = addr_to_region_idx(end);
500 HeapWord* addr = beg;
501 while (cur_region < end_region) {
502 _region_data[cur_region].set_destination(addr);
503 _region_data[cur_region].set_destination_count(0);
504 _region_data[cur_region].set_source_region(cur_region);
505 _region_data[cur_region].set_data_location(addr);
507 // Update live_obj_size so the region appears completely full.
508 size_t live_size = RegionSize - _region_data[cur_region].partial_obj_size();
509 _region_data[cur_region].set_live_obj_size(live_size);
511 ++cur_region;
512 addr += RegionSize;
513 }
514 }
516 // Find the point at which a space can be split and, if necessary, record the
517 // split point.
518 //
519 // If the current src region (which overflowed the destination space) doesn't
520 // have a partial object, the split point is at the beginning of the current src
521 // region (an "easy" split, no extra bookkeeping required).
522 //
523 // If the current src region has a partial object, the split point is in the
524 // region where that partial object starts (call it the split_region). If
525 // split_region has a partial object, then the split point is just after that
526 // partial object (a "hard" split where we have to record the split data and
527 // zero the partial_obj_size field). With a "hard" split, we know that the
528 // partial_obj ends within split_region because the partial object that caused
529 // the overflow starts in split_region. If split_region doesn't have a partial
530 // obj, then the split is at the beginning of split_region (another "easy"
531 // split).
532 HeapWord*
533 ParallelCompactData::summarize_split_space(size_t src_region,
534 SplitInfo& split_info,
535 HeapWord* destination,
536 HeapWord* target_end,
537 HeapWord** target_next)
538 {
539 assert(destination <= target_end, "sanity");
540 assert(destination + _region_data[src_region].data_size() > target_end,
541 "region should not fit into target space");
542 assert(is_region_aligned(target_end), "sanity");
544 size_t split_region = src_region;
545 HeapWord* split_destination = destination;
546 size_t partial_obj_size = _region_data[src_region].partial_obj_size();
548 if (destination + partial_obj_size > target_end) {
549 // The split point is just after the partial object (if any) in the
550 // src_region that contains the start of the object that overflowed the
551 // destination space.
552 //
553 // Find the start of the "overflow" object and set split_region to the
554 // region containing it.
555 HeapWord* const overflow_obj = _region_data[src_region].partial_obj_addr();
556 split_region = addr_to_region_idx(overflow_obj);
558 // Clear the source_region field of all destination regions whose first word
559 // came from data after the split point (a non-null source_region field
560 // implies a region must be filled).
561 //
562 // An alternative to the simple loop below: clear during post_compact(),
563 // which uses memcpy instead of individual stores, and is easy to
564 // parallelize. (The downside is that it clears the entire RegionData
565 // object as opposed to just one field.)
566 //
567 // post_compact() would have to clear the summary data up to the highest
568 // address that was written during the summary phase, which would be
569 //
570 // max(top, max(new_top, clear_top))
571 //
572 // where clear_top is a new field in SpaceInfo. Would have to set clear_top
573 // to target_end.
574 const RegionData* const sr = region(split_region);
575 const size_t beg_idx =
576 addr_to_region_idx(region_align_up(sr->destination() +
577 sr->partial_obj_size()));
578 const size_t end_idx = addr_to_region_idx(target_end);
580 if (TraceParallelOldGCSummaryPhase) {
581 gclog_or_tty->print_cr("split: clearing source_region field in ["
582 SIZE_FORMAT ", " SIZE_FORMAT ")",
583 beg_idx, end_idx);
584 }
585 for (size_t idx = beg_idx; idx < end_idx; ++idx) {
586 _region_data[idx].set_source_region(0);
587 }
589 // Set split_destination and partial_obj_size to reflect the split region.
590 split_destination = sr->destination();
591 partial_obj_size = sr->partial_obj_size();
592 }
594 // The split is recorded only if a partial object extends onto the region.
595 if (partial_obj_size != 0) {
596 _region_data[split_region].set_partial_obj_size(0);
597 split_info.record(split_region, partial_obj_size, split_destination);
598 }
600 // Setup the continuation addresses.
601 *target_next = split_destination + partial_obj_size;
602 HeapWord* const source_next = region_to_addr(split_region) + partial_obj_size;
604 if (TraceParallelOldGCSummaryPhase) {
605 const char * split_type = partial_obj_size == 0 ? "easy" : "hard";
606 gclog_or_tty->print_cr("%s split: src=" PTR_FORMAT " src_c=" SIZE_FORMAT
607 " pos=" SIZE_FORMAT,
608 split_type, source_next, split_region,
609 partial_obj_size);
610 gclog_or_tty->print_cr("%s split: dst=" PTR_FORMAT " dst_c=" SIZE_FORMAT
611 " tn=" PTR_FORMAT,
612 split_type, split_destination,
613 addr_to_region_idx(split_destination),
614 *target_next);
616 if (partial_obj_size != 0) {
617 HeapWord* const po_beg = split_info.destination();
618 HeapWord* const po_end = po_beg + split_info.partial_obj_size();
619 gclog_or_tty->print_cr("%s split: "
620 "po_beg=" PTR_FORMAT " " SIZE_FORMAT " "
621 "po_end=" PTR_FORMAT " " SIZE_FORMAT,
622 split_type,
623 po_beg, addr_to_region_idx(po_beg),
624 po_end, addr_to_region_idx(po_end));
625 }
626 }
628 return source_next;
629 }
631 bool ParallelCompactData::summarize(SplitInfo& split_info,
632 HeapWord* source_beg, HeapWord* source_end,
633 HeapWord** source_next,
634 HeapWord* target_beg, HeapWord* target_end,
635 HeapWord** target_next)
636 {
637 if (TraceParallelOldGCSummaryPhase) {
638 HeapWord* const source_next_val = source_next == NULL ? NULL : *source_next;
639 tty->print_cr("sb=" PTR_FORMAT " se=" PTR_FORMAT " sn=" PTR_FORMAT
640 "tb=" PTR_FORMAT " te=" PTR_FORMAT " tn=" PTR_FORMAT,
641 source_beg, source_end, source_next_val,
642 target_beg, target_end, *target_next);
643 }
645 size_t cur_region = addr_to_region_idx(source_beg);
646 const size_t end_region = addr_to_region_idx(region_align_up(source_end));
648 HeapWord *dest_addr = target_beg;
649 while (cur_region < end_region) {
650 // The destination must be set even if the region has no data.
651 _region_data[cur_region].set_destination(dest_addr);
653 size_t words = _region_data[cur_region].data_size();
654 if (words > 0) {
655 // If cur_region does not fit entirely into the target space, find a point
656 // at which the source space can be 'split' so that part is copied to the
657 // target space and the rest is copied elsewhere.
658 if (dest_addr + words > target_end) {
659 assert(source_next != NULL, "source_next is NULL when splitting");
660 *source_next = summarize_split_space(cur_region, split_info, dest_addr,
661 target_end, target_next);
662 return false;
663 }
665 // Compute the destination_count for cur_region, and if necessary, update
666 // source_region for a destination region. The source_region field is
667 // updated if cur_region is the first (left-most) region to be copied to a
668 // destination region.
669 //
670 // The destination_count calculation is a bit subtle. A region that has
671 // data that compacts into itself does not count itself as a destination.
672 // This maintains the invariant that a zero count means the region is
673 // available and can be claimed and then filled.
674 uint destination_count = 0;
675 if (split_info.is_split(cur_region)) {
676 // The current region has been split: the partial object will be copied
677 // to one destination space and the remaining data will be copied to
678 // another destination space. Adjust the initial destination_count and,
679 // if necessary, set the source_region field if the partial object will
680 // cross a destination region boundary.
681 destination_count = split_info.destination_count();
682 if (destination_count == 2) {
683 size_t dest_idx = addr_to_region_idx(split_info.dest_region_addr());
684 _region_data[dest_idx].set_source_region(cur_region);
685 }
686 }
688 HeapWord* const last_addr = dest_addr + words - 1;
689 const size_t dest_region_1 = addr_to_region_idx(dest_addr);
690 const size_t dest_region_2 = addr_to_region_idx(last_addr);
692 // Initially assume that the destination regions will be the same and
693 // adjust the value below if necessary. Under this assumption, if
694 // cur_region == dest_region_2, then cur_region will be compacted
695 // completely into itself.
696 destination_count += cur_region == dest_region_2 ? 0 : 1;
697 if (dest_region_1 != dest_region_2) {
698 // Destination regions differ; adjust destination_count.
699 destination_count += 1;
700 // Data from cur_region will be copied to the start of dest_region_2.
701 _region_data[dest_region_2].set_source_region(cur_region);
702 } else if (region_offset(dest_addr) == 0) {
703 // Data from cur_region will be copied to the start of the destination
704 // region.
705 _region_data[dest_region_1].set_source_region(cur_region);
706 }
708 _region_data[cur_region].set_destination_count(destination_count);
709 _region_data[cur_region].set_data_location(region_to_addr(cur_region));
710 dest_addr += words;
711 }
713 ++cur_region;
714 }
716 *target_next = dest_addr;
717 return true;
718 }
720 HeapWord* ParallelCompactData::calc_new_pointer(HeapWord* addr) {
721 assert(addr != NULL, "Should detect NULL oop earlier");
722 assert(PSParallelCompact::gc_heap()->is_in(addr), "addr not in heap");
723 #ifdef ASSERT
724 if (PSParallelCompact::mark_bitmap()->is_unmarked(addr)) {
725 gclog_or_tty->print_cr("calc_new_pointer:: addr " PTR_FORMAT, addr);
726 }
727 #endif
728 assert(PSParallelCompact::mark_bitmap()->is_marked(addr), "obj not marked");
730 // Region covering the object.
731 size_t region_index = addr_to_region_idx(addr);
732 const RegionData* const region_ptr = region(region_index);
733 HeapWord* const region_addr = region_align_down(addr);
735 assert(addr < region_addr + RegionSize, "Region does not cover object");
736 assert(addr_to_region_ptr(region_addr) == region_ptr, "sanity check");
738 HeapWord* result = region_ptr->destination();
740 // If all the data in the region is live, then the new location of the object
741 // can be calculated from the destination of the region plus the offset of the
742 // object in the region.
743 if (region_ptr->data_size() == RegionSize) {
744 result += pointer_delta(addr, region_addr);
745 DEBUG_ONLY(PSParallelCompact::check_new_location(addr, result);)
746 return result;
747 }
749 // The new location of the object is
750 // region destination +
751 // size of the partial object extending onto the region +
752 // sizes of the live objects in the Region that are to the left of addr
753 const size_t partial_obj_size = region_ptr->partial_obj_size();
754 HeapWord* const search_start = region_addr + partial_obj_size;
756 const ParMarkBitMap* bitmap = PSParallelCompact::mark_bitmap();
757 size_t live_to_left = bitmap->live_words_in_range(search_start, oop(addr));
759 result += partial_obj_size + live_to_left;
760 DEBUG_ONLY(PSParallelCompact::check_new_location(addr, result);)
761 return result;
762 }
764 klassOop ParallelCompactData::calc_new_klass(klassOop old_klass) {
765 klassOop updated_klass;
766 if (PSParallelCompact::should_update_klass(old_klass)) {
767 updated_klass = (klassOop) calc_new_pointer(old_klass);
768 } else {
769 updated_klass = old_klass;
770 }
772 return updated_klass;
773 }
775 #ifdef ASSERT
776 void ParallelCompactData::verify_clear(const PSVirtualSpace* vspace)
777 {
778 const size_t* const beg = (const size_t*)vspace->committed_low_addr();
779 const size_t* const end = (const size_t*)vspace->committed_high_addr();
780 for (const size_t* p = beg; p < end; ++p) {
781 assert(*p == 0, "not zero");
782 }
783 }
785 void ParallelCompactData::verify_clear()
786 {
787 verify_clear(_region_vspace);
788 }
789 #endif // #ifdef ASSERT
791 #ifdef NOT_PRODUCT
792 ParallelCompactData::RegionData* debug_region(size_t region_index) {
793 ParallelCompactData& sd = PSParallelCompact::summary_data();
794 return sd.region(region_index);
795 }
796 #endif
798 elapsedTimer PSParallelCompact::_accumulated_time;
799 unsigned int PSParallelCompact::_total_invocations = 0;
800 unsigned int PSParallelCompact::_maximum_compaction_gc_num = 0;
801 jlong PSParallelCompact::_time_of_last_gc = 0;
802 CollectorCounters* PSParallelCompact::_counters = NULL;
803 ParMarkBitMap PSParallelCompact::_mark_bitmap;
804 ParallelCompactData PSParallelCompact::_summary_data;
806 PSParallelCompact::IsAliveClosure PSParallelCompact::_is_alive_closure;
808 void PSParallelCompact::IsAliveClosure::do_object(oop p) { ShouldNotReachHere(); }
809 bool PSParallelCompact::IsAliveClosure::do_object_b(oop p) { return mark_bitmap()->is_marked(p); }
811 void PSParallelCompact::KeepAliveClosure::do_oop(oop* p) { PSParallelCompact::KeepAliveClosure::do_oop_work(p); }
812 void PSParallelCompact::KeepAliveClosure::do_oop(narrowOop* p) { PSParallelCompact::KeepAliveClosure::do_oop_work(p); }
814 PSParallelCompact::AdjustPointerClosure PSParallelCompact::_adjust_root_pointer_closure(true);
815 PSParallelCompact::AdjustPointerClosure PSParallelCompact::_adjust_pointer_closure(false);
817 void PSParallelCompact::AdjustPointerClosure::do_oop(oop* p) { adjust_pointer(p, _is_root); }
818 void PSParallelCompact::AdjustPointerClosure::do_oop(narrowOop* p) { adjust_pointer(p, _is_root); }
820 void PSParallelCompact::FollowStackClosure::do_void() { _compaction_manager->follow_marking_stacks(); }
822 void PSParallelCompact::MarkAndPushClosure::do_oop(oop* p) { mark_and_push(_compaction_manager, p); }
823 void PSParallelCompact::MarkAndPushClosure::do_oop(narrowOop* p) { mark_and_push(_compaction_manager, p); }
825 void PSParallelCompact::post_initialize() {
826 ParallelScavengeHeap* heap = gc_heap();
827 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity");
829 MemRegion mr = heap->reserved_region();
830 _ref_processor =
831 new ReferenceProcessor(mr, // span
832 ParallelRefProcEnabled && (ParallelGCThreads > 1), // mt processing
833 (int) ParallelGCThreads, // mt processing degree
834 true, // mt discovery
835 (int) ParallelGCThreads, // mt discovery degree
836 true, // atomic_discovery
837 &_is_alive_closure, // non-header is alive closure
838 false); // write barrier for next field updates
839 _counters = new CollectorCounters("PSParallelCompact", 1);
841 // Initialize static fields in ParCompactionManager.
842 ParCompactionManager::initialize(mark_bitmap());
843 }
845 bool PSParallelCompact::initialize() {
846 ParallelScavengeHeap* heap = gc_heap();
847 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity");
848 MemRegion mr = heap->reserved_region();
850 // Was the old gen get allocated successfully?
851 if (!heap->old_gen()->is_allocated()) {
852 return false;
853 }
855 initialize_space_info();
856 initialize_dead_wood_limiter();
858 if (!_mark_bitmap.initialize(mr)) {
859 vm_shutdown_during_initialization("Unable to allocate bit map for "
860 "parallel garbage collection for the requested heap size.");
861 return false;
862 }
864 if (!_summary_data.initialize(mr)) {
865 vm_shutdown_during_initialization("Unable to allocate tables for "
866 "parallel garbage collection for the requested heap size.");
867 return false;
868 }
870 return true;
871 }
873 void PSParallelCompact::initialize_space_info()
874 {
875 memset(&_space_info, 0, sizeof(_space_info));
877 ParallelScavengeHeap* heap = gc_heap();
878 PSYoungGen* young_gen = heap->young_gen();
879 MutableSpace* perm_space = heap->perm_gen()->object_space();
881 _space_info[perm_space_id].set_space(perm_space);
882 _space_info[old_space_id].set_space(heap->old_gen()->object_space());
883 _space_info[eden_space_id].set_space(young_gen->eden_space());
884 _space_info[from_space_id].set_space(young_gen->from_space());
885 _space_info[to_space_id].set_space(young_gen->to_space());
887 _space_info[perm_space_id].set_start_array(heap->perm_gen()->start_array());
888 _space_info[old_space_id].set_start_array(heap->old_gen()->start_array());
890 _space_info[perm_space_id].set_min_dense_prefix(perm_space->top());
891 if (TraceParallelOldGCDensePrefix) {
892 tty->print_cr("perm min_dense_prefix=" PTR_FORMAT,
893 _space_info[perm_space_id].min_dense_prefix());
894 }
895 }
897 void PSParallelCompact::initialize_dead_wood_limiter()
898 {
899 const size_t max = 100;
900 _dwl_mean = double(MIN2(ParallelOldDeadWoodLimiterMean, max)) / 100.0;
901 _dwl_std_dev = double(MIN2(ParallelOldDeadWoodLimiterStdDev, max)) / 100.0;
902 _dwl_first_term = 1.0 / (sqrt(2.0 * M_PI) * _dwl_std_dev);
903 DEBUG_ONLY(_dwl_initialized = true;)
904 _dwl_adjustment = normal_distribution(1.0);
905 }
907 // Simple class for storing info about the heap at the start of GC, to be used
908 // after GC for comparison/printing.
909 class PreGCValues {
910 public:
911 PreGCValues() { }
912 PreGCValues(ParallelScavengeHeap* heap) { fill(heap); }
914 void fill(ParallelScavengeHeap* heap) {
915 _heap_used = heap->used();
916 _young_gen_used = heap->young_gen()->used_in_bytes();
917 _old_gen_used = heap->old_gen()->used_in_bytes();
918 _perm_gen_used = heap->perm_gen()->used_in_bytes();
919 };
921 size_t heap_used() const { return _heap_used; }
922 size_t young_gen_used() const { return _young_gen_used; }
923 size_t old_gen_used() const { return _old_gen_used; }
924 size_t perm_gen_used() const { return _perm_gen_used; }
926 private:
927 size_t _heap_used;
928 size_t _young_gen_used;
929 size_t _old_gen_used;
930 size_t _perm_gen_used;
931 };
933 void
934 PSParallelCompact::clear_data_covering_space(SpaceId id)
935 {
936 // At this point, top is the value before GC, new_top() is the value that will
937 // be set at the end of GC. The marking bitmap is cleared to top; nothing
938 // should be marked above top. The summary data is cleared to the larger of
939 // top & new_top.
940 MutableSpace* const space = _space_info[id].space();
941 HeapWord* const bot = space->bottom();
942 HeapWord* const top = space->top();
943 HeapWord* const max_top = MAX2(top, _space_info[id].new_top());
945 const idx_t beg_bit = _mark_bitmap.addr_to_bit(bot);
946 const idx_t end_bit = BitMap::word_align_up(_mark_bitmap.addr_to_bit(top));
947 _mark_bitmap.clear_range(beg_bit, end_bit);
949 const size_t beg_region = _summary_data.addr_to_region_idx(bot);
950 const size_t end_region =
951 _summary_data.addr_to_region_idx(_summary_data.region_align_up(max_top));
952 _summary_data.clear_range(beg_region, end_region);
954 // Clear the data used to 'split' regions.
955 SplitInfo& split_info = _space_info[id].split_info();
956 if (split_info.is_valid()) {
957 split_info.clear();
958 }
959 DEBUG_ONLY(split_info.verify_clear();)
960 }
962 void PSParallelCompact::pre_compact(PreGCValues* pre_gc_values)
963 {
964 // Update the from & to space pointers in space_info, since they are swapped
965 // at each young gen gc. Do the update unconditionally (even though a
966 // promotion failure does not swap spaces) because an unknown number of minor
967 // collections will have swapped the spaces an unknown number of times.
968 TraceTime tm("pre compact", print_phases(), true, gclog_or_tty);
969 ParallelScavengeHeap* heap = gc_heap();
970 _space_info[from_space_id].set_space(heap->young_gen()->from_space());
971 _space_info[to_space_id].set_space(heap->young_gen()->to_space());
973 pre_gc_values->fill(heap);
975 ParCompactionManager::reset();
976 NOT_PRODUCT(_mark_bitmap.reset_counters());
977 DEBUG_ONLY(add_obj_count = add_obj_size = 0;)
978 DEBUG_ONLY(mark_bitmap_count = mark_bitmap_size = 0;)
980 // Increment the invocation count
981 heap->increment_total_collections(true);
983 // We need to track unique mark sweep invocations as well.
984 _total_invocations++;
986 heap->print_heap_before_gc();
988 // Fill in TLABs
989 heap->accumulate_statistics_all_tlabs();
990 heap->ensure_parsability(true); // retire TLABs
992 if (VerifyBeforeGC && heap->total_collections() >= VerifyGCStartAt) {
993 HandleMark hm; // Discard invalid handles created during verification
994 gclog_or_tty->print(" VerifyBeforeGC:");
995 Universe::verify(true);
996 }
998 // Verify object start arrays
999 if (VerifyObjectStartArray &&
1000 VerifyBeforeGC) {
1001 heap->old_gen()->verify_object_start_array();
1002 heap->perm_gen()->verify_object_start_array();
1003 }
1005 DEBUG_ONLY(mark_bitmap()->verify_clear();)
1006 DEBUG_ONLY(summary_data().verify_clear();)
1008 // Have worker threads release resources the next time they run a task.
1009 gc_task_manager()->release_all_resources();
1010 }
1012 void PSParallelCompact::post_compact()
1013 {
1014 TraceTime tm("post compact", print_phases(), true, gclog_or_tty);
1016 for (unsigned int id = perm_space_id; id < last_space_id; ++id) {
1017 // Clear the marking bitmap, summary data and split info.
1018 clear_data_covering_space(SpaceId(id));
1019 // Update top(). Must be done after clearing the bitmap and summary data.
1020 _space_info[id].publish_new_top();
1021 }
1023 MutableSpace* const eden_space = _space_info[eden_space_id].space();
1024 MutableSpace* const from_space = _space_info[from_space_id].space();
1025 MutableSpace* const to_space = _space_info[to_space_id].space();
1027 ParallelScavengeHeap* heap = gc_heap();
1028 bool eden_empty = eden_space->is_empty();
1029 if (!eden_empty) {
1030 eden_empty = absorb_live_data_from_eden(heap->size_policy(),
1031 heap->young_gen(), heap->old_gen());
1032 }
1034 // Update heap occupancy information which is used as input to the soft ref
1035 // clearing policy at the next gc.
1036 Universe::update_heap_info_at_gc();
1038 bool young_gen_empty = eden_empty && from_space->is_empty() &&
1039 to_space->is_empty();
1041 BarrierSet* bs = heap->barrier_set();
1042 if (bs->is_a(BarrierSet::ModRef)) {
1043 ModRefBarrierSet* modBS = (ModRefBarrierSet*)bs;
1044 MemRegion old_mr = heap->old_gen()->reserved();
1045 MemRegion perm_mr = heap->perm_gen()->reserved();
1046 assert(perm_mr.end() <= old_mr.start(), "Generations out of order");
1048 if (young_gen_empty) {
1049 modBS->clear(MemRegion(perm_mr.start(), old_mr.end()));
1050 } else {
1051 modBS->invalidate(MemRegion(perm_mr.start(), old_mr.end()));
1052 }
1053 }
1055 Threads::gc_epilogue();
1056 CodeCache::gc_epilogue();
1057 JvmtiExport::gc_epilogue();
1059 COMPILER2_PRESENT(DerivedPointerTable::update_pointers());
1061 ref_processor()->enqueue_discovered_references(NULL);
1063 if (ZapUnusedHeapArea) {
1064 heap->gen_mangle_unused_area();
1065 }
1067 // Update time of last GC
1068 reset_millis_since_last_gc();
1069 }
1071 HeapWord*
1072 PSParallelCompact::compute_dense_prefix_via_density(const SpaceId id,
1073 bool maximum_compaction)
1074 {
1075 const size_t region_size = ParallelCompactData::RegionSize;
1076 const ParallelCompactData& sd = summary_data();
1078 const MutableSpace* const space = _space_info[id].space();
1079 HeapWord* const top_aligned_up = sd.region_align_up(space->top());
1080 const RegionData* const beg_cp = sd.addr_to_region_ptr(space->bottom());
1081 const RegionData* const end_cp = sd.addr_to_region_ptr(top_aligned_up);
1083 // Skip full regions at the beginning of the space--they are necessarily part
1084 // of the dense prefix.
1085 size_t full_count = 0;
1086 const RegionData* cp;
1087 for (cp = beg_cp; cp < end_cp && cp->data_size() == region_size; ++cp) {
1088 ++full_count;
1089 }
1091 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity");
1092 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num;
1093 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval;
1094 if (maximum_compaction || cp == end_cp || interval_ended) {
1095 _maximum_compaction_gc_num = total_invocations();
1096 return sd.region_to_addr(cp);
1097 }
1099 HeapWord* const new_top = _space_info[id].new_top();
1100 const size_t space_live = pointer_delta(new_top, space->bottom());
1101 const size_t space_used = space->used_in_words();
1102 const size_t space_capacity = space->capacity_in_words();
1104 const double cur_density = double(space_live) / space_capacity;
1105 const double deadwood_density =
1106 (1.0 - cur_density) * (1.0 - cur_density) * cur_density * cur_density;
1107 const size_t deadwood_goal = size_t(space_capacity * deadwood_density);
1109 if (TraceParallelOldGCDensePrefix) {
1110 tty->print_cr("cur_dens=%5.3f dw_dens=%5.3f dw_goal=" SIZE_FORMAT,
1111 cur_density, deadwood_density, deadwood_goal);
1112 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " "
1113 "space_cap=" SIZE_FORMAT,
1114 space_live, space_used,
1115 space_capacity);
1116 }
1118 // XXX - Use binary search?
1119 HeapWord* dense_prefix = sd.region_to_addr(cp);
1120 const RegionData* full_cp = cp;
1121 const RegionData* const top_cp = sd.addr_to_region_ptr(space->top() - 1);
1122 while (cp < end_cp) {
1123 HeapWord* region_destination = cp->destination();
1124 const size_t cur_deadwood = pointer_delta(dense_prefix, region_destination);
1125 if (TraceParallelOldGCDensePrefix && Verbose) {
1126 tty->print_cr("c#=" SIZE_FORMAT_W(4) " dst=" PTR_FORMAT " "
1127 "dp=" SIZE_FORMAT_W(8) " " "cdw=" SIZE_FORMAT_W(8),
1128 sd.region(cp), region_destination,
1129 dense_prefix, cur_deadwood);
1130 }
1132 if (cur_deadwood >= deadwood_goal) {
1133 // Found the region that has the correct amount of deadwood to the left.
1134 // This typically occurs after crossing a fairly sparse set of regions, so
1135 // iterate backwards over those sparse regions, looking for the region
1136 // that has the lowest density of live objects 'to the right.'
1137 size_t space_to_left = sd.region(cp) * region_size;
1138 size_t live_to_left = space_to_left - cur_deadwood;
1139 size_t space_to_right = space_capacity - space_to_left;
1140 size_t live_to_right = space_live - live_to_left;
1141 double density_to_right = double(live_to_right) / space_to_right;
1142 while (cp > full_cp) {
1143 --cp;
1144 const size_t prev_region_live_to_right = live_to_right -
1145 cp->data_size();
1146 const size_t prev_region_space_to_right = space_to_right + region_size;
1147 double prev_region_density_to_right =
1148 double(prev_region_live_to_right) / prev_region_space_to_right;
1149 if (density_to_right <= prev_region_density_to_right) {
1150 return dense_prefix;
1151 }
1152 if (TraceParallelOldGCDensePrefix && Verbose) {
1153 tty->print_cr("backing up from c=" SIZE_FORMAT_W(4) " d2r=%10.8f "
1154 "pc_d2r=%10.8f", sd.region(cp), density_to_right,
1155 prev_region_density_to_right);
1156 }
1157 dense_prefix -= region_size;
1158 live_to_right = prev_region_live_to_right;
1159 space_to_right = prev_region_space_to_right;
1160 density_to_right = prev_region_density_to_right;
1161 }
1162 return dense_prefix;
1163 }
1165 dense_prefix += region_size;
1166 ++cp;
1167 }
1169 return dense_prefix;
1170 }
1172 #ifndef PRODUCT
1173 void PSParallelCompact::print_dense_prefix_stats(const char* const algorithm,
1174 const SpaceId id,
1175 const bool maximum_compaction,
1176 HeapWord* const addr)
1177 {
1178 const size_t region_idx = summary_data().addr_to_region_idx(addr);
1179 RegionData* const cp = summary_data().region(region_idx);
1180 const MutableSpace* const space = _space_info[id].space();
1181 HeapWord* const new_top = _space_info[id].new_top();
1183 const size_t space_live = pointer_delta(new_top, space->bottom());
1184 const size_t dead_to_left = pointer_delta(addr, cp->destination());
1185 const size_t space_cap = space->capacity_in_words();
1186 const double dead_to_left_pct = double(dead_to_left) / space_cap;
1187 const size_t live_to_right = new_top - cp->destination();
1188 const size_t dead_to_right = space->top() - addr - live_to_right;
1190 tty->print_cr("%s=" PTR_FORMAT " dpc=" SIZE_FORMAT_W(5) " "
1191 "spl=" SIZE_FORMAT " "
1192 "d2l=" SIZE_FORMAT " d2l%%=%6.4f "
1193 "d2r=" SIZE_FORMAT " l2r=" SIZE_FORMAT
1194 " ratio=%10.8f",
1195 algorithm, addr, region_idx,
1196 space_live,
1197 dead_to_left, dead_to_left_pct,
1198 dead_to_right, live_to_right,
1199 double(dead_to_right) / live_to_right);
1200 }
1201 #endif // #ifndef PRODUCT
1203 // Return a fraction indicating how much of the generation can be treated as
1204 // "dead wood" (i.e., not reclaimed). The function uses a normal distribution
1205 // based on the density of live objects in the generation to determine a limit,
1206 // which is then adjusted so the return value is min_percent when the density is
1207 // 1.
1208 //
1209 // The following table shows some return values for a different values of the
1210 // standard deviation (ParallelOldDeadWoodLimiterStdDev); the mean is 0.5 and
1211 // min_percent is 1.
1212 //
1213 // fraction allowed as dead wood
1214 // -----------------------------------------------------------------
1215 // density std_dev=70 std_dev=75 std_dev=80 std_dev=85 std_dev=90 std_dev=95
1216 // ------- ---------- ---------- ---------- ---------- ---------- ----------
1217 // 0.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000
1218 // 0.05000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941
1219 // 0.10000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272
1220 // 0.15000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066
1221 // 0.20000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975
1222 // 0.25000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313
1223 // 0.30000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132
1224 // 0.35000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289
1225 // 0.40000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500
1226 // 0.45000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386
1227 // 0.50000 0.13832410 0.11599237 0.09847664 0.08456518 0.07338887 0.06431510
1228 // 0.55000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386
1229 // 0.60000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500
1230 // 0.65000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289
1231 // 0.70000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132
1232 // 0.75000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313
1233 // 0.80000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975
1234 // 0.85000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066
1235 // 0.90000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272
1236 // 0.95000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941
1237 // 1.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000
1239 double PSParallelCompact::dead_wood_limiter(double density, size_t min_percent)
1240 {
1241 assert(_dwl_initialized, "uninitialized");
1243 // The raw limit is the value of the normal distribution at x = density.
1244 const double raw_limit = normal_distribution(density);
1246 // Adjust the raw limit so it becomes the minimum when the density is 1.
1247 //
1248 // First subtract the adjustment value (which is simply the precomputed value
1249 // normal_distribution(1.0)); this yields a value of 0 when the density is 1.
1250 // Then add the minimum value, so the minimum is returned when the density is
1251 // 1. Finally, prevent negative values, which occur when the mean is not 0.5.
1252 const double min = double(min_percent) / 100.0;
1253 const double limit = raw_limit - _dwl_adjustment + min;
1254 return MAX2(limit, 0.0);
1255 }
1257 ParallelCompactData::RegionData*
1258 PSParallelCompact::first_dead_space_region(const RegionData* beg,
1259 const RegionData* end)
1260 {
1261 const size_t region_size = ParallelCompactData::RegionSize;
1262 ParallelCompactData& sd = summary_data();
1263 size_t left = sd.region(beg);
1264 size_t right = end > beg ? sd.region(end) - 1 : left;
1266 // Binary search.
1267 while (left < right) {
1268 // Equivalent to (left + right) / 2, but does not overflow.
1269 const size_t middle = left + (right - left) / 2;
1270 RegionData* const middle_ptr = sd.region(middle);
1271 HeapWord* const dest = middle_ptr->destination();
1272 HeapWord* const addr = sd.region_to_addr(middle);
1273 assert(dest != NULL, "sanity");
1274 assert(dest <= addr, "must move left");
1276 if (middle > left && dest < addr) {
1277 right = middle - 1;
1278 } else if (middle < right && middle_ptr->data_size() == region_size) {
1279 left = middle + 1;
1280 } else {
1281 return middle_ptr;
1282 }
1283 }
1284 return sd.region(left);
1285 }
1287 ParallelCompactData::RegionData*
1288 PSParallelCompact::dead_wood_limit_region(const RegionData* beg,
1289 const RegionData* end,
1290 size_t dead_words)
1291 {
1292 ParallelCompactData& sd = summary_data();
1293 size_t left = sd.region(beg);
1294 size_t right = end > beg ? sd.region(end) - 1 : left;
1296 // Binary search.
1297 while (left < right) {
1298 // Equivalent to (left + right) / 2, but does not overflow.
1299 const size_t middle = left + (right - left) / 2;
1300 RegionData* const middle_ptr = sd.region(middle);
1301 HeapWord* const dest = middle_ptr->destination();
1302 HeapWord* const addr = sd.region_to_addr(middle);
1303 assert(dest != NULL, "sanity");
1304 assert(dest <= addr, "must move left");
1306 const size_t dead_to_left = pointer_delta(addr, dest);
1307 if (middle > left && dead_to_left > dead_words) {
1308 right = middle - 1;
1309 } else if (middle < right && dead_to_left < dead_words) {
1310 left = middle + 1;
1311 } else {
1312 return middle_ptr;
1313 }
1314 }
1315 return sd.region(left);
1316 }
1318 // The result is valid during the summary phase, after the initial summarization
1319 // of each space into itself, and before final summarization.
1320 inline double
1321 PSParallelCompact::reclaimed_ratio(const RegionData* const cp,
1322 HeapWord* const bottom,
1323 HeapWord* const top,
1324 HeapWord* const new_top)
1325 {
1326 ParallelCompactData& sd = summary_data();
1328 assert(cp != NULL, "sanity");
1329 assert(bottom != NULL, "sanity");
1330 assert(top != NULL, "sanity");
1331 assert(new_top != NULL, "sanity");
1332 assert(top >= new_top, "summary data problem?");
1333 assert(new_top > bottom, "space is empty; should not be here");
1334 assert(new_top >= cp->destination(), "sanity");
1335 assert(top >= sd.region_to_addr(cp), "sanity");
1337 HeapWord* const destination = cp->destination();
1338 const size_t dense_prefix_live = pointer_delta(destination, bottom);
1339 const size_t compacted_region_live = pointer_delta(new_top, destination);
1340 const size_t compacted_region_used = pointer_delta(top,
1341 sd.region_to_addr(cp));
1342 const size_t reclaimable = compacted_region_used - compacted_region_live;
1344 const double divisor = dense_prefix_live + 1.25 * compacted_region_live;
1345 return double(reclaimable) / divisor;
1346 }
1348 // Return the address of the end of the dense prefix, a.k.a. the start of the
1349 // compacted region. The address is always on a region boundary.
1350 //
1351 // Completely full regions at the left are skipped, since no compaction can
1352 // occur in those regions. Then the maximum amount of dead wood to allow is
1353 // computed, based on the density (amount live / capacity) of the generation;
1354 // the region with approximately that amount of dead space to the left is
1355 // identified as the limit region. Regions between the last completely full
1356 // region and the limit region are scanned and the one that has the best
1357 // (maximum) reclaimed_ratio() is selected.
1358 HeapWord*
1359 PSParallelCompact::compute_dense_prefix(const SpaceId id,
1360 bool maximum_compaction)
1361 {
1362 if (ParallelOldGCSplitALot) {
1363 if (_space_info[id].dense_prefix() != _space_info[id].space()->bottom()) {
1364 // The value was chosen to provoke splitting a young gen space; use it.
1365 return _space_info[id].dense_prefix();
1366 }
1367 }
1369 const size_t region_size = ParallelCompactData::RegionSize;
1370 const ParallelCompactData& sd = summary_data();
1372 const MutableSpace* const space = _space_info[id].space();
1373 HeapWord* const top = space->top();
1374 HeapWord* const top_aligned_up = sd.region_align_up(top);
1375 HeapWord* const new_top = _space_info[id].new_top();
1376 HeapWord* const new_top_aligned_up = sd.region_align_up(new_top);
1377 HeapWord* const bottom = space->bottom();
1378 const RegionData* const beg_cp = sd.addr_to_region_ptr(bottom);
1379 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up);
1380 const RegionData* const new_top_cp =
1381 sd.addr_to_region_ptr(new_top_aligned_up);
1383 // Skip full regions at the beginning of the space--they are necessarily part
1384 // of the dense prefix.
1385 const RegionData* const full_cp = first_dead_space_region(beg_cp, new_top_cp);
1386 assert(full_cp->destination() == sd.region_to_addr(full_cp) ||
1387 space->is_empty(), "no dead space allowed to the left");
1388 assert(full_cp->data_size() < region_size || full_cp == new_top_cp - 1,
1389 "region must have dead space");
1391 // The gc number is saved whenever a maximum compaction is done, and used to
1392 // determine when the maximum compaction interval has expired. This avoids
1393 // successive max compactions for different reasons.
1394 assert(total_invocations() >= _maximum_compaction_gc_num, "sanity");
1395 const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num;
1396 const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval ||
1397 total_invocations() == HeapFirstMaximumCompactionCount;
1398 if (maximum_compaction || full_cp == top_cp || interval_ended) {
1399 _maximum_compaction_gc_num = total_invocations();
1400 return sd.region_to_addr(full_cp);
1401 }
1403 const size_t space_live = pointer_delta(new_top, bottom);
1404 const size_t space_used = space->used_in_words();
1405 const size_t space_capacity = space->capacity_in_words();
1407 const double density = double(space_live) / double(space_capacity);
1408 const size_t min_percent_free =
1409 id == perm_space_id ? PermMarkSweepDeadRatio : MarkSweepDeadRatio;
1410 const double limiter = dead_wood_limiter(density, min_percent_free);
1411 const size_t dead_wood_max = space_used - space_live;
1412 const size_t dead_wood_limit = MIN2(size_t(space_capacity * limiter),
1413 dead_wood_max);
1415 if (TraceParallelOldGCDensePrefix) {
1416 tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " "
1417 "space_cap=" SIZE_FORMAT,
1418 space_live, space_used,
1419 space_capacity);
1420 tty->print_cr("dead_wood_limiter(%6.4f, %d)=%6.4f "
1421 "dead_wood_max=" SIZE_FORMAT " dead_wood_limit=" SIZE_FORMAT,
1422 density, min_percent_free, limiter,
1423 dead_wood_max, dead_wood_limit);
1424 }
1426 // Locate the region with the desired amount of dead space to the left.
1427 const RegionData* const limit_cp =
1428 dead_wood_limit_region(full_cp, top_cp, dead_wood_limit);
1430 // Scan from the first region with dead space to the limit region and find the
1431 // one with the best (largest) reclaimed ratio.
1432 double best_ratio = 0.0;
1433 const RegionData* best_cp = full_cp;
1434 for (const RegionData* cp = full_cp; cp < limit_cp; ++cp) {
1435 double tmp_ratio = reclaimed_ratio(cp, bottom, top, new_top);
1436 if (tmp_ratio > best_ratio) {
1437 best_cp = cp;
1438 best_ratio = tmp_ratio;
1439 }
1440 }
1442 #if 0
1443 // Something to consider: if the region with the best ratio is 'close to' the
1444 // first region w/free space, choose the first region with free space
1445 // ("first-free"). The first-free region is usually near the start of the
1446 // heap, which means we are copying most of the heap already, so copy a bit
1447 // more to get complete compaction.
1448 if (pointer_delta(best_cp, full_cp, sizeof(RegionData)) < 4) {
1449 _maximum_compaction_gc_num = total_invocations();
1450 best_cp = full_cp;
1451 }
1452 #endif // #if 0
1454 return sd.region_to_addr(best_cp);
1455 }
1457 #ifndef PRODUCT
1458 void
1459 PSParallelCompact::fill_with_live_objects(SpaceId id, HeapWord* const start,
1460 size_t words)
1461 {
1462 if (TraceParallelOldGCSummaryPhase) {
1463 tty->print_cr("fill_with_live_objects [" PTR_FORMAT " " PTR_FORMAT ") "
1464 SIZE_FORMAT, start, start + words, words);
1465 }
1467 ObjectStartArray* const start_array = _space_info[id].start_array();
1468 CollectedHeap::fill_with_objects(start, words);
1469 for (HeapWord* p = start; p < start + words; p += oop(p)->size()) {
1470 _mark_bitmap.mark_obj(p, words);
1471 _summary_data.add_obj(p, words);
1472 start_array->allocate_block(p);
1473 }
1474 }
1476 void
1477 PSParallelCompact::summarize_new_objects(SpaceId id, HeapWord* start)
1478 {
1479 ParallelCompactData& sd = summary_data();
1480 MutableSpace* space = _space_info[id].space();
1482 // Find the source and destination start addresses.
1483 HeapWord* const src_addr = sd.region_align_down(start);
1484 HeapWord* dst_addr;
1485 if (src_addr < start) {
1486 dst_addr = sd.addr_to_region_ptr(src_addr)->destination();
1487 } else if (src_addr > space->bottom()) {
1488 // The start (the original top() value) is aligned to a region boundary so
1489 // the associated region does not have a destination. Compute the
1490 // destination from the previous region.
1491 RegionData* const cp = sd.addr_to_region_ptr(src_addr) - 1;
1492 dst_addr = cp->destination() + cp->data_size();
1493 } else {
1494 // Filling the entire space.
1495 dst_addr = space->bottom();
1496 }
1497 assert(dst_addr != NULL, "sanity");
1499 // Update the summary data.
1500 bool result = _summary_data.summarize(_space_info[id].split_info(),
1501 src_addr, space->top(), NULL,
1502 dst_addr, space->end(),
1503 _space_info[id].new_top_addr());
1504 assert(result, "should not fail: bad filler object size");
1505 }
1507 void
1508 PSParallelCompact::provoke_split_fill_survivor(SpaceId id)
1509 {
1510 if (total_invocations() % (ParallelOldGCSplitInterval * 3) != 0) {
1511 return;
1512 }
1514 MutableSpace* const space = _space_info[id].space();
1515 if (space->is_empty()) {
1516 HeapWord* b = space->bottom();
1517 HeapWord* t = b + space->capacity_in_words() / 2;
1518 space->set_top(t);
1519 if (ZapUnusedHeapArea) {
1520 space->set_top_for_allocations();
1521 }
1523 size_t min_size = CollectedHeap::min_fill_size();
1524 size_t obj_len = min_size;
1525 while (b + obj_len <= t) {
1526 CollectedHeap::fill_with_object(b, obj_len);
1527 mark_bitmap()->mark_obj(b, obj_len);
1528 summary_data().add_obj(b, obj_len);
1529 b += obj_len;
1530 obj_len = (obj_len & (min_size*3)) + min_size; // 8 16 24 32 8 16 24 32 ...
1531 }
1532 if (b < t) {
1533 // The loop didn't completely fill to t (top); adjust top downward.
1534 space->set_top(b);
1535 if (ZapUnusedHeapArea) {
1536 space->set_top_for_allocations();
1537 }
1538 }
1540 HeapWord** nta = _space_info[id].new_top_addr();
1541 bool result = summary_data().summarize(_space_info[id].split_info(),
1542 space->bottom(), space->top(), NULL,
1543 space->bottom(), space->end(), nta);
1544 assert(result, "space must fit into itself");
1545 }
1546 }
1548 void
1549 PSParallelCompact::provoke_split(bool & max_compaction)
1550 {
1551 if (total_invocations() % ParallelOldGCSplitInterval != 0) {
1552 return;
1553 }
1555 const size_t region_size = ParallelCompactData::RegionSize;
1556 ParallelCompactData& sd = summary_data();
1558 MutableSpace* const eden_space = _space_info[eden_space_id].space();
1559 MutableSpace* const from_space = _space_info[from_space_id].space();
1560 const size_t eden_live = pointer_delta(eden_space->top(),
1561 _space_info[eden_space_id].new_top());
1562 const size_t from_live = pointer_delta(from_space->top(),
1563 _space_info[from_space_id].new_top());
1565 const size_t min_fill_size = CollectedHeap::min_fill_size();
1566 const size_t eden_free = pointer_delta(eden_space->end(), eden_space->top());
1567 const size_t eden_fillable = eden_free >= min_fill_size ? eden_free : 0;
1568 const size_t from_free = pointer_delta(from_space->end(), from_space->top());
1569 const size_t from_fillable = from_free >= min_fill_size ? from_free : 0;
1571 // Choose the space to split; need at least 2 regions live (or fillable).
1572 SpaceId id;
1573 MutableSpace* space;
1574 size_t live_words;
1575 size_t fill_words;
1576 if (eden_live + eden_fillable >= region_size * 2) {
1577 id = eden_space_id;
1578 space = eden_space;
1579 live_words = eden_live;
1580 fill_words = eden_fillable;
1581 } else if (from_live + from_fillable >= region_size * 2) {
1582 id = from_space_id;
1583 space = from_space;
1584 live_words = from_live;
1585 fill_words = from_fillable;
1586 } else {
1587 return; // Give up.
1588 }
1589 assert(fill_words == 0 || fill_words >= min_fill_size, "sanity");
1591 if (live_words < region_size * 2) {
1592 // Fill from top() to end() w/live objects of mixed sizes.
1593 HeapWord* const fill_start = space->top();
1594 live_words += fill_words;
1596 space->set_top(fill_start + fill_words);
1597 if (ZapUnusedHeapArea) {
1598 space->set_top_for_allocations();
1599 }
1601 HeapWord* cur_addr = fill_start;
1602 while (fill_words > 0) {
1603 const size_t r = (size_t)os::random() % (region_size / 2) + min_fill_size;
1604 size_t cur_size = MIN2(align_object_size_(r), fill_words);
1605 if (fill_words - cur_size < min_fill_size) {
1606 cur_size = fill_words; // Avoid leaving a fragment too small to fill.
1607 }
1609 CollectedHeap::fill_with_object(cur_addr, cur_size);
1610 mark_bitmap()->mark_obj(cur_addr, cur_size);
1611 sd.add_obj(cur_addr, cur_size);
1613 cur_addr += cur_size;
1614 fill_words -= cur_size;
1615 }
1617 summarize_new_objects(id, fill_start);
1618 }
1620 max_compaction = false;
1622 // Manipulate the old gen so that it has room for about half of the live data
1623 // in the target young gen space (live_words / 2).
1624 id = old_space_id;
1625 space = _space_info[id].space();
1626 const size_t free_at_end = space->free_in_words();
1627 const size_t free_target = align_object_size(live_words / 2);
1628 const size_t dead = pointer_delta(space->top(), _space_info[id].new_top());
1630 if (free_at_end >= free_target + min_fill_size) {
1631 // Fill space above top() and set the dense prefix so everything survives.
1632 HeapWord* const fill_start = space->top();
1633 const size_t fill_size = free_at_end - free_target;
1634 space->set_top(space->top() + fill_size);
1635 if (ZapUnusedHeapArea) {
1636 space->set_top_for_allocations();
1637 }
1638 fill_with_live_objects(id, fill_start, fill_size);
1639 summarize_new_objects(id, fill_start);
1640 _space_info[id].set_dense_prefix(sd.region_align_down(space->top()));
1641 } else if (dead + free_at_end > free_target) {
1642 // Find a dense prefix that makes the right amount of space available.
1643 HeapWord* cur = sd.region_align_down(space->top());
1644 HeapWord* cur_destination = sd.addr_to_region_ptr(cur)->destination();
1645 size_t dead_to_right = pointer_delta(space->end(), cur_destination);
1646 while (dead_to_right < free_target) {
1647 cur -= region_size;
1648 cur_destination = sd.addr_to_region_ptr(cur)->destination();
1649 dead_to_right = pointer_delta(space->end(), cur_destination);
1650 }
1651 _space_info[id].set_dense_prefix(cur);
1652 }
1653 }
1654 #endif // #ifndef PRODUCT
1656 void PSParallelCompact::summarize_spaces_quick()
1657 {
1658 for (unsigned int i = 0; i < last_space_id; ++i) {
1659 const MutableSpace* space = _space_info[i].space();
1660 HeapWord** nta = _space_info[i].new_top_addr();
1661 bool result = _summary_data.summarize(_space_info[i].split_info(),
1662 space->bottom(), space->top(), NULL,
1663 space->bottom(), space->end(), nta);
1664 assert(result, "space must fit into itself");
1665 _space_info[i].set_dense_prefix(space->bottom());
1666 }
1668 #ifndef PRODUCT
1669 if (ParallelOldGCSplitALot) {
1670 provoke_split_fill_survivor(to_space_id);
1671 }
1672 #endif // #ifndef PRODUCT
1673 }
1675 void PSParallelCompact::fill_dense_prefix_end(SpaceId id)
1676 {
1677 HeapWord* const dense_prefix_end = dense_prefix(id);
1678 const RegionData* region = _summary_data.addr_to_region_ptr(dense_prefix_end);
1679 const idx_t dense_prefix_bit = _mark_bitmap.addr_to_bit(dense_prefix_end);
1680 if (dead_space_crosses_boundary(region, dense_prefix_bit)) {
1681 // Only enough dead space is filled so that any remaining dead space to the
1682 // left is larger than the minimum filler object. (The remainder is filled
1683 // during the copy/update phase.)
1684 //
1685 // The size of the dead space to the right of the boundary is not a
1686 // concern, since compaction will be able to use whatever space is
1687 // available.
1688 //
1689 // Here '||' is the boundary, 'x' represents a don't care bit and a box
1690 // surrounds the space to be filled with an object.
1691 //
1692 // In the 32-bit VM, each bit represents two 32-bit words:
1693 // +---+
1694 // a) beg_bits: ... x x x | 0 | || 0 x x ...
1695 // end_bits: ... x x x | 0 | || 0 x x ...
1696 // +---+
1697 //
1698 // In the 64-bit VM, each bit represents one 64-bit word:
1699 // +------------+
1700 // b) beg_bits: ... x x x | 0 || 0 | x x ...
1701 // end_bits: ... x x 1 | 0 || 0 | x x ...
1702 // +------------+
1703 // +-------+
1704 // c) beg_bits: ... x x | 0 0 | || 0 x x ...
1705 // end_bits: ... x 1 | 0 0 | || 0 x x ...
1706 // +-------+
1707 // +-----------+
1708 // d) beg_bits: ... x | 0 0 0 | || 0 x x ...
1709 // end_bits: ... 1 | 0 0 0 | || 0 x x ...
1710 // +-----------+
1711 // +-------+
1712 // e) beg_bits: ... 0 0 | 0 0 | || 0 x x ...
1713 // end_bits: ... 0 0 | 0 0 | || 0 x x ...
1714 // +-------+
1716 // Initially assume case a, c or e will apply.
1717 size_t obj_len = CollectedHeap::min_fill_size();
1718 HeapWord* obj_beg = dense_prefix_end - obj_len;
1720 #ifdef _LP64
1721 if (MinObjAlignment > 1) { // object alignment > heap word size
1722 // Cases a, c or e.
1723 } else if (_mark_bitmap.is_obj_end(dense_prefix_bit - 2)) {
1724 // Case b above.
1725 obj_beg = dense_prefix_end - 1;
1726 } else if (!_mark_bitmap.is_obj_end(dense_prefix_bit - 3) &&
1727 _mark_bitmap.is_obj_end(dense_prefix_bit - 4)) {
1728 // Case d above.
1729 obj_beg = dense_prefix_end - 3;
1730 obj_len = 3;
1731 }
1732 #endif // #ifdef _LP64
1734 CollectedHeap::fill_with_object(obj_beg, obj_len);
1735 _mark_bitmap.mark_obj(obj_beg, obj_len);
1736 _summary_data.add_obj(obj_beg, obj_len);
1737 assert(start_array(id) != NULL, "sanity");
1738 start_array(id)->allocate_block(obj_beg);
1739 }
1740 }
1742 void
1743 PSParallelCompact::clear_source_region(HeapWord* beg_addr, HeapWord* end_addr)
1744 {
1745 RegionData* const beg_ptr = _summary_data.addr_to_region_ptr(beg_addr);
1746 HeapWord* const end_aligned_up = _summary_data.region_align_up(end_addr);
1747 RegionData* const end_ptr = _summary_data.addr_to_region_ptr(end_aligned_up);
1748 for (RegionData* cur = beg_ptr; cur < end_ptr; ++cur) {
1749 cur->set_source_region(0);
1750 }
1751 }
1753 void
1754 PSParallelCompact::summarize_space(SpaceId id, bool maximum_compaction)
1755 {
1756 assert(id < last_space_id, "id out of range");
1757 assert(_space_info[id].dense_prefix() == _space_info[id].space()->bottom() ||
1758 ParallelOldGCSplitALot && id == old_space_id,
1759 "should have been reset in summarize_spaces_quick()");
1761 const MutableSpace* space = _space_info[id].space();
1762 if (_space_info[id].new_top() != space->bottom()) {
1763 HeapWord* dense_prefix_end = compute_dense_prefix(id, maximum_compaction);
1764 _space_info[id].set_dense_prefix(dense_prefix_end);
1766 #ifndef PRODUCT
1767 if (TraceParallelOldGCDensePrefix) {
1768 print_dense_prefix_stats("ratio", id, maximum_compaction,
1769 dense_prefix_end);
1770 HeapWord* addr = compute_dense_prefix_via_density(id, maximum_compaction);
1771 print_dense_prefix_stats("density", id, maximum_compaction, addr);
1772 }
1773 #endif // #ifndef PRODUCT
1775 // Recompute the summary data, taking into account the dense prefix. If
1776 // every last byte will be reclaimed, then the existing summary data which
1777 // compacts everything can be left in place.
1778 if (!maximum_compaction && dense_prefix_end != space->bottom()) {
1779 // If dead space crosses the dense prefix boundary, it is (at least
1780 // partially) filled with a dummy object, marked live and added to the
1781 // summary data. This simplifies the copy/update phase and must be done
1782 // before the final locations of objects are determined, to prevent
1783 // leaving a fragment of dead space that is too small to fill.
1784 fill_dense_prefix_end(id);
1786 // Compute the destination of each Region, and thus each object.
1787 _summary_data.summarize_dense_prefix(space->bottom(), dense_prefix_end);
1788 _summary_data.summarize(_space_info[id].split_info(),
1789 dense_prefix_end, space->top(), NULL,
1790 dense_prefix_end, space->end(),
1791 _space_info[id].new_top_addr());
1792 }
1793 }
1795 if (TraceParallelOldGCSummaryPhase) {
1796 const size_t region_size = ParallelCompactData::RegionSize;
1797 HeapWord* const dense_prefix_end = _space_info[id].dense_prefix();
1798 const size_t dp_region = _summary_data.addr_to_region_idx(dense_prefix_end);
1799 const size_t dp_words = pointer_delta(dense_prefix_end, space->bottom());
1800 HeapWord* const new_top = _space_info[id].new_top();
1801 const HeapWord* nt_aligned_up = _summary_data.region_align_up(new_top);
1802 const size_t cr_words = pointer_delta(nt_aligned_up, dense_prefix_end);
1803 tty->print_cr("id=%d cap=" SIZE_FORMAT " dp=" PTR_FORMAT " "
1804 "dp_region=" SIZE_FORMAT " " "dp_count=" SIZE_FORMAT " "
1805 "cr_count=" SIZE_FORMAT " " "nt=" PTR_FORMAT,
1806 id, space->capacity_in_words(), dense_prefix_end,
1807 dp_region, dp_words / region_size,
1808 cr_words / region_size, new_top);
1809 }
1810 }
1812 #ifndef PRODUCT
1813 void PSParallelCompact::summary_phase_msg(SpaceId dst_space_id,
1814 HeapWord* dst_beg, HeapWord* dst_end,
1815 SpaceId src_space_id,
1816 HeapWord* src_beg, HeapWord* src_end)
1817 {
1818 if (TraceParallelOldGCSummaryPhase) {
1819 tty->print_cr("summarizing %d [%s] into %d [%s]: "
1820 "src=" PTR_FORMAT "-" PTR_FORMAT " "
1821 SIZE_FORMAT "-" SIZE_FORMAT " "
1822 "dst=" PTR_FORMAT "-" PTR_FORMAT " "
1823 SIZE_FORMAT "-" SIZE_FORMAT,
1824 src_space_id, space_names[src_space_id],
1825 dst_space_id, space_names[dst_space_id],
1826 src_beg, src_end,
1827 _summary_data.addr_to_region_idx(src_beg),
1828 _summary_data.addr_to_region_idx(src_end),
1829 dst_beg, dst_end,
1830 _summary_data.addr_to_region_idx(dst_beg),
1831 _summary_data.addr_to_region_idx(dst_end));
1832 }
1833 }
1834 #endif // #ifndef PRODUCT
1836 void PSParallelCompact::summary_phase(ParCompactionManager* cm,
1837 bool maximum_compaction)
1838 {
1839 TraceTime tm("summary phase", print_phases(), true, gclog_or_tty);
1840 // trace("2");
1842 #ifdef ASSERT
1843 if (TraceParallelOldGCMarkingPhase) {
1844 tty->print_cr("add_obj_count=" SIZE_FORMAT " "
1845 "add_obj_bytes=" SIZE_FORMAT,
1846 add_obj_count, add_obj_size * HeapWordSize);
1847 tty->print_cr("mark_bitmap_count=" SIZE_FORMAT " "
1848 "mark_bitmap_bytes=" SIZE_FORMAT,
1849 mark_bitmap_count, mark_bitmap_size * HeapWordSize);
1850 }
1851 #endif // #ifdef ASSERT
1853 // Quick summarization of each space into itself, to see how much is live.
1854 summarize_spaces_quick();
1856 if (TraceParallelOldGCSummaryPhase) {
1857 tty->print_cr("summary_phase: after summarizing each space to self");
1858 Universe::print();
1859 NOT_PRODUCT(print_region_ranges());
1860 if (Verbose) {
1861 NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info));
1862 }
1863 }
1865 // The amount of live data that will end up in old space (assuming it fits).
1866 size_t old_space_total_live = 0;
1867 assert(perm_space_id < old_space_id, "should not count perm data here");
1868 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
1869 old_space_total_live += pointer_delta(_space_info[id].new_top(),
1870 _space_info[id].space()->bottom());
1871 }
1873 MutableSpace* const old_space = _space_info[old_space_id].space();
1874 const size_t old_capacity = old_space->capacity_in_words();
1875 if (old_space_total_live > old_capacity) {
1876 // XXX - should also try to expand
1877 maximum_compaction = true;
1878 }
1879 #ifndef PRODUCT
1880 if (ParallelOldGCSplitALot && old_space_total_live < old_capacity) {
1881 provoke_split(maximum_compaction);
1882 }
1883 #endif // #ifndef PRODUCT
1885 // Permanent and Old generations.
1886 summarize_space(perm_space_id, maximum_compaction);
1887 summarize_space(old_space_id, maximum_compaction);
1889 // Summarize the remaining spaces in the young gen. The initial target space
1890 // is the old gen. If a space does not fit entirely into the target, then the
1891 // remainder is compacted into the space itself and that space becomes the new
1892 // target.
1893 SpaceId dst_space_id = old_space_id;
1894 HeapWord* dst_space_end = old_space->end();
1895 HeapWord** new_top_addr = _space_info[dst_space_id].new_top_addr();
1896 for (unsigned int id = eden_space_id; id < last_space_id; ++id) {
1897 const MutableSpace* space = _space_info[id].space();
1898 const size_t live = pointer_delta(_space_info[id].new_top(),
1899 space->bottom());
1900 const size_t available = pointer_delta(dst_space_end, *new_top_addr);
1902 NOT_PRODUCT(summary_phase_msg(dst_space_id, *new_top_addr, dst_space_end,
1903 SpaceId(id), space->bottom(), space->top());)
1904 if (live > 0 && live <= available) {
1905 // All the live data will fit.
1906 bool done = _summary_data.summarize(_space_info[id].split_info(),
1907 space->bottom(), space->top(),
1908 NULL,
1909 *new_top_addr, dst_space_end,
1910 new_top_addr);
1911 assert(done, "space must fit into old gen");
1913 // Reset the new_top value for the space.
1914 _space_info[id].set_new_top(space->bottom());
1915 } else if (live > 0) {
1916 // Attempt to fit part of the source space into the target space.
1917 HeapWord* next_src_addr = NULL;
1918 bool done = _summary_data.summarize(_space_info[id].split_info(),
1919 space->bottom(), space->top(),
1920 &next_src_addr,
1921 *new_top_addr, dst_space_end,
1922 new_top_addr);
1923 assert(!done, "space should not fit into old gen");
1924 assert(next_src_addr != NULL, "sanity");
1926 // The source space becomes the new target, so the remainder is compacted
1927 // within the space itself.
1928 dst_space_id = SpaceId(id);
1929 dst_space_end = space->end();
1930 new_top_addr = _space_info[id].new_top_addr();
1931 NOT_PRODUCT(summary_phase_msg(dst_space_id,
1932 space->bottom(), dst_space_end,
1933 SpaceId(id), next_src_addr, space->top());)
1934 done = _summary_data.summarize(_space_info[id].split_info(),
1935 next_src_addr, space->top(),
1936 NULL,
1937 space->bottom(), dst_space_end,
1938 new_top_addr);
1939 assert(done, "space must fit when compacted into itself");
1940 assert(*new_top_addr <= space->top(), "usage should not grow");
1941 }
1942 }
1944 if (TraceParallelOldGCSummaryPhase) {
1945 tty->print_cr("summary_phase: after final summarization");
1946 Universe::print();
1947 NOT_PRODUCT(print_region_ranges());
1948 if (Verbose) {
1949 NOT_PRODUCT(print_generic_summary_data(_summary_data, _space_info));
1950 }
1951 }
1952 }
1954 // This method should contain all heap-specific policy for invoking a full
1955 // collection. invoke_no_policy() will only attempt to compact the heap; it
1956 // will do nothing further. If we need to bail out for policy reasons, scavenge
1957 // before full gc, or any other specialized behavior, it needs to be added here.
1958 //
1959 // Note that this method should only be called from the vm_thread while at a
1960 // safepoint.
1961 //
1962 // Note that the all_soft_refs_clear flag in the collector policy
1963 // may be true because this method can be called without intervening
1964 // activity. For example when the heap space is tight and full measure
1965 // are being taken to free space.
1966 void PSParallelCompact::invoke(bool maximum_heap_compaction) {
1967 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
1968 assert(Thread::current() == (Thread*)VMThread::vm_thread(),
1969 "should be in vm thread");
1971 ParallelScavengeHeap* heap = gc_heap();
1972 GCCause::Cause gc_cause = heap->gc_cause();
1973 assert(!heap->is_gc_active(), "not reentrant");
1975 PSAdaptiveSizePolicy* policy = heap->size_policy();
1976 IsGCActiveMark mark;
1978 if (ScavengeBeforeFullGC) {
1979 PSScavenge::invoke_no_policy();
1980 }
1982 const bool clear_all_soft_refs =
1983 heap->collector_policy()->should_clear_all_soft_refs();
1985 PSParallelCompact::invoke_no_policy(clear_all_soft_refs ||
1986 maximum_heap_compaction);
1987 }
1989 bool ParallelCompactData::region_contains(size_t region_index, HeapWord* addr) {
1990 size_t addr_region_index = addr_to_region_idx(addr);
1991 return region_index == addr_region_index;
1992 }
1994 // This method contains no policy. You should probably
1995 // be calling invoke() instead.
1996 bool PSParallelCompact::invoke_no_policy(bool maximum_heap_compaction) {
1997 assert(SafepointSynchronize::is_at_safepoint(), "must be at a safepoint");
1998 assert(ref_processor() != NULL, "Sanity");
2000 if (GC_locker::check_active_before_gc()) {
2001 return false;
2002 }
2004 TimeStamp marking_start;
2005 TimeStamp compaction_start;
2006 TimeStamp collection_exit;
2008 ParallelScavengeHeap* heap = gc_heap();
2009 GCCause::Cause gc_cause = heap->gc_cause();
2010 PSYoungGen* young_gen = heap->young_gen();
2011 PSOldGen* old_gen = heap->old_gen();
2012 PSPermGen* perm_gen = heap->perm_gen();
2013 PSAdaptiveSizePolicy* size_policy = heap->size_policy();
2015 // The scope of casr should end after code that can change
2016 // CollectorPolicy::_should_clear_all_soft_refs.
2017 ClearedAllSoftRefs casr(maximum_heap_compaction,
2018 heap->collector_policy());
2020 if (ZapUnusedHeapArea) {
2021 // Save information needed to minimize mangling
2022 heap->record_gen_tops_before_GC();
2023 }
2025 heap->pre_full_gc_dump();
2027 _print_phases = PrintGCDetails && PrintParallelOldGCPhaseTimes;
2029 // Make sure data structures are sane, make the heap parsable, and do other
2030 // miscellaneous bookkeeping.
2031 PreGCValues pre_gc_values;
2032 pre_compact(&pre_gc_values);
2034 // Get the compaction manager reserved for the VM thread.
2035 ParCompactionManager* const vmthread_cm =
2036 ParCompactionManager::manager_array(gc_task_manager()->workers());
2038 // Place after pre_compact() where the number of invocations is incremented.
2039 AdaptiveSizePolicyOutput(size_policy, heap->total_collections());
2041 {
2042 ResourceMark rm;
2043 HandleMark hm;
2045 // Set the number of GC threads to be used in this collection
2046 gc_task_manager()->set_active_gang();
2047 gc_task_manager()->task_idle_workers();
2048 heap->set_par_threads(gc_task_manager()->active_workers());
2050 const bool is_system_gc = gc_cause == GCCause::_java_lang_system_gc;
2052 // This is useful for debugging but don't change the output the
2053 // the customer sees.
2054 const char* gc_cause_str = "Full GC";
2055 if (is_system_gc && PrintGCDetails) {
2056 gc_cause_str = "Full GC (System)";
2057 }
2058 gclog_or_tty->date_stamp(PrintGC && PrintGCDateStamps);
2059 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
2060 TraceTime t1(gc_cause_str, PrintGC, !PrintGCDetails, gclog_or_tty);
2061 TraceCollectorStats tcs(counters());
2062 TraceMemoryManagerStats tms(true /* Full GC */,gc_cause);
2064 if (TraceGen1Time) accumulated_time()->start();
2066 // Let the size policy know we're starting
2067 size_policy->major_collection_begin();
2069 // When collecting the permanent generation methodOops may be moving,
2070 // so we either have to flush all bcp data or convert it into bci.
2071 CodeCache::gc_prologue();
2072 Threads::gc_prologue();
2074 COMPILER2_PRESENT(DerivedPointerTable::clear());
2076 ref_processor()->enable_discovery(true /*verify_disabled*/, true /*verify_no_refs*/);
2077 ref_processor()->setup_policy(maximum_heap_compaction);
2079 bool marked_for_unloading = false;
2081 marking_start.update();
2082 marking_phase(vmthread_cm, maximum_heap_compaction);
2084 #ifndef PRODUCT
2085 if (TraceParallelOldGCMarkingPhase) {
2086 gclog_or_tty->print_cr("marking_phase: cas_tries %d cas_retries %d "
2087 "cas_by_another %d",
2088 mark_bitmap()->cas_tries(), mark_bitmap()->cas_retries(),
2089 mark_bitmap()->cas_by_another());
2090 }
2091 #endif // #ifndef PRODUCT
2093 bool max_on_system_gc = UseMaximumCompactionOnSystemGC && is_system_gc;
2094 summary_phase(vmthread_cm, maximum_heap_compaction || max_on_system_gc);
2096 COMPILER2_PRESENT(assert(DerivedPointerTable::is_active(), "Sanity"));
2097 COMPILER2_PRESENT(DerivedPointerTable::set_active(false));
2099 // adjust_roots() updates Universe::_intArrayKlassObj which is
2100 // needed by the compaction for filling holes in the dense prefix.
2101 adjust_roots();
2103 compaction_start.update();
2104 // Does the perm gen always have to be done serially because
2105 // klasses are used in the update of an object?
2106 compact_perm(vmthread_cm);
2108 compact();
2110 // Reset the mark bitmap, summary data, and do other bookkeeping. Must be
2111 // done before resizing.
2112 post_compact();
2114 // Let the size policy know we're done
2115 size_policy->major_collection_end(old_gen->used_in_bytes(), gc_cause);
2117 if (UseAdaptiveSizePolicy) {
2118 if (PrintAdaptiveSizePolicy) {
2119 gclog_or_tty->print("AdaptiveSizeStart: ");
2120 gclog_or_tty->stamp();
2121 gclog_or_tty->print_cr(" collection: %d ",
2122 heap->total_collections());
2123 if (Verbose) {
2124 gclog_or_tty->print("old_gen_capacity: %d young_gen_capacity: %d"
2125 " perm_gen_capacity: %d ",
2126 old_gen->capacity_in_bytes(), young_gen->capacity_in_bytes(),
2127 perm_gen->capacity_in_bytes());
2128 }
2129 }
2131 // Don't check if the size_policy is ready here. Let
2132 // the size_policy check that internally.
2133 if (UseAdaptiveGenerationSizePolicyAtMajorCollection &&
2134 ((gc_cause != GCCause::_java_lang_system_gc) ||
2135 UseAdaptiveSizePolicyWithSystemGC)) {
2136 // Calculate optimal free space amounts
2137 assert(young_gen->max_size() >
2138 young_gen->from_space()->capacity_in_bytes() +
2139 young_gen->to_space()->capacity_in_bytes(),
2140 "Sizes of space in young gen are out-of-bounds");
2141 size_t max_eden_size = young_gen->max_size() -
2142 young_gen->from_space()->capacity_in_bytes() -
2143 young_gen->to_space()->capacity_in_bytes();
2144 size_policy->compute_generation_free_space(
2145 young_gen->used_in_bytes(),
2146 young_gen->eden_space()->used_in_bytes(),
2147 old_gen->used_in_bytes(),
2148 perm_gen->used_in_bytes(),
2149 young_gen->eden_space()->capacity_in_bytes(),
2150 old_gen->max_gen_size(),
2151 max_eden_size,
2152 true /* full gc*/,
2153 gc_cause,
2154 heap->collector_policy());
2156 heap->resize_old_gen(
2157 size_policy->calculated_old_free_size_in_bytes());
2159 // Don't resize the young generation at an major collection. A
2160 // desired young generation size may have been calculated but
2161 // resizing the young generation complicates the code because the
2162 // resizing of the old generation may have moved the boundary
2163 // between the young generation and the old generation. Let the
2164 // young generation resizing happen at the minor collections.
2165 }
2166 if (PrintAdaptiveSizePolicy) {
2167 gclog_or_tty->print_cr("AdaptiveSizeStop: collection: %d ",
2168 heap->total_collections());
2169 }
2170 }
2172 if (UsePerfData) {
2173 PSGCAdaptivePolicyCounters* const counters = heap->gc_policy_counters();
2174 counters->update_counters();
2175 counters->update_old_capacity(old_gen->capacity_in_bytes());
2176 counters->update_young_capacity(young_gen->capacity_in_bytes());
2177 }
2179 heap->resize_all_tlabs();
2181 // We collected the perm gen, so we'll resize it here.
2182 perm_gen->compute_new_size(pre_gc_values.perm_gen_used());
2184 if (TraceGen1Time) accumulated_time()->stop();
2186 if (PrintGC) {
2187 if (PrintGCDetails) {
2188 // No GC timestamp here. This is after GC so it would be confusing.
2189 young_gen->print_used_change(pre_gc_values.young_gen_used());
2190 old_gen->print_used_change(pre_gc_values.old_gen_used());
2191 heap->print_heap_change(pre_gc_values.heap_used());
2192 // Print perm gen last (print_heap_change() excludes the perm gen).
2193 perm_gen->print_used_change(pre_gc_values.perm_gen_used());
2194 } else {
2195 heap->print_heap_change(pre_gc_values.heap_used());
2196 }
2197 }
2199 // Track memory usage and detect low memory
2200 MemoryService::track_memory_usage();
2201 heap->update_counters();
2202 gc_task_manager()->release_idle_workers();
2203 }
2205 #ifdef ASSERT
2206 for (size_t i = 0; i < ParallelGCThreads + 1; ++i) {
2207 ParCompactionManager* const cm =
2208 ParCompactionManager::manager_array(int(i));
2209 assert(cm->marking_stack()->is_empty(), "should be empty");
2210 assert(ParCompactionManager::region_list(int(i))->is_empty(), "should be empty");
2211 assert(cm->revisit_klass_stack()->is_empty(), "should be empty");
2212 }
2213 #endif // ASSERT
2215 if (VerifyAfterGC && heap->total_collections() >= VerifyGCStartAt) {
2216 HandleMark hm; // Discard invalid handles created during verification
2217 gclog_or_tty->print(" VerifyAfterGC:");
2218 Universe::verify(false);
2219 }
2221 // Re-verify object start arrays
2222 if (VerifyObjectStartArray &&
2223 VerifyAfterGC) {
2224 old_gen->verify_object_start_array();
2225 perm_gen->verify_object_start_array();
2226 }
2228 if (ZapUnusedHeapArea) {
2229 old_gen->object_space()->check_mangled_unused_area_complete();
2230 perm_gen->object_space()->check_mangled_unused_area_complete();
2231 }
2233 NOT_PRODUCT(ref_processor()->verify_no_references_recorded());
2235 collection_exit.update();
2237 heap->print_heap_after_gc();
2238 if (PrintGCTaskTimeStamps) {
2239 gclog_or_tty->print_cr("VM-Thread " INT64_FORMAT " " INT64_FORMAT " "
2240 INT64_FORMAT,
2241 marking_start.ticks(), compaction_start.ticks(),
2242 collection_exit.ticks());
2243 gc_task_manager()->print_task_time_stamps();
2244 }
2246 heap->post_full_gc_dump();
2248 #ifdef TRACESPINNING
2249 ParallelTaskTerminator::print_termination_counts();
2250 #endif
2252 return true;
2253 }
2255 bool PSParallelCompact::absorb_live_data_from_eden(PSAdaptiveSizePolicy* size_policy,
2256 PSYoungGen* young_gen,
2257 PSOldGen* old_gen) {
2258 MutableSpace* const eden_space = young_gen->eden_space();
2259 assert(!eden_space->is_empty(), "eden must be non-empty");
2260 assert(young_gen->virtual_space()->alignment() ==
2261 old_gen->virtual_space()->alignment(), "alignments do not match");
2263 if (!(UseAdaptiveSizePolicy && UseAdaptiveGCBoundary)) {
2264 return false;
2265 }
2267 // Both generations must be completely committed.
2268 if (young_gen->virtual_space()->uncommitted_size() != 0) {
2269 return false;
2270 }
2271 if (old_gen->virtual_space()->uncommitted_size() != 0) {
2272 return false;
2273 }
2275 // Figure out how much to take from eden. Include the average amount promoted
2276 // in the total; otherwise the next young gen GC will simply bail out to a
2277 // full GC.
2278 const size_t alignment = old_gen->virtual_space()->alignment();
2279 const size_t eden_used = eden_space->used_in_bytes();
2280 const size_t promoted = (size_t)size_policy->avg_promoted()->padded_average();
2281 const size_t absorb_size = align_size_up(eden_used + promoted, alignment);
2282 const size_t eden_capacity = eden_space->capacity_in_bytes();
2284 if (absorb_size >= eden_capacity) {
2285 return false; // Must leave some space in eden.
2286 }
2288 const size_t new_young_size = young_gen->capacity_in_bytes() - absorb_size;
2289 if (new_young_size < young_gen->min_gen_size()) {
2290 return false; // Respect young gen minimum size.
2291 }
2293 if (TraceAdaptiveGCBoundary && Verbose) {
2294 gclog_or_tty->print(" absorbing " SIZE_FORMAT "K: "
2295 "eden " SIZE_FORMAT "K->" SIZE_FORMAT "K "
2296 "from " SIZE_FORMAT "K, to " SIZE_FORMAT "K "
2297 "young_gen " SIZE_FORMAT "K->" SIZE_FORMAT "K ",
2298 absorb_size / K,
2299 eden_capacity / K, (eden_capacity - absorb_size) / K,
2300 young_gen->from_space()->used_in_bytes() / K,
2301 young_gen->to_space()->used_in_bytes() / K,
2302 young_gen->capacity_in_bytes() / K, new_young_size / K);
2303 }
2305 // Fill the unused part of the old gen.
2306 MutableSpace* const old_space = old_gen->object_space();
2307 HeapWord* const unused_start = old_space->top();
2308 size_t const unused_words = pointer_delta(old_space->end(), unused_start);
2310 if (unused_words > 0) {
2311 if (unused_words < CollectedHeap::min_fill_size()) {
2312 return false; // If the old gen cannot be filled, must give up.
2313 }
2314 CollectedHeap::fill_with_objects(unused_start, unused_words);
2315 }
2317 // Take the live data from eden and set both top and end in the old gen to
2318 // eden top. (Need to set end because reset_after_change() mangles the region
2319 // from end to virtual_space->high() in debug builds).
2320 HeapWord* const new_top = eden_space->top();
2321 old_gen->virtual_space()->expand_into(young_gen->virtual_space(),
2322 absorb_size);
2323 young_gen->reset_after_change();
2324 old_space->set_top(new_top);
2325 old_space->set_end(new_top);
2326 old_gen->reset_after_change();
2328 // Update the object start array for the filler object and the data from eden.
2329 ObjectStartArray* const start_array = old_gen->start_array();
2330 for (HeapWord* p = unused_start; p < new_top; p += oop(p)->size()) {
2331 start_array->allocate_block(p);
2332 }
2334 // Could update the promoted average here, but it is not typically updated at
2335 // full GCs and the value to use is unclear. Something like
2336 //
2337 // cur_promoted_avg + absorb_size / number_of_scavenges_since_last_full_gc.
2339 size_policy->set_bytes_absorbed_from_eden(absorb_size);
2340 return true;
2341 }
2343 GCTaskManager* const PSParallelCompact::gc_task_manager() {
2344 assert(ParallelScavengeHeap::gc_task_manager() != NULL,
2345 "shouldn't return NULL");
2346 return ParallelScavengeHeap::gc_task_manager();
2347 }
2349 void PSParallelCompact::marking_phase(ParCompactionManager* cm,
2350 bool maximum_heap_compaction) {
2351 // Recursively traverse all live objects and mark them
2352 TraceTime tm("marking phase", print_phases(), true, gclog_or_tty);
2354 ParallelScavengeHeap* heap = gc_heap();
2355 uint parallel_gc_threads = heap->gc_task_manager()->workers();
2356 uint active_gc_threads = heap->gc_task_manager()->active_workers();
2357 TaskQueueSetSuper* qset = ParCompactionManager::region_array();
2358 ParallelTaskTerminator terminator(active_gc_threads, qset);
2360 PSParallelCompact::MarkAndPushClosure mark_and_push_closure(cm);
2361 PSParallelCompact::FollowStackClosure follow_stack_closure(cm);
2363 {
2364 TraceTime tm_m("par mark", print_phases(), true, gclog_or_tty);
2365 ParallelScavengeHeap::ParStrongRootsScope psrs;
2367 GCTaskQueue* q = GCTaskQueue::create();
2369 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::universe));
2370 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jni_handles));
2371 // We scan the thread roots in parallel
2372 Threads::create_thread_roots_marking_tasks(q);
2373 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::object_synchronizer));
2374 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::flat_profiler));
2375 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::management));
2376 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::system_dictionary));
2377 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jvmti));
2378 q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::code_cache));
2380 if (active_gc_threads > 1) {
2381 for (uint j = 0; j < active_gc_threads; j++) {
2382 q->enqueue(new StealMarkingTask(&terminator));
2383 }
2384 }
2386 gc_task_manager()->execute_and_wait(q);
2387 }
2389 // Process reference objects found during marking
2390 {
2391 TraceTime tm_r("reference processing", print_phases(), true, gclog_or_tty);
2392 if (ref_processor()->processing_is_mt()) {
2393 RefProcTaskExecutor task_executor;
2394 ref_processor()->process_discovered_references(
2395 is_alive_closure(), &mark_and_push_closure, &follow_stack_closure,
2396 &task_executor);
2397 } else {
2398 ref_processor()->process_discovered_references(
2399 is_alive_closure(), &mark_and_push_closure, &follow_stack_closure, NULL);
2400 }
2401 }
2403 TraceTime tm_c("class unloading", print_phases(), true, gclog_or_tty);
2404 // Follow system dictionary roots and unload classes.
2405 bool purged_class = SystemDictionary::do_unloading(is_alive_closure());
2407 // Follow code cache roots.
2408 CodeCache::do_unloading(is_alive_closure(), &mark_and_push_closure,
2409 purged_class);
2410 cm->follow_marking_stacks(); // Flush marking stack.
2412 // Update subklass/sibling/implementor links of live klasses
2413 // revisit_klass_stack is used in follow_weak_klass_links().
2414 follow_weak_klass_links();
2416 // Revisit memoized MDO's and clear any unmarked weak refs
2417 follow_mdo_weak_refs();
2419 // Visit interned string tables and delete unmarked oops
2420 StringTable::unlink(is_alive_closure());
2421 // Clean up unreferenced symbols in symbol table.
2422 SymbolTable::unlink();
2424 assert(cm->marking_stacks_empty(), "marking stacks should be empty");
2425 }
2427 // This should be moved to the shared markSweep code!
2428 class PSAlwaysTrueClosure: public BoolObjectClosure {
2429 public:
2430 void do_object(oop p) { ShouldNotReachHere(); }
2431 bool do_object_b(oop p) { return true; }
2432 };
2433 static PSAlwaysTrueClosure always_true;
2435 void PSParallelCompact::adjust_roots() {
2436 // Adjust the pointers to reflect the new locations
2437 TraceTime tm("adjust roots", print_phases(), true, gclog_or_tty);
2439 // General strong roots.
2440 Universe::oops_do(adjust_root_pointer_closure());
2441 JNIHandles::oops_do(adjust_root_pointer_closure()); // Global (strong) JNI handles
2442 Threads::oops_do(adjust_root_pointer_closure(), NULL);
2443 ObjectSynchronizer::oops_do(adjust_root_pointer_closure());
2444 FlatProfiler::oops_do(adjust_root_pointer_closure());
2445 Management::oops_do(adjust_root_pointer_closure());
2446 JvmtiExport::oops_do(adjust_root_pointer_closure());
2447 // SO_AllClasses
2448 SystemDictionary::oops_do(adjust_root_pointer_closure());
2450 // Now adjust pointers in remaining weak roots. (All of which should
2451 // have been cleared if they pointed to non-surviving objects.)
2452 // Global (weak) JNI handles
2453 JNIHandles::weak_oops_do(&always_true, adjust_root_pointer_closure());
2455 CodeCache::oops_do(adjust_pointer_closure());
2456 StringTable::oops_do(adjust_root_pointer_closure());
2457 ref_processor()->weak_oops_do(adjust_root_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(
2463 adjust_root_pointer_closure());
2464 }
2466 void PSParallelCompact::compact_perm(ParCompactionManager* cm) {
2467 TraceTime tm("compact perm gen", print_phases(), true, gclog_or_tty);
2468 // trace("4");
2470 gc_heap()->perm_gen()->start_array()->reset();
2471 move_and_update(cm, perm_space_id);
2472 }
2474 void PSParallelCompact::enqueue_region_draining_tasks(GCTaskQueue* q,
2475 uint parallel_gc_threads)
2476 {
2477 TraceTime tm("drain task setup", print_phases(), true, gclog_or_tty);
2479 // Find the threads that are active
2480 unsigned int which = 0;
2482 const uint task_count = MAX2(parallel_gc_threads, 1U);
2483 for (uint j = 0; j < task_count; j++) {
2484 q->enqueue(new DrainStacksCompactionTask(j));
2485 ParCompactionManager::verify_region_list_empty(j);
2486 // Set the region stacks variables to "no" region stack values
2487 // so that they will be recognized and needing a region stack
2488 // in the stealing tasks if they do not get one by executing
2489 // a draining stack.
2490 ParCompactionManager* cm = ParCompactionManager::manager_array(j);
2491 cm->set_region_stack(NULL);
2492 cm->set_region_stack_index((uint)max_uintx);
2493 }
2494 ParCompactionManager::reset_recycled_stack_index();
2496 // Find all regions that are available (can be filled immediately) and
2497 // distribute them to the thread stacks. The iteration is done in reverse
2498 // order (high to low) so the regions will be removed in ascending order.
2500 const ParallelCompactData& sd = PSParallelCompact::summary_data();
2502 size_t fillable_regions = 0; // A count for diagnostic purposes.
2503 // A region index which corresponds to the tasks created above.
2504 // "which" must be 0 <= which < task_count
2506 which = 0;
2507 for (unsigned int id = to_space_id; id > perm_space_id; --id) {
2508 SpaceInfo* const space_info = _space_info + id;
2509 MutableSpace* const space = space_info->space();
2510 HeapWord* const new_top = space_info->new_top();
2512 const size_t beg_region = sd.addr_to_region_idx(space_info->dense_prefix());
2513 const size_t end_region =
2514 sd.addr_to_region_idx(sd.region_align_up(new_top));
2515 assert(end_region > 0, "perm gen cannot be empty");
2517 for (size_t cur = end_region - 1; cur >= beg_region; --cur) {
2518 if (sd.region(cur)->claim_unsafe()) {
2519 ParCompactionManager::region_list_push(which, cur);
2521 if (TraceParallelOldGCCompactionPhase && Verbose) {
2522 const size_t count_mod_8 = fillable_regions & 7;
2523 if (count_mod_8 == 0) gclog_or_tty->print("fillable: ");
2524 gclog_or_tty->print(" " SIZE_FORMAT_W(7), cur);
2525 if (count_mod_8 == 7) gclog_or_tty->cr();
2526 }
2528 NOT_PRODUCT(++fillable_regions;)
2530 // Assign regions to tasks in round-robin fashion.
2531 if (++which == task_count) {
2532 assert(which <= parallel_gc_threads,
2533 "Inconsistent number of workers");
2534 which = 0;
2535 }
2536 }
2537 }
2538 }
2540 if (TraceParallelOldGCCompactionPhase) {
2541 if (Verbose && (fillable_regions & 7) != 0) gclog_or_tty->cr();
2542 gclog_or_tty->print_cr("%u initially fillable regions", fillable_regions);
2543 }
2544 }
2546 #define PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING 4
2548 void PSParallelCompact::enqueue_dense_prefix_tasks(GCTaskQueue* q,
2549 uint parallel_gc_threads) {
2550 TraceTime tm("dense prefix task setup", print_phases(), true, gclog_or_tty);
2552 ParallelCompactData& sd = PSParallelCompact::summary_data();
2554 // Iterate over all the spaces adding tasks for updating
2555 // regions in the dense prefix. Assume that 1 gc thread
2556 // will work on opening the gaps and the remaining gc threads
2557 // will work on the dense prefix.
2558 unsigned int space_id;
2559 for (space_id = old_space_id; space_id < last_space_id; ++ space_id) {
2560 HeapWord* const dense_prefix_end = _space_info[space_id].dense_prefix();
2561 const MutableSpace* const space = _space_info[space_id].space();
2563 if (dense_prefix_end == space->bottom()) {
2564 // There is no dense prefix for this space.
2565 continue;
2566 }
2568 // The dense prefix is before this region.
2569 size_t region_index_end_dense_prefix =
2570 sd.addr_to_region_idx(dense_prefix_end);
2571 RegionData* const dense_prefix_cp =
2572 sd.region(region_index_end_dense_prefix);
2573 assert(dense_prefix_end == space->end() ||
2574 dense_prefix_cp->available() ||
2575 dense_prefix_cp->claimed(),
2576 "The region after the dense prefix should always be ready to fill");
2578 size_t region_index_start = sd.addr_to_region_idx(space->bottom());
2580 // Is there dense prefix work?
2581 size_t total_dense_prefix_regions =
2582 region_index_end_dense_prefix - region_index_start;
2583 // How many regions of the dense prefix should be given to
2584 // each thread?
2585 if (total_dense_prefix_regions > 0) {
2586 uint tasks_for_dense_prefix = 1;
2587 if (total_dense_prefix_regions <=
2588 (parallel_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING)) {
2589 // Don't over partition. This assumes that
2590 // PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING is a small integer value
2591 // so there are not many regions to process.
2592 tasks_for_dense_prefix = parallel_gc_threads;
2593 } else {
2594 // Over partition
2595 tasks_for_dense_prefix = parallel_gc_threads *
2596 PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING;
2597 }
2598 size_t regions_per_thread = total_dense_prefix_regions /
2599 tasks_for_dense_prefix;
2600 // Give each thread at least 1 region.
2601 if (regions_per_thread == 0) {
2602 regions_per_thread = 1;
2603 }
2605 for (uint k = 0; k < tasks_for_dense_prefix; k++) {
2606 if (region_index_start >= region_index_end_dense_prefix) {
2607 break;
2608 }
2609 // region_index_end is not processed
2610 size_t region_index_end = MIN2(region_index_start + regions_per_thread,
2611 region_index_end_dense_prefix);
2612 q->enqueue(new UpdateDensePrefixTask(SpaceId(space_id),
2613 region_index_start,
2614 region_index_end));
2615 region_index_start = region_index_end;
2616 }
2617 }
2618 // This gets any part of the dense prefix that did not
2619 // fit evenly.
2620 if (region_index_start < region_index_end_dense_prefix) {
2621 q->enqueue(new UpdateDensePrefixTask(SpaceId(space_id),
2622 region_index_start,
2623 region_index_end_dense_prefix));
2624 }
2625 }
2626 }
2628 void PSParallelCompact::enqueue_region_stealing_tasks(
2629 GCTaskQueue* q,
2630 ParallelTaskTerminator* terminator_ptr,
2631 uint parallel_gc_threads) {
2632 TraceTime tm("steal task setup", print_phases(), true, gclog_or_tty);
2634 // Once a thread has drained it's stack, it should try to steal regions from
2635 // other threads.
2636 if (parallel_gc_threads > 1) {
2637 for (uint j = 0; j < parallel_gc_threads; j++) {
2638 q->enqueue(new StealRegionCompactionTask(terminator_ptr));
2639 }
2640 }
2641 }
2643 void PSParallelCompact::compact() {
2644 // trace("5");
2645 TraceTime tm("compaction phase", print_phases(), true, gclog_or_tty);
2647 ParallelScavengeHeap* heap = (ParallelScavengeHeap*)Universe::heap();
2648 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity");
2649 PSOldGen* old_gen = heap->old_gen();
2650 old_gen->start_array()->reset();
2651 uint parallel_gc_threads = heap->gc_task_manager()->workers();
2652 uint active_gc_threads = heap->gc_task_manager()->active_workers();
2653 TaskQueueSetSuper* qset = ParCompactionManager::region_array();
2654 ParallelTaskTerminator terminator(active_gc_threads, qset);
2656 GCTaskQueue* q = GCTaskQueue::create();
2657 enqueue_region_draining_tasks(q, active_gc_threads);
2658 enqueue_dense_prefix_tasks(q, active_gc_threads);
2659 enqueue_region_stealing_tasks(q, &terminator, active_gc_threads);
2661 {
2662 TraceTime tm_pc("par compact", print_phases(), true, gclog_or_tty);
2664 gc_task_manager()->execute_and_wait(q);
2666 #ifdef ASSERT
2667 // Verify that all regions have been processed before the deferred updates.
2668 // Note that perm_space_id is skipped; this type of verification is not
2669 // valid until the perm gen is compacted by regions.
2670 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2671 verify_complete(SpaceId(id));
2672 }
2673 #endif
2674 }
2676 {
2677 // Update the deferred objects, if any. Any compaction manager can be used.
2678 TraceTime tm_du("deferred updates", print_phases(), true, gclog_or_tty);
2679 ParCompactionManager* cm = ParCompactionManager::manager_array(0);
2680 for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2681 update_deferred_objects(cm, SpaceId(id));
2682 }
2683 }
2684 }
2686 #ifdef ASSERT
2687 void PSParallelCompact::verify_complete(SpaceId space_id) {
2688 // All Regions between space bottom() to new_top() should be marked as filled
2689 // and all Regions between new_top() and top() should be available (i.e.,
2690 // should have been emptied).
2691 ParallelCompactData& sd = summary_data();
2692 SpaceInfo si = _space_info[space_id];
2693 HeapWord* new_top_addr = sd.region_align_up(si.new_top());
2694 HeapWord* old_top_addr = sd.region_align_up(si.space()->top());
2695 const size_t beg_region = sd.addr_to_region_idx(si.space()->bottom());
2696 const size_t new_top_region = sd.addr_to_region_idx(new_top_addr);
2697 const size_t old_top_region = sd.addr_to_region_idx(old_top_addr);
2699 bool issued_a_warning = false;
2701 size_t cur_region;
2702 for (cur_region = beg_region; cur_region < new_top_region; ++cur_region) {
2703 const RegionData* const c = sd.region(cur_region);
2704 if (!c->completed()) {
2705 warning("region " SIZE_FORMAT " not filled: "
2706 "destination_count=" SIZE_FORMAT,
2707 cur_region, c->destination_count());
2708 issued_a_warning = true;
2709 }
2710 }
2712 for (cur_region = new_top_region; cur_region < old_top_region; ++cur_region) {
2713 const RegionData* const c = sd.region(cur_region);
2714 if (!c->available()) {
2715 warning("region " SIZE_FORMAT " not empty: "
2716 "destination_count=" SIZE_FORMAT,
2717 cur_region, c->destination_count());
2718 issued_a_warning = true;
2719 }
2720 }
2722 if (issued_a_warning) {
2723 print_region_ranges();
2724 }
2725 }
2726 #endif // #ifdef ASSERT
2728 void
2729 PSParallelCompact::follow_weak_klass_links() {
2730 // All klasses on the revisit stack are marked at this point.
2731 // Update and follow all subklass, sibling and implementor links.
2732 // Check all the stacks here even if not all the workers are active.
2733 // There is no accounting which indicates which stacks might have
2734 // contents to be followed.
2735 if (PrintRevisitStats) {
2736 gclog_or_tty->print_cr("#classes in system dictionary = %d",
2737 SystemDictionary::number_of_classes());
2738 }
2739 for (uint i = 0; i < ParallelGCThreads + 1; i++) {
2740 ParCompactionManager* cm = ParCompactionManager::manager_array(i);
2741 KeepAliveClosure keep_alive_closure(cm);
2742 Stack<Klass*>* const rks = cm->revisit_klass_stack();
2743 if (PrintRevisitStats) {
2744 gclog_or_tty->print_cr("Revisit klass stack[%u] length = " SIZE_FORMAT,
2745 i, rks->size());
2746 }
2747 while (!rks->is_empty()) {
2748 Klass* const k = rks->pop();
2749 k->follow_weak_klass_links(is_alive_closure(), &keep_alive_closure);
2750 }
2752 cm->follow_marking_stacks();
2753 }
2754 }
2756 void
2757 PSParallelCompact::revisit_weak_klass_link(ParCompactionManager* cm, Klass* k) {
2758 cm->revisit_klass_stack()->push(k);
2759 }
2761 void PSParallelCompact::revisit_mdo(ParCompactionManager* cm, DataLayout* p) {
2762 cm->revisit_mdo_stack()->push(p);
2763 }
2765 void PSParallelCompact::follow_mdo_weak_refs() {
2766 // All strongly reachable oops have been marked at this point;
2767 // we can visit and clear any weak references from MDO's which
2768 // we memoized during the strong marking phase.
2769 if (PrintRevisitStats) {
2770 gclog_or_tty->print_cr("#classes in system dictionary = %d",
2771 SystemDictionary::number_of_classes());
2772 }
2773 for (uint i = 0; i < ParallelGCThreads + 1; i++) {
2774 ParCompactionManager* cm = ParCompactionManager::manager_array(i);
2775 Stack<DataLayout*>* rms = cm->revisit_mdo_stack();
2776 if (PrintRevisitStats) {
2777 gclog_or_tty->print_cr("Revisit MDO stack[%u] size = " SIZE_FORMAT,
2778 i, rms->size());
2779 }
2780 while (!rms->is_empty()) {
2781 rms->pop()->follow_weak_refs(is_alive_closure());
2782 }
2784 cm->follow_marking_stacks();
2785 }
2786 }
2789 #ifdef VALIDATE_MARK_SWEEP
2791 void PSParallelCompact::track_adjusted_pointer(void* p, bool isroot) {
2792 if (!ValidateMarkSweep)
2793 return;
2795 if (!isroot) {
2796 if (_pointer_tracking) {
2797 guarantee(_adjusted_pointers->contains(p), "should have seen this pointer");
2798 _adjusted_pointers->remove(p);
2799 }
2800 } else {
2801 ptrdiff_t index = _root_refs_stack->find(p);
2802 if (index != -1) {
2803 int l = _root_refs_stack->length();
2804 if (l > 0 && l - 1 != index) {
2805 void* last = _root_refs_stack->pop();
2806 assert(last != p, "should be different");
2807 _root_refs_stack->at_put(index, last);
2808 } else {
2809 _root_refs_stack->remove(p);
2810 }
2811 }
2812 }
2813 }
2816 void PSParallelCompact::check_adjust_pointer(void* p) {
2817 _adjusted_pointers->push(p);
2818 }
2821 class AdjusterTracker: public OopClosure {
2822 public:
2823 AdjusterTracker() {};
2824 void do_oop(oop* o) { PSParallelCompact::check_adjust_pointer(o); }
2825 void do_oop(narrowOop* o) { PSParallelCompact::check_adjust_pointer(o); }
2826 };
2829 void PSParallelCompact::track_interior_pointers(oop obj) {
2830 if (ValidateMarkSweep) {
2831 _adjusted_pointers->clear();
2832 _pointer_tracking = true;
2834 AdjusterTracker checker;
2835 obj->oop_iterate(&checker);
2836 }
2837 }
2840 void PSParallelCompact::check_interior_pointers() {
2841 if (ValidateMarkSweep) {
2842 _pointer_tracking = false;
2843 guarantee(_adjusted_pointers->length() == 0, "should have processed the same pointers");
2844 }
2845 }
2848 void PSParallelCompact::reset_live_oop_tracking(bool at_perm) {
2849 if (ValidateMarkSweep) {
2850 guarantee((size_t)_live_oops->length() == _live_oops_index, "should be at end of live oops");
2851 _live_oops_index = at_perm ? _live_oops_index_at_perm : 0;
2852 }
2853 }
2856 void PSParallelCompact::register_live_oop(oop p, size_t size) {
2857 if (ValidateMarkSweep) {
2858 _live_oops->push(p);
2859 _live_oops_size->push(size);
2860 _live_oops_index++;
2861 }
2862 }
2864 void PSParallelCompact::validate_live_oop(oop p, size_t size) {
2865 if (ValidateMarkSweep) {
2866 oop obj = _live_oops->at((int)_live_oops_index);
2867 guarantee(obj == p, "should be the same object");
2868 guarantee(_live_oops_size->at((int)_live_oops_index) == size, "should be the same size");
2869 _live_oops_index++;
2870 }
2871 }
2873 void PSParallelCompact::live_oop_moved_to(HeapWord* q, size_t size,
2874 HeapWord* compaction_top) {
2875 assert(oop(q)->forwardee() == NULL || oop(q)->forwardee() == oop(compaction_top),
2876 "should be moved to forwarded location");
2877 if (ValidateMarkSweep) {
2878 PSParallelCompact::validate_live_oop(oop(q), size);
2879 _live_oops_moved_to->push(oop(compaction_top));
2880 }
2881 if (RecordMarkSweepCompaction) {
2882 _cur_gc_live_oops->push(q);
2883 _cur_gc_live_oops_moved_to->push(compaction_top);
2884 _cur_gc_live_oops_size->push(size);
2885 }
2886 }
2889 void PSParallelCompact::compaction_complete() {
2890 if (RecordMarkSweepCompaction) {
2891 GrowableArray<HeapWord*>* _tmp_live_oops = _cur_gc_live_oops;
2892 GrowableArray<HeapWord*>* _tmp_live_oops_moved_to = _cur_gc_live_oops_moved_to;
2893 GrowableArray<size_t> * _tmp_live_oops_size = _cur_gc_live_oops_size;
2895 _cur_gc_live_oops = _last_gc_live_oops;
2896 _cur_gc_live_oops_moved_to = _last_gc_live_oops_moved_to;
2897 _cur_gc_live_oops_size = _last_gc_live_oops_size;
2898 _last_gc_live_oops = _tmp_live_oops;
2899 _last_gc_live_oops_moved_to = _tmp_live_oops_moved_to;
2900 _last_gc_live_oops_size = _tmp_live_oops_size;
2901 }
2902 }
2905 void PSParallelCompact::print_new_location_of_heap_address(HeapWord* q) {
2906 if (!RecordMarkSweepCompaction) {
2907 tty->print_cr("Requires RecordMarkSweepCompaction to be enabled");
2908 return;
2909 }
2911 if (_last_gc_live_oops == NULL) {
2912 tty->print_cr("No compaction information gathered yet");
2913 return;
2914 }
2916 for (int i = 0; i < _last_gc_live_oops->length(); i++) {
2917 HeapWord* old_oop = _last_gc_live_oops->at(i);
2918 size_t sz = _last_gc_live_oops_size->at(i);
2919 if (old_oop <= q && q < (old_oop + sz)) {
2920 HeapWord* new_oop = _last_gc_live_oops_moved_to->at(i);
2921 size_t offset = (q - old_oop);
2922 tty->print_cr("Address " PTR_FORMAT, q);
2923 tty->print_cr(" Was in oop " PTR_FORMAT ", size %d, at offset %d", old_oop, sz, offset);
2924 tty->print_cr(" Now in oop " PTR_FORMAT ", actual address " PTR_FORMAT, new_oop, new_oop + offset);
2925 return;
2926 }
2927 }
2929 tty->print_cr("Address " PTR_FORMAT " not found in live oop information from last GC", q);
2930 }
2931 #endif //VALIDATE_MARK_SWEEP
2933 // Update interior oops in the ranges of regions [beg_region, end_region).
2934 void
2935 PSParallelCompact::update_and_deadwood_in_dense_prefix(ParCompactionManager* cm,
2936 SpaceId space_id,
2937 size_t beg_region,
2938 size_t end_region) {
2939 ParallelCompactData& sd = summary_data();
2940 ParMarkBitMap* const mbm = mark_bitmap();
2942 HeapWord* beg_addr = sd.region_to_addr(beg_region);
2943 HeapWord* const end_addr = sd.region_to_addr(end_region);
2944 assert(beg_region <= end_region, "bad region range");
2945 assert(end_addr <= dense_prefix(space_id), "not in the dense prefix");
2947 #ifdef ASSERT
2948 // Claim the regions to avoid triggering an assert when they are marked as
2949 // filled.
2950 for (size_t claim_region = beg_region; claim_region < end_region; ++claim_region) {
2951 assert(sd.region(claim_region)->claim_unsafe(), "claim() failed");
2952 }
2953 #endif // #ifdef ASSERT
2955 if (beg_addr != space(space_id)->bottom()) {
2956 // Find the first live object or block of dead space that *starts* in this
2957 // range of regions. If a partial object crosses onto the region, skip it;
2958 // it will be marked for 'deferred update' when the object head is
2959 // processed. If dead space crosses onto the region, it is also skipped; it
2960 // will be filled when the prior region is processed. If neither of those
2961 // apply, the first word in the region is the start of a live object or dead
2962 // space.
2963 assert(beg_addr > space(space_id)->bottom(), "sanity");
2964 const RegionData* const cp = sd.region(beg_region);
2965 if (cp->partial_obj_size() != 0) {
2966 beg_addr = sd.partial_obj_end(beg_region);
2967 } else if (dead_space_crosses_boundary(cp, mbm->addr_to_bit(beg_addr))) {
2968 beg_addr = mbm->find_obj_beg(beg_addr, end_addr);
2969 }
2970 }
2972 if (beg_addr < end_addr) {
2973 // A live object or block of dead space starts in this range of Regions.
2974 HeapWord* const dense_prefix_end = dense_prefix(space_id);
2976 // Create closures and iterate.
2977 UpdateOnlyClosure update_closure(mbm, cm, space_id);
2978 FillClosure fill_closure(cm, space_id);
2979 ParMarkBitMap::IterationStatus status;
2980 status = mbm->iterate(&update_closure, &fill_closure, beg_addr, end_addr,
2981 dense_prefix_end);
2982 if (status == ParMarkBitMap::incomplete) {
2983 update_closure.do_addr(update_closure.source());
2984 }
2985 }
2987 // Mark the regions as filled.
2988 RegionData* const beg_cp = sd.region(beg_region);
2989 RegionData* const end_cp = sd.region(end_region);
2990 for (RegionData* cp = beg_cp; cp < end_cp; ++cp) {
2991 cp->set_completed();
2992 }
2993 }
2995 // Return the SpaceId for the space containing addr. If addr is not in the
2996 // heap, last_space_id is returned. In debug mode it expects the address to be
2997 // in the heap and asserts such.
2998 PSParallelCompact::SpaceId PSParallelCompact::space_id(HeapWord* addr) {
2999 assert(Universe::heap()->is_in_reserved(addr), "addr not in the heap");
3001 for (unsigned int id = perm_space_id; id < last_space_id; ++id) {
3002 if (_space_info[id].space()->contains(addr)) {
3003 return SpaceId(id);
3004 }
3005 }
3007 assert(false, "no space contains the addr");
3008 return last_space_id;
3009 }
3011 void PSParallelCompact::update_deferred_objects(ParCompactionManager* cm,
3012 SpaceId id) {
3013 assert(id < last_space_id, "bad space id");
3015 ParallelCompactData& sd = summary_data();
3016 const SpaceInfo* const space_info = _space_info + id;
3017 ObjectStartArray* const start_array = space_info->start_array();
3019 const MutableSpace* const space = space_info->space();
3020 assert(space_info->dense_prefix() >= space->bottom(), "dense_prefix not set");
3021 HeapWord* const beg_addr = space_info->dense_prefix();
3022 HeapWord* const end_addr = sd.region_align_up(space_info->new_top());
3024 const RegionData* const beg_region = sd.addr_to_region_ptr(beg_addr);
3025 const RegionData* const end_region = sd.addr_to_region_ptr(end_addr);
3026 const RegionData* cur_region;
3027 for (cur_region = beg_region; cur_region < end_region; ++cur_region) {
3028 HeapWord* const addr = cur_region->deferred_obj_addr();
3029 if (addr != NULL) {
3030 if (start_array != NULL) {
3031 start_array->allocate_block(addr);
3032 }
3033 oop(addr)->update_contents(cm);
3034 assert(oop(addr)->is_oop_or_null(), "should be an oop now");
3035 }
3036 }
3037 }
3039 // Skip over count live words starting from beg, and return the address of the
3040 // next live word. Unless marked, the word corresponding to beg is assumed to
3041 // be dead. Callers must either ensure beg does not correspond to the middle of
3042 // an object, or account for those live words in some other way. Callers must
3043 // also ensure that there are enough live words in the range [beg, end) to skip.
3044 HeapWord*
3045 PSParallelCompact::skip_live_words(HeapWord* beg, HeapWord* end, size_t count)
3046 {
3047 assert(count > 0, "sanity");
3049 ParMarkBitMap* m = mark_bitmap();
3050 idx_t bits_to_skip = m->words_to_bits(count);
3051 idx_t cur_beg = m->addr_to_bit(beg);
3052 const idx_t search_end = BitMap::word_align_up(m->addr_to_bit(end));
3054 do {
3055 cur_beg = m->find_obj_beg(cur_beg, search_end);
3056 idx_t cur_end = m->find_obj_end(cur_beg, search_end);
3057 const size_t obj_bits = cur_end - cur_beg + 1;
3058 if (obj_bits > bits_to_skip) {
3059 return m->bit_to_addr(cur_beg + bits_to_skip);
3060 }
3061 bits_to_skip -= obj_bits;
3062 cur_beg = cur_end + 1;
3063 } while (bits_to_skip > 0);
3065 // Skipping the desired number of words landed just past the end of an object.
3066 // Find the start of the next object.
3067 cur_beg = m->find_obj_beg(cur_beg, search_end);
3068 assert(cur_beg < m->addr_to_bit(end), "not enough live words to skip");
3069 return m->bit_to_addr(cur_beg);
3070 }
3072 HeapWord* PSParallelCompact::first_src_addr(HeapWord* const dest_addr,
3073 SpaceId src_space_id,
3074 size_t src_region_idx)
3075 {
3076 assert(summary_data().is_region_aligned(dest_addr), "not aligned");
3078 const SplitInfo& split_info = _space_info[src_space_id].split_info();
3079 if (split_info.dest_region_addr() == dest_addr) {
3080 // The partial object ending at the split point contains the first word to
3081 // be copied to dest_addr.
3082 return split_info.first_src_addr();
3083 }
3085 const ParallelCompactData& sd = summary_data();
3086 ParMarkBitMap* const bitmap = mark_bitmap();
3087 const size_t RegionSize = ParallelCompactData::RegionSize;
3089 assert(sd.is_region_aligned(dest_addr), "not aligned");
3090 const RegionData* const src_region_ptr = sd.region(src_region_idx);
3091 const size_t partial_obj_size = src_region_ptr->partial_obj_size();
3092 HeapWord* const src_region_destination = src_region_ptr->destination();
3094 assert(dest_addr >= src_region_destination, "wrong src region");
3095 assert(src_region_ptr->data_size() > 0, "src region cannot be empty");
3097 HeapWord* const src_region_beg = sd.region_to_addr(src_region_idx);
3098 HeapWord* const src_region_end = src_region_beg + RegionSize;
3100 HeapWord* addr = src_region_beg;
3101 if (dest_addr == src_region_destination) {
3102 // Return the first live word in the source region.
3103 if (partial_obj_size == 0) {
3104 addr = bitmap->find_obj_beg(addr, src_region_end);
3105 assert(addr < src_region_end, "no objects start in src region");
3106 }
3107 return addr;
3108 }
3110 // Must skip some live data.
3111 size_t words_to_skip = dest_addr - src_region_destination;
3112 assert(src_region_ptr->data_size() > words_to_skip, "wrong src region");
3114 if (partial_obj_size >= words_to_skip) {
3115 // All the live words to skip are part of the partial object.
3116 addr += words_to_skip;
3117 if (partial_obj_size == words_to_skip) {
3118 // Find the first live word past the partial object.
3119 addr = bitmap->find_obj_beg(addr, src_region_end);
3120 assert(addr < src_region_end, "wrong src region");
3121 }
3122 return addr;
3123 }
3125 // Skip over the partial object (if any).
3126 if (partial_obj_size != 0) {
3127 words_to_skip -= partial_obj_size;
3128 addr += partial_obj_size;
3129 }
3131 // Skip over live words due to objects that start in the region.
3132 addr = skip_live_words(addr, src_region_end, words_to_skip);
3133 assert(addr < src_region_end, "wrong src region");
3134 return addr;
3135 }
3137 void PSParallelCompact::decrement_destination_counts(ParCompactionManager* cm,
3138 SpaceId src_space_id,
3139 size_t beg_region,
3140 HeapWord* end_addr)
3141 {
3142 ParallelCompactData& sd = summary_data();
3144 #ifdef ASSERT
3145 MutableSpace* const src_space = _space_info[src_space_id].space();
3146 HeapWord* const beg_addr = sd.region_to_addr(beg_region);
3147 assert(src_space->contains(beg_addr) || beg_addr == src_space->end(),
3148 "src_space_id does not match beg_addr");
3149 assert(src_space->contains(end_addr) || end_addr == src_space->end(),
3150 "src_space_id does not match end_addr");
3151 #endif // #ifdef ASSERT
3153 RegionData* const beg = sd.region(beg_region);
3154 RegionData* const end = sd.addr_to_region_ptr(sd.region_align_up(end_addr));
3156 // Regions up to new_top() are enqueued if they become available.
3157 HeapWord* const new_top = _space_info[src_space_id].new_top();
3158 RegionData* const enqueue_end =
3159 sd.addr_to_region_ptr(sd.region_align_up(new_top));
3161 for (RegionData* cur = beg; cur < end; ++cur) {
3162 assert(cur->data_size() > 0, "region must have live data");
3163 cur->decrement_destination_count();
3164 if (cur < enqueue_end && cur->available() && cur->claim()) {
3165 cm->push_region(sd.region(cur));
3166 }
3167 }
3168 }
3170 size_t PSParallelCompact::next_src_region(MoveAndUpdateClosure& closure,
3171 SpaceId& src_space_id,
3172 HeapWord*& src_space_top,
3173 HeapWord* end_addr)
3174 {
3175 typedef ParallelCompactData::RegionData RegionData;
3177 ParallelCompactData& sd = PSParallelCompact::summary_data();
3178 const size_t region_size = ParallelCompactData::RegionSize;
3180 size_t src_region_idx = 0;
3182 // Skip empty regions (if any) up to the top of the space.
3183 HeapWord* const src_aligned_up = sd.region_align_up(end_addr);
3184 RegionData* src_region_ptr = sd.addr_to_region_ptr(src_aligned_up);
3185 HeapWord* const top_aligned_up = sd.region_align_up(src_space_top);
3186 const RegionData* const top_region_ptr =
3187 sd.addr_to_region_ptr(top_aligned_up);
3188 while (src_region_ptr < top_region_ptr && src_region_ptr->data_size() == 0) {
3189 ++src_region_ptr;
3190 }
3192 if (src_region_ptr < top_region_ptr) {
3193 // The next source region is in the current space. Update src_region_idx
3194 // and the source address to match src_region_ptr.
3195 src_region_idx = sd.region(src_region_ptr);
3196 HeapWord* const src_region_addr = sd.region_to_addr(src_region_idx);
3197 if (src_region_addr > closure.source()) {
3198 closure.set_source(src_region_addr);
3199 }
3200 return src_region_idx;
3201 }
3203 // Switch to a new source space and find the first non-empty region.
3204 unsigned int space_id = src_space_id + 1;
3205 assert(space_id < last_space_id, "not enough spaces");
3207 HeapWord* const destination = closure.destination();
3209 do {
3210 MutableSpace* space = _space_info[space_id].space();
3211 HeapWord* const bottom = space->bottom();
3212 const RegionData* const bottom_cp = sd.addr_to_region_ptr(bottom);
3214 // Iterate over the spaces that do not compact into themselves.
3215 if (bottom_cp->destination() != bottom) {
3216 HeapWord* const top_aligned_up = sd.region_align_up(space->top());
3217 const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up);
3219 for (const RegionData* src_cp = bottom_cp; src_cp < top_cp; ++src_cp) {
3220 if (src_cp->live_obj_size() > 0) {
3221 // Found it.
3222 assert(src_cp->destination() == destination,
3223 "first live obj in the space must match the destination");
3224 assert(src_cp->partial_obj_size() == 0,
3225 "a space cannot begin with a partial obj");
3227 src_space_id = SpaceId(space_id);
3228 src_space_top = space->top();
3229 const size_t src_region_idx = sd.region(src_cp);
3230 closure.set_source(sd.region_to_addr(src_region_idx));
3231 return src_region_idx;
3232 } else {
3233 assert(src_cp->data_size() == 0, "sanity");
3234 }
3235 }
3236 }
3237 } while (++space_id < last_space_id);
3239 assert(false, "no source region was found");
3240 return 0;
3241 }
3243 void PSParallelCompact::fill_region(ParCompactionManager* cm, size_t region_idx)
3244 {
3245 typedef ParMarkBitMap::IterationStatus IterationStatus;
3246 const size_t RegionSize = ParallelCompactData::RegionSize;
3247 ParMarkBitMap* const bitmap = mark_bitmap();
3248 ParallelCompactData& sd = summary_data();
3249 RegionData* const region_ptr = sd.region(region_idx);
3251 // Get the items needed to construct the closure.
3252 HeapWord* dest_addr = sd.region_to_addr(region_idx);
3253 SpaceId dest_space_id = space_id(dest_addr);
3254 ObjectStartArray* start_array = _space_info[dest_space_id].start_array();
3255 HeapWord* new_top = _space_info[dest_space_id].new_top();
3256 assert(dest_addr < new_top, "sanity");
3257 const size_t words = MIN2(pointer_delta(new_top, dest_addr), RegionSize);
3259 // Get the source region and related info.
3260 size_t src_region_idx = region_ptr->source_region();
3261 SpaceId src_space_id = space_id(sd.region_to_addr(src_region_idx));
3262 HeapWord* src_space_top = _space_info[src_space_id].space()->top();
3264 MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words);
3265 closure.set_source(first_src_addr(dest_addr, src_space_id, src_region_idx));
3267 // Adjust src_region_idx to prepare for decrementing destination counts (the
3268 // destination count is not decremented when a region is copied to itself).
3269 if (src_region_idx == region_idx) {
3270 src_region_idx += 1;
3271 }
3273 if (bitmap->is_unmarked(closure.source())) {
3274 // The first source word is in the middle of an object; copy the remainder
3275 // of the object or as much as will fit. The fact that pointer updates were
3276 // deferred will be noted when the object header is processed.
3277 HeapWord* const old_src_addr = closure.source();
3278 closure.copy_partial_obj();
3279 if (closure.is_full()) {
3280 decrement_destination_counts(cm, src_space_id, src_region_idx,
3281 closure.source());
3282 region_ptr->set_deferred_obj_addr(NULL);
3283 region_ptr->set_completed();
3284 return;
3285 }
3287 HeapWord* const end_addr = sd.region_align_down(closure.source());
3288 if (sd.region_align_down(old_src_addr) != end_addr) {
3289 // The partial object was copied from more than one source region.
3290 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr);
3292 // Move to the next source region, possibly switching spaces as well. All
3293 // args except end_addr may be modified.
3294 src_region_idx = next_src_region(closure, src_space_id, src_space_top,
3295 end_addr);
3296 }
3297 }
3299 do {
3300 HeapWord* const cur_addr = closure.source();
3301 HeapWord* const end_addr = MIN2(sd.region_align_up(cur_addr + 1),
3302 src_space_top);
3303 IterationStatus status = bitmap->iterate(&closure, cur_addr, end_addr);
3305 if (status == ParMarkBitMap::incomplete) {
3306 // The last obj that starts in the source region does not end in the
3307 // region.
3308 assert(closure.source() < end_addr, "sanity");
3309 HeapWord* const obj_beg = closure.source();
3310 HeapWord* const range_end = MIN2(obj_beg + closure.words_remaining(),
3311 src_space_top);
3312 HeapWord* const obj_end = bitmap->find_obj_end(obj_beg, range_end);
3313 if (obj_end < range_end) {
3314 // The end was found; the entire object will fit.
3315 status = closure.do_addr(obj_beg, bitmap->obj_size(obj_beg, obj_end));
3316 assert(status != ParMarkBitMap::would_overflow, "sanity");
3317 } else {
3318 // The end was not found; the object will not fit.
3319 assert(range_end < src_space_top, "obj cannot cross space boundary");
3320 status = ParMarkBitMap::would_overflow;
3321 }
3322 }
3324 if (status == ParMarkBitMap::would_overflow) {
3325 // The last object did not fit. Note that interior oop updates were
3326 // deferred, then copy enough of the object to fill the region.
3327 region_ptr->set_deferred_obj_addr(closure.destination());
3328 status = closure.copy_until_full(); // copies from closure.source()
3330 decrement_destination_counts(cm, src_space_id, src_region_idx,
3331 closure.source());
3332 region_ptr->set_completed();
3333 return;
3334 }
3336 if (status == ParMarkBitMap::full) {
3337 decrement_destination_counts(cm, src_space_id, src_region_idx,
3338 closure.source());
3339 region_ptr->set_deferred_obj_addr(NULL);
3340 region_ptr->set_completed();
3341 return;
3342 }
3344 decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr);
3346 // Move to the next source region, possibly switching spaces as well. All
3347 // args except end_addr may be modified.
3348 src_region_idx = next_src_region(closure, src_space_id, src_space_top,
3349 end_addr);
3350 } while (true);
3351 }
3353 void
3354 PSParallelCompact::move_and_update(ParCompactionManager* cm, SpaceId space_id) {
3355 const MutableSpace* sp = space(space_id);
3356 if (sp->is_empty()) {
3357 return;
3358 }
3360 ParallelCompactData& sd = PSParallelCompact::summary_data();
3361 ParMarkBitMap* const bitmap = mark_bitmap();
3362 HeapWord* const dp_addr = dense_prefix(space_id);
3363 HeapWord* beg_addr = sp->bottom();
3364 HeapWord* end_addr = sp->top();
3366 assert(beg_addr <= dp_addr && dp_addr <= end_addr, "bad dense prefix");
3368 const size_t beg_region = sd.addr_to_region_idx(beg_addr);
3369 const size_t dp_region = sd.addr_to_region_idx(dp_addr);
3370 if (beg_region < dp_region) {
3371 update_and_deadwood_in_dense_prefix(cm, space_id, beg_region, dp_region);
3372 }
3374 // The destination of the first live object that starts in the region is one
3375 // past the end of the partial object entering the region (if any).
3376 HeapWord* const dest_addr = sd.partial_obj_end(dp_region);
3377 HeapWord* const new_top = _space_info[space_id].new_top();
3378 assert(new_top >= dest_addr, "bad new_top value");
3379 const size_t words = pointer_delta(new_top, dest_addr);
3381 if (words > 0) {
3382 ObjectStartArray* start_array = _space_info[space_id].start_array();
3383 MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words);
3385 ParMarkBitMap::IterationStatus status;
3386 status = bitmap->iterate(&closure, dest_addr, end_addr);
3387 assert(status == ParMarkBitMap::full, "iteration not complete");
3388 assert(bitmap->find_obj_beg(closure.source(), end_addr) == end_addr,
3389 "live objects skipped because closure is full");
3390 }
3391 }
3393 jlong PSParallelCompact::millis_since_last_gc() {
3394 // We need a monotonically non-deccreasing time in ms but
3395 // os::javaTimeMillis() does not guarantee monotonicity.
3396 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
3397 jlong ret_val = now - _time_of_last_gc;
3398 // XXX See note in genCollectedHeap::millis_since_last_gc().
3399 if (ret_val < 0) {
3400 NOT_PRODUCT(warning("time warp: "INT64_FORMAT, ret_val);)
3401 return 0;
3402 }
3403 return ret_val;
3404 }
3406 void PSParallelCompact::reset_millis_since_last_gc() {
3407 // We need a monotonically non-deccreasing time in ms but
3408 // os::javaTimeMillis() does not guarantee monotonicity.
3409 _time_of_last_gc = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
3410 }
3412 ParMarkBitMap::IterationStatus MoveAndUpdateClosure::copy_until_full()
3413 {
3414 if (source() != destination()) {
3415 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3416 Copy::aligned_conjoint_words(source(), destination(), words_remaining());
3417 }
3418 update_state(words_remaining());
3419 assert(is_full(), "sanity");
3420 return ParMarkBitMap::full;
3421 }
3423 void MoveAndUpdateClosure::copy_partial_obj()
3424 {
3425 size_t words = words_remaining();
3427 HeapWord* const range_end = MIN2(source() + words, bitmap()->region_end());
3428 HeapWord* const end_addr = bitmap()->find_obj_end(source(), range_end);
3429 if (end_addr < range_end) {
3430 words = bitmap()->obj_size(source(), end_addr);
3431 }
3433 // This test is necessary; if omitted, the pointer updates to a partial object
3434 // that crosses the dense prefix boundary could be overwritten.
3435 if (source() != destination()) {
3436 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3437 Copy::aligned_conjoint_words(source(), destination(), words);
3438 }
3439 update_state(words);
3440 }
3442 ParMarkBitMapClosure::IterationStatus
3443 MoveAndUpdateClosure::do_addr(HeapWord* addr, size_t words) {
3444 assert(destination() != NULL, "sanity");
3445 assert(bitmap()->obj_size(addr) == words, "bad size");
3447 _source = addr;
3448 assert(PSParallelCompact::summary_data().calc_new_pointer(source()) ==
3449 destination(), "wrong destination");
3451 if (words > words_remaining()) {
3452 return ParMarkBitMap::would_overflow;
3453 }
3455 // The start_array must be updated even if the object is not moving.
3456 if (_start_array != NULL) {
3457 _start_array->allocate_block(destination());
3458 }
3460 if (destination() != source()) {
3461 DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3462 Copy::aligned_conjoint_words(source(), destination(), words);
3463 }
3465 oop moved_oop = (oop) destination();
3466 moved_oop->update_contents(compaction_manager());
3467 assert(moved_oop->is_oop_or_null(), "Object should be whole at this point");
3469 update_state(words);
3470 assert(destination() == (HeapWord*)moved_oop + moved_oop->size(), "sanity");
3471 return is_full() ? ParMarkBitMap::full : ParMarkBitMap::incomplete;
3472 }
3474 UpdateOnlyClosure::UpdateOnlyClosure(ParMarkBitMap* mbm,
3475 ParCompactionManager* cm,
3476 PSParallelCompact::SpaceId space_id) :
3477 ParMarkBitMapClosure(mbm, cm),
3478 _space_id(space_id),
3479 _start_array(PSParallelCompact::start_array(space_id))
3480 {
3481 }
3483 // Updates the references in the object to their new values.
3484 ParMarkBitMapClosure::IterationStatus
3485 UpdateOnlyClosure::do_addr(HeapWord* addr, size_t words) {
3486 do_addr(addr);
3487 return ParMarkBitMap::incomplete;
3488 }
3490 // Prepare for compaction. This method is executed once
3491 // (i.e., by a single thread) before compaction.
3492 // Save the updated location of the intArrayKlassObj for
3493 // filling holes in the dense prefix.
3494 void PSParallelCompact::compact_prologue() {
3495 _updated_int_array_klass_obj = (klassOop)
3496 summary_data().calc_new_pointer(Universe::intArrayKlassObj());
3497 }