1 /*
   2  * Copyright (c) 2005, 2026, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #include "classfile/classLoaderDataGraph.hpp"
  26 #include "classfile/javaClasses.inline.hpp"
  27 #include "classfile/stringTable.hpp"
  28 #include "classfile/symbolTable.hpp"
  29 #include "classfile/systemDictionary.hpp"
  30 #include "code/codeCache.hpp"
  31 #include "code/nmethod.hpp"
  32 #include "compiler/oopMap.hpp"
  33 #include "gc/parallel/objectStartArray.inline.hpp"
  34 #include "gc/parallel/parallelArguments.hpp"
  35 #include "gc/parallel/parallelScavengeHeap.inline.hpp"
  36 #include "gc/parallel/parMarkBitMap.inline.hpp"
  37 #include "gc/parallel/psAdaptiveSizePolicy.hpp"
  38 #include "gc/parallel/psCompactionManager.inline.hpp"
  39 #include "gc/parallel/psOldGen.hpp"
  40 #include "gc/parallel/psParallelCompact.inline.hpp"
  41 #include "gc/parallel/psPromotionManager.inline.hpp"
  42 #include "gc/parallel/psRootType.hpp"
  43 #include "gc/parallel/psScavenge.hpp"
  44 #include "gc/parallel/psStringDedup.hpp"
  45 #include "gc/parallel/psYoungGen.hpp"
  46 #include "gc/shared/classUnloadingContext.hpp"
  47 #include "gc/shared/collectedHeap.inline.hpp"
  48 #include "gc/shared/fullGCForwarding.inline.hpp"
  49 #include "gc/shared/gcCause.hpp"
  50 #include "gc/shared/gcHeapSummary.hpp"
  51 #include "gc/shared/gcId.hpp"
  52 #include "gc/shared/gcLocker.hpp"
  53 #include "gc/shared/gcTimer.hpp"
  54 #include "gc/shared/gcTrace.hpp"
  55 #include "gc/shared/gcTraceTime.inline.hpp"
  56 #include "gc/shared/gcVMOperations.hpp"
  57 #include "gc/shared/isGCActiveMark.hpp"
  58 #include "gc/shared/oopStorage.inline.hpp"
  59 #include "gc/shared/oopStorageSet.inline.hpp"
  60 #include "gc/shared/oopStorageSetParState.inline.hpp"
  61 #include "gc/shared/parallelCleaning.hpp"
  62 #include "gc/shared/preservedMarks.inline.hpp"
  63 #include "gc/shared/referencePolicy.hpp"
  64 #include "gc/shared/referenceProcessor.hpp"
  65 #include "gc/shared/referenceProcessorPhaseTimes.hpp"
  66 #include "gc/shared/spaceDecorator.hpp"
  67 #include "gc/shared/taskTerminator.hpp"
  68 #include "gc/shared/weakProcessor.inline.hpp"
  69 #include "gc/shared/workerPolicy.hpp"
  70 #include "gc/shared/workerThread.hpp"
  71 #include "gc/shared/workerUtils.hpp"
  72 #include "logging/log.hpp"
  73 #include "memory/iterator.inline.hpp"
  74 #include "memory/memoryReserver.hpp"
  75 #include "memory/metaspaceUtils.hpp"
  76 #include "memory/resourceArea.hpp"
  77 #include "memory/universe.hpp"
  78 #include "nmt/memTracker.hpp"
  79 #include "oops/access.inline.hpp"
  80 #include "oops/instanceClassLoaderKlass.inline.hpp"
  81 #include "oops/instanceKlass.inline.hpp"
  82 #include "oops/instanceMirrorKlass.inline.hpp"
  83 #include "oops/methodData.hpp"
  84 #include "oops/objArrayKlass.inline.hpp"
  85 #include "oops/oop.inline.hpp"
  86 #include "runtime/atomicAccess.hpp"
  87 #include "runtime/handles.inline.hpp"
  88 #include "runtime/java.hpp"
  89 #include "runtime/safepoint.hpp"
  90 #include "runtime/threads.hpp"
  91 #include "runtime/vmThread.hpp"
  92 #include "services/memoryService.hpp"
  93 #include "utilities/align.hpp"
  94 #include "utilities/debug.hpp"
  95 #include "utilities/events.hpp"
  96 #include "utilities/formatBuffer.hpp"
  97 #include "utilities/macros.hpp"
  98 #include "utilities/stack.inline.hpp"
  99 #if INCLUDE_JVMCI
 100 #include "jvmci/jvmci.hpp"
 101 #endif
 102 
 103 #include <math.h>
 104 
 105 // All sizes are in HeapWords.
 106 const size_t ParallelCompactData::Log2RegionSize  = 16; // 64K words
 107 const size_t ParallelCompactData::RegionSize      = (size_t)1 << Log2RegionSize;
 108 static_assert(ParallelCompactData::RegionSize >= BitsPerWord, "region-start bit word-aligned");
 109 const size_t ParallelCompactData::RegionSizeBytes =
 110   RegionSize << LogHeapWordSize;
 111 const size_t ParallelCompactData::RegionSizeOffsetMask = RegionSize - 1;
 112 const size_t ParallelCompactData::RegionAddrOffsetMask = RegionSizeBytes - 1;
 113 const size_t ParallelCompactData::RegionAddrMask       = ~RegionAddrOffsetMask;
 114 
 115 const ParallelCompactData::RegionData::region_sz_t
 116 ParallelCompactData::RegionData::dc_shift = 27;
 117 
 118 const ParallelCompactData::RegionData::region_sz_t
 119 ParallelCompactData::RegionData::dc_mask = ~0U << dc_shift;
 120 
 121 const ParallelCompactData::RegionData::region_sz_t
 122 ParallelCompactData::RegionData::dc_one = 0x1U << dc_shift;
 123 
 124 const ParallelCompactData::RegionData::region_sz_t
 125 ParallelCompactData::RegionData::los_mask = ~dc_mask;
 126 
 127 const ParallelCompactData::RegionData::region_sz_t
 128 ParallelCompactData::RegionData::dc_claimed = 0x8U << dc_shift;
 129 
 130 const ParallelCompactData::RegionData::region_sz_t
 131 ParallelCompactData::RegionData::dc_completed = 0xcU << dc_shift;
 132 
 133 bool ParallelCompactData::RegionData::is_clear() {
 134   return (_destination == nullptr) &&
 135          (_source_region == 0) &&
 136          (_partial_obj_addr == nullptr) &&
 137          (_partial_obj_size == 0) &&
 138          (_dc_and_los == 0) &&
 139          (_shadow_state == 0);
 140 }
 141 
 142 #ifdef ASSERT
 143 void ParallelCompactData::RegionData::verify_clear() {
 144   assert(_destination == nullptr, "inv");
 145   assert(_source_region == 0, "inv");
 146   assert(_partial_obj_addr == nullptr, "inv");
 147   assert(_partial_obj_size == 0, "inv");
 148   assert(_dc_and_los == 0, "inv");
 149   assert(_shadow_state == 0, "inv");
 150 }
 151 #endif
 152 
 153 SpaceInfo PSParallelCompact::_space_info[PSParallelCompact::last_space_id];
 154 
 155 SpanSubjectToDiscoveryClosure PSParallelCompact::_span_based_discoverer;
 156 ReferenceProcessor* PSParallelCompact::_ref_processor = nullptr;
 157 
 158 void SplitInfo::record(size_t split_region_idx, HeapWord* split_point, size_t preceding_live_words) {
 159   assert(split_region_idx != 0, "precondition");
 160 
 161   // Obj denoted by split_point will be deferred to the next space.
 162   assert(split_point != nullptr, "precondition");
 163 
 164   const ParallelCompactData& sd = PSParallelCompact::summary_data();
 165 
 166   PSParallelCompact::RegionData* split_region_ptr = sd.region(split_region_idx);
 167   assert(preceding_live_words < split_region_ptr->data_size(), "inv");
 168 
 169   HeapWord* preceding_destination = split_region_ptr->destination();
 170   assert(preceding_destination != nullptr, "inv");
 171 
 172   // How many regions does the preceding part occupy
 173   uint preceding_destination_count;
 174   if (preceding_live_words == 0) {
 175     preceding_destination_count = 0;
 176   } else {
 177     // -1 so that the ending address doesn't fall on the region-boundary
 178     if (sd.region_align_down(preceding_destination) ==
 179         sd.region_align_down(preceding_destination + preceding_live_words - 1)) {
 180       preceding_destination_count = 1;
 181     } else {
 182       preceding_destination_count = 2;
 183     }
 184   }
 185 
 186   _split_region_idx = split_region_idx;
 187   _split_point = split_point;
 188   _preceding_live_words = preceding_live_words;
 189   _preceding_destination = preceding_destination;
 190   _preceding_destination_count = preceding_destination_count;
 191 }
 192 
 193 void SplitInfo::clear()
 194 {
 195   _split_region_idx = 0;
 196   _split_point = nullptr;
 197   _preceding_live_words = 0;
 198   _preceding_destination = nullptr;
 199   _preceding_destination_count = 0;
 200   assert(!is_valid(), "sanity");
 201 }
 202 
 203 #ifdef  ASSERT
 204 void SplitInfo::verify_clear()
 205 {
 206   assert(_split_region_idx == 0, "not clear");
 207   assert(_split_point == nullptr, "not clear");
 208   assert(_preceding_live_words == 0, "not clear");
 209   assert(_preceding_destination == nullptr, "not clear");
 210   assert(_preceding_destination_count == 0, "not clear");
 211 }
 212 #endif  // #ifdef ASSERT
 213 
 214 
 215 void PSParallelCompact::print_on(outputStream* st) {
 216   _mark_bitmap.print_on(st);
 217 }
 218 
 219 ParallelCompactData::ParallelCompactData() :
 220   _heap_start(nullptr),
 221   DEBUG_ONLY(_heap_end(nullptr) COMMA)
 222   _region_vspace(nullptr),
 223   _reserved_byte_size(0),
 224   _region_data(nullptr),
 225   _region_count(0) {}
 226 
 227 bool ParallelCompactData::initialize(MemRegion reserved_heap)
 228 {
 229   _heap_start = reserved_heap.start();
 230   const size_t heap_size = reserved_heap.word_size();
 231   DEBUG_ONLY(_heap_end = _heap_start + heap_size;)
 232 
 233   assert(region_align_down(_heap_start) == _heap_start,
 234          "region start not aligned");
 235 
 236   return initialize_region_data(heap_size);
 237 }
 238 
 239 PSVirtualSpace*
 240 ParallelCompactData::create_vspace(size_t count, size_t element_size)
 241 {
 242   const size_t raw_bytes = count * element_size;
 243   const size_t page_sz = os::page_size_for_region_aligned(raw_bytes, 10);
 244   const size_t granularity = os::vm_allocation_granularity();
 245   const size_t rs_align = MAX2(page_sz, granularity);
 246 
 247   _reserved_byte_size = align_up(raw_bytes, rs_align);
 248 
 249   ReservedSpace rs = MemoryReserver::reserve(_reserved_byte_size,
 250                                              rs_align,
 251                                              page_sz,
 252                                              mtGC);
 253 
 254   if (!rs.is_reserved()) {
 255     // Failed to reserve memory.
 256     return nullptr;
 257   }
 258 
 259   os::trace_page_sizes("Parallel Compact Data", raw_bytes, raw_bytes, rs.base(),
 260                        rs.size(), page_sz);
 261 
 262   MemTracker::record_virtual_memory_tag(rs, mtGC);
 263 
 264   PSVirtualSpace* vspace = new PSVirtualSpace(rs, page_sz);
 265 
 266   if (!vspace->expand_by(_reserved_byte_size)) {
 267     // Failed to commit memory.
 268 
 269     delete vspace;
 270 
 271     // Release memory reserved in the space.
 272     MemoryReserver::release(rs);
 273 
 274     return nullptr;
 275   }
 276 
 277   return vspace;
 278 }
 279 
 280 bool ParallelCompactData::initialize_region_data(size_t heap_size)
 281 {
 282   assert(is_aligned(heap_size, RegionSize), "precondition");
 283 
 284   const size_t count = heap_size >> Log2RegionSize;
 285   _region_vspace = create_vspace(count, sizeof(RegionData));
 286   if (_region_vspace != nullptr) {
 287     _region_data = (RegionData*)_region_vspace->reserved_low_addr();
 288     _region_count = count;
 289     return true;
 290   }
 291   return false;
 292 }
 293 
 294 void ParallelCompactData::clear_range(size_t beg_region, size_t end_region) {
 295   assert(beg_region <= _region_count, "beg_region out of range");
 296   assert(end_region <= _region_count, "end_region out of range");
 297 
 298   const size_t region_cnt = end_region - beg_region;
 299   memset(_region_data + beg_region, 0, region_cnt * sizeof(RegionData));
 300 }
 301 
 302 // The total live words on src_region would overflow the target space, so find
 303 // the overflowing object and record the split point. The invariant is that an
 304 // obj should not cross space boundary.
 305 HeapWord* ParallelCompactData::summarize_split_space(size_t src_region,
 306                                                      SplitInfo& split_info,
 307                                                      HeapWord* const destination,
 308                                                      HeapWord* const target_end,
 309                                                      HeapWord** target_next) {
 310   assert(destination <= target_end, "sanity");
 311   assert(destination + _region_data[src_region].data_size() > target_end,
 312     "region should not fit into target space");
 313   assert(is_region_aligned(target_end), "sanity");
 314 
 315   size_t partial_obj_size = _region_data[src_region].partial_obj_size();
 316 
 317   if (destination + partial_obj_size > target_end) {
 318     assert(partial_obj_size > 0, "inv");
 319     // The overflowing obj is from a previous region.
 320     //
 321     // source-regions:
 322     //
 323     // ***************
 324     // |     A|AA    |
 325     // ***************
 326     //       ^
 327     //       | split-point
 328     //
 329     // dest-region:
 330     //
 331     // ********
 332     // |~~~~A |
 333     // ********
 334     //       ^^
 335     //       || target-space-end
 336     //       |
 337     //       | destination
 338     //
 339     // AAA would overflow target-space.
 340     //
 341     HeapWord* overflowing_obj = _region_data[src_region].partial_obj_addr();
 342     size_t split_region = addr_to_region_idx(overflowing_obj);
 343 
 344     // The number of live words before the overflowing object on this split region
 345     size_t preceding_live_words;
 346     if (is_region_aligned(overflowing_obj)) {
 347       preceding_live_words = 0;
 348     } else {
 349       // Words accounted by the overflowing object on the split region
 350       size_t overflowing_size = pointer_delta(region_align_up(overflowing_obj), overflowing_obj);
 351       preceding_live_words = region(split_region)->data_size() - overflowing_size;
 352     }
 353 
 354     split_info.record(split_region, overflowing_obj, preceding_live_words);
 355 
 356     // The [overflowing_obj, src_region_start) part has been accounted for, so
 357     // must move back the new_top, now that this overflowing obj is deferred.
 358     HeapWord* new_top = destination - pointer_delta(region_to_addr(src_region), overflowing_obj);
 359 
 360     // If the overflowing obj was relocated to its original destination,
 361     // those destination regions would have their source_region set. Now that
 362     // this overflowing obj is relocated somewhere else, reset the
 363     // source_region.
 364     {
 365       size_t range_start = addr_to_region_idx(region_align_up(new_top));
 366       size_t range_end = addr_to_region_idx(region_align_up(destination));
 367       for (size_t i = range_start; i < range_end; ++i) {
 368         region(i)->set_source_region(0);
 369       }
 370     }
 371 
 372     // Update new top of target space
 373     *target_next = new_top;
 374 
 375     return overflowing_obj;
 376   }
 377 
 378   // Obj-iteration to locate the overflowing obj
 379   HeapWord* region_start = region_to_addr(src_region);
 380   HeapWord* region_end = region_start + RegionSize;
 381   HeapWord* cur_addr = region_start + partial_obj_size;
 382   size_t live_words = partial_obj_size;
 383 
 384   while (true) {
 385     assert(cur_addr < region_end, "inv");
 386     cur_addr = PSParallelCompact::mark_bitmap()->find_obj_beg(cur_addr, region_end);
 387     // There must be an overflowing obj in this region
 388     assert(cur_addr < region_end, "inv");
 389 
 390     oop obj = cast_to_oop(cur_addr);
 391     size_t obj_size = obj->size();
 392     if (destination + live_words + obj_size > target_end) {
 393       // Found the overflowing obj
 394       split_info.record(src_region, cur_addr, live_words);
 395       *target_next = destination + live_words;
 396       return cur_addr;
 397     }
 398 
 399     live_words += obj_size;
 400     cur_addr += obj_size;
 401   }
 402 }
 403 
 404 size_t ParallelCompactData::live_words_in_space(const MutableSpace* space,
 405                                                 HeapWord** full_region_prefix_end) {
 406   size_t cur_region = addr_to_region_idx(space->bottom());
 407   const size_t end_region = addr_to_region_idx(region_align_up(space->top()));
 408   size_t live_words = 0;
 409   if (full_region_prefix_end == nullptr) {
 410     for (/* empty */; cur_region < end_region; ++cur_region) {
 411       live_words += _region_data[cur_region].data_size();
 412     }
 413   } else {
 414     bool first_set = false;
 415     for (/* empty */; cur_region < end_region; ++cur_region) {
 416       size_t live_words_in_region = _region_data[cur_region].data_size();
 417       if (!first_set && live_words_in_region < RegionSize) {
 418         *full_region_prefix_end = region_to_addr(cur_region);
 419         first_set = true;
 420       }
 421       live_words += live_words_in_region;
 422     }
 423     if (!first_set) {
 424       // All regions are full of live objs.
 425       assert(is_region_aligned(space->top()), "inv");
 426       *full_region_prefix_end = space->top();
 427     }
 428     assert(*full_region_prefix_end != nullptr, "postcondition");
 429     assert(is_region_aligned(*full_region_prefix_end), "inv");
 430     assert(*full_region_prefix_end >= space->bottom(), "in-range");
 431     assert(*full_region_prefix_end <= space->top(), "in-range");
 432   }
 433   return live_words;
 434 }
 435 
 436 bool ParallelCompactData::summarize(SplitInfo& split_info,
 437                                     HeapWord* source_beg, HeapWord* source_end,
 438                                     HeapWord** source_next,
 439                                     HeapWord* target_beg, HeapWord* target_end,
 440                                     HeapWord** target_next)
 441 {
 442   HeapWord* const source_next_val = source_next == nullptr ? nullptr : *source_next;
 443   log_develop_trace(gc, compaction)(
 444       "sb=" PTR_FORMAT " se=" PTR_FORMAT " sn=" PTR_FORMAT
 445       "tb=" PTR_FORMAT " te=" PTR_FORMAT " tn=" PTR_FORMAT,
 446       p2i(source_beg), p2i(source_end), p2i(source_next_val),
 447       p2i(target_beg), p2i(target_end), p2i(*target_next));
 448 
 449   size_t cur_region = addr_to_region_idx(source_beg);
 450   const size_t end_region = addr_to_region_idx(region_align_up(source_end));
 451 
 452   HeapWord *dest_addr = target_beg;
 453   for (/* empty */; cur_region < end_region; cur_region++) {
 454     size_t words = _region_data[cur_region].data_size();
 455 
 456     // Skip empty ones
 457     if (words == 0) {
 458       continue;
 459     }
 460 
 461     if (split_info.is_split(cur_region)) {
 462       assert(words > split_info.preceding_live_words(), "inv");
 463       words -= split_info.preceding_live_words();
 464     }
 465 
 466     _region_data[cur_region].set_destination(dest_addr);
 467 
 468     // If cur_region does not fit entirely into the target space, find a point
 469     // at which the source space can be 'split' so that part is copied to the
 470     // target space and the rest is copied elsewhere.
 471     if (dest_addr + words > target_end) {
 472       assert(source_next != nullptr, "source_next is null when splitting");
 473       *source_next = summarize_split_space(cur_region, split_info, dest_addr,
 474                                            target_end, target_next);
 475       return false;
 476     }
 477 
 478     uint destination_count = split_info.is_split(cur_region)
 479                              ? split_info.preceding_destination_count()
 480                              : 0;
 481 
 482     HeapWord* const last_addr = dest_addr + words - 1;
 483     const size_t dest_region_1 = addr_to_region_idx(dest_addr);
 484     const size_t dest_region_2 = addr_to_region_idx(last_addr);
 485 
 486     // Initially assume that the destination regions will be the same and
 487     // adjust the value below if necessary.  Under this assumption, if
 488     // cur_region == dest_region_2, then cur_region will be compacted
 489     // completely into itself.
 490     destination_count += cur_region == dest_region_2 ? 0 : 1;
 491     if (dest_region_1 != dest_region_2) {
 492       // Destination regions differ; adjust destination_count.
 493       destination_count += 1;
 494       // Data from cur_region will be copied to the start of dest_region_2.
 495       _region_data[dest_region_2].set_source_region(cur_region);
 496     } else if (is_region_aligned(dest_addr)) {
 497       // Data from cur_region will be copied to the start of the destination
 498       // region.
 499       _region_data[dest_region_1].set_source_region(cur_region);
 500     }
 501 
 502     _region_data[cur_region].set_destination_count(destination_count);
 503     dest_addr += words;
 504   }
 505 
 506   *target_next = dest_addr;
 507   return true;
 508 }
 509 
 510 #ifdef ASSERT
 511 void ParallelCompactData::verify_clear() {
 512   for (uint cur_idx = 0; cur_idx < region_count(); ++cur_idx) {
 513     if (!region(cur_idx)->is_clear()) {
 514       log_warning(gc)("Uncleared Region: %u", cur_idx);
 515       region(cur_idx)->verify_clear();
 516     }
 517   }
 518 }
 519 #endif  // #ifdef ASSERT
 520 
 521 STWGCTimer          PSParallelCompact::_gc_timer;
 522 ParallelOldTracer   PSParallelCompact::_gc_tracer;
 523 elapsedTimer        PSParallelCompact::_accumulated_time;
 524 unsigned int        PSParallelCompact::_maximum_compaction_gc_num = 0;
 525 CollectorCounters*  PSParallelCompact::_counters = nullptr;
 526 ParMarkBitMap       PSParallelCompact::_mark_bitmap;
 527 ParallelCompactData PSParallelCompact::_summary_data;
 528 
 529 PSParallelCompact::IsAliveClosure PSParallelCompact::_is_alive_closure;
 530 
 531 class PCAdjustPointerClosure: public BasicOopIterateClosure {
 532   template <typename T>
 533   void do_oop_work(T* p) { PSParallelCompact::adjust_pointer(p); }
 534 
 535 public:
 536   virtual void do_oop(oop* p)                { do_oop_work(p); }
 537   virtual void do_oop(narrowOop* p)          { do_oop_work(p); }
 538 
 539   virtual ReferenceIterationMode reference_iteration_mode() { return DO_FIELDS; }
 540 };
 541 
 542 static PCAdjustPointerClosure pc_adjust_pointer_closure;
 543 
 544 bool PSParallelCompact::IsAliveClosure::do_object_b(oop p) { return mark_bitmap()->is_marked(p); }
 545 
 546 void PSParallelCompact::post_initialize() {
 547   ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
 548   _span_based_discoverer.set_span(heap->reserved_region());
 549   _ref_processor =
 550     new ReferenceProcessor(&_span_based_discoverer,
 551                            ParallelGCThreads,   // mt processing degree
 552                            ParallelGCThreads,   // mt discovery degree
 553                            false,               // concurrent_discovery
 554                            &_is_alive_closure); // non-header is alive closure
 555 
 556   _counters = new CollectorCounters("Parallel full collection pauses", 1);
 557 
 558   // Initialize static fields in ParCompactionManager.
 559   ParCompactionManager::initialize(mark_bitmap());
 560 }
 561 
 562 bool PSParallelCompact::initialize_aux_data() {
 563   ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
 564   MemRegion mr = heap->reserved_region();
 565   assert(mr.byte_size() != 0, "heap should be reserved");
 566 
 567   initialize_space_info();
 568 
 569   if (!_mark_bitmap.initialize(mr)) {
 570     vm_shutdown_during_initialization(
 571       err_msg("Unable to allocate %zuKB bitmaps for parallel "
 572       "garbage collection for the requested %zuKB heap.",
 573       _mark_bitmap.reserved_byte_size()/K, mr.byte_size()/K));
 574     return false;
 575   }
 576 
 577   if (!_summary_data.initialize(mr)) {
 578     vm_shutdown_during_initialization(
 579       err_msg("Unable to allocate %zuKB card tables for parallel "
 580       "garbage collection for the requested %zuKB heap.",
 581       _summary_data.reserved_byte_size()/K, mr.byte_size()/K));
 582     return false;
 583   }
 584 
 585   return true;
 586 }
 587 
 588 void PSParallelCompact::initialize_space_info()
 589 {
 590   memset(&_space_info, 0, sizeof(_space_info));
 591 
 592   ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
 593   PSYoungGen* young_gen = heap->young_gen();
 594 
 595   _space_info[old_space_id].set_space(heap->old_gen()->object_space());
 596   _space_info[eden_space_id].set_space(young_gen->eden_space());
 597   _space_info[from_space_id].set_space(young_gen->from_space());
 598   _space_info[to_space_id].set_space(young_gen->to_space());
 599 
 600   _space_info[old_space_id].set_start_array(heap->old_gen()->start_array());
 601 }
 602 
 603 void
 604 PSParallelCompact::clear_data_covering_space(SpaceId id)
 605 {
 606   // At this point, top is the value before GC, new_top() is the value that will
 607   // be set at the end of GC.  The marking bitmap is cleared to top; nothing
 608   // should be marked above top.  The summary data is cleared to the larger of
 609   // top & new_top.
 610   MutableSpace* const space = _space_info[id].space();
 611   HeapWord* const bot = space->bottom();
 612   HeapWord* const top = space->top();
 613   HeapWord* const max_top = MAX2(top, _space_info[id].new_top());
 614 
 615   _mark_bitmap.clear_range(bot, top);
 616 
 617   const size_t beg_region = _summary_data.addr_to_region_idx(bot);
 618   const size_t end_region =
 619     _summary_data.addr_to_region_idx(_summary_data.region_align_up(max_top));
 620   _summary_data.clear_range(beg_region, end_region);
 621 
 622   // Clear the data used to 'split' regions.
 623   SplitInfo& split_info = _space_info[id].split_info();
 624   if (split_info.is_valid()) {
 625     split_info.clear();
 626   }
 627   DEBUG_ONLY(split_info.verify_clear();)
 628 }
 629 
 630 void PSParallelCompact::pre_compact()
 631 {
 632   // Update the from & to space pointers in space_info, since they are swapped
 633   // at each young gen gc.  Do the update unconditionally (even though a
 634   // promotion failure does not swap spaces) because an unknown number of young
 635   // collections will have swapped the spaces an unknown number of times.
 636   GCTraceTime(Debug, gc, phases) tm("Pre Compact", &_gc_timer);
 637   ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
 638   _space_info[from_space_id].set_space(heap->young_gen()->from_space());
 639   _space_info[to_space_id].set_space(heap->young_gen()->to_space());
 640 
 641   heap->increment_total_collections(true);
 642 
 643   CodeCache::on_gc_marking_cycle_start();
 644 
 645   heap->print_before_gc();
 646   heap->trace_heap_before_gc(&_gc_tracer);
 647 
 648   // Fill in TLABs
 649   heap->ensure_parsability(true);  // retire TLABs
 650 
 651   if (VerifyBeforeGC && heap->total_collections() >= VerifyGCStartAt) {
 652     Universe::verify("Before GC");
 653   }
 654 
 655   DEBUG_ONLY(mark_bitmap()->verify_clear();)
 656   DEBUG_ONLY(summary_data().verify_clear();)
 657 }
 658 
 659 void PSParallelCompact::post_compact()
 660 {
 661   GCTraceTime(Info, gc, phases) tm("Post Compact", &_gc_timer);
 662   ParCompactionManager::remove_all_shadow_regions();
 663 
 664   CodeCache::on_gc_marking_cycle_finish();
 665   CodeCache::arm_all_nmethods();
 666 
 667   // Need to clear claim bits for the next full-gc (marking and adjust-pointers).
 668   ClassLoaderDataGraph::clear_claimed_marks();
 669 
 670   for (unsigned int id = old_space_id; id < last_space_id; ++id) {
 671     // Clear the marking bitmap, summary data and split info.
 672     clear_data_covering_space(SpaceId(id));
 673     {
 674       MutableSpace* space = _space_info[id].space();
 675       HeapWord* top = space->top();
 676       HeapWord* new_top = _space_info[id].new_top();
 677       if (ZapUnusedHeapArea && new_top < top) {
 678         space->mangle_region(MemRegion(new_top, top));
 679       }
 680       // Update top().  Must be done after clearing the bitmap and summary data.
 681       space->set_top(new_top);
 682     }
 683   }
 684 
 685 #ifdef ASSERT
 686   {
 687     mark_bitmap()->verify_clear();
 688     summary_data().verify_clear();
 689   }
 690 #endif
 691 
 692   ParCompactionManager::flush_all_string_dedup_requests();
 693 
 694   MutableSpace* const eden_space = _space_info[eden_space_id].space();
 695   MutableSpace* const from_space = _space_info[from_space_id].space();
 696   MutableSpace* const to_space   = _space_info[to_space_id].space();
 697 
 698   ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
 699   bool eden_empty = eden_space->is_empty();
 700 
 701   // Update heap occupancy information which is used as input to the soft ref
 702   // clearing policy at the next gc.
 703   Universe::heap()->update_capacity_and_used_at_gc();
 704 
 705   bool young_gen_empty = eden_empty && from_space->is_empty() &&
 706     to_space->is_empty();
 707 
 708   PSCardTable* ct = heap->card_table();
 709   MemRegion old_mr = heap->old_gen()->committed();
 710   if (young_gen_empty) {
 711     ct->clear_MemRegion(old_mr);
 712   } else {
 713     ct->dirty_MemRegion(old_mr);
 714   }
 715 
 716   heap->prune_scavengable_nmethods();
 717 
 718 #if COMPILER2_OR_JVMCI
 719   DerivedPointerTable::update_pointers();
 720 #endif
 721 
 722   // Signal that we have completed a visit to all live objects.
 723   Universe::heap()->record_whole_heap_examined_timestamp();
 724 }
 725 
 726 HeapWord* PSParallelCompact::compute_dense_prefix_for_old_space(MutableSpace* old_space,
 727                                                                 HeapWord* full_region_prefix_end) {
 728   const size_t region_size = ParallelCompactData::RegionSize;
 729   const ParallelCompactData& sd = summary_data();
 730 
 731   // Iteration starts with the region *after* the full-region-prefix-end.
 732   const RegionData* const start_region = sd.addr_to_region_ptr(full_region_prefix_end);
 733   // If final region is not full, iteration stops before that region,
 734   // because fill_dense_prefix_end assumes that prefix_end <= top.
 735   const RegionData* const end_region = sd.addr_to_region_ptr(old_space->top());
 736   assert(start_region <= end_region, "inv");
 737 
 738   size_t max_waste = old_space->capacity_in_words() * (MarkSweepDeadRatio / 100.0);
 739   const RegionData* cur_region = start_region;
 740   for (/* empty */; cur_region < end_region; ++cur_region) {
 741     assert(region_size >= cur_region->data_size(), "inv");
 742     size_t dead_size = region_size - cur_region->data_size();
 743     if (max_waste < dead_size) {
 744       break;
 745     }
 746     max_waste -= dead_size;
 747   }
 748 
 749   HeapWord* const prefix_end = sd.region_to_addr(cur_region);
 750   assert(sd.is_region_aligned(prefix_end), "postcondition");
 751   assert(prefix_end >= full_region_prefix_end, "in-range");
 752   assert(prefix_end <= old_space->top(), "in-range");
 753   return prefix_end;
 754 }
 755 
 756 void PSParallelCompact::fill_dense_prefix_end(SpaceId id) {
 757   // Comparing two sizes to decide if filling is required:
 758   //
 759   // The size of the filler (min-obj-size) is 2 heap words with the default
 760   // MinObjAlignment, since both markword and klass take 1 heap word.
 761   // With +UseCompactObjectHeaders, the minimum filler size is only one word,
 762   // because the Klass* gets encoded in the mark-word.
 763   //
 764   // The size of the gap (if any) right before dense-prefix-end is
 765   // MinObjAlignment.
 766   //
 767   // Need to fill in the gap only if it's smaller than min-obj-size, and the
 768   // filler obj will extend to next region.
 769 
 770   if (MinObjAlignment >= checked_cast<int>(CollectedHeap::min_fill_size())) {
 771     return;
 772   }
 773 
 774   assert(!UseCompactObjectHeaders, "Compact headers can allocate small objects");
 775   assert(CollectedHeap::min_fill_size() == 2, "inv");
 776   HeapWord* const dense_prefix_end = dense_prefix(id);
 777   assert(_summary_data.is_region_aligned(dense_prefix_end), "precondition");
 778   assert(dense_prefix_end <= space(id)->top(), "precondition");
 779   if (dense_prefix_end == space(id)->top()) {
 780     // Must not have single-word gap right before prefix-end/top.
 781     return;
 782   }
 783   RegionData* const region_after_dense_prefix = _summary_data.addr_to_region_ptr(dense_prefix_end);
 784 
 785   if (region_after_dense_prefix->partial_obj_size() != 0 ||
 786       _mark_bitmap.is_marked(dense_prefix_end)) {
 787     // The region after the dense prefix starts with live bytes.
 788     return;
 789   }
 790 
 791   HeapWord* block_start = start_array(id)->block_start_reaching_into_card(dense_prefix_end);
 792   if (block_start == dense_prefix_end - 1) {
 793     assert(!_mark_bitmap.is_marked(block_start), "inv");
 794     // There is exactly one heap word gap right before the dense prefix end, so we need a filler object.
 795     // The filler object will extend into region_after_dense_prefix.
 796     const size_t obj_len = 2; // min-fill-size
 797     HeapWord* const obj_beg = dense_prefix_end - 1;
 798     CollectedHeap::fill_with_object(obj_beg, obj_len);
 799     _mark_bitmap.mark_obj(obj_beg);
 800     _summary_data.addr_to_region_ptr(obj_beg)->add_live_obj(1);
 801     region_after_dense_prefix->set_partial_obj_size(1);
 802     region_after_dense_prefix->set_partial_obj_addr(obj_beg);
 803     assert(start_array(id) != nullptr, "sanity");
 804     start_array(id)->update_for_block(obj_beg, obj_beg + obj_len);
 805   }
 806 }
 807 
 808 bool PSParallelCompact::check_maximum_compaction(bool should_do_max_compaction,
 809                                                  size_t total_live_words,
 810                                                  MutableSpace* const old_space,
 811                                                  HeapWord* full_region_prefix_end) {
 812 
 813   ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
 814 
 815   // Check System.GC
 816   bool is_max_on_system_gc = UseMaximumCompactionOnSystemGC
 817                           && GCCause::is_user_requested_gc(heap->gc_cause());
 818 
 819   // Check if all live objs are too much for old-gen.
 820   const bool is_old_gen_too_full = (total_live_words >= old_space->capacity_in_words());
 821 
 822   // If all regions in old-gen are full
 823   const bool is_region_full =
 824     full_region_prefix_end >= _summary_data.region_align_down(old_space->top());
 825 
 826   return should_do_max_compaction
 827       || is_max_on_system_gc
 828       || is_old_gen_too_full
 829       || is_region_full;
 830 }
 831 
 832 void PSParallelCompact::summary_phase(bool should_do_max_compaction)
 833 {
 834   GCTraceTime(Info, gc, phases) tm("Summary Phase", &_gc_timer);
 835 
 836   MutableSpace* const old_space = _space_info[old_space_id].space();
 837   {
 838     size_t total_live_words = 0;
 839     HeapWord* full_region_prefix_end = nullptr;
 840     {
 841       // old-gen
 842       size_t live_words = _summary_data.live_words_in_space(old_space,
 843                                                             &full_region_prefix_end);
 844       total_live_words += live_words;
 845     }
 846     // young-gen
 847     for (uint i = eden_space_id; i < last_space_id; ++i) {
 848       const MutableSpace* space = _space_info[i].space();
 849       size_t live_words = _summary_data.live_words_in_space(space);
 850       total_live_words += live_words;
 851       _space_info[i].set_new_top(space->bottom() + live_words);
 852       _space_info[i].set_dense_prefix(space->bottom());
 853     }
 854 
 855     should_do_max_compaction = check_maximum_compaction(should_do_max_compaction,
 856                                                         total_live_words,
 857                                                         old_space,
 858                                                         full_region_prefix_end);
 859     {
 860       GCTraceTime(Info, gc, phases) tm("Summary Phase: expand", &_gc_timer);
 861       // Try to expand old-gen in order to fit all live objs and waste.
 862       size_t target_capacity_bytes = total_live_words * HeapWordSize
 863                                    + old_space->capacity_in_bytes() * (MarkSweepDeadRatio / 100);
 864       ParallelScavengeHeap::heap()->old_gen()->try_expand_till_size(target_capacity_bytes);
 865     }
 866 
 867     HeapWord* dense_prefix_end = should_do_max_compaction
 868                                  ? full_region_prefix_end
 869                                  : compute_dense_prefix_for_old_space(old_space,
 870                                                                       full_region_prefix_end);
 871     SpaceId id = old_space_id;
 872     _space_info[id].set_dense_prefix(dense_prefix_end);
 873 
 874     if (dense_prefix_end != old_space->bottom()) {
 875       fill_dense_prefix_end(id);
 876     }
 877 
 878     // Compacting objs in [dense_prefix_end, old_space->top())
 879     _summary_data.summarize(_space_info[id].split_info(),
 880                             dense_prefix_end, old_space->top(), nullptr,
 881                             dense_prefix_end, old_space->end(),
 882                             _space_info[id].new_top_addr());
 883   }
 884 
 885   // Summarize the remaining spaces in the young gen.  The initial target space
 886   // is the old gen.  If a space does not fit entirely into the target, then the
 887   // remainder is compacted into the space itself and that space becomes the new
 888   // target.
 889   SpaceId dst_space_id = old_space_id;
 890   HeapWord* dst_space_end = old_space->end();
 891   HeapWord** new_top_addr = _space_info[dst_space_id].new_top_addr();
 892   for (unsigned int id = eden_space_id; id < last_space_id; ++id) {
 893     const MutableSpace* space = _space_info[id].space();
 894     const size_t live = pointer_delta(_space_info[id].new_top(),
 895                                       space->bottom());
 896     const size_t available = pointer_delta(dst_space_end, *new_top_addr);
 897 
 898     if (live > 0 && live <= available) {
 899       // All the live data will fit.
 900       bool done = _summary_data.summarize(_space_info[id].split_info(),
 901                                           space->bottom(), space->top(),
 902                                           nullptr,
 903                                           *new_top_addr, dst_space_end,
 904                                           new_top_addr);
 905       assert(done, "space must fit into old gen");
 906 
 907       // Reset the new_top value for the space.
 908       _space_info[id].set_new_top(space->bottom());
 909     } else if (live > 0) {
 910       // Attempt to fit part of the source space into the target space.
 911       HeapWord* next_src_addr = nullptr;
 912       bool done = _summary_data.summarize(_space_info[id].split_info(),
 913                                           space->bottom(), space->top(),
 914                                           &next_src_addr,
 915                                           *new_top_addr, dst_space_end,
 916                                           new_top_addr);
 917       assert(!done, "space should not fit into old gen");
 918       assert(next_src_addr != nullptr, "sanity");
 919 
 920       // The source space becomes the new target, so the remainder is compacted
 921       // within the space itself.
 922       dst_space_id = SpaceId(id);
 923       dst_space_end = space->end();
 924       new_top_addr = _space_info[id].new_top_addr();
 925       done = _summary_data.summarize(_space_info[id].split_info(),
 926                                      next_src_addr, space->top(),
 927                                      nullptr,
 928                                      space->bottom(), dst_space_end,
 929                                      new_top_addr);
 930       assert(done, "space must fit when compacted into itself");
 931       assert(*new_top_addr <= space->top(), "usage should not grow");
 932     }
 933   }
 934 }
 935 
 936 void PSParallelCompact::report_object_count_after_gc() {
 937   GCTraceTime(Debug, gc, phases) tm("Report Object Count", &_gc_timer);
 938   // The heap is compacted, all objects are iterable. However there may be
 939   // filler objects in the heap which we should ignore.
 940   class SkipFillerObjectClosure : public BoolObjectClosure {
 941   public:
 942     bool do_object_b(oop obj) override { return !CollectedHeap::is_filler_object(obj); }
 943   } cl;
 944   _gc_tracer.report_object_count_after_gc(&cl, &ParallelScavengeHeap::heap()->workers());
 945 }
 946 
 947 bool PSParallelCompact::invoke(bool clear_all_soft_refs, bool should_do_max_compaction) {
 948   assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
 949   assert(Thread::current() == (Thread*)VMThread::vm_thread(),
 950          "should be in vm thread");
 951   assert(ref_processor() != nullptr, "Sanity");
 952 
 953   SvcGCMarker sgcm(SvcGCMarker::FULL);
 954   IsSTWGCActiveMark mark;
 955 
 956   ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
 957 
 958   GCIdMark gc_id_mark;
 959   _gc_timer.register_gc_start();
 960   _gc_tracer.report_gc_start(heap->gc_cause(), _gc_timer.gc_start());
 961 
 962   GCCause::Cause gc_cause = heap->gc_cause();
 963   PSOldGen* old_gen = heap->old_gen();
 964   PSAdaptiveSizePolicy* size_policy = heap->size_policy();
 965 
 966   // Make sure data structures are sane, make the heap parsable, and do other
 967   // miscellaneous bookkeeping.
 968   pre_compact();
 969 
 970   const PreGenGCValues pre_gc_values = heap->get_pre_gc_values();
 971 
 972   {
 973     const uint active_workers =
 974       WorkerPolicy::calc_active_workers(ParallelScavengeHeap::heap()->workers().max_workers(),
 975                                         ParallelScavengeHeap::heap()->workers().active_workers(),
 976                                         Threads::number_of_non_daemon_threads());
 977     ParallelScavengeHeap::heap()->workers().set_active_workers(active_workers);
 978 
 979     GCTraceCPUTime tcpu(&_gc_tracer);
 980     GCTraceTime(Info, gc) tm("Pause Full", nullptr, gc_cause, true);
 981 
 982     heap->pre_full_gc_dump(&_gc_timer);
 983 
 984     TraceCollectorStats tcs(counters());
 985     TraceMemoryManagerStats tms(heap->old_gc_manager(), gc_cause, "end of major GC");
 986 
 987     if (log_is_enabled(Debug, gc, heap, exit)) {
 988       accumulated_time()->start();
 989     }
 990 
 991     // Let the size policy know we're starting
 992     size_policy->major_collection_begin();
 993 
 994 #if COMPILER2_OR_JVMCI
 995     DerivedPointerTable::clear();
 996 #endif
 997 
 998     ref_processor()->start_discovery(clear_all_soft_refs);
 999 
1000     marking_phase(&_gc_tracer);
1001 
1002     summary_phase(should_do_max_compaction);
1003 
1004 #if COMPILER2_OR_JVMCI
1005     assert(DerivedPointerTable::is_active(), "Sanity");
1006     DerivedPointerTable::set_active(false);
1007 #endif
1008 
1009     forward_to_new_addr();
1010 
1011     adjust_pointers();
1012 
1013     compact();
1014 
1015     ParCompactionManager::_preserved_marks_set->restore(&ParallelScavengeHeap::heap()->workers());
1016 
1017     ParCompactionManager::verify_all_region_stack_empty();
1018 
1019     // Reset the mark bitmap, summary data, and do other bookkeeping.  Must be
1020     // done before resizing.
1021     post_compact();
1022 
1023     size_policy->major_collection_end();
1024 
1025     size_policy->sample_old_gen_used_bytes(MAX2(pre_gc_values.old_gen_used(), old_gen->used_in_bytes()));
1026 
1027     if (UseAdaptiveSizePolicy) {
1028       heap->resize_after_full_gc();
1029     }
1030 
1031     heap->resize_all_tlabs();
1032 
1033     // Resize the metaspace capacity after a collection
1034     MetaspaceGC::compute_new_size();
1035 
1036     if (log_is_enabled(Debug, gc, heap, exit)) {
1037       accumulated_time()->stop();
1038     }
1039 
1040     heap->print_heap_change(pre_gc_values);
1041 
1042     report_object_count_after_gc();
1043 
1044     // Track memory usage and detect low memory
1045     MemoryService::track_memory_usage();
1046     heap->update_counters();
1047 
1048     heap->post_full_gc_dump(&_gc_timer);
1049 
1050     size_policy->record_gc_pause_end_instant();
1051   }
1052 
1053   heap->gc_epilogue(true);
1054 
1055   if (VerifyAfterGC && heap->total_collections() >= VerifyGCStartAt) {
1056     Universe::verify("After GC");
1057   }
1058 
1059   heap->print_after_gc();
1060   heap->trace_heap_after_gc(&_gc_tracer);
1061 
1062   _gc_timer.register_gc_end();
1063 
1064   _gc_tracer.report_dense_prefix(dense_prefix(old_space_id));
1065   _gc_tracer.report_gc_end(_gc_timer.gc_end(), _gc_timer.time_partitions());
1066 
1067   return true;
1068 }
1069 
1070 class PCAddThreadRootsMarkingTaskClosure : public ThreadClosure {
1071   ParCompactionManager* _cm;
1072 
1073 public:
1074   PCAddThreadRootsMarkingTaskClosure(ParCompactionManager* cm) : _cm(cm) { }
1075   void do_thread(Thread* thread) {
1076     ResourceMark rm;
1077 
1078     MarkingNMethodClosure mark_and_push_in_blobs(&_cm->_mark_and_push_closure);
1079 
1080     thread->oops_do(&_cm->_mark_and_push_closure, &mark_and_push_in_blobs);
1081 
1082     // Do the real work
1083     _cm->follow_marking_stacks();
1084   }
1085 };
1086 
1087 void steal_marking_work(TaskTerminator& terminator, uint worker_id) {
1088   assert(ParallelScavengeHeap::heap()->is_stw_gc_active(), "called outside gc");
1089 
1090   ParCompactionManager* cm =
1091     ParCompactionManager::gc_thread_compaction_manager(worker_id);
1092 
1093   do {
1094     ScannerTask task;
1095     if (ParCompactionManager::steal(worker_id, task)) {
1096       cm->follow_contents(task, true);
1097     }
1098     cm->follow_marking_stacks();
1099   } while (!terminator.offer_termination());
1100 }
1101 
1102 class MarkFromRootsTask : public WorkerTask {
1103   NMethodMarkingScope _nmethod_marking_scope;
1104   ThreadsClaimTokenScope _threads_claim_token_scope;
1105   OopStorageSetStrongParState<false /* concurrent */, false /* is_const */> _oop_storage_set_par_state;
1106   TaskTerminator _terminator;
1107   uint _active_workers;
1108 
1109 public:
1110   MarkFromRootsTask(uint active_workers) :
1111       WorkerTask("MarkFromRootsTask"),
1112       _nmethod_marking_scope(),
1113       _threads_claim_token_scope(),
1114       _terminator(active_workers, ParCompactionManager::marking_stacks()),
1115       _active_workers(active_workers) {}
1116 
1117   virtual void work(uint worker_id) {
1118     ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(worker_id);
1119     cm->create_marking_stats_cache();
1120     {
1121       CLDToOopClosure cld_closure(&cm->_mark_and_push_closure, ClassLoaderData::_claim_stw_fullgc_mark);
1122       ClassLoaderDataGraph::always_strong_cld_do(&cld_closure);
1123 
1124       // Do the real work
1125       cm->follow_marking_stacks();
1126     }
1127 
1128     {
1129       PCAddThreadRootsMarkingTaskClosure closure(cm);
1130       Threads::possibly_parallel_threads_do(_active_workers > 1 /* is_par */, &closure);
1131     }
1132 
1133     // Mark from OopStorages
1134     {
1135       _oop_storage_set_par_state.oops_do(&cm->_mark_and_push_closure);
1136       // Do the real work
1137       cm->follow_marking_stacks();
1138     }
1139 
1140     if (_active_workers > 1) {
1141       steal_marking_work(_terminator, worker_id);
1142     }
1143   }
1144 };
1145 
1146 class ParallelCompactRefProcProxyTask : public RefProcProxyTask {
1147   TaskTerminator _terminator;
1148 
1149 public:
1150   ParallelCompactRefProcProxyTask(uint max_workers)
1151     : RefProcProxyTask("ParallelCompactRefProcProxyTask", max_workers),
1152       _terminator(_max_workers, ParCompactionManager::marking_stacks()) {}
1153 
1154   void work(uint worker_id) override {
1155     assert(worker_id < _max_workers, "sanity");
1156     ParCompactionManager* cm = (_tm == RefProcThreadModel::Single) ? ParCompactionManager::get_vmthread_cm() : ParCompactionManager::gc_thread_compaction_manager(worker_id);
1157     BarrierEnqueueDiscoveredFieldClosure enqueue;
1158     ParCompactionManager::FollowStackClosure complete_gc(cm, (_tm == RefProcThreadModel::Single) ? nullptr : &_terminator, worker_id);
1159     _rp_task->rp_work(worker_id, PSParallelCompact::is_alive_closure(), &cm->_mark_and_push_closure, &enqueue, &complete_gc);
1160   }
1161 
1162   void prepare_run_task_hook() override {
1163     _terminator.reset_for_reuse(_queue_count);
1164   }
1165 };
1166 
1167 static void flush_marking_stats_cache(const uint num_workers) {
1168   for (uint i = 0; i < num_workers; ++i) {
1169     ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(i);
1170     cm->flush_and_destroy_marking_stats_cache();
1171   }
1172 }
1173 
1174 class PSParallelCleaningTask : public WorkerTask {
1175   bool                    _unloading_occurred;
1176   CodeCacheUnloadingTask  _code_cache_task;
1177   // Prune dead klasses from subklass/sibling/implementor lists.
1178   KlassCleaningTask       _klass_cleaning_task;
1179 
1180 public:
1181   PSParallelCleaningTask(bool unloading_occurred) :
1182     WorkerTask("PS Parallel Cleaning"),
1183     _unloading_occurred(unloading_occurred),
1184     _code_cache_task(unloading_occurred),
1185     _klass_cleaning_task() {}
1186 
1187   void work(uint worker_id) {
1188 #if INCLUDE_JVMCI
1189     if (EnableJVMCI && worker_id == 0) {
1190       // Serial work; only first worker.
1191       // Clean JVMCI metadata handles.
1192       JVMCI::do_unloading(_unloading_occurred);
1193     }
1194 #endif
1195 
1196     // Do first pass of code cache cleaning.
1197     _code_cache_task.work(worker_id);
1198 
1199     // Clean all klasses that were not unloaded.
1200     // The weak metadata in klass doesn't need to be
1201     // processed if there was no unloading.
1202     if (_unloading_occurred) {
1203       _klass_cleaning_task.work();
1204     }
1205   }
1206 };
1207 
1208 void PSParallelCompact::marking_phase(ParallelOldTracer *gc_tracer) {
1209   // Recursively traverse all live objects and mark them
1210   GCTraceTime(Info, gc, phases) tm("Marking Phase", &_gc_timer);
1211 
1212   uint active_gc_threads = ParallelScavengeHeap::heap()->workers().active_workers();
1213 
1214   ClassLoaderDataGraph::verify_claimed_marks_cleared(ClassLoaderData::_claim_stw_fullgc_mark);
1215   {
1216     GCTraceTime(Debug, gc, phases) tm("Par Mark", &_gc_timer);
1217 
1218     MarkFromRootsTask task(active_gc_threads);
1219     ParallelScavengeHeap::heap()->workers().run_task(&task);
1220   }
1221 
1222   // Process reference objects found during marking
1223   {
1224     GCTraceTime(Debug, gc, phases) tm("Reference Processing", &_gc_timer);
1225 
1226     ReferenceProcessorStats stats;
1227     ReferenceProcessorPhaseTimes pt(&_gc_timer, ref_processor()->max_num_queues());
1228 
1229     ParallelCompactRefProcProxyTask task(ref_processor()->max_num_queues());
1230     stats = ref_processor()->process_discovered_references(task, &ParallelScavengeHeap::heap()->workers(), pt);
1231 
1232     gc_tracer->report_gc_reference_stats(stats);
1233     pt.print_all_references();
1234   }
1235 
1236   {
1237     GCTraceTime(Debug, gc, phases) tm("Flush Marking Stats", &_gc_timer);
1238 
1239     flush_marking_stats_cache(active_gc_threads);
1240   }
1241 
1242   // This is the point where the entire marking should have completed.
1243   ParCompactionManager::verify_all_marking_stack_empty();
1244 
1245   {
1246     GCTraceTime(Debug, gc, phases) tm("Weak Processing", &_gc_timer);
1247     WeakProcessor::weak_oops_do(&ParallelScavengeHeap::heap()->workers(),
1248                                 is_alive_closure(),
1249                                 &do_nothing_cl,
1250                                 1);
1251   }
1252 
1253   {
1254     GCTraceTime(Debug, gc, phases) tm_m("Class Unloading", &_gc_timer);
1255 
1256     ClassUnloadingContext ctx(active_gc_threads /* num_nmethod_unlink_workers */,
1257                               false /* unregister_nmethods_during_purge */,
1258                               false /* lock_nmethod_free_separately */);
1259 
1260     {
1261       CodeCache::UnlinkingScope scope(is_alive_closure());
1262 
1263       // Follow system dictionary roots and unload classes.
1264       bool unloading_occurred = SystemDictionary::do_unloading(&_gc_timer);
1265 
1266       PSParallelCleaningTask task{unloading_occurred};
1267       ParallelScavengeHeap::heap()->workers().run_task(&task);
1268     }
1269 
1270     {
1271       GCTraceTime(Debug, gc, phases) t("Purge Unlinked NMethods", gc_timer());
1272       // Release unloaded nmethod's memory.
1273       ctx.purge_nmethods();
1274     }
1275     {
1276       GCTraceTime(Debug, gc, phases) ur("Unregister NMethods", &_gc_timer);
1277       ParallelScavengeHeap::heap()->prune_unlinked_nmethods();
1278     }
1279     {
1280       GCTraceTime(Debug, gc, phases) t("Free Code Blobs", gc_timer());
1281       ctx.free_nmethods();
1282     }
1283     {
1284       // Delete metaspaces for unloaded class loaders and clean up loader_data graph
1285       GCTraceTime(Debug, gc, phases) t("Purge Class Loader Data", gc_timer());
1286       ClassLoaderDataGraph::purge(true /* at_safepoint */);
1287       DEBUG_ONLY(MetaspaceUtils::verify();)
1288     }
1289   }
1290 
1291 #if TASKQUEUE_STATS
1292   ParCompactionManager::print_and_reset_taskqueue_stats();
1293 #endif
1294 }
1295 
1296 template<typename Func>
1297 void PSParallelCompact::adjust_in_space_helper(SpaceId id, volatile uint* claim_counter, Func&& on_stripe) {
1298   MutableSpace* sp = PSParallelCompact::space(id);
1299   HeapWord* const bottom = sp->bottom();
1300   HeapWord* const top = sp->top();
1301   if (bottom == top) {
1302     return;
1303   }
1304 
1305   const uint num_regions_per_stripe = 2;
1306   const size_t region_size = ParallelCompactData::RegionSize;
1307   const size_t stripe_size = num_regions_per_stripe * region_size;
1308 
1309   while (true) {
1310     uint counter = AtomicAccess::fetch_then_add(claim_counter, num_regions_per_stripe);
1311     HeapWord* cur_stripe = bottom + counter * region_size;
1312     if (cur_stripe >= top) {
1313       break;
1314     }
1315     HeapWord* stripe_end = MIN2(cur_stripe + stripe_size, top);
1316     on_stripe(cur_stripe, stripe_end);
1317   }
1318 }
1319 
1320 void PSParallelCompact::adjust_in_old_space(volatile uint* claim_counter) {
1321   // Regions in old-space shouldn't be split.
1322   assert(!_space_info[old_space_id].split_info().is_valid(), "inv");
1323 
1324   auto scan_obj_with_limit = [&] (HeapWord* obj_start, HeapWord* left, HeapWord* right) {
1325     assert(mark_bitmap()->is_marked(obj_start), "inv");
1326     oop obj = cast_to_oop(obj_start);
1327     return obj->oop_iterate_size(&pc_adjust_pointer_closure, MemRegion(left, right));
1328   };
1329 
1330   adjust_in_space_helper(old_space_id, claim_counter, [&] (HeapWord* stripe_start, HeapWord* stripe_end) {
1331     assert(_summary_data.is_region_aligned(stripe_start), "inv");
1332     RegionData* cur_region = _summary_data.addr_to_region_ptr(stripe_start);
1333     HeapWord* obj_start;
1334     if (cur_region->partial_obj_size() != 0) {
1335       obj_start = cur_region->partial_obj_addr();
1336       obj_start += scan_obj_with_limit(obj_start, stripe_start, stripe_end);
1337     } else {
1338       obj_start = stripe_start;
1339     }
1340 
1341     while (obj_start < stripe_end) {
1342       obj_start = mark_bitmap()->find_obj_beg(obj_start, stripe_end);
1343       if (obj_start >= stripe_end) {
1344         break;
1345       }
1346       obj_start += scan_obj_with_limit(obj_start, stripe_start, stripe_end);
1347     }
1348   });
1349 }
1350 
1351 void PSParallelCompact::adjust_in_young_space(SpaceId id, volatile uint* claim_counter) {
1352   adjust_in_space_helper(id, claim_counter, [](HeapWord* stripe_start, HeapWord* stripe_end) {
1353     HeapWord* obj_start = stripe_start;
1354     while (obj_start < stripe_end) {
1355       obj_start = mark_bitmap()->find_obj_beg(obj_start, stripe_end);
1356       if (obj_start >= stripe_end) {
1357         break;
1358       }
1359       oop obj = cast_to_oop(obj_start);
1360       obj_start += obj->oop_iterate_size(&pc_adjust_pointer_closure);
1361     }
1362   });
1363 }
1364 
1365 void PSParallelCompact::adjust_pointers_in_spaces(uint worker_id, volatile uint* claim_counters) {
1366   auto start_time = Ticks::now();
1367   adjust_in_old_space(&claim_counters[0]);
1368   for (uint id = eden_space_id; id < last_space_id; ++id) {
1369     adjust_in_young_space(SpaceId(id), &claim_counters[id]);
1370   }
1371   log_trace(gc, phases)("adjust_pointers_in_spaces worker %u: %.3f ms", worker_id, (Ticks::now() - start_time).seconds() * 1000);
1372 }
1373 
1374 class PSAdjustTask final : public WorkerTask {
1375   ThreadsClaimTokenScope                     _threads_claim_token_scope;
1376   WeakProcessor::Task                        _weak_proc_task;
1377   OopStorageSetStrongParState<false, false>  _oop_storage_iter;
1378   uint                                       _nworkers;
1379   volatile bool                              _code_cache_claimed;
1380   volatile uint _claim_counters[PSParallelCompact::last_space_id] = {};
1381 
1382   bool try_claim_code_cache_task() {
1383     return AtomicAccess::load(&_code_cache_claimed) == false
1384         && AtomicAccess::cmpxchg(&_code_cache_claimed, false, true) == false;
1385   }
1386 
1387 public:
1388   PSAdjustTask(uint nworkers) :
1389     WorkerTask("PSAdjust task"),
1390     _threads_claim_token_scope(),
1391     _weak_proc_task(nworkers),
1392     _oop_storage_iter(),
1393     _nworkers(nworkers),
1394     _code_cache_claimed(false) {
1395 
1396     ClassLoaderDataGraph::verify_claimed_marks_cleared(ClassLoaderData::_claim_stw_fullgc_adjust);
1397   }
1398 
1399   void work(uint worker_id) {
1400     {
1401       // Pointers in heap.
1402       ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(worker_id);
1403       cm->preserved_marks()->adjust_during_full_gc();
1404 
1405       PSParallelCompact::adjust_pointers_in_spaces(worker_id, _claim_counters);
1406     }
1407 
1408     {
1409       // All (strong and weak) CLDs.
1410       CLDToOopClosure cld_closure(&pc_adjust_pointer_closure, ClassLoaderData::_claim_stw_fullgc_adjust);
1411       ClassLoaderDataGraph::cld_do(&cld_closure);
1412     }
1413 
1414     {
1415       // Threads stack frames. No need to visit on-stack nmethods, because all
1416       // nmethods are visited in one go via CodeCache::nmethods_do.
1417       ResourceMark rm;
1418       Threads::possibly_parallel_oops_do(_nworkers > 1, &pc_adjust_pointer_closure, nullptr);
1419       if (try_claim_code_cache_task()) {
1420         NMethodToOopClosure adjust_code(&pc_adjust_pointer_closure, NMethodToOopClosure::FixRelocations);
1421         CodeCache::nmethods_do(&adjust_code);
1422       }
1423     }
1424 
1425     {
1426       // VM internal strong and weak roots.
1427       _oop_storage_iter.oops_do(&pc_adjust_pointer_closure);
1428       AlwaysTrueClosure always_alive;
1429       _weak_proc_task.work(worker_id, &always_alive, &pc_adjust_pointer_closure);
1430     }
1431   }
1432 };
1433 
1434 void PSParallelCompact::adjust_pointers() {
1435   // Adjust the pointers to reflect the new locations
1436   GCTraceTime(Info, gc, phases) tm("Adjust Pointers", &_gc_timer);
1437   uint nworkers = ParallelScavengeHeap::heap()->workers().active_workers();
1438   PSAdjustTask task(nworkers);
1439   ParallelScavengeHeap::heap()->workers().run_task(&task);
1440 }
1441 
1442 // Split [start, end) evenly for a number of workers and return the
1443 // range for worker_id.
1444 static void split_regions_for_worker(size_t start, size_t end,
1445                                      uint worker_id, uint num_workers,
1446                                      size_t* worker_start, size_t* worker_end) {
1447   assert(start < end, "precondition");
1448   assert(num_workers > 0, "precondition");
1449   assert(worker_id < num_workers, "precondition");
1450 
1451   size_t num_regions = end - start;
1452   size_t num_regions_per_worker = num_regions / num_workers;
1453   size_t remainder = num_regions % num_workers;
1454   // The first few workers will get one extra.
1455   *worker_start = start + worker_id * num_regions_per_worker
1456                   + MIN2(checked_cast<size_t>(worker_id), remainder);
1457   *worker_end = *worker_start + num_regions_per_worker
1458                 + (worker_id < remainder ? 1 : 0);
1459 }
1460 
1461 void PSParallelCompact::forward_to_new_addr() {
1462   GCTraceTime(Info, gc, phases) tm("Forward", &_gc_timer);
1463   uint nworkers = ParallelScavengeHeap::heap()->workers().active_workers();
1464 
1465   struct ForwardTask final : public WorkerTask {
1466     uint _num_workers;
1467 
1468     explicit ForwardTask(uint num_workers) :
1469       WorkerTask("PSForward task"),
1470       _num_workers(num_workers) {}
1471 
1472     static void forward_objs_in_range(ParCompactionManager* cm,
1473                                       HeapWord* start,
1474                                       HeapWord* end,
1475                                       HeapWord* destination) {
1476       HeapWord* cur_addr = start;
1477       HeapWord* new_addr = destination;
1478 
1479       while (cur_addr < end) {
1480         cur_addr = mark_bitmap()->find_obj_beg(cur_addr, end);
1481         if (cur_addr >= end) {
1482           return;
1483         }
1484         assert(mark_bitmap()->is_marked(cur_addr), "inv");
1485         oop obj = cast_to_oop(cur_addr);
1486         if (new_addr != cur_addr) {
1487           cm->preserved_marks()->push_if_necessary(obj, obj->mark());
1488           FullGCForwarding::forward_to(obj, cast_to_oop(new_addr));
1489         }
1490         size_t obj_size = obj->size();
1491         new_addr += obj_size;
1492         cur_addr += obj_size;
1493       }
1494     }
1495 
1496     void work(uint worker_id) override {
1497       ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(worker_id);
1498       for (uint id = old_space_id; id < last_space_id; ++id) {
1499         MutableSpace* sp = PSParallelCompact::space(SpaceId(id));
1500         HeapWord* dense_prefix_addr = dense_prefix(SpaceId(id));
1501         HeapWord* top = sp->top();
1502 
1503         if (dense_prefix_addr == top) {
1504           // Empty space
1505           continue;
1506         }
1507 
1508         const SplitInfo& split_info = _space_info[SpaceId(id)].split_info();
1509         size_t dense_prefix_region = _summary_data.addr_to_region_idx(dense_prefix_addr);
1510         size_t top_region = _summary_data.addr_to_region_idx(_summary_data.region_align_up(top));
1511         size_t start_region;
1512         size_t end_region;
1513         split_regions_for_worker(dense_prefix_region, top_region,
1514                                  worker_id, _num_workers,
1515                                  &start_region, &end_region);
1516         for (size_t cur_region = start_region; cur_region < end_region; ++cur_region) {
1517           RegionData* region_ptr = _summary_data.region(cur_region);
1518           size_t partial_obj_size = region_ptr->partial_obj_size();
1519 
1520           if (partial_obj_size == ParallelCompactData::RegionSize) {
1521             // No obj-start
1522             continue;
1523           }
1524 
1525           HeapWord* region_start = _summary_data.region_to_addr(cur_region);
1526           HeapWord* region_end = region_start + ParallelCompactData::RegionSize;
1527 
1528           if (split_info.is_split(cur_region)) {
1529             // Part 1: will be relocated to space-1
1530             HeapWord* preceding_destination = split_info.preceding_destination();
1531             HeapWord* split_point = split_info.split_point();
1532             forward_objs_in_range(cm, region_start + partial_obj_size, split_point, preceding_destination + partial_obj_size);
1533 
1534             // Part 2: will be relocated to space-2
1535             HeapWord* destination = region_ptr->destination();
1536             forward_objs_in_range(cm, split_point, region_end, destination);
1537           } else {
1538             HeapWord* destination = region_ptr->destination();
1539             forward_objs_in_range(cm, region_start + partial_obj_size, region_end, destination + partial_obj_size);
1540           }
1541         }
1542       }
1543     }
1544   } task(nworkers);
1545 
1546   ParallelScavengeHeap::heap()->workers().run_task(&task);
1547   DEBUG_ONLY(verify_forward();)
1548 }
1549 
1550 #ifdef ASSERT
1551 void PSParallelCompact::verify_forward() {
1552   HeapWord* const old_dense_prefix_addr = dense_prefix(SpaceId(old_space_id));
1553   // The destination addr for the first live obj after dense-prefix.
1554   HeapWord* bump_ptr = old_dense_prefix_addr
1555                      + _summary_data.addr_to_region_ptr(old_dense_prefix_addr)->partial_obj_size();
1556   SpaceId bump_ptr_space = old_space_id;
1557 
1558   for (uint id = old_space_id; id < last_space_id; ++id) {
1559     MutableSpace* sp = PSParallelCompact::space(SpaceId(id));
1560     // Only verify objs after dense-prefix, because those before dense-prefix are not moved (forwarded).
1561     HeapWord* cur_addr = dense_prefix(SpaceId(id));
1562     HeapWord* top = sp->top();
1563 
1564     while (cur_addr < top) {
1565       cur_addr = mark_bitmap()->find_obj_beg(cur_addr, top);
1566       if (cur_addr >= top) {
1567         break;
1568       }
1569       assert(mark_bitmap()->is_marked(cur_addr), "inv");
1570       assert(bump_ptr <= _space_info[bump_ptr_space].new_top(), "inv");
1571       // Move to the space containing cur_addr
1572       if (bump_ptr == _space_info[bump_ptr_space].new_top()) {
1573         bump_ptr = space(space_id(cur_addr))->bottom();
1574         bump_ptr_space = space_id(bump_ptr);
1575       }
1576       oop obj = cast_to_oop(cur_addr);
1577       if (cur_addr == bump_ptr) {
1578         assert(!FullGCForwarding::is_forwarded(obj), "inv");
1579       } else {
1580         assert(FullGCForwarding::forwardee(obj) == cast_to_oop(bump_ptr), "inv");
1581       }
1582       bump_ptr += obj->size();
1583       cur_addr += obj->size();
1584     }
1585   }
1586 }
1587 #endif
1588 
1589 // Helper class to print 8 region numbers per line and then print the total at the end.
1590 class FillableRegionLogger : public StackObj {
1591 private:
1592   Log(gc, compaction) log;
1593   static const int LineLength = 8;
1594   size_t _regions[LineLength];
1595   int _next_index;
1596   bool _enabled;
1597   size_t _total_regions;
1598 public:
1599   FillableRegionLogger() : _next_index(0), _enabled(log_develop_is_enabled(Trace, gc, compaction)), _total_regions(0) { }
1600   ~FillableRegionLogger() {
1601     log.trace("%zu initially fillable regions", _total_regions);
1602   }
1603 
1604   void print_line() {
1605     if (!_enabled || _next_index == 0) {
1606       return;
1607     }
1608     FormatBuffer<> line("Fillable: ");
1609     for (int i = 0; i < _next_index; i++) {
1610       line.append(" %7zu", _regions[i]);
1611     }
1612     log.trace("%s", line.buffer());
1613     _next_index = 0;
1614   }
1615 
1616   void handle(size_t region) {
1617     if (!_enabled) {
1618       return;
1619     }
1620     _regions[_next_index++] = region;
1621     if (_next_index == LineLength) {
1622       print_line();
1623     }
1624     _total_regions++;
1625   }
1626 };
1627 
1628 void PSParallelCompact::prepare_region_draining_tasks(uint parallel_gc_threads)
1629 {
1630   GCTraceTime(Trace, gc, phases) tm("Drain Task Setup", &_gc_timer);
1631 
1632   // Find the threads that are active
1633   uint worker_id = 0;
1634 
1635   // Find all regions that are available (can be filled immediately) and
1636   // distribute them to the thread stacks.  The iteration is done in reverse
1637   // order (high to low) so the regions will be removed in ascending order.
1638 
1639   const ParallelCompactData& sd = PSParallelCompact::summary_data();
1640 
1641   // id + 1 is used to test termination so unsigned  can
1642   // be used with an old_space_id == 0.
1643   FillableRegionLogger region_logger;
1644   for (unsigned int id = last_space_id - 1; id + 1 > old_space_id; --id) {
1645     SpaceInfo* const space_info = _space_info + id;
1646     HeapWord* const new_top = space_info->new_top();
1647 
1648     const size_t beg_region = sd.addr_to_region_idx(space_info->dense_prefix());
1649     const size_t end_region =
1650       sd.addr_to_region_idx(sd.region_align_up(new_top));
1651 
1652     for (size_t cur = end_region - 1; cur + 1 > beg_region; --cur) {
1653       if (sd.region(cur)->claim_unsafe()) {
1654         ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(worker_id);
1655         bool result = sd.region(cur)->mark_normal();
1656         assert(result, "Must succeed at this point.");
1657         cm->region_stack()->push(cur);
1658         region_logger.handle(cur);
1659         // Assign regions to tasks in round-robin fashion.
1660         if (++worker_id == parallel_gc_threads) {
1661           worker_id = 0;
1662         }
1663       }
1664     }
1665     region_logger.print_line();
1666   }
1667 }
1668 
1669 static void compaction_with_stealing_work(TaskTerminator* terminator, uint worker_id) {
1670   assert(ParallelScavengeHeap::heap()->is_stw_gc_active(), "called outside gc");
1671 
1672   ParCompactionManager* cm =
1673     ParCompactionManager::gc_thread_compaction_manager(worker_id);
1674 
1675   // Drain the stacks that have been preloaded with regions
1676   // that are ready to fill.
1677 
1678   cm->drain_region_stacks();
1679 
1680   guarantee(cm->region_stack()->is_empty(), "Not empty");
1681 
1682   size_t region_index = 0;
1683 
1684   while (true) {
1685     if (ParCompactionManager::steal(worker_id, region_index)) {
1686       PSParallelCompact::fill_and_update_region(cm, region_index);
1687       cm->drain_region_stacks();
1688     } else if (PSParallelCompact::steal_unavailable_region(cm, region_index)) {
1689       // Fill and update an unavailable region with the help of a shadow region
1690       PSParallelCompact::fill_and_update_shadow_region(cm, region_index);
1691       cm->drain_region_stacks();
1692     } else {
1693       if (terminator->offer_termination()) {
1694         break;
1695       }
1696       // Go around again.
1697     }
1698   }
1699 }
1700 
1701 class FillDensePrefixAndCompactionTask: public WorkerTask {
1702   TaskTerminator _terminator;
1703 
1704 public:
1705   FillDensePrefixAndCompactionTask(uint active_workers) :
1706       WorkerTask("FillDensePrefixAndCompactionTask"),
1707       _terminator(active_workers, ParCompactionManager::region_task_queues()) {
1708   }
1709 
1710   virtual void work(uint worker_id) {
1711     if (worker_id == 0) {
1712       auto start = Ticks::now();
1713       PSParallelCompact::fill_dead_objs_in_dense_prefix();
1714       log_trace(gc, phases)("Fill dense prefix by worker 0: %.3f ms", (Ticks::now() - start).seconds() * 1000);
1715     }
1716     compaction_with_stealing_work(&_terminator, worker_id);
1717   }
1718 };
1719 
1720 void PSParallelCompact::fill_range_in_dense_prefix(HeapWord* start, HeapWord* end) {
1721 #ifdef ASSERT
1722   {
1723     assert(start < end, "precondition");
1724     assert(mark_bitmap()->find_obj_beg(start, end) == end, "precondition");
1725     HeapWord* bottom = _space_info[old_space_id].space()->bottom();
1726     if (start != bottom) {
1727       // The preceding live obj.
1728       HeapWord* obj_start = mark_bitmap()->find_obj_beg_reverse(bottom, start);
1729       HeapWord* obj_end = obj_start + cast_to_oop(obj_start)->size();
1730       assert(obj_end == start, "precondition");
1731     }
1732   }
1733 #endif
1734 
1735   CollectedHeap::fill_with_objects(start, pointer_delta(end, start));
1736   HeapWord* addr = start;
1737   do {
1738     size_t size = cast_to_oop(addr)->size();
1739     start_array(old_space_id)->update_for_block(addr, addr + size);
1740     addr += size;
1741   } while (addr < end);
1742 }
1743 
1744 void PSParallelCompact::fill_dead_objs_in_dense_prefix() {
1745   ParMarkBitMap* bitmap = mark_bitmap();
1746 
1747   HeapWord* const bottom = _space_info[old_space_id].space()->bottom();
1748   HeapWord* const prefix_end = dense_prefix(old_space_id);
1749 
1750   const size_t region_size = ParallelCompactData::RegionSize;
1751 
1752   // Fill dead space in [start_addr, end_addr)
1753   HeapWord* const start_addr = bottom;
1754   HeapWord* const end_addr   = prefix_end;
1755 
1756   for (HeapWord* cur_addr = start_addr; cur_addr < end_addr; /* empty */) {
1757     RegionData* cur_region_ptr = _summary_data.addr_to_region_ptr(cur_addr);
1758     if (cur_region_ptr->data_size() == region_size) {
1759       // Full; no dead space. Next region.
1760       if (_summary_data.is_region_aligned(cur_addr)) {
1761         cur_addr += region_size;
1762       } else {
1763         cur_addr = _summary_data.region_align_up(cur_addr);
1764       }
1765       continue;
1766     }
1767 
1768     // Fill dead space inside cur_region.
1769     if (_summary_data.is_region_aligned(cur_addr)) {
1770       cur_addr += cur_region_ptr->partial_obj_size();
1771     }
1772 
1773     HeapWord* region_end_addr = _summary_data.region_align_up(cur_addr + 1);
1774     assert(region_end_addr <= end_addr, "inv");
1775     while (cur_addr < region_end_addr) {
1776       // Use end_addr to allow filler-obj to cross region boundary.
1777       HeapWord* live_start = bitmap->find_obj_beg(cur_addr, end_addr);
1778       if (cur_addr != live_start) {
1779         // Found dead space [cur_addr, live_start).
1780         fill_range_in_dense_prefix(cur_addr, live_start);
1781       }
1782       if (live_start >= region_end_addr) {
1783         cur_addr = live_start;
1784         break;
1785       }
1786       assert(bitmap->is_marked(live_start), "inv");
1787       cur_addr = live_start + cast_to_oop(live_start)->size();
1788     }
1789   }
1790 }
1791 
1792 void PSParallelCompact::compact() {
1793   GCTraceTime(Info, gc, phases) tm("Compaction Phase", &_gc_timer);
1794 
1795   uint active_gc_threads = ParallelScavengeHeap::heap()->workers().active_workers();
1796 
1797   initialize_shadow_regions(active_gc_threads);
1798   prepare_region_draining_tasks(active_gc_threads);
1799 
1800   {
1801     GCTraceTime(Trace, gc, phases) tm("Par Compact", &_gc_timer);
1802 
1803     FillDensePrefixAndCompactionTask task(active_gc_threads);
1804     ParallelScavengeHeap::heap()->workers().run_task(&task);
1805 
1806 #ifdef  ASSERT
1807     verify_filler_in_dense_prefix();
1808 
1809     // Verify that all regions have been processed.
1810     for (unsigned int id = old_space_id; id < last_space_id; ++id) {
1811       verify_complete(SpaceId(id));
1812     }
1813 #endif
1814   }
1815 }
1816 
1817 #ifdef  ASSERT
1818 void PSParallelCompact::verify_filler_in_dense_prefix() {
1819   HeapWord* bottom = _space_info[old_space_id].space()->bottom();
1820   HeapWord* dense_prefix_end = dense_prefix(old_space_id);
1821 
1822   const size_t region_size = ParallelCompactData::RegionSize;
1823 
1824   for (HeapWord* cur_addr = bottom; cur_addr < dense_prefix_end; /* empty */) {
1825     RegionData* cur_region_ptr = _summary_data.addr_to_region_ptr(cur_addr);
1826     if (cur_region_ptr->data_size() == region_size) {
1827       // Full; no dead space. Next region.
1828       if (_summary_data.is_region_aligned(cur_addr)) {
1829         cur_addr += region_size;
1830       } else {
1831         cur_addr = _summary_data.region_align_up(cur_addr);
1832       }
1833       continue;
1834     }
1835 
1836     // This region contains filler objs.
1837     if (_summary_data.is_region_aligned(cur_addr)) {
1838       cur_addr += cur_region_ptr->partial_obj_size();
1839     }
1840 
1841     HeapWord* region_end_addr = _summary_data.region_align_up(cur_addr + 1);
1842     assert(region_end_addr <= dense_prefix_end, "inv");
1843 
1844     while (cur_addr < region_end_addr) {
1845       oop obj = cast_to_oop(cur_addr);
1846       oopDesc::verify(obj);
1847       if (!mark_bitmap()->is_marked(cur_addr)) {
1848         assert(CollectedHeap::is_filler_object(cast_to_oop(cur_addr)), "inv");
1849       }
1850       cur_addr += obj->size();
1851     }
1852   }
1853 }
1854 
1855 void PSParallelCompact::verify_complete(SpaceId space_id) {
1856   // All Regions served as compaction targets, from dense_prefix() to
1857   // new_top(), should be marked as filled and all Regions between new_top()
1858   // and top() should be available (i.e., should have been emptied).
1859   ParallelCompactData& sd = summary_data();
1860   SpaceInfo si = _space_info[space_id];
1861   HeapWord* new_top_addr = sd.region_align_up(si.new_top());
1862   HeapWord* old_top_addr = sd.region_align_up(si.space()->top());
1863   const size_t beg_region = sd.addr_to_region_idx(si.dense_prefix());
1864   const size_t new_top_region = sd.addr_to_region_idx(new_top_addr);
1865   const size_t old_top_region = sd.addr_to_region_idx(old_top_addr);
1866 
1867   size_t cur_region;
1868   for (cur_region = beg_region; cur_region < new_top_region; ++cur_region) {
1869     const RegionData* const c = sd.region(cur_region);
1870     assert(c->completed(), "region %zu not filled: destination_count=%u",
1871            cur_region, c->destination_count());
1872   }
1873 
1874   for (cur_region = new_top_region; cur_region < old_top_region; ++cur_region) {
1875     const RegionData* const c = sd.region(cur_region);
1876     assert(c->available(), "region %zu not empty: destination_count=%u",
1877            cur_region, c->destination_count());
1878   }
1879 }
1880 #endif  // #ifdef ASSERT
1881 
1882 // Return the SpaceId for the space containing addr.  If addr is not in the
1883 // heap, last_space_id is returned.  In debug mode it expects the address to be
1884 // in the heap and asserts such.
1885 PSParallelCompact::SpaceId PSParallelCompact::space_id(HeapWord* addr) {
1886   assert(ParallelScavengeHeap::heap()->is_in_reserved(addr), "addr not in the heap");
1887 
1888   for (unsigned int id = old_space_id; id < last_space_id; ++id) {
1889     if (_space_info[id].space()->contains(addr)) {
1890       return SpaceId(id);
1891     }
1892   }
1893 
1894   assert(false, "no space contains the addr");
1895   return last_space_id;
1896 }
1897 
1898 // Skip over count live words starting from beg, and return the address of the
1899 // next live word. Callers must also ensure that there are enough live words in
1900 // the range [beg, end) to skip.
1901 HeapWord* PSParallelCompact::skip_live_words(HeapWord* beg, HeapWord* end, size_t count)
1902 {
1903   ParMarkBitMap* m = mark_bitmap();
1904   HeapWord* cur_addr = beg;
1905   while (true) {
1906     cur_addr = m->find_obj_beg(cur_addr, end);
1907     assert(cur_addr < end, "inv");
1908     size_t obj_size = cast_to_oop(cur_addr)->size();
1909     // Strictly greater-than
1910     if (obj_size > count) {
1911       return cur_addr + count;
1912     }
1913     count -= obj_size;
1914     cur_addr += obj_size;
1915   }
1916 }
1917 
1918 // On starting to fill a destination region (dest-region), we need to know the
1919 // location of the word that will be at the start of the dest-region after
1920 // compaction. A dest-region can have one or more source regions, but only the
1921 // first source-region contains this location. This location is retrieved by
1922 // calling `first_src_addr` on a dest-region.
1923 // Conversely, a source-region has a dest-region which holds the destination of
1924 // the first live word on this source-region, based on which the destination
1925 // for the rest of live words can be derived.
1926 //
1927 // Note:
1928 // There is some complication due to space-boundary-fragmentation (an obj can't
1929 // cross space-boundary) -- a source-region may be split and behave like two
1930 // distinct regions with their own dest-region, as depicted below.
1931 //
1932 // source-region: region-n
1933 //
1934 // **********************
1935 // |     A|A~~~~B|B     |
1936 // **********************
1937 //    n-1     n     n+1
1938 //
1939 // AA, BB denote two live objs. ~~~~ denotes unknown number of live objs.
1940 //
1941 // Assuming the dest-region for region-n is the final region before
1942 // old-space-end and its first-live-word is the middle of AA, the heap content
1943 // will look like the following after compaction:
1944 //
1945 // **************                  *************
1946 //      A|A~~~~ |                  |BB    |
1947 // **************                  *************
1948 //              ^                  ^
1949 //              | old-space-end    | eden-space-start
1950 //
1951 // Therefore, in this example, region-n will have two dest-regions:
1952 // 1. the final region in old-space
1953 // 2. the first region in eden-space.
1954 // To handle this special case, we introduce the concept of split-region, whose
1955 // contents are relocated to two spaces. `SplitInfo` captures all necessary
1956 // info about the split, the first part, spliting-point, and the second part.
1957 HeapWord* PSParallelCompact::first_src_addr(HeapWord* const dest_addr,
1958                                             SpaceId src_space_id,
1959                                             size_t src_region_idx)
1960 {
1961   const size_t RegionSize = ParallelCompactData::RegionSize;
1962   const ParallelCompactData& sd = summary_data();
1963   assert(sd.is_region_aligned(dest_addr), "precondition");
1964 
1965   const RegionData* const src_region_ptr = sd.region(src_region_idx);
1966   assert(src_region_ptr->data_size() > 0, "src region cannot be empty");
1967 
1968   const size_t partial_obj_size = src_region_ptr->partial_obj_size();
1969   HeapWord* const src_region_destination = src_region_ptr->destination();
1970 
1971   HeapWord* const region_start = sd.region_to_addr(src_region_idx);
1972   HeapWord* const region_end = sd.region_to_addr(src_region_idx) + RegionSize;
1973 
1974   // Identify the actual destination for the first live words on this region,
1975   // taking split-region into account.
1976   HeapWord* region_start_destination;
1977   const SplitInfo& split_info = _space_info[src_space_id].split_info();
1978   if (split_info.is_split(src_region_idx)) {
1979     // The second part of this split region; use the recorded split point.
1980     if (dest_addr == src_region_destination) {
1981       return split_info.split_point();
1982     }
1983     region_start_destination = split_info.preceding_destination();
1984   } else {
1985     region_start_destination = src_region_destination;
1986   }
1987 
1988   // Calculate the offset to be skipped
1989   size_t words_to_skip = pointer_delta(dest_addr, region_start_destination);
1990 
1991   HeapWord* result;
1992   if (partial_obj_size > words_to_skip) {
1993     result = region_start + words_to_skip;
1994   } else {
1995     words_to_skip -= partial_obj_size;
1996     result = skip_live_words(region_start + partial_obj_size, region_end, words_to_skip);
1997   }
1998 
1999   if (split_info.is_split(src_region_idx)) {
2000     assert(result < split_info.split_point(), "postcondition");
2001   } else {
2002     assert(result < region_end, "postcondition");
2003   }
2004 
2005   return result;
2006 }
2007 
2008 void PSParallelCompact::decrement_destination_counts(ParCompactionManager* cm,
2009                                                      SpaceId src_space_id,
2010                                                      size_t beg_region,
2011                                                      HeapWord* end_addr)
2012 {
2013   ParallelCompactData& sd = summary_data();
2014 
2015 #ifdef ASSERT
2016   MutableSpace* const src_space = _space_info[src_space_id].space();
2017   HeapWord* const beg_addr = sd.region_to_addr(beg_region);
2018   assert(src_space->contains(beg_addr) || beg_addr == src_space->end(),
2019          "src_space_id does not match beg_addr");
2020   assert(src_space->contains(end_addr) || end_addr == src_space->end(),
2021          "src_space_id does not match end_addr");
2022 #endif // #ifdef ASSERT
2023 
2024   RegionData* const beg = sd.region(beg_region);
2025   RegionData* const end = sd.addr_to_region_ptr(sd.region_align_up(end_addr));
2026 
2027   // Regions up to new_top() are enqueued if they become available.
2028   HeapWord* const new_top = _space_info[src_space_id].new_top();
2029   RegionData* const enqueue_end =
2030     sd.addr_to_region_ptr(sd.region_align_up(new_top));
2031 
2032   for (RegionData* cur = beg; cur < end; ++cur) {
2033     assert(cur->data_size() > 0, "region must have live data");
2034     cur->decrement_destination_count();
2035     if (cur < enqueue_end && cur->available() && cur->claim()) {
2036       if (cur->mark_normal()) {
2037         cm->push_region(sd.region(cur));
2038       } else if (cur->mark_copied()) {
2039         // Try to copy the content of the shadow region back to its corresponding
2040         // heap region if the shadow region is filled. Otherwise, the GC thread
2041         // fills the shadow region will copy the data back (see
2042         // MoveAndUpdateShadowClosure::complete_region).
2043         copy_back(sd.region_to_addr(cur->shadow_region()), sd.region_to_addr(cur));
2044         ParCompactionManager::push_shadow_region_mt_safe(cur->shadow_region());
2045         cur->set_completed();
2046       }
2047     }
2048   }
2049 }
2050 
2051 size_t PSParallelCompact::next_src_region(MoveAndUpdateClosure& closure,
2052                                           SpaceId& src_space_id,
2053                                           HeapWord*& src_space_top,
2054                                           HeapWord* end_addr)
2055 {
2056   ParallelCompactData& sd = PSParallelCompact::summary_data();
2057 
2058   size_t src_region_idx = 0;
2059 
2060   // Skip empty regions (if any) up to the top of the space.
2061   HeapWord* const src_aligned_up = sd.region_align_up(end_addr);
2062   RegionData* src_region_ptr = sd.addr_to_region_ptr(src_aligned_up);
2063   HeapWord* const top_aligned_up = sd.region_align_up(src_space_top);
2064   const RegionData* const top_region_ptr = sd.addr_to_region_ptr(top_aligned_up);
2065 
2066   while (src_region_ptr < top_region_ptr && src_region_ptr->data_size() == 0) {
2067     ++src_region_ptr;
2068   }
2069 
2070   if (src_region_ptr < top_region_ptr) {
2071     // Found the first non-empty region in the same space.
2072     src_region_idx = sd.region(src_region_ptr);
2073     closure.set_source(sd.region_to_addr(src_region_idx));
2074     return src_region_idx;
2075   }
2076 
2077   // Switch to a new source space and find the first non-empty region.
2078   uint space_id = src_space_id + 1;
2079   assert(space_id < last_space_id, "not enough spaces");
2080 
2081   for (/* empty */; space_id < last_space_id; ++space_id) {
2082     HeapWord* bottom = _space_info[space_id].space()->bottom();
2083     HeapWord* top = _space_info[space_id].space()->top();
2084     // Skip empty space
2085     if (bottom == top) {
2086       continue;
2087     }
2088 
2089     // Identify the first region that contains live words in this space
2090     size_t cur_region = sd.addr_to_region_idx(bottom);
2091     size_t end_region = sd.addr_to_region_idx(sd.region_align_up(top));
2092 
2093     for (/* empty */ ; cur_region < end_region; ++cur_region) {
2094       RegionData* cur = sd.region(cur_region);
2095       if (cur->live_obj_size() > 0) {
2096         HeapWord* region_start_addr = sd.region_to_addr(cur_region);
2097 
2098         src_space_id = SpaceId(space_id);
2099         src_space_top = top;
2100         closure.set_source(region_start_addr);
2101         return cur_region;
2102       }
2103     }
2104   }
2105 
2106   ShouldNotReachHere();
2107 }
2108 
2109 HeapWord* PSParallelCompact::partial_obj_end(HeapWord* region_start_addr) {
2110   ParallelCompactData& sd = summary_data();
2111   assert(sd.is_region_aligned(region_start_addr), "precondition");
2112 
2113   // Use per-region partial_obj_size to locate the end of the obj, that extends
2114   // to region_start_addr.
2115   size_t start_region_idx = sd.addr_to_region_idx(region_start_addr);
2116   size_t end_region_idx = sd.region_count();
2117   size_t accumulated_size = 0;
2118   for (size_t region_idx = start_region_idx; region_idx < end_region_idx; ++region_idx) {
2119     size_t cur_partial_obj_size = sd.region(region_idx)->partial_obj_size();
2120     accumulated_size += cur_partial_obj_size;
2121     if (cur_partial_obj_size != ParallelCompactData::RegionSize) {
2122       break;
2123     }
2124   }
2125   return region_start_addr + accumulated_size;
2126 }
2127 
2128 // Use region_idx as the destination region, and evacuate all live objs on its
2129 // source regions to this destination region.
2130 void PSParallelCompact::fill_region(ParCompactionManager* cm, MoveAndUpdateClosure& closure, size_t region_idx)
2131 {
2132   ParMarkBitMap* const bitmap = mark_bitmap();
2133   ParallelCompactData& sd = summary_data();
2134   RegionData* const region_ptr = sd.region(region_idx);
2135 
2136   // Get the source region and related info.
2137   size_t src_region_idx = region_ptr->source_region();
2138   SpaceId src_space_id = space_id(sd.region_to_addr(src_region_idx));
2139   HeapWord* src_space_top = _space_info[src_space_id].space()->top();
2140   HeapWord* dest_addr = sd.region_to_addr(region_idx);
2141 
2142   closure.set_source(first_src_addr(dest_addr, src_space_id, src_region_idx));
2143 
2144   // Adjust src_region_idx to prepare for decrementing destination counts (the
2145   // destination count is not decremented when a region is copied to itself).
2146   if (src_region_idx == region_idx) {
2147     src_region_idx += 1;
2148   }
2149 
2150   // source-region:
2151   //
2152   // **********
2153   // |   ~~~  |
2154   // **********
2155   //      ^
2156   //      |-- closure.source() / first_src_addr
2157   //
2158   //
2159   // ~~~ : live words
2160   //
2161   // destination-region:
2162   //
2163   // **********
2164   // |        |
2165   // **********
2166   // ^
2167   // |-- region-start
2168   if (bitmap->is_unmarked(closure.source())) {
2169     // An object overflows the previous destination region, so this
2170     // destination region should copy the remainder of the object or as much as
2171     // will fit.
2172     HeapWord* const old_src_addr = closure.source();
2173     {
2174       HeapWord* region_start = sd.region_align_down(closure.source());
2175       HeapWord* obj_start = bitmap->find_obj_beg_reverse(region_start, closure.source());
2176       HeapWord* obj_end;
2177       if (obj_start != closure.source()) {
2178         assert(bitmap->is_marked(obj_start), "inv");
2179         // Found the actual obj-start, try to find the obj-end using either
2180         // size() if this obj is completely contained in the current region.
2181         HeapWord* next_region_start = region_start + ParallelCompactData::RegionSize;
2182         HeapWord* partial_obj_start = (next_region_start >= src_space_top)
2183                                       ? nullptr
2184                                       : sd.addr_to_region_ptr(next_region_start)->partial_obj_addr();
2185         // This obj extends to next region iff partial_obj_addr of the *next*
2186         // region is the same as obj-start.
2187         if (partial_obj_start == obj_start) {
2188           // This obj extends to next region.
2189           obj_end = partial_obj_end(next_region_start);
2190         } else {
2191           // Completely contained in this region; safe to use size().
2192           obj_end = obj_start + cast_to_oop(obj_start)->size();
2193         }
2194       } else {
2195         // This obj extends to current region.
2196         obj_end = partial_obj_end(region_start);
2197       }
2198       size_t partial_obj_size = pointer_delta(obj_end, closure.source());
2199       closure.copy_partial_obj(partial_obj_size);
2200     }
2201 
2202     if (closure.is_full()) {
2203       decrement_destination_counts(cm, src_space_id, src_region_idx, closure.source());
2204       closure.complete_region(dest_addr, region_ptr);
2205       return;
2206     }
2207 
2208     // Finished copying without using up the current destination-region
2209     HeapWord* const end_addr = sd.region_align_down(closure.source());
2210     if (sd.region_align_down(old_src_addr) != end_addr) {
2211       assert(sd.region_align_up(old_src_addr) == end_addr, "only one region");
2212       // The partial object was copied from more than one source region.
2213       decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr);
2214 
2215       // Move to the next source region, possibly switching spaces as well.  All
2216       // args except end_addr may be modified.
2217       src_region_idx = next_src_region(closure, src_space_id, src_space_top, end_addr);
2218     }
2219   }
2220 
2221   // Handle the rest obj-by-obj, where we know obj-start.
2222   do {
2223     HeapWord* cur_addr = closure.source();
2224     HeapWord* const end_addr = MIN2(sd.region_align_up(cur_addr + 1),
2225                                     src_space_top);
2226     // To handle the case where the final obj in source region extends to next region.
2227     HeapWord* final_obj_start = (end_addr == src_space_top)
2228                                 ? nullptr
2229                                 : sd.addr_to_region_ptr(end_addr)->partial_obj_addr();
2230     // Apply closure on objs inside [cur_addr, end_addr)
2231     do {
2232       cur_addr = bitmap->find_obj_beg(cur_addr, end_addr);
2233       if (cur_addr == end_addr) {
2234         break;
2235       }
2236       size_t obj_size;
2237       if (final_obj_start == cur_addr) {
2238         obj_size = pointer_delta(partial_obj_end(end_addr), cur_addr);
2239       } else {
2240         // This obj doesn't extend into next region; size() is safe to use.
2241         obj_size = cast_to_oop(cur_addr)->size();
2242       }
2243       closure.do_addr(cur_addr, obj_size);
2244       cur_addr += obj_size;
2245     } while (cur_addr < end_addr && !closure.is_full());
2246 
2247     if (closure.is_full()) {
2248       decrement_destination_counts(cm, src_space_id, src_region_idx, closure.source());
2249       closure.complete_region(dest_addr, region_ptr);
2250       return;
2251     }
2252 
2253     decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr);
2254 
2255     // Move to the next source region, possibly switching spaces as well.  All
2256     // args except end_addr may be modified.
2257     src_region_idx = next_src_region(closure, src_space_id, src_space_top, end_addr);
2258   } while (true);
2259 }
2260 
2261 void PSParallelCompact::fill_and_update_region(ParCompactionManager* cm, size_t region_idx)
2262 {
2263   MoveAndUpdateClosure cl(mark_bitmap(), region_idx);
2264   fill_region(cm, cl, region_idx);
2265 }
2266 
2267 void PSParallelCompact::fill_and_update_shadow_region(ParCompactionManager* cm, size_t region_idx)
2268 {
2269   // Get a shadow region first
2270   ParallelCompactData& sd = summary_data();
2271   RegionData* const region_ptr = sd.region(region_idx);
2272   size_t shadow_region = ParCompactionManager::pop_shadow_region_mt_safe(region_ptr);
2273   // The InvalidShadow return value indicates the corresponding heap region is available,
2274   // so use MoveAndUpdateClosure to fill the normal region. Otherwise, use
2275   // MoveAndUpdateShadowClosure to fill the acquired shadow region.
2276   if (shadow_region == ParCompactionManager::InvalidShadow) {
2277     MoveAndUpdateClosure cl(mark_bitmap(), region_idx);
2278     region_ptr->shadow_to_normal();
2279     return fill_region(cm, cl, region_idx);
2280   } else {
2281     MoveAndUpdateShadowClosure cl(mark_bitmap(), region_idx, shadow_region);
2282     return fill_region(cm, cl, region_idx);
2283   }
2284 }
2285 
2286 void PSParallelCompact::copy_back(HeapWord *shadow_addr, HeapWord *region_addr)
2287 {
2288   Copy::aligned_conjoint_words(shadow_addr, region_addr, _summary_data.RegionSize);
2289 }
2290 
2291 bool PSParallelCompact::steal_unavailable_region(ParCompactionManager* cm, size_t &region_idx)
2292 {
2293   size_t next = cm->next_shadow_region();
2294   ParallelCompactData& sd = summary_data();
2295   size_t old_new_top = sd.addr_to_region_idx(_space_info[old_space_id].new_top());
2296   uint active_gc_threads = ParallelScavengeHeap::heap()->workers().active_workers();
2297 
2298   while (next < old_new_top) {
2299     if (sd.region(next)->mark_shadow()) {
2300       region_idx = next;
2301       return true;
2302     }
2303     next = cm->move_next_shadow_region_by(active_gc_threads);
2304   }
2305 
2306   return false;
2307 }
2308 
2309 // The shadow region is an optimization to address region dependencies in full GC. The basic
2310 // idea is making more regions available by temporally storing their live objects in empty
2311 // shadow regions to resolve dependencies between them and the destination regions. Therefore,
2312 // GC threads need not wait destination regions to be available before processing sources.
2313 //
2314 // A typical workflow would be:
2315 // After draining its own stack and failing to steal from others, a GC worker would pick an
2316 // unavailable region (destination count > 0) and get a shadow region. Then the worker fills
2317 // the shadow region by copying live objects from source regions of the unavailable one. Once
2318 // the unavailable region becomes available, the data in the shadow region will be copied back.
2319 // Shadow regions are empty regions in the to-space and regions between top and end of other spaces.
2320 void PSParallelCompact::initialize_shadow_regions(uint parallel_gc_threads)
2321 {
2322   const ParallelCompactData& sd = PSParallelCompact::summary_data();
2323 
2324   for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2325     SpaceInfo* const space_info = _space_info + id;
2326     MutableSpace* const space = space_info->space();
2327 
2328     const size_t beg_region =
2329       sd.addr_to_region_idx(sd.region_align_up(MAX2(space_info->new_top(), space->top())));
2330     const size_t end_region =
2331       sd.addr_to_region_idx(sd.region_align_down(space->end()));
2332 
2333     for (size_t cur = beg_region; cur < end_region; ++cur) {
2334       ParCompactionManager::push_shadow_region(cur);
2335     }
2336   }
2337 
2338   size_t beg_region = sd.addr_to_region_idx(_space_info[old_space_id].dense_prefix());
2339   for (uint i = 0; i < parallel_gc_threads; i++) {
2340     ParCompactionManager *cm = ParCompactionManager::gc_thread_compaction_manager(i);
2341     cm->set_next_shadow_region(beg_region + i);
2342   }
2343 }
2344 
2345 void MoveAndUpdateClosure::copy_partial_obj(size_t partial_obj_size)
2346 {
2347   size_t words = MIN2(partial_obj_size, words_remaining());
2348 
2349   // This test is necessary; if omitted, the pointer updates to a partial object
2350   // that crosses the dense prefix boundary could be overwritten.
2351   if (source() != copy_destination()) {
2352     DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
2353     Copy::aligned_conjoint_words(source(), copy_destination(), words);
2354   }
2355   update_state(words);
2356 }
2357 
2358 void MoveAndUpdateClosure::complete_region(HeapWord* dest_addr, PSParallelCompact::RegionData* region_ptr) {
2359   assert(region_ptr->shadow_state() == ParallelCompactData::RegionData::NormalRegion, "Region should be finished");
2360   region_ptr->set_completed();
2361 }
2362 
2363 void MoveAndUpdateClosure::do_addr(HeapWord* addr, size_t words) {
2364   assert(destination() != nullptr, "sanity");
2365   _source = addr;
2366 
2367   // The start_array must be updated even if the object is not moving.
2368   if (_start_array != nullptr) {
2369     _start_array->update_for_block(destination(), destination() + words);
2370   }
2371 
2372   // Avoid overflow
2373   words = MIN2(words, words_remaining());
2374   assert(words > 0, "inv");
2375 
2376   if (copy_destination() != source()) {
2377     DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
2378     assert(source() != destination(), "inv");
2379     assert(FullGCForwarding::is_forwarded(cast_to_oop(source())), "inv");
2380     assert(FullGCForwarding::forwardee(cast_to_oop(source())) == cast_to_oop(destination()), "inv");
2381     Copy::aligned_conjoint_words(source(), copy_destination(), words);
2382     cast_to_oop(copy_destination())->init_mark();
2383   }
2384 
2385   update_state(words);
2386 }
2387 
2388 void MoveAndUpdateShadowClosure::complete_region(HeapWord* dest_addr, PSParallelCompact::RegionData* region_ptr) {
2389   assert(region_ptr->shadow_state() == ParallelCompactData::RegionData::ShadowRegion, "Region should be shadow");
2390   // Record the shadow region index
2391   region_ptr->set_shadow_region(_shadow);
2392   // Mark the shadow region as filled to indicate the data is ready to be
2393   // copied back
2394   region_ptr->mark_filled();
2395   // Try to copy the content of the shadow region back to its corresponding
2396   // heap region if available; the GC thread that decreases the destination
2397   // count to zero will do the copying otherwise (see
2398   // PSParallelCompact::decrement_destination_counts).
2399   if (((region_ptr->available() && region_ptr->claim()) || region_ptr->claimed()) && region_ptr->mark_copied()) {
2400     region_ptr->set_completed();
2401     PSParallelCompact::copy_back(PSParallelCompact::summary_data().region_to_addr(_shadow), dest_addr);
2402     ParCompactionManager::push_shadow_region_mt_safe(_shadow);
2403   }
2404 }