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