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