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