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/flatArrayKlass.inline.hpp"
  82 #include "oops/instanceClassLoaderKlass.inline.hpp"
  83 #include "oops/instanceKlass.inline.hpp"
  84 #include "oops/instanceMirrorKlass.inline.hpp"
  85 #include "oops/methodData.hpp"
  86 #include "oops/objArrayKlass.inline.hpp"
  87 #include "oops/oop.inline.hpp"
  88 #include "runtime/arguments.hpp"
  89 #include "runtime/handles.inline.hpp"
  90 #include "runtime/java.hpp"
  91 #include "runtime/safepoint.hpp"
  92 #include "runtime/threads.hpp"
  93 #include "runtime/vmThread.hpp"
  94 #include "services/memoryService.hpp"
  95 #include "utilities/align.hpp"
  96 #include "utilities/debug.hpp"
  97 #include "utilities/events.hpp"
  98 #include "utilities/formatBuffer.hpp"
  99 #include "utilities/macros.hpp"
 100 #include "utilities/stack.inline.hpp"
 101 #if INCLUDE_JVMCI
 102 #include "jvmci/jvmci.hpp"
 103 #endif
 104 
 105 #include <math.h>
 106 
 107 // All sizes are in HeapWords.
 108 const size_t ParallelCompactData::Log2RegionSize  = 16; // 64K words
 109 const size_t ParallelCompactData::RegionSize      = (size_t)1 << Log2RegionSize;
 110 static_assert(ParallelCompactData::RegionSize >= BitsPerWord, "region-start bit word-aligned");
 111 const size_t ParallelCompactData::RegionSizeBytes =
 112   RegionSize << LogHeapWordSize;
 113 const size_t ParallelCompactData::RegionSizeOffsetMask = RegionSize - 1;
 114 const size_t ParallelCompactData::RegionAddrOffsetMask = RegionSizeBytes - 1;
 115 const size_t ParallelCompactData::RegionAddrMask       = ~RegionAddrOffsetMask;
 116 
 117 const ParallelCompactData::RegionData::region_sz_t
 118 ParallelCompactData::RegionData::dc_shift = 27;
 119 
 120 const ParallelCompactData::RegionData::region_sz_t
 121 ParallelCompactData::RegionData::dc_mask = ~0U << dc_shift;
 122 
 123 const ParallelCompactData::RegionData::region_sz_t
 124 ParallelCompactData::RegionData::dc_one = 0x1U << dc_shift;
 125 
 126 const ParallelCompactData::RegionData::region_sz_t
 127 ParallelCompactData::RegionData::los_mask = ~dc_mask;
 128 
 129 const ParallelCompactData::RegionData::region_sz_t
 130 ParallelCompactData::RegionData::dc_claimed = 0x8U << dc_shift;
 131 
 132 const ParallelCompactData::RegionData::region_sz_t
 133 ParallelCompactData::RegionData::dc_completed = 0xcU << dc_shift;
 134 
 135 bool ParallelCompactData::RegionData::is_clear() {
 136   return (_destination == nullptr) &&
 137          (_source_region == 0) &&
 138          (_partial_obj_addr == nullptr) &&
 139          (_partial_obj_size == 0) &&
 140          (dc_and_los() == 0) &&
 141          (shadow_state() == 0);
 142 }
 143 
 144 #ifdef ASSERT
 145 void ParallelCompactData::RegionData::verify_clear() {
 146   assert(_destination == nullptr, "inv");
 147   assert(_source_region == 0, "inv");
 148   assert(_partial_obj_addr == nullptr, "inv");
 149   assert(_partial_obj_size == 0, "inv");
 150   assert(dc_and_los() == 0, "inv");
 151   assert(shadow_state() == 0, "inv");
 152 }
 153 #endif
 154 
 155 SpaceInfo PSParallelCompact::_space_info[PSParallelCompact::last_space_id];
 156 
 157 SpanSubjectToDiscoveryClosure PSParallelCompact::_span_based_discoverer;
 158 ReferenceProcessor* PSParallelCompact::_ref_processor = nullptr;
 159 
 160 void SplitInfo::record(size_t split_region_idx, HeapWord* split_point, size_t preceding_live_words) {
 161   assert(split_region_idx != 0, "precondition");
 162 
 163   // Obj denoted by split_point will be deferred to the next space.
 164   assert(split_point != nullptr, "precondition");
 165 
 166   const ParallelCompactData& sd = PSParallelCompact::summary_data();
 167 
 168   PSParallelCompact::RegionData* split_region_ptr = sd.region(split_region_idx);
 169   assert(preceding_live_words < split_region_ptr->data_size(), "inv");
 170 
 171   HeapWord* preceding_destination = split_region_ptr->destination();
 172   assert(preceding_destination != nullptr, "inv");
 173 
 174   // How many regions does the preceding part occupy
 175   uint preceding_destination_count;
 176   if (preceding_live_words == 0) {
 177     preceding_destination_count = 0;
 178   } else {
 179     // -1 so that the ending address doesn't fall on the region-boundary
 180     if (sd.region_align_down(preceding_destination) ==
 181         sd.region_align_down(preceding_destination + preceding_live_words - 1)) {
 182       preceding_destination_count = 1;
 183     } else {
 184       preceding_destination_count = 2;
 185     }
 186   }
 187 
 188   _split_region_idx = split_region_idx;
 189   _split_point = split_point;
 190   _preceding_live_words = preceding_live_words;
 191   _preceding_destination = preceding_destination;
 192   _preceding_destination_count = preceding_destination_count;
 193 }
 194 
 195 void SplitInfo::clear()
 196 {
 197   _split_region_idx = 0;
 198   _split_point = nullptr;
 199   _preceding_live_words = 0;
 200   _preceding_destination = nullptr;
 201   _preceding_destination_count = 0;
 202   assert(!is_valid(), "sanity");
 203 }
 204 
 205 #ifdef  ASSERT
 206 void SplitInfo::verify_clear()
 207 {
 208   assert(_split_region_idx == 0, "not clear");
 209   assert(_split_point == nullptr, "not clear");
 210   assert(_preceding_live_words == 0, "not clear");
 211   assert(_preceding_destination == nullptr, "not clear");
 212   assert(_preceding_destination_count == 0, "not clear");
 213 }
 214 #endif  // #ifdef ASSERT
 215 
 216 
 217 void PSParallelCompact::print_on(outputStream* st) {
 218   _mark_bitmap.print_on(st);
 219 }
 220 
 221 ParallelCompactData::ParallelCompactData() :
 222   _heap_start(nullptr),
 223   DEBUG_ONLY(_heap_end(nullptr) COMMA)
 224   _region_vspace(nullptr),
 225   _reserved_byte_size(0),
 226   _region_data(nullptr),
 227   _region_count(0) {}
 228 
 229 bool ParallelCompactData::initialize(MemRegion reserved_heap)
 230 {
 231   _heap_start = reserved_heap.start();
 232   const size_t heap_size = reserved_heap.word_size();
 233   DEBUG_ONLY(_heap_end = _heap_start + heap_size;)
 234 
 235   assert(region_align_down(_heap_start) == _heap_start,
 236          "region start not aligned");
 237   assert(is_aligned(heap_size, RegionSize), "precondition");
 238 
 239   const size_t count = heap_size >> Log2RegionSize;
 240   const size_t raw_bytes = count * sizeof(RegionData);
 241   const size_t page_sz = os::page_size_for_region_aligned(raw_bytes, 10);
 242   const size_t granularity = os::vm_allocation_granularity();
 243   const size_t rs_align = MAX2(page_sz, granularity);
 244 
 245   _reserved_byte_size = align_up(raw_bytes, rs_align);
 246 
 247   ReservedSpace rs = MemoryReserver::reserve(_reserved_byte_size,
 248                                              rs_align,
 249                                              page_sz,
 250                                              mtGC);
 251 
 252   if (!rs.is_reserved()) {
 253     // Failed to reserve memory.
 254     return false;
 255   }
 256 
 257   os::trace_page_sizes("Parallel Compact Data", raw_bytes, raw_bytes, rs.base(),
 258                        rs.size(), page_sz);
 259 
 260   MemTracker::record_virtual_memory_tag(rs, mtGC);
 261 
 262   PSVirtualSpace* region_vspace = new PSVirtualSpace(rs, page_sz);
 263 
 264   if (!region_vspace->expand_by(_reserved_byte_size)) {
 265     // Failed to commit memory.
 266 
 267     delete region_vspace;
 268 
 269     // Release memory reserved in the space.
 270     MemoryReserver::release(rs);
 271 
 272     return false;
 273   }
 274 
 275   _region_vspace = region_vspace;
 276   _region_data = (RegionData*)_region_vspace->reserved_low_addr();
 277   _region_count = count;
 278   return true;
 279 }
 280 
 281 void ParallelCompactData::clear_range(size_t beg_region, size_t end_region) {
 282   assert(beg_region <= _region_count, "beg_region out of range");
 283   assert(end_region <= _region_count, "end_region out of range");
 284 
 285   const size_t region_cnt = end_region - beg_region;
 286   for (size_t i = beg_region; i < end_region; i++) {
 287     ::new (&_region_data[i]) RegionData{};
 288   }
 289 }
 290 
 291 // The total live words on src_region would overflow the target space, so find
 292 // the overflowing object and record the split point. The invariant is that an
 293 // obj should not cross space boundary.
 294 HeapWord* ParallelCompactData::summarize_split_space(size_t src_region,
 295                                                      SplitInfo& split_info,
 296                                                      HeapWord* const destination,
 297                                                      HeapWord* const target_end,
 298                                                      HeapWord** target_next) {
 299   assert(destination <= target_end, "sanity");
 300   assert(destination + _region_data[src_region].data_size() > target_end,
 301     "region should not fit into target space");
 302   assert(is_region_aligned(target_end), "sanity");
 303 
 304   size_t partial_obj_size = _region_data[src_region].partial_obj_size();
 305 
 306   if (destination + partial_obj_size > target_end) {
 307     assert(partial_obj_size > 0, "inv");
 308     // The overflowing obj is from a previous region.
 309     //
 310     // source-regions:
 311     //
 312     // ***************
 313     // |     A|AA    |
 314     // ***************
 315     //       ^
 316     //       | split-point
 317     //
 318     // dest-region:
 319     //
 320     // ********
 321     // |~~~~A |
 322     // ********
 323     //       ^^
 324     //       || target-space-end
 325     //       |
 326     //       | destination
 327     //
 328     // AAA would overflow target-space.
 329     //
 330     HeapWord* overflowing_obj = _region_data[src_region].partial_obj_addr();
 331     size_t split_region = addr_to_region_idx(overflowing_obj);
 332 
 333     // The number of live words before the overflowing object on this split region
 334     size_t preceding_live_words;
 335     if (is_region_aligned(overflowing_obj)) {
 336       preceding_live_words = 0;
 337     } else {
 338       // Words accounted by the overflowing object on the split region
 339       size_t overflowing_size = pointer_delta(region_align_up(overflowing_obj), overflowing_obj);
 340       preceding_live_words = region(split_region)->data_size() - overflowing_size;
 341     }
 342 
 343     split_info.record(split_region, overflowing_obj, preceding_live_words);
 344 
 345     // The [overflowing_obj, src_region_start) part has been accounted for, so
 346     // must move back the new_top, now that this overflowing obj is deferred.
 347     HeapWord* new_top = destination - pointer_delta(region_to_addr(src_region), overflowing_obj);
 348 
 349     // If the overflowing obj was relocated to its original destination,
 350     // those destination regions would have their source_region set. Now that
 351     // this overflowing obj is relocated somewhere else, reset the
 352     // source_region.
 353     {
 354       size_t range_start = addr_to_region_idx(region_align_up(new_top));
 355       size_t range_end = addr_to_region_idx(region_align_up(destination));
 356       for (size_t i = range_start; i < range_end; ++i) {
 357         region(i)->set_source_region(0);
 358       }
 359     }
 360 
 361     // Update new top of target space
 362     *target_next = new_top;
 363 
 364     return overflowing_obj;
 365   }
 366 
 367   // Obj-iteration to locate the overflowing obj
 368   HeapWord* region_start = region_to_addr(src_region);
 369   HeapWord* region_end = region_start + RegionSize;
 370   HeapWord* cur_addr = region_start + partial_obj_size;
 371   size_t live_words = partial_obj_size;
 372 
 373   while (true) {
 374     assert(cur_addr < region_end, "inv");
 375     cur_addr = PSParallelCompact::mark_bitmap()->find_obj_beg(cur_addr, region_end);
 376     // There must be an overflowing obj in this region
 377     assert(cur_addr < region_end, "inv");
 378 
 379     oop obj = cast_to_oop(cur_addr);
 380     size_t obj_size = obj->size();
 381     if (destination + live_words + obj_size > target_end) {
 382       // Found the overflowing obj
 383       split_info.record(src_region, cur_addr, live_words);
 384       *target_next = destination + live_words;
 385       return cur_addr;
 386     }
 387 
 388     live_words += obj_size;
 389     cur_addr += obj_size;
 390   }
 391 }
 392 
 393 size_t ParallelCompactData::live_words_in_space(const MutableSpace* space,
 394                                                 HeapWord** full_region_prefix_end) {
 395   size_t cur_region = addr_to_region_idx(space->bottom());
 396   const size_t end_region = addr_to_region_idx(region_align_up(space->top()));
 397   size_t live_words = 0;
 398   if (full_region_prefix_end == nullptr) {
 399     for (/* empty */; cur_region < end_region; ++cur_region) {
 400       live_words += _region_data[cur_region].data_size();
 401     }
 402   } else {
 403     bool first_set = false;
 404     for (/* empty */; cur_region < end_region; ++cur_region) {
 405       size_t live_words_in_region = _region_data[cur_region].data_size();
 406       if (!first_set && live_words_in_region < RegionSize) {
 407         *full_region_prefix_end = region_to_addr(cur_region);
 408         first_set = true;
 409       }
 410       live_words += live_words_in_region;
 411     }
 412     if (!first_set) {
 413       // All regions are full of live objs.
 414       assert(is_region_aligned(space->top()), "inv");
 415       *full_region_prefix_end = space->top();
 416     }
 417     assert(*full_region_prefix_end != nullptr, "postcondition");
 418     assert(is_region_aligned(*full_region_prefix_end), "inv");
 419     assert(*full_region_prefix_end >= space->bottom(), "in-range");
 420     assert(*full_region_prefix_end <= space->top(), "in-range");
 421   }
 422   return live_words;
 423 }
 424 
 425 bool ParallelCompactData::summarize(SplitInfo& split_info,
 426                                     HeapWord* source_beg, HeapWord* source_end,
 427                                     HeapWord** source_next,
 428                                     HeapWord* target_beg, HeapWord* target_end,
 429                                     HeapWord** target_next)
 430 {
 431   HeapWord* const source_next_val = source_next == nullptr ? nullptr : *source_next;
 432   log_develop_trace(gc, compaction)(
 433       "sb=" PTR_FORMAT " se=" PTR_FORMAT " sn=" PTR_FORMAT
 434       "tb=" PTR_FORMAT " te=" PTR_FORMAT " tn=" PTR_FORMAT,
 435       p2i(source_beg), p2i(source_end), p2i(source_next_val),
 436       p2i(target_beg), p2i(target_end), p2i(*target_next));
 437 
 438   size_t cur_region = addr_to_region_idx(source_beg);
 439   const size_t end_region = addr_to_region_idx(region_align_up(source_end));
 440 
 441   HeapWord *dest_addr = target_beg;
 442   for (/* empty */; cur_region < end_region; cur_region++) {
 443     size_t words = _region_data[cur_region].data_size();
 444 
 445     // Skip empty ones
 446     if (words == 0) {
 447       continue;
 448     }
 449 
 450     if (split_info.is_split(cur_region)) {
 451       assert(words > split_info.preceding_live_words(), "inv");
 452       words -= split_info.preceding_live_words();
 453     }
 454 
 455     _region_data[cur_region].set_destination(dest_addr);
 456 
 457     // If cur_region does not fit entirely into the target space, find a point
 458     // at which the source space can be 'split' so that part is copied to the
 459     // target space and the rest is copied elsewhere.
 460     if (dest_addr + words > target_end) {
 461       assert(source_next != nullptr, "source_next is null when splitting");
 462       *source_next = summarize_split_space(cur_region, split_info, dest_addr,
 463                                            target_end, target_next);
 464       return false;
 465     }
 466 
 467     uint destination_count = split_info.is_split(cur_region)
 468                              ? split_info.preceding_destination_count()
 469                              : 0;
 470 
 471     HeapWord* const last_addr = dest_addr + words - 1;
 472     const size_t dest_region_1 = addr_to_region_idx(dest_addr);
 473     const size_t dest_region_2 = addr_to_region_idx(last_addr);
 474 
 475     // Initially assume that the destination regions will be the same and
 476     // adjust the value below if necessary.  Under this assumption, if
 477     // cur_region == dest_region_2, then cur_region will be compacted
 478     // completely into itself.
 479     destination_count += cur_region == dest_region_2 ? 0 : 1;
 480     if (dest_region_1 != dest_region_2) {
 481       // Destination regions differ; adjust destination_count.
 482       destination_count += 1;
 483       // Data from cur_region will be copied to the start of dest_region_2.
 484       _region_data[dest_region_2].set_source_region(cur_region);
 485     } else if (is_region_aligned(dest_addr)) {
 486       // Data from cur_region will be copied to the start of the destination
 487       // region.
 488       _region_data[dest_region_1].set_source_region(cur_region);
 489     }
 490 
 491     _region_data[cur_region].set_destination_count(destination_count);
 492     dest_addr += words;
 493   }
 494 
 495   *target_next = dest_addr;
 496   return true;
 497 }
 498 
 499 #ifdef ASSERT
 500 void ParallelCompactData::verify_clear() {
 501   for (uint cur_idx = 0; cur_idx < region_count(); ++cur_idx) {
 502     if (!region(cur_idx)->is_clear()) {
 503       log_warning(gc)("Uncleared Region: %u", cur_idx);
 504       region(cur_idx)->verify_clear();
 505     }
 506   }
 507 }
 508 #endif  // #ifdef ASSERT
 509 
 510 STWGCTimer          PSParallelCompact::_gc_timer;
 511 ParallelOldTracer   PSParallelCompact::_gc_tracer;
 512 elapsedTimer        PSParallelCompact::_accumulated_time;
 513 unsigned int        PSParallelCompact::_maximum_compaction_gc_num = 0;
 514 CollectorCounters*  PSParallelCompact::_counters = nullptr;
 515 ParMarkBitMap       PSParallelCompact::_mark_bitmap;
 516 ParallelCompactData PSParallelCompact::_summary_data;
 517 
 518 PSParallelCompact::IsAliveClosure PSParallelCompact::_is_alive_closure;
 519 
 520 class PCAdjustPointerClosure: public BasicOopIterateClosure {
 521   template <typename T>
 522   void do_oop_work(T* p) { PSParallelCompact::adjust_pointer(p); }
 523 
 524 public:
 525   virtual void do_oop(oop* p)                { do_oop_work(p); }
 526   virtual void do_oop(narrowOop* p)          { do_oop_work(p); }
 527 
 528   virtual ReferenceIterationMode reference_iteration_mode() { return DO_FIELDS; }
 529 };
 530 
 531 static PCAdjustPointerClosure pc_adjust_pointer_closure;
 532 
 533 bool PSParallelCompact::IsAliveClosure::do_object_b(oop p) { return mark_bitmap()->is_marked(p); }
 534 
 535 void PSParallelCompact::post_initialize() {
 536   ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
 537   _span_based_discoverer.set_span(heap->reserved_region());
 538   _ref_processor =
 539     new ReferenceProcessor(&_span_based_discoverer,
 540                            ParallelGCThreads,   // mt processing degree
 541                            ParallelGCThreads,   // mt discovery degree
 542                            false,               // concurrent_discovery
 543                            &_is_alive_closure); // non-header is alive closure
 544 
 545   _counters = new CollectorCounters("Parallel full collection pauses", 1);
 546 
 547   // Initialize static fields in ParCompactionManager.
 548   ParCompactionManager::initialize(mark_bitmap());
 549 }
 550 
 551 bool PSParallelCompact::initialize_aux_data() {
 552   ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
 553   MemRegion mr = heap->reserved_region();
 554   assert(mr.byte_size() != 0, "heap should be reserved");
 555 
 556   initialize_space_info();
 557 
 558   if (!_mark_bitmap.initialize(mr)) {
 559     vm_shutdown_during_initialization(
 560       err_msg("Unable to allocate %zuKB bitmaps for parallel "
 561       "garbage collection for the requested %zuKB heap.",
 562       _mark_bitmap.reserved_byte_size()/K, mr.byte_size()/K));
 563     return false;
 564   }
 565 
 566   if (!_summary_data.initialize(mr)) {
 567     vm_shutdown_during_initialization(
 568       err_msg("Unable to allocate %zuKB card tables for parallel "
 569       "garbage collection for the requested %zuKB heap.",
 570       _summary_data.reserved_byte_size()/K, mr.byte_size()/K));
 571     return false;
 572   }
 573 
 574   return true;
 575 }
 576 
 577 void PSParallelCompact::initialize_space_info()
 578 {
 579   memset(&_space_info, 0, sizeof(_space_info));
 580 
 581   ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
 582   PSYoungGen* young_gen = heap->young_gen();
 583 
 584   _space_info[old_space_id].set_space(heap->old_gen()->object_space());
 585   _space_info[eden_space_id].set_space(young_gen->eden_space());
 586   _space_info[from_space_id].set_space(young_gen->from_space());
 587   _space_info[to_space_id].set_space(young_gen->to_space());
 588 
 589   _space_info[old_space_id].set_start_array(heap->old_gen()->start_array());
 590 }
 591 
 592 void
 593 PSParallelCompact::clear_data_covering_space(SpaceId id)
 594 {
 595   // At this point, top is the value before GC, new_top() is the value that will
 596   // be set at the end of GC.  The marking bitmap is cleared to top; nothing
 597   // should be marked above top.  The summary data is cleared to the larger of
 598   // top & new_top.
 599   MutableSpace* const space = _space_info[id].space();
 600   HeapWord* const bot = space->bottom();
 601   HeapWord* const top = space->top();
 602   HeapWord* const max_top = MAX2(top, _space_info[id].new_top());
 603 
 604   _mark_bitmap.clear_range(bot, top);
 605 
 606   const size_t beg_region = _summary_data.addr_to_region_idx(bot);
 607   const size_t end_region =
 608     _summary_data.addr_to_region_idx(_summary_data.region_align_up(max_top));
 609   _summary_data.clear_range(beg_region, end_region);
 610 
 611   // Clear the data used to 'split' regions.
 612   SplitInfo& split_info = _space_info[id].split_info();
 613   if (split_info.is_valid()) {
 614     split_info.clear();
 615   }
 616   DEBUG_ONLY(split_info.verify_clear();)
 617 }
 618 
 619 void PSParallelCompact::pre_compact()
 620 {
 621   // Update the from & to space pointers in space_info, since they are swapped
 622   // at each young gen gc.  Do the update unconditionally (even though a
 623   // promotion failure does not swap spaces) because an unknown number of young
 624   // collections will have swapped the spaces an unknown number of times.
 625   GCTraceTime(Debug, gc, phases) tm("Pre Compact", &_gc_timer);
 626   ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
 627   _space_info[from_space_id].set_space(heap->young_gen()->from_space());
 628   _space_info[to_space_id].set_space(heap->young_gen()->to_space());
 629 
 630   heap->increment_total_collections(true);
 631 
 632   CodeCache::on_gc_marking_cycle_start();
 633 
 634   heap->print_before_gc();
 635   heap->trace_heap_before_gc(&_gc_tracer);
 636 
 637   // Fill in TLABs
 638   heap->ensure_parsability(true);  // retire TLABs
 639 
 640   if (VerifyBeforeGC && heap->total_collections() >= VerifyGCStartAt) {
 641     Universe::verify("Before GC");
 642   }
 643 
 644   DEBUG_ONLY(mark_bitmap()->verify_clear();)
 645   DEBUG_ONLY(summary_data().verify_clear();)
 646 }
 647 
 648 void PSParallelCompact::post_compact()
 649 {
 650   GCTraceTime(Info, gc, phases) tm("Post Compact", &_gc_timer);
 651   ParCompactionManager::remove_all_shadow_regions();
 652 
 653   CodeCache::on_gc_marking_cycle_finish();
 654   CodeCache::arm_all_nmethods();
 655 
 656   // Need to clear claim bits for the next full-gc (marking and adjust-pointers).
 657   ClassLoaderDataGraph::clear_claimed_marks();
 658 
 659   for (unsigned int id = old_space_id; id < last_space_id; ++id) {
 660     // Clear the marking bitmap, summary data and split info.
 661     clear_data_covering_space(SpaceId(id));
 662     {
 663       MutableSpace* space = _space_info[id].space();
 664       HeapWord* top = space->top();
 665       HeapWord* new_top = _space_info[id].new_top();
 666       if (ZapUnusedHeapArea && new_top < top) {
 667         space->mangle_region(MemRegion(new_top, top));
 668       }
 669       // Update top().  Must be done after clearing the bitmap and summary data.
 670       space->set_top(new_top);
 671     }
 672   }
 673 
 674 #ifdef ASSERT
 675   {
 676     mark_bitmap()->verify_clear();
 677     summary_data().verify_clear();
 678   }
 679 #endif
 680 
 681   ParCompactionManager::flush_all_string_dedup_requests();
 682 
 683   MutableSpace* const eden_space = _space_info[eden_space_id].space();
 684   MutableSpace* const from_space = _space_info[from_space_id].space();
 685   MutableSpace* const to_space   = _space_info[to_space_id].space();
 686 
 687   ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
 688   bool eden_empty = eden_space->is_empty();
 689 
 690   // Update heap occupancy information which is used as input to the soft ref
 691   // clearing policy at the next gc.
 692   Universe::heap()->update_capacity_and_used_at_gc();
 693 
 694   bool young_gen_empty = eden_empty && from_space->is_empty() &&
 695     to_space->is_empty();
 696 
 697   PSCardTable* ct = heap->card_table();
 698   MemRegion old_mr = heap->old_gen()->committed();
 699   if (young_gen_empty) {
 700     ct->clear_MemRegion(old_mr);
 701   } else {
 702     ct->dirty_MemRegion(old_mr);
 703   }
 704 
 705   heap->prune_scavengable_nmethods();
 706 
 707 #if COMPILER2_OR_JVMCI
 708   DerivedPointerTable::update_pointers();
 709 #endif
 710 
 711   // Signal that we have completed a visit to all live objects.
 712   Universe::heap()->record_whole_heap_examined_timestamp();
 713 }
 714 
 715 HeapWord* PSParallelCompact::compute_dense_prefix_for_old_space(MutableSpace* old_space,
 716                                                                 HeapWord* full_region_prefix_end) {
 717   const size_t region_size = ParallelCompactData::RegionSize;
 718   const ParallelCompactData& sd = summary_data();
 719 
 720   // Iteration starts with the region *after* the full-region-prefix-end.
 721   const RegionData* const start_region = sd.addr_to_region_ptr(full_region_prefix_end);
 722   // If final region is not full, iteration stops before that region,
 723   // because fill_dense_prefix_end assumes that prefix_end <= top.
 724   const RegionData* const end_region = sd.addr_to_region_ptr(old_space->top());
 725   assert(start_region <= end_region, "inv");
 726 
 727   size_t max_waste = old_space->capacity_in_words() * (MarkSweepDeadRatio / 100.0);
 728   const RegionData* cur_region = start_region;
 729   for (/* empty */; cur_region < end_region; ++cur_region) {
 730     assert(region_size >= cur_region->data_size(), "inv");
 731     size_t dead_size = region_size - cur_region->data_size();
 732     if (max_waste < dead_size) {
 733       break;
 734     }
 735     max_waste -= dead_size;
 736   }
 737 
 738   HeapWord* const prefix_end = sd.region_to_addr(cur_region);
 739   assert(sd.is_region_aligned(prefix_end), "postcondition");
 740   assert(prefix_end >= full_region_prefix_end, "in-range");
 741   assert(prefix_end <= old_space->top(), "in-range");
 742   return prefix_end;
 743 }
 744 
 745 void PSParallelCompact::fill_dense_prefix_end(SpaceId id) {
 746   // Comparing two sizes to decide if filling is required:
 747   //
 748   // The size of the filler (min-obj-size) is 2 heap words with the default
 749   // MinObjAlignment, since both markword and klass take 1 heap word.
 750   // With +UseCompactObjectHeaders, the minimum filler size is only one word,
 751   // because the Klass* gets encoded in the mark-word.
 752   //
 753   // The size of the gap (if any) right before dense-prefix-end is
 754   // MinObjAlignment.
 755   //
 756   // Need to fill in the gap only if it's smaller than min-obj-size, and the
 757   // filler obj will extend to next region.
 758 
 759   if (MinObjAlignment >= checked_cast<int>(CollectedHeap::min_fill_size())) {
 760     return;
 761   }
 762 
 763   assert(!UseCompactObjectHeaders, "Compact headers can allocate small objects");
 764   assert(CollectedHeap::min_fill_size() == 2, "inv");
 765   HeapWord* const dense_prefix_end = dense_prefix(id);
 766   assert(_summary_data.is_region_aligned(dense_prefix_end), "precondition");
 767   assert(dense_prefix_end <= space(id)->top(), "precondition");
 768   if (dense_prefix_end == space(id)->top()) {
 769     // Must not have single-word gap right before prefix-end/top.
 770     return;
 771   }
 772   RegionData* const region_after_dense_prefix = _summary_data.addr_to_region_ptr(dense_prefix_end);
 773 
 774   if (region_after_dense_prefix->partial_obj_size() != 0 ||
 775       _mark_bitmap.is_marked(dense_prefix_end)) {
 776     // The region after the dense prefix starts with live bytes.
 777     return;
 778   }
 779 
 780   HeapWord* block_start = start_array(id)->block_start_reaching_into_card(dense_prefix_end);
 781   if (block_start == dense_prefix_end - 1) {
 782     assert(!_mark_bitmap.is_marked(block_start), "inv");
 783     // There is exactly one heap word gap right before the dense prefix end, so we need a filler object.
 784     // The filler object will extend into region_after_dense_prefix.
 785     const size_t obj_len = 2; // min-fill-size
 786     HeapWord* const obj_beg = dense_prefix_end - 1;
 787     CollectedHeap::fill_with_object(obj_beg, obj_len);
 788     _mark_bitmap.mark_obj(obj_beg);
 789     _summary_data.addr_to_region_ptr(obj_beg)->add_live_obj(1);
 790     region_after_dense_prefix->set_partial_obj_size(1);
 791     region_after_dense_prefix->set_partial_obj_addr(obj_beg);
 792     assert(start_array(id) != nullptr, "sanity");
 793     start_array(id)->update_for_block(obj_beg, obj_beg + obj_len);
 794   }
 795 }
 796 
 797 bool PSParallelCompact::check_maximum_compaction(bool should_do_max_compaction,
 798                                                  size_t total_live_words,
 799                                                  MutableSpace* const old_space,
 800                                                  HeapWord* full_region_prefix_end) {
 801 
 802   ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
 803 
 804   // Check System.GC
 805   bool is_max_on_system_gc = UseMaximumCompactionOnSystemGC
 806                           && GCCause::is_user_requested_gc(heap->gc_cause());
 807 
 808   // Check if all live objs are too much for old-gen.
 809   const bool is_old_gen_too_full = (total_live_words >= old_space->capacity_in_words());
 810 
 811   // If all regions in old-gen are full
 812   const bool is_region_full =
 813     full_region_prefix_end >= _summary_data.region_align_down(old_space->top());
 814 
 815   return should_do_max_compaction
 816       || is_max_on_system_gc
 817       || is_old_gen_too_full
 818       || is_region_full;
 819 }
 820 
 821 void PSParallelCompact::summary_phase(bool should_do_max_compaction)
 822 {
 823   GCTraceTime(Info, gc, phases) tm("Summary Phase", &_gc_timer);
 824 
 825   MutableSpace* const old_space = _space_info[old_space_id].space();
 826   {
 827     size_t total_live_words = 0;
 828     HeapWord* full_region_prefix_end = nullptr;
 829     {
 830       // old-gen
 831       size_t live_words = _summary_data.live_words_in_space(old_space,
 832                                                             &full_region_prefix_end);
 833       total_live_words += live_words;
 834     }
 835     // young-gen
 836     for (uint i = eden_space_id; i < last_space_id; ++i) {
 837       const MutableSpace* space = _space_info[i].space();
 838       size_t live_words = _summary_data.live_words_in_space(space);
 839       total_live_words += live_words;
 840       _space_info[i].set_new_top(space->bottom() + live_words);
 841       _space_info[i].set_dense_prefix(space->bottom());
 842     }
 843 
 844     should_do_max_compaction = check_maximum_compaction(should_do_max_compaction,
 845                                                         total_live_words,
 846                                                         old_space,
 847                                                         full_region_prefix_end);
 848     {
 849       GCTraceTime(Info, gc, phases) tm("Summary Phase: expand", &_gc_timer);
 850       // Try to expand old-gen in order to fit all live objs and waste.
 851       size_t target_capacity_bytes = total_live_words * HeapWordSize
 852                                    + old_space->capacity_in_bytes() * (MarkSweepDeadRatio / 100);
 853       ParallelScavengeHeap::heap()->old_gen()->try_expand_till_size(target_capacity_bytes);
 854     }
 855 
 856     HeapWord* dense_prefix_end = should_do_max_compaction
 857                                  ? full_region_prefix_end
 858                                  : compute_dense_prefix_for_old_space(old_space,
 859                                                                       full_region_prefix_end);
 860     SpaceId id = old_space_id;
 861     _space_info[id].set_dense_prefix(dense_prefix_end);
 862 
 863     if (dense_prefix_end != old_space->bottom()) {
 864       fill_dense_prefix_end(id);
 865     }
 866 
 867     // Compacting objs in [dense_prefix_end, old_space->top())
 868     _summary_data.summarize(_space_info[id].split_info(),
 869                             dense_prefix_end, old_space->top(), nullptr,
 870                             dense_prefix_end, old_space->end(),
 871                             _space_info[id].new_top_addr());
 872   }
 873 
 874   // Summarize the remaining spaces in the young gen.  The initial target space
 875   // is the old gen.  If a space does not fit entirely into the target, then the
 876   // remainder is compacted into the space itself and that space becomes the new
 877   // target.
 878   SpaceId dst_space_id = old_space_id;
 879   HeapWord* dst_space_end = old_space->end();
 880   HeapWord** new_top_addr = _space_info[dst_space_id].new_top_addr();
 881   for (unsigned int id = eden_space_id; id < last_space_id; ++id) {
 882     const MutableSpace* space = _space_info[id].space();
 883     const size_t live = pointer_delta(_space_info[id].new_top(),
 884                                       space->bottom());
 885     const size_t available = pointer_delta(dst_space_end, *new_top_addr);
 886 
 887     if (live > 0 && live <= available) {
 888       // All the live data will fit.
 889       bool done = _summary_data.summarize(_space_info[id].split_info(),
 890                                           space->bottom(), space->top(),
 891                                           nullptr,
 892                                           *new_top_addr, dst_space_end,
 893                                           new_top_addr);
 894       assert(done, "space must fit into old gen");
 895 
 896       // Reset the new_top value for the space.
 897       _space_info[id].set_new_top(space->bottom());
 898     } else if (live > 0) {
 899       // Attempt to fit part of the source space into the target space.
 900       HeapWord* next_src_addr = nullptr;
 901       bool done = _summary_data.summarize(_space_info[id].split_info(),
 902                                           space->bottom(), space->top(),
 903                                           &next_src_addr,
 904                                           *new_top_addr, dst_space_end,
 905                                           new_top_addr);
 906       assert(!done, "space should not fit into old gen");
 907       assert(next_src_addr != nullptr, "sanity");
 908 
 909       // The source space becomes the new target, so the remainder is compacted
 910       // within the space itself.
 911       dst_space_id = SpaceId(id);
 912       dst_space_end = space->end();
 913       new_top_addr = _space_info[id].new_top_addr();
 914       done = _summary_data.summarize(_space_info[id].split_info(),
 915                                      next_src_addr, space->top(),
 916                                      nullptr,
 917                                      space->bottom(), dst_space_end,
 918                                      new_top_addr);
 919       assert(done, "space must fit when compacted into itself");
 920       assert(*new_top_addr <= space->top(), "usage should not grow");
 921     }
 922   }
 923 }
 924 
 925 void PSParallelCompact::report_object_count_after_gc() {
 926   GCTraceTime(Debug, gc, phases) tm("Report Object Count", &_gc_timer);
 927   // The heap is compacted, all objects are iterable. However there may be
 928   // filler objects in the heap which we should ignore.
 929   class SkipFillerObjectClosure : public BoolObjectClosure {
 930   public:
 931     bool do_object_b(oop obj) override { return !CollectedHeap::is_filler_object(obj); }
 932   } cl;
 933   _gc_tracer.report_object_count_after_gc(&cl, &ParallelScavengeHeap::heap()->workers());
 934 }
 935 
 936 bool PSParallelCompact::invoke(bool clear_all_soft_refs, bool should_do_max_compaction) {
 937   assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
 938   assert(Thread::current() == (Thread*)VMThread::vm_thread(),
 939          "should be in vm thread");
 940   assert(ref_processor() != nullptr, "Sanity");
 941 
 942   SvcGCMarker sgcm(SvcGCMarker::FULL);
 943   IsSTWGCActiveMark mark;
 944 
 945   ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
 946 
 947   GCIdMark gc_id_mark;
 948   _gc_timer.register_gc_start();
 949   _gc_tracer.report_gc_start(heap->gc_cause(), _gc_timer.gc_start());
 950 
 951   GCCause::Cause gc_cause = heap->gc_cause();
 952   PSOldGen* old_gen = heap->old_gen();
 953   PSAdaptiveSizePolicy* size_policy = heap->size_policy();
 954 
 955   // Make sure data structures are sane, make the heap parsable, and do other
 956   // miscellaneous bookkeeping.
 957   pre_compact();
 958 
 959   const PreGenGCValues pre_gc_values = heap->get_pre_gc_values();
 960 
 961   {
 962     const uint active_workers =
 963       WorkerPolicy::calc_active_workers(ParallelScavengeHeap::heap()->workers().max_workers(),
 964                                         ParallelScavengeHeap::heap()->workers().active_workers(),
 965                                         Threads::number_of_non_daemon_threads());
 966     ParallelScavengeHeap::heap()->workers().set_active_workers(active_workers);
 967 
 968     GCTraceCPUTime tcpu(&_gc_tracer);
 969     GCTraceTime(Info, gc) tm("Pause Full", nullptr, gc_cause, true);
 970 
 971     heap->pre_full_gc_dump(&_gc_timer);
 972 
 973     TraceCollectorStats tcs(counters());
 974     TraceMemoryManagerStats tms(heap->old_gc_manager(), gc_cause, "end of major GC");
 975 
 976     if (log_is_enabled(Debug, gc, heap, exit)) {
 977       accumulated_time()->start();
 978     }
 979 
 980     // Let the size policy know we're starting
 981     size_policy->major_collection_begin();
 982 
 983 #if COMPILER2_OR_JVMCI
 984     DerivedPointerTable::clear();
 985 #endif
 986 
 987     ref_processor()->start_discovery(clear_all_soft_refs);
 988 
 989     marking_phase(&_gc_tracer);
 990 
 991     summary_phase(should_do_max_compaction);
 992 
 993 #if COMPILER2_OR_JVMCI
 994     assert(DerivedPointerTable::is_active(), "Sanity");
 995     DerivedPointerTable::set_active(false);
 996 #endif
 997 
 998     forward_to_new_addr();
 999 
1000     adjust_pointers();
1001 
1002     compact();
1003 
1004     ParCompactionManager::_preserved_marks_set->restore(&ParallelScavengeHeap::heap()->workers());
1005 
1006     ParCompactionManager::verify_all_region_stack_empty();
1007 
1008     // Reset the mark bitmap, summary data, and do other bookkeeping.  Must be
1009     // done before resizing.
1010     post_compact();
1011 
1012     size_policy->major_collection_end();
1013 
1014     size_policy->sample_old_gen_used_bytes(MAX2(pre_gc_values.old_gen_used(), old_gen->used_in_bytes()));
1015 
1016     if (UseAdaptiveSizePolicy) {
1017       heap->resize_after_full_gc();
1018     }
1019 
1020     heap->resize_all_tlabs();
1021 
1022     // Resize the metaspace capacity after a collection
1023     MetaspaceGC::compute_new_size();
1024 
1025     if (log_is_enabled(Debug, gc, heap, exit)) {
1026       accumulated_time()->stop();
1027     }
1028 
1029     heap->print_heap_change(pre_gc_values);
1030 
1031     report_object_count_after_gc();
1032 
1033     // Track memory usage and detect low memory
1034     MemoryService::track_memory_usage();
1035     heap->update_counters();
1036 
1037     heap->post_full_gc_dump(&_gc_timer);
1038 
1039     size_policy->record_gc_pause_end_instant();
1040   }
1041 
1042   heap->gc_epilogue(true);
1043 
1044   if (VerifyAfterGC && heap->total_collections() >= VerifyGCStartAt) {
1045     Universe::verify("After GC");
1046   }
1047 
1048   heap->print_after_gc();
1049   heap->trace_heap_after_gc(&_gc_tracer);
1050 
1051   _gc_timer.register_gc_end();
1052 
1053   _gc_tracer.report_dense_prefix(dense_prefix(old_space_id));
1054   _gc_tracer.report_gc_end(_gc_timer.gc_end(), _gc_timer.time_partitions());
1055 
1056   return true;
1057 }
1058 
1059 class PCAddThreadRootsMarkingTaskClosure : public ThreadClosure {
1060   ParCompactionManager* _cm;
1061 
1062 public:
1063   PCAddThreadRootsMarkingTaskClosure(ParCompactionManager* cm) : _cm(cm) { }
1064   void do_thread(Thread* thread) {
1065     ResourceMark rm;
1066 
1067     MarkingNMethodClosure mark_and_push_in_blobs(&_cm->_mark_and_push_closure);
1068 
1069     thread->oops_do(&_cm->_mark_and_push_closure, &mark_and_push_in_blobs);
1070 
1071     // Do the real work
1072     _cm->follow_marking_stacks();
1073   }
1074 };
1075 
1076 void steal_marking_work(TaskTerminator& terminator, uint worker_id) {
1077   assert(ParallelScavengeHeap::heap()->is_stw_gc_active(), "called outside gc");
1078 
1079   ParCompactionManager* cm =
1080     ParCompactionManager::gc_thread_compaction_manager(worker_id);
1081 
1082   do {
1083     ScannerTask task;
1084     if (ParCompactionManager::steal(worker_id, task)) {
1085       cm->follow_contents(task, true);
1086     }
1087     cm->follow_marking_stacks();
1088   } while (!terminator.offer_termination());
1089 }
1090 
1091 class MarkFromRootsTask : public WorkerTask {
1092   NMethodMarkingScope _nmethod_marking_scope;
1093   ThreadsClaimTokenScope _threads_claim_token_scope;
1094   OopStorageSetStrongParState<false /* concurrent */, false /* is_const */> _oop_storage_set_par_state;
1095   TaskTerminator _terminator;
1096   uint _active_workers;
1097 
1098 public:
1099   MarkFromRootsTask(uint active_workers) :
1100       WorkerTask("MarkFromRootsTask"),
1101       _nmethod_marking_scope(),
1102       _threads_claim_token_scope(),
1103       _terminator(active_workers, ParCompactionManager::marking_stacks()),
1104       _active_workers(active_workers) {}
1105 
1106   virtual void work(uint worker_id) {
1107     ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(worker_id);
1108     cm->create_marking_stats_cache();
1109     {
1110       CLDToOopClosure cld_closure(&cm->_mark_and_push_closure, ClassLoaderData::_claim_stw_fullgc_mark);
1111       ClassLoaderDataGraph::always_strong_cld_do(&cld_closure);
1112 
1113       // Do the real work
1114       cm->follow_marking_stacks();
1115     }
1116 
1117     {
1118       PCAddThreadRootsMarkingTaskClosure closure(cm);
1119       Threads::possibly_parallel_threads_do(_active_workers > 1 /* is_par */, &closure);
1120     }
1121 
1122     // Mark from OopStorages
1123     {
1124       _oop_storage_set_par_state.oops_do(&cm->_mark_and_push_closure);
1125       // Do the real work
1126       cm->follow_marking_stacks();
1127     }
1128 
1129     if (_active_workers > 1) {
1130       steal_marking_work(_terminator, worker_id);
1131     }
1132   }
1133 };
1134 
1135 class ParallelCompactRefProcProxyTask : public RefProcProxyTask {
1136   TaskTerminator _terminator;
1137 
1138 public:
1139   ParallelCompactRefProcProxyTask(uint max_workers)
1140     : RefProcProxyTask("ParallelCompactRefProcProxyTask", max_workers),
1141       _terminator(_max_workers, ParCompactionManager::marking_stacks()) {}
1142 
1143   void work(uint worker_id) override {
1144     assert(worker_id < _max_workers, "sanity");
1145     ParCompactionManager* cm = (_tm == RefProcThreadModel::Single) ? ParCompactionManager::get_vmthread_cm() : ParCompactionManager::gc_thread_compaction_manager(worker_id);
1146     BarrierEnqueueDiscoveredFieldClosure enqueue;
1147     ParCompactionManager::FollowStackClosure complete_gc(cm, (_tm == RefProcThreadModel::Single) ? nullptr : &_terminator, worker_id);
1148     _rp_task->rp_work(worker_id, PSParallelCompact::is_alive_closure(), &cm->_mark_and_push_closure, &enqueue, &complete_gc);
1149   }
1150 
1151   void prepare_run_task_hook() override {
1152     _terminator.reset_for_reuse(_queue_count);
1153   }
1154 };
1155 
1156 static void flush_marking_stats_cache(const uint num_workers) {
1157   for (uint i = 0; i < num_workers; ++i) {
1158     ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(i);
1159     cm->flush_and_destroy_marking_stats_cache();
1160   }
1161 }
1162 
1163 class PSParallelCleaningTask : public WorkerTask {
1164   bool                    _unloading_occurred;
1165   CodeCacheUnloadingTask  _code_cache_task;
1166   // Prune dead klasses from subklass/sibling/implementor lists.
1167   KlassCleaningTask       _klass_cleaning_task;
1168 
1169 public:
1170   PSParallelCleaningTask(bool unloading_occurred) :
1171     WorkerTask("PS Parallel Cleaning"),
1172     _unloading_occurred(unloading_occurred),
1173     _code_cache_task(unloading_occurred),
1174     _klass_cleaning_task() {}
1175 
1176   void work(uint worker_id) {
1177 #if INCLUDE_JVMCI
1178     if (EnableJVMCI && worker_id == 0) {
1179       // Serial work; only first worker.
1180       // Clean JVMCI metadata handles.
1181       JVMCI::do_unloading(_unloading_occurred);
1182     }
1183 #endif
1184 
1185     // Do first pass of code cache cleaning.
1186     _code_cache_task.work(worker_id);
1187 
1188     // Clean all klasses that were not unloaded.
1189     // The weak metadata in klass doesn't need to be
1190     // processed if there was no unloading.
1191     if (_unloading_occurred) {
1192       _klass_cleaning_task.work();
1193     }
1194   }
1195 };
1196 
1197 void PSParallelCompact::marking_phase(ParallelOldTracer *gc_tracer) {
1198   // Recursively traverse all live objects and mark them
1199   GCTraceTime(Info, gc, phases) tm("Marking Phase", &_gc_timer);
1200 
1201   uint active_gc_threads = ParallelScavengeHeap::heap()->workers().active_workers();
1202 
1203   ClassLoaderDataGraph::verify_claimed_marks_cleared(ClassLoaderData::_claim_stw_fullgc_mark);
1204   {
1205     GCTraceTime(Debug, gc, phases) tm("Par Mark", &_gc_timer);
1206 
1207     MarkFromRootsTask task(active_gc_threads);
1208     ParallelScavengeHeap::heap()->workers().run_task(&task);
1209   }
1210 
1211   // Process reference objects found during marking
1212   {
1213     GCTraceTime(Debug, gc, phases) tm("Reference Processing", &_gc_timer);
1214 
1215     ReferenceProcessorStats stats;
1216     ReferenceProcessorPhaseTimes pt(&_gc_timer, ref_processor()->max_num_queues());
1217 
1218     ParallelCompactRefProcProxyTask task(ref_processor()->max_num_queues());
1219     stats = ref_processor()->process_discovered_references(task, &ParallelScavengeHeap::heap()->workers(), pt);
1220 
1221     gc_tracer->report_gc_reference_stats(stats);
1222     pt.print_all_references();
1223   }
1224 
1225   {
1226     GCTraceTime(Debug, gc, phases) tm("Flush Marking Stats", &_gc_timer);
1227 
1228     flush_marking_stats_cache(active_gc_threads);
1229   }
1230 
1231   // This is the point where the entire marking should have completed.
1232   ParCompactionManager::verify_all_marking_stack_empty();
1233 
1234   {
1235     GCTraceTime(Debug, gc, phases) tm("Weak Processing", &_gc_timer);
1236     WeakProcessor::weak_oops_do(&ParallelScavengeHeap::heap()->workers(),
1237                                 is_alive_closure(),
1238                                 &do_nothing_cl,
1239                                 1);
1240   }
1241 
1242   {
1243     GCTraceTime(Debug, gc, phases) tm_m("Class Unloading", &_gc_timer);
1244 
1245     ClassUnloadingContext ctx(active_gc_threads /* num_nmethod_unlink_workers */,
1246                               false /* unregister_nmethods_during_purge */,
1247                               false /* lock_nmethod_free_separately */);
1248 
1249     {
1250       CodeCache::UnlinkingScope scope(is_alive_closure());
1251 
1252       // Follow system dictionary roots and unload classes.
1253       bool unloading_occurred = SystemDictionary::do_unloading(&_gc_timer);
1254 
1255       PSParallelCleaningTask task{unloading_occurred};
1256       ParallelScavengeHeap::heap()->workers().run_task(&task);
1257     }
1258 
1259     {
1260       GCTraceTime(Debug, gc, phases) t("Purge Unlinked NMethods", gc_timer());
1261       // Release unloaded nmethod's memory.
1262       ctx.purge_nmethods();
1263     }
1264     {
1265       GCTraceTime(Debug, gc, phases) ur("Unregister NMethods", &_gc_timer);
1266       ParallelScavengeHeap::heap()->prune_unlinked_nmethods();
1267     }
1268     {
1269       GCTraceTime(Debug, gc, phases) t("Free Code Blobs", gc_timer());
1270       ctx.free_nmethods();
1271     }
1272     {
1273       // Delete metaspaces for unloaded class loaders and clean up loader_data graph
1274       GCTraceTime(Debug, gc, phases) t("Purge Class Loader Data", gc_timer());
1275       ClassLoaderDataGraph::purge(true /* at_safepoint */);
1276       DEBUG_ONLY(MetaspaceUtils::verify();)
1277     }
1278   }
1279 
1280 #if TASKQUEUE_STATS
1281   ParCompactionManager::print_and_reset_taskqueue_stats();
1282 #endif
1283 }
1284 
1285 template<typename Func>
1286 void PSParallelCompact::adjust_in_space_helper(SpaceId id, Atomic<uint>* claim_counter, Func&& on_stripe) {
1287   MutableSpace* sp = PSParallelCompact::space(id);
1288   HeapWord* const bottom = sp->bottom();
1289   HeapWord* const top = sp->top();
1290   if (bottom == top) {
1291     return;
1292   }
1293 
1294   const uint num_regions_per_stripe = 2;
1295   const size_t region_size = ParallelCompactData::RegionSize;
1296   const size_t stripe_size = num_regions_per_stripe * region_size;
1297 
1298   while (true) {
1299     uint counter = claim_counter->fetch_then_add(num_regions_per_stripe);
1300     HeapWord* cur_stripe = bottom + counter * region_size;
1301     if (cur_stripe >= top) {
1302       break;
1303     }
1304     HeapWord* stripe_end = MIN2(cur_stripe + stripe_size, top);
1305     on_stripe(cur_stripe, stripe_end);
1306   }
1307 }
1308 
1309 void PSParallelCompact::adjust_in_old_space(Atomic<uint>* claim_counter) {
1310   // Regions in old-space shouldn't be split.
1311   assert(!_space_info[old_space_id].split_info().is_valid(), "inv");
1312 
1313   auto scan_obj_with_limit = [&] (HeapWord* obj_start, HeapWord* left, HeapWord* right) {
1314     assert(mark_bitmap()->is_marked(obj_start), "inv");
1315     oop obj = cast_to_oop(obj_start);
1316     return obj->oop_iterate_size(&pc_adjust_pointer_closure, MemRegion(left, right));
1317   };
1318 
1319   adjust_in_space_helper(old_space_id, claim_counter, [&] (HeapWord* stripe_start, HeapWord* stripe_end) {
1320     assert(_summary_data.is_region_aligned(stripe_start), "inv");
1321     RegionData* cur_region = _summary_data.addr_to_region_ptr(stripe_start);
1322     HeapWord* obj_start;
1323     if (cur_region->partial_obj_size() != 0) {
1324       obj_start = cur_region->partial_obj_addr();
1325       obj_start += scan_obj_with_limit(obj_start, stripe_start, stripe_end);
1326     } else {
1327       obj_start = stripe_start;
1328     }
1329 
1330     while (obj_start < stripe_end) {
1331       obj_start = mark_bitmap()->find_obj_beg(obj_start, stripe_end);
1332       if (obj_start >= stripe_end) {
1333         break;
1334       }
1335       obj_start += scan_obj_with_limit(obj_start, stripe_start, stripe_end);
1336     }
1337   });
1338 }
1339 
1340 void PSParallelCompact::adjust_in_young_space(SpaceId id, Atomic<uint>* claim_counter) {
1341   adjust_in_space_helper(id, claim_counter, [](HeapWord* stripe_start, HeapWord* stripe_end) {
1342     HeapWord* obj_start = stripe_start;
1343     while (obj_start < stripe_end) {
1344       obj_start = mark_bitmap()->find_obj_beg(obj_start, stripe_end);
1345       if (obj_start >= stripe_end) {
1346         break;
1347       }
1348       oop obj = cast_to_oop(obj_start);
1349       obj_start += obj->oop_iterate_size(&pc_adjust_pointer_closure);
1350     }
1351   });
1352 }
1353 
1354 void PSParallelCompact::adjust_pointers_in_spaces(uint worker_id, Atomic<uint>* claim_counters) {
1355   auto start_time = Ticks::now();
1356   adjust_in_old_space(&claim_counters[0]);
1357   for (uint id = eden_space_id; id < last_space_id; ++id) {
1358     adjust_in_young_space(SpaceId(id), &claim_counters[id]);
1359   }
1360   log_trace(gc, phases)("adjust_pointers_in_spaces worker %u: %.3f ms", worker_id, (Ticks::now() - start_time).seconds() * 1000);
1361 }
1362 
1363 class PSAdjustTask final : public WorkerTask {
1364   ThreadsClaimTokenScope                     _threads_claim_token_scope;
1365   WeakProcessor::Task                        _weak_proc_task;
1366   OopStorageSetStrongParState<false, false>  _oop_storage_iter;
1367   uint                                       _nworkers;
1368   Atomic<bool>                               _code_cache_claimed;
1369   Atomic<uint> _claim_counters[PSParallelCompact::last_space_id];
1370 
1371   bool try_claim_code_cache_task() {
1372     return _code_cache_claimed.load_relaxed() == false
1373         && _code_cache_claimed.compare_set(false, true);
1374   }
1375 
1376 public:
1377   PSAdjustTask(uint nworkers) :
1378     WorkerTask("PSAdjust task"),
1379     _threads_claim_token_scope(),
1380     _weak_proc_task(nworkers),
1381     _oop_storage_iter(),
1382     _nworkers(nworkers),
1383     _code_cache_claimed(false) {
1384 
1385     for (unsigned int i = PSParallelCompact::old_space_id; i < PSParallelCompact::last_space_id; ++i) {
1386       ::new (&_claim_counters[i]) Atomic<uint>{};
1387     }
1388     ClassLoaderDataGraph::verify_claimed_marks_cleared(ClassLoaderData::_claim_stw_fullgc_adjust);
1389   }
1390 
1391   void work(uint worker_id) {
1392     {
1393       // Pointers in heap.
1394       ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(worker_id);
1395       cm->preserved_marks()->adjust_during_full_gc();
1396 
1397       PSParallelCompact::adjust_pointers_in_spaces(worker_id, _claim_counters);
1398     }
1399 
1400     {
1401       // All (strong and weak) CLDs.
1402       CLDToOopClosure cld_closure(&pc_adjust_pointer_closure, ClassLoaderData::_claim_stw_fullgc_adjust);
1403       ClassLoaderDataGraph::cld_do(&cld_closure);
1404     }
1405 
1406     {
1407       // Threads stack frames. No need to visit on-stack nmethods, because all
1408       // nmethods are visited in one go via CodeCache::nmethods_do.
1409       ResourceMark rm;
1410       Threads::possibly_parallel_oops_do(_nworkers > 1, &pc_adjust_pointer_closure, nullptr);
1411       if (try_claim_code_cache_task()) {
1412         NMethodToOopClosure adjust_code(&pc_adjust_pointer_closure, NMethodToOopClosure::FixRelocations);
1413         CodeCache::nmethods_do(&adjust_code);
1414       }
1415     }
1416 
1417     {
1418       // VM internal strong and weak roots.
1419       _oop_storage_iter.oops_do(&pc_adjust_pointer_closure);
1420       AlwaysTrueClosure always_alive;
1421       _weak_proc_task.work(worker_id, &always_alive, &pc_adjust_pointer_closure);
1422     }
1423   }
1424 };
1425 
1426 void PSParallelCompact::adjust_pointers() {
1427   // Adjust the pointers to reflect the new locations
1428   GCTraceTime(Info, gc, phases) tm("Adjust Pointers", &_gc_timer);
1429   uint nworkers = ParallelScavengeHeap::heap()->workers().active_workers();
1430   PSAdjustTask task(nworkers);
1431   ParallelScavengeHeap::heap()->workers().run_task(&task);
1432 }
1433 
1434 // Split [start, end) evenly for a number of workers and return the
1435 // range for worker_id.
1436 static void split_regions_for_worker(size_t start, size_t end,
1437                                      uint worker_id, uint num_workers,
1438                                      size_t* worker_start, size_t* worker_end) {
1439   assert(start < end, "precondition");
1440   assert(num_workers > 0, "precondition");
1441   assert(worker_id < num_workers, "precondition");
1442 
1443   size_t num_regions = end - start;
1444   size_t num_regions_per_worker = num_regions / num_workers;
1445   size_t remainder = num_regions % num_workers;
1446   // The first few workers will get one extra.
1447   *worker_start = start + worker_id * num_regions_per_worker
1448                   + MIN2(checked_cast<size_t>(worker_id), remainder);
1449   *worker_end = *worker_start + num_regions_per_worker
1450                 + (worker_id < remainder ? 1 : 0);
1451 }
1452 
1453 static bool safe_to_read_header(size_t words) {
1454   precond(words > 0);
1455 
1456   // Safe to read if we have enough words for the full header, i.e., both
1457   // markWord and Klass pointer.
1458   const bool safe = words >= (size_t)oopDesc::header_size();
1459 
1460   // If using Compact Object Headers, the full header is inside the markWord,
1461   // so will always be safe to read
1462   assert(!UseCompactObjectHeaders || safe, "Compact Object Headers should always be safe");
1463 
1464   return safe;
1465 }
1466 
1467 void PSParallelCompact::forward_to_new_addr() {
1468   GCTraceTime(Info, gc, phases) tm("Forward", &_gc_timer);
1469   uint nworkers = ParallelScavengeHeap::heap()->workers().active_workers();
1470 
1471   struct ForwardTask final : public WorkerTask {
1472     uint _num_workers;
1473 
1474     explicit ForwardTask(uint num_workers) :
1475       WorkerTask("PSForward task"),
1476       _num_workers(num_workers) {}
1477 
1478     static bool should_preserve_mark(oop obj, HeapWord* end_addr) {
1479       size_t remaining_words = pointer_delta(end_addr, cast_from_oop<HeapWord*>(obj));
1480 
1481       if (Arguments::is_valhalla_enabled() && !safe_to_read_header(remaining_words)) {
1482         // When using Valhalla, it might be necessary to preserve the Valhalla-
1483         // specific bits in the markWord. If the entire object header is
1484         // copied, the correct markWord (with the appropriate Valhalla bits)
1485         // can be safely read from the Klass. However, if the full header is
1486         // not copied, we cannot safely read the Klass to obtain this information.
1487         // In such cases, we always preserve the markWord to ensure that all
1488         // relevant bits, including Valhalla-specific ones, are retained.
1489         return true;
1490       } else {
1491         return obj->mark().must_be_preserved();
1492       }
1493     }
1494 
1495     static void forward_objs_in_range(ParCompactionManager* cm,
1496                                       HeapWord* start,
1497                                       HeapWord* end,
1498                                       HeapWord* destination) {
1499       HeapWord* cur_addr = start;
1500       HeapWord* new_addr = destination;
1501 
1502       while (cur_addr < end) {
1503         cur_addr = mark_bitmap()->find_obj_beg(cur_addr, end);
1504         if (cur_addr >= end) {
1505           return;
1506         }
1507         assert(mark_bitmap()->is_marked(cur_addr), "inv");
1508         oop obj = cast_to_oop(cur_addr);
1509 
1510         if (new_addr != cur_addr) {
1511           if (should_preserve_mark(obj, end)) {
1512             cm->preserved_marks()->push_always(obj, obj->mark());
1513           }
1514 
1515           FullGCForwarding::forward_to(obj, cast_to_oop(new_addr));
1516         }
1517         size_t obj_size = obj->size();
1518         new_addr += obj_size;
1519         cur_addr += obj_size;
1520       }
1521     }
1522 
1523     void work(uint worker_id) override {
1524       ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(worker_id);
1525       for (uint id = old_space_id; id < last_space_id; ++id) {
1526         MutableSpace* sp = PSParallelCompact::space(SpaceId(id));
1527         HeapWord* dense_prefix_addr = dense_prefix(SpaceId(id));
1528         HeapWord* top = sp->top();
1529 
1530         if (dense_prefix_addr == top) {
1531           // Empty space
1532           continue;
1533         }
1534 
1535         const SplitInfo& split_info = _space_info[SpaceId(id)].split_info();
1536         size_t dense_prefix_region = _summary_data.addr_to_region_idx(dense_prefix_addr);
1537         size_t top_region = _summary_data.addr_to_region_idx(_summary_data.region_align_up(top));
1538         size_t start_region;
1539         size_t end_region;
1540         split_regions_for_worker(dense_prefix_region, top_region,
1541                                  worker_id, _num_workers,
1542                                  &start_region, &end_region);
1543         for (size_t cur_region = start_region; cur_region < end_region; ++cur_region) {
1544           RegionData* region_ptr = _summary_data.region(cur_region);
1545           size_t partial_obj_size = region_ptr->partial_obj_size();
1546 
1547           if (partial_obj_size == ParallelCompactData::RegionSize) {
1548             // No obj-start
1549             continue;
1550           }
1551 
1552           HeapWord* region_start = _summary_data.region_to_addr(cur_region);
1553           HeapWord* region_end = region_start + ParallelCompactData::RegionSize;
1554 
1555           if (split_info.is_split(cur_region)) {
1556             // Part 1: will be relocated to space-1
1557             HeapWord* preceding_destination = split_info.preceding_destination();
1558             HeapWord* split_point = split_info.split_point();
1559             forward_objs_in_range(cm, region_start + partial_obj_size, split_point, preceding_destination + partial_obj_size);
1560 
1561             // Part 2: will be relocated to space-2
1562             HeapWord* destination = region_ptr->destination();
1563             forward_objs_in_range(cm, split_point, region_end, destination);
1564           } else {
1565             HeapWord* destination = region_ptr->destination();
1566             forward_objs_in_range(cm, region_start + partial_obj_size, region_end, destination + partial_obj_size);
1567           }
1568         }
1569       }
1570     }
1571   } task(nworkers);
1572 
1573   ParallelScavengeHeap::heap()->workers().run_task(&task);
1574   DEBUG_ONLY(verify_forward();)
1575 }
1576 
1577 #ifdef ASSERT
1578 void PSParallelCompact::verify_forward() {
1579   HeapWord* const old_dense_prefix_addr = dense_prefix(SpaceId(old_space_id));
1580   // The destination addr for the first live obj after dense-prefix.
1581   HeapWord* bump_ptr = old_dense_prefix_addr
1582                      + _summary_data.addr_to_region_ptr(old_dense_prefix_addr)->partial_obj_size();
1583   SpaceId bump_ptr_space = old_space_id;
1584 
1585   for (uint id = old_space_id; id < last_space_id; ++id) {
1586     MutableSpace* sp = PSParallelCompact::space(SpaceId(id));
1587     // Only verify objs after dense-prefix, because those before dense-prefix are not moved (forwarded).
1588     HeapWord* cur_addr = dense_prefix(SpaceId(id));
1589     HeapWord* top = sp->top();
1590 
1591     while (cur_addr < top) {
1592       cur_addr = mark_bitmap()->find_obj_beg(cur_addr, top);
1593       if (cur_addr >= top) {
1594         break;
1595       }
1596       assert(mark_bitmap()->is_marked(cur_addr), "inv");
1597       assert(bump_ptr <= _space_info[bump_ptr_space].new_top(), "inv");
1598       // Move to the space containing cur_addr
1599       if (bump_ptr == _space_info[bump_ptr_space].new_top()) {
1600         bump_ptr = space(space_id(cur_addr))->bottom();
1601         bump_ptr_space = space_id(bump_ptr);
1602       }
1603       oop obj = cast_to_oop(cur_addr);
1604       if (cur_addr == bump_ptr) {
1605         assert(!FullGCForwarding::is_forwarded(obj), "inv");
1606       } else {
1607         assert(FullGCForwarding::forwardee(obj) == cast_to_oop(bump_ptr), "inv");
1608       }
1609       bump_ptr += obj->size();
1610       cur_addr += obj->size();
1611     }
1612   }
1613 }
1614 #endif
1615 
1616 // Helper class to print 8 region numbers per line and then print the total at the end.
1617 class FillableRegionLogger : public StackObj {
1618 private:
1619   Log(gc, compaction) log;
1620   static const int LineLength = 8;
1621   size_t _regions[LineLength];
1622   int _next_index;
1623   bool _enabled;
1624   size_t _total_regions;
1625 public:
1626   FillableRegionLogger() : _next_index(0), _enabled(log_develop_is_enabled(Trace, gc, compaction)), _total_regions(0) { }
1627   ~FillableRegionLogger() {
1628     log.trace("%zu initially fillable regions", _total_regions);
1629   }
1630 
1631   void print_line() {
1632     if (!_enabled || _next_index == 0) {
1633       return;
1634     }
1635     FormatBuffer<> line("Fillable: ");
1636     for (int i = 0; i < _next_index; i++) {
1637       line.append(" %7zu", _regions[i]);
1638     }
1639     log.trace("%s", line.buffer());
1640     _next_index = 0;
1641   }
1642 
1643   void handle(size_t region) {
1644     if (!_enabled) {
1645       return;
1646     }
1647     _regions[_next_index++] = region;
1648     if (_next_index == LineLength) {
1649       print_line();
1650     }
1651     _total_regions++;
1652   }
1653 };
1654 
1655 void PSParallelCompact::prepare_region_draining_tasks(uint parallel_gc_threads)
1656 {
1657   GCTraceTime(Trace, gc, phases) tm("Drain Task Setup", &_gc_timer);
1658 
1659   // Find the threads that are active
1660   uint worker_id = 0;
1661 
1662   // Find all regions that are available (can be filled immediately) and
1663   // distribute them to the thread stacks.  The iteration is done in reverse
1664   // order (high to low) so the regions will be removed in ascending order.
1665 
1666   const ParallelCompactData& sd = PSParallelCompact::summary_data();
1667 
1668   // id + 1 is used to test termination so unsigned  can
1669   // be used with an old_space_id == 0.
1670   FillableRegionLogger region_logger;
1671   for (unsigned int id = last_space_id - 1; id + 1 > old_space_id; --id) {
1672     SpaceInfo* const space_info = _space_info + id;
1673     HeapWord* const new_top = space_info->new_top();
1674 
1675     const size_t beg_region = sd.addr_to_region_idx(space_info->dense_prefix());
1676     const size_t end_region =
1677       sd.addr_to_region_idx(sd.region_align_up(new_top));
1678 
1679     for (size_t cur = end_region - 1; cur + 1 > beg_region; --cur) {
1680       if (sd.region(cur)->claim_unsafe()) {
1681         ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(worker_id);
1682         bool result = sd.region(cur)->mark_normal();
1683         assert(result, "Must succeed at this point.");
1684         cm->region_stack()->push(cur);
1685         region_logger.handle(cur);
1686         // Assign regions to tasks in round-robin fashion.
1687         if (++worker_id == parallel_gc_threads) {
1688           worker_id = 0;
1689         }
1690       }
1691     }
1692     region_logger.print_line();
1693   }
1694 }
1695 
1696 static void compaction_with_stealing_work(TaskTerminator* terminator, uint worker_id) {
1697   assert(ParallelScavengeHeap::heap()->is_stw_gc_active(), "called outside gc");
1698 
1699   ParCompactionManager* cm =
1700     ParCompactionManager::gc_thread_compaction_manager(worker_id);
1701 
1702   // Drain the stacks that have been preloaded with regions
1703   // that are ready to fill.
1704 
1705   cm->drain_region_stacks();
1706 
1707   guarantee(cm->region_stack()->is_empty(), "Not empty");
1708 
1709   size_t region_index = 0;
1710 
1711   while (true) {
1712     if (ParCompactionManager::steal(worker_id, region_index)) {
1713       PSParallelCompact::fill_and_update_region(cm, region_index);
1714       cm->drain_region_stacks();
1715     } else if (PSParallelCompact::steal_unavailable_region(cm, region_index)) {
1716       // Fill and update an unavailable region with the help of a shadow region
1717       PSParallelCompact::fill_and_update_shadow_region(cm, region_index);
1718       cm->drain_region_stacks();
1719     } else {
1720       if (terminator->offer_termination()) {
1721         break;
1722       }
1723       // Go around again.
1724     }
1725   }
1726 }
1727 
1728 class FillDensePrefixAndCompactionTask: public WorkerTask {
1729   TaskTerminator _terminator;
1730 
1731 public:
1732   FillDensePrefixAndCompactionTask(uint active_workers) :
1733       WorkerTask("FillDensePrefixAndCompactionTask"),
1734       _terminator(active_workers, ParCompactionManager::region_task_queues()) {
1735   }
1736 
1737   virtual void work(uint worker_id) {
1738     if (worker_id == 0) {
1739       auto start = Ticks::now();
1740       PSParallelCompact::fill_dead_objs_in_dense_prefix();
1741       log_trace(gc, phases)("Fill dense prefix by worker 0: %.3f ms", (Ticks::now() - start).seconds() * 1000);
1742     }
1743     compaction_with_stealing_work(&_terminator, worker_id);
1744   }
1745 };
1746 
1747 void PSParallelCompact::fill_range_in_dense_prefix(HeapWord* start, HeapWord* end) {
1748 #ifdef ASSERT
1749   {
1750     assert(start < end, "precondition");
1751     assert(mark_bitmap()->find_obj_beg(start, end) == end, "precondition");
1752     HeapWord* bottom = _space_info[old_space_id].space()->bottom();
1753     if (start != bottom) {
1754       // The preceding live obj.
1755       HeapWord* obj_start = mark_bitmap()->find_obj_beg_reverse(bottom, start);
1756       HeapWord* obj_end = obj_start + cast_to_oop(obj_start)->size();
1757       assert(obj_end == start, "precondition");
1758     }
1759   }
1760 #endif
1761 
1762   CollectedHeap::fill_with_objects(start, pointer_delta(end, start));
1763   HeapWord* addr = start;
1764   do {
1765     size_t size = cast_to_oop(addr)->size();
1766     start_array(old_space_id)->update_for_block(addr, addr + size);
1767     addr += size;
1768   } while (addr < end);
1769 }
1770 
1771 void PSParallelCompact::fill_dead_objs_in_dense_prefix() {
1772   ParMarkBitMap* bitmap = mark_bitmap();
1773 
1774   HeapWord* const bottom = _space_info[old_space_id].space()->bottom();
1775   HeapWord* const prefix_end = dense_prefix(old_space_id);
1776 
1777   const size_t region_size = ParallelCompactData::RegionSize;
1778 
1779   // Fill dead space in [start_addr, end_addr)
1780   HeapWord* const start_addr = bottom;
1781   HeapWord* const end_addr   = prefix_end;
1782 
1783   for (HeapWord* cur_addr = start_addr; cur_addr < end_addr; /* empty */) {
1784     RegionData* cur_region_ptr = _summary_data.addr_to_region_ptr(cur_addr);
1785     if (cur_region_ptr->data_size() == region_size) {
1786       // Full; no dead space. Next region.
1787       if (_summary_data.is_region_aligned(cur_addr)) {
1788         cur_addr += region_size;
1789       } else {
1790         cur_addr = _summary_data.region_align_up(cur_addr);
1791       }
1792       continue;
1793     }
1794 
1795     // Fill dead space inside cur_region.
1796     if (_summary_data.is_region_aligned(cur_addr)) {
1797       cur_addr += cur_region_ptr->partial_obj_size();
1798     }
1799 
1800     HeapWord* region_end_addr = _summary_data.region_align_up(cur_addr + 1);
1801     assert(region_end_addr <= end_addr, "inv");
1802     while (cur_addr < region_end_addr) {
1803       // Use end_addr to allow filler-obj to cross region boundary.
1804       HeapWord* live_start = bitmap->find_obj_beg(cur_addr, end_addr);
1805       if (cur_addr != live_start) {
1806         // Found dead space [cur_addr, live_start).
1807         fill_range_in_dense_prefix(cur_addr, live_start);
1808       }
1809       if (live_start >= region_end_addr) {
1810         cur_addr = live_start;
1811         break;
1812       }
1813       assert(bitmap->is_marked(live_start), "inv");
1814       cur_addr = live_start + cast_to_oop(live_start)->size();
1815     }
1816   }
1817 }
1818 
1819 void PSParallelCompact::compact() {
1820   GCTraceTime(Info, gc, phases) tm("Compaction Phase", &_gc_timer);
1821 
1822   uint active_gc_threads = ParallelScavengeHeap::heap()->workers().active_workers();
1823 
1824   initialize_shadow_regions(active_gc_threads);
1825   prepare_region_draining_tasks(active_gc_threads);
1826 
1827   {
1828     GCTraceTime(Trace, gc, phases) tm("Par Compact", &_gc_timer);
1829 
1830     FillDensePrefixAndCompactionTask task(active_gc_threads);
1831     ParallelScavengeHeap::heap()->workers().run_task(&task);
1832 
1833 #ifdef  ASSERT
1834     verify_filler_in_dense_prefix();
1835 
1836     // Verify that all regions have been processed.
1837     for (unsigned int id = old_space_id; id < last_space_id; ++id) {
1838       verify_complete(SpaceId(id));
1839     }
1840 #endif
1841   }
1842 }
1843 
1844 #ifdef  ASSERT
1845 void PSParallelCompact::verify_filler_in_dense_prefix() {
1846   HeapWord* bottom = _space_info[old_space_id].space()->bottom();
1847   HeapWord* dense_prefix_end = dense_prefix(old_space_id);
1848 
1849   const size_t region_size = ParallelCompactData::RegionSize;
1850 
1851   for (HeapWord* cur_addr = bottom; cur_addr < dense_prefix_end; /* empty */) {
1852     RegionData* cur_region_ptr = _summary_data.addr_to_region_ptr(cur_addr);
1853     if (cur_region_ptr->data_size() == region_size) {
1854       // Full; no dead space. Next region.
1855       if (_summary_data.is_region_aligned(cur_addr)) {
1856         cur_addr += region_size;
1857       } else {
1858         cur_addr = _summary_data.region_align_up(cur_addr);
1859       }
1860       continue;
1861     }
1862 
1863     // This region contains filler objs.
1864     if (_summary_data.is_region_aligned(cur_addr)) {
1865       cur_addr += cur_region_ptr->partial_obj_size();
1866     }
1867 
1868     HeapWord* region_end_addr = _summary_data.region_align_up(cur_addr + 1);
1869     assert(region_end_addr <= dense_prefix_end, "inv");
1870 
1871     while (cur_addr < region_end_addr) {
1872       oop obj = cast_to_oop(cur_addr);
1873       oopDesc::verify(obj);
1874       if (!mark_bitmap()->is_marked(cur_addr)) {
1875         assert(CollectedHeap::is_filler_object(cast_to_oop(cur_addr)), "inv");
1876       }
1877       cur_addr += obj->size();
1878     }
1879   }
1880 }
1881 
1882 void PSParallelCompact::verify_complete(SpaceId space_id) {
1883   // All Regions served as compaction targets, from dense_prefix() to
1884   // new_top(), should be marked as filled and all Regions between new_top()
1885   // and top() should be available (i.e., should have been emptied).
1886   ParallelCompactData& sd = summary_data();
1887   SpaceInfo si = _space_info[space_id];
1888   HeapWord* new_top_addr = sd.region_align_up(si.new_top());
1889   HeapWord* old_top_addr = sd.region_align_up(si.space()->top());
1890   const size_t beg_region = sd.addr_to_region_idx(si.dense_prefix());
1891   const size_t new_top_region = sd.addr_to_region_idx(new_top_addr);
1892   const size_t old_top_region = sd.addr_to_region_idx(old_top_addr);
1893 
1894   size_t cur_region;
1895   for (cur_region = beg_region; cur_region < new_top_region; ++cur_region) {
1896     const RegionData* const c = sd.region(cur_region);
1897     assert(c->completed(), "region %zu not filled: destination_count=%u",
1898            cur_region, c->destination_count());
1899   }
1900 
1901   for (cur_region = new_top_region; cur_region < old_top_region; ++cur_region) {
1902     const RegionData* const c = sd.region(cur_region);
1903     assert(c->available(), "region %zu not empty: destination_count=%u",
1904            cur_region, c->destination_count());
1905   }
1906 }
1907 #endif  // #ifdef ASSERT
1908 
1909 // Return the SpaceId for the space containing addr.  If addr is not in the
1910 // heap, last_space_id is returned.  In debug mode it expects the address to be
1911 // in the heap and asserts such.
1912 PSParallelCompact::SpaceId PSParallelCompact::space_id(HeapWord* addr) {
1913   assert(ParallelScavengeHeap::heap()->is_in_reserved(addr), "addr not in the heap");
1914 
1915   for (unsigned int id = old_space_id; id < last_space_id; ++id) {
1916     if (_space_info[id].space()->contains(addr)) {
1917       return SpaceId(id);
1918     }
1919   }
1920 
1921   assert(false, "no space contains the addr");
1922   return last_space_id;
1923 }
1924 
1925 // Skip over count live words starting from beg, and return the address of the
1926 // next live word. Callers must also ensure that there are enough live words in
1927 // the range [beg, end) to skip.
1928 HeapWord* PSParallelCompact::skip_live_words(HeapWord* beg, HeapWord* end, size_t count)
1929 {
1930   ParMarkBitMap* m = mark_bitmap();
1931   HeapWord* cur_addr = beg;
1932   while (true) {
1933     cur_addr = m->find_obj_beg(cur_addr, end);
1934     assert(cur_addr < end, "inv");
1935     size_t obj_size = cast_to_oop(cur_addr)->size();
1936     // Strictly greater-than
1937     if (obj_size > count) {
1938       return cur_addr + count;
1939     }
1940     count -= obj_size;
1941     cur_addr += obj_size;
1942   }
1943 }
1944 
1945 // On starting to fill a destination region (dest-region), we need to know the
1946 // location of the word that will be at the start of the dest-region after
1947 // compaction. A dest-region can have one or more source regions, but only the
1948 // first source-region contains this location. This location is retrieved by
1949 // calling `first_src_addr` on a dest-region.
1950 // Conversely, a source-region has a dest-region which holds the destination of
1951 // the first live word on this source-region, based on which the destination
1952 // for the rest of live words can be derived.
1953 //
1954 // Note:
1955 // There is some complication due to space-boundary-fragmentation (an obj can't
1956 // cross space-boundary) -- a source-region may be split and behave like two
1957 // distinct regions with their own dest-region, as depicted below.
1958 //
1959 // source-region: region-n
1960 //
1961 // **********************
1962 // |     A|A~~~~B|B     |
1963 // **********************
1964 //    n-1     n     n+1
1965 //
1966 // AA, BB denote two live objs. ~~~~ denotes unknown number of live objs.
1967 //
1968 // Assuming the dest-region for region-n is the final region before
1969 // old-space-end and its first-live-word is the middle of AA, the heap content
1970 // will look like the following after compaction:
1971 //
1972 // **************                  *************
1973 //      A|A~~~~ |                  |BB    |
1974 // **************                  *************
1975 //              ^                  ^
1976 //              | old-space-end    | eden-space-start
1977 //
1978 // Therefore, in this example, region-n will have two dest-regions:
1979 // 1. the final region in old-space
1980 // 2. the first region in eden-space.
1981 // To handle this special case, we introduce the concept of split-region, whose
1982 // contents are relocated to two spaces. `SplitInfo` captures all necessary
1983 // info about the split, the first part, spliting-point, and the second part.
1984 HeapWord* PSParallelCompact::first_src_addr(HeapWord* const dest_addr,
1985                                             SpaceId src_space_id,
1986                                             size_t src_region_idx)
1987 {
1988   const size_t RegionSize = ParallelCompactData::RegionSize;
1989   const ParallelCompactData& sd = summary_data();
1990   assert(sd.is_region_aligned(dest_addr), "precondition");
1991 
1992   const RegionData* const src_region_ptr = sd.region(src_region_idx);
1993   assert(src_region_ptr->data_size() > 0, "src region cannot be empty");
1994 
1995   const size_t partial_obj_size = src_region_ptr->partial_obj_size();
1996   HeapWord* const src_region_destination = src_region_ptr->destination();
1997 
1998   HeapWord* const region_start = sd.region_to_addr(src_region_idx);
1999   HeapWord* const region_end = sd.region_to_addr(src_region_idx) + RegionSize;
2000 
2001   // Identify the actual destination for the first live words on this region,
2002   // taking split-region into account.
2003   HeapWord* region_start_destination;
2004   const SplitInfo& split_info = _space_info[src_space_id].split_info();
2005   if (split_info.is_split(src_region_idx)) {
2006     // The second part of this split region; use the recorded split point.
2007     if (dest_addr == src_region_destination) {
2008       return split_info.split_point();
2009     }
2010     region_start_destination = split_info.preceding_destination();
2011   } else {
2012     region_start_destination = src_region_destination;
2013   }
2014 
2015   // Calculate the offset to be skipped
2016   size_t words_to_skip = pointer_delta(dest_addr, region_start_destination);
2017 
2018   HeapWord* result;
2019   if (partial_obj_size > words_to_skip) {
2020     result = region_start + words_to_skip;
2021   } else {
2022     words_to_skip -= partial_obj_size;
2023     result = skip_live_words(region_start + partial_obj_size, region_end, words_to_skip);
2024   }
2025 
2026   if (split_info.is_split(src_region_idx)) {
2027     assert(result < split_info.split_point(), "postcondition");
2028   } else {
2029     assert(result < region_end, "postcondition");
2030   }
2031 
2032   return result;
2033 }
2034 
2035 void PSParallelCompact::decrement_destination_counts(ParCompactionManager* cm,
2036                                                      SpaceId src_space_id,
2037                                                      size_t beg_region,
2038                                                      HeapWord* end_addr)
2039 {
2040   ParallelCompactData& sd = summary_data();
2041 
2042 #ifdef ASSERT
2043   MutableSpace* const src_space = _space_info[src_space_id].space();
2044   HeapWord* const beg_addr = sd.region_to_addr(beg_region);
2045   assert(src_space->contains(beg_addr) || beg_addr == src_space->end(),
2046          "src_space_id does not match beg_addr");
2047   assert(src_space->contains(end_addr) || end_addr == src_space->end(),
2048          "src_space_id does not match end_addr");
2049 #endif // #ifdef ASSERT
2050 
2051   RegionData* const beg = sd.region(beg_region);
2052   RegionData* const end = sd.addr_to_region_ptr(sd.region_align_up(end_addr));
2053 
2054   // Regions up to new_top() are enqueued if they become available.
2055   HeapWord* const new_top = _space_info[src_space_id].new_top();
2056   RegionData* const enqueue_end =
2057     sd.addr_to_region_ptr(sd.region_align_up(new_top));
2058 
2059   for (RegionData* cur = beg; cur < end; ++cur) {
2060     assert(cur->data_size() > 0, "region must have live data");
2061     cur->decrement_destination_count();
2062     if (cur < enqueue_end && cur->available() && cur->claim()) {
2063       if (cur->mark_normal()) {
2064         cm->push_region(sd.region(cur));
2065       } else if (cur->mark_copied()) {
2066         // Try to copy the content of the shadow region back to its corresponding
2067         // heap region if the shadow region is filled. Otherwise, the GC thread
2068         // fills the shadow region will copy the data back (see
2069         // MoveAndUpdateShadowClosure::complete_region).
2070         copy_back(sd.region_to_addr(cur->shadow_region()), sd.region_to_addr(cur));
2071         ParCompactionManager::push_shadow_region_mt_safe(cur->shadow_region());
2072         cur->set_completed();
2073       }
2074     }
2075   }
2076 }
2077 
2078 size_t PSParallelCompact::next_src_region(MoveAndUpdateClosure& closure,
2079                                           SpaceId& src_space_id,
2080                                           HeapWord*& src_space_top,
2081                                           HeapWord* end_addr)
2082 {
2083   ParallelCompactData& sd = PSParallelCompact::summary_data();
2084 
2085   size_t src_region_idx = 0;
2086 
2087   // Skip empty regions (if any) up to the top of the space.
2088   HeapWord* const src_aligned_up = sd.region_align_up(end_addr);
2089   RegionData* src_region_ptr = sd.addr_to_region_ptr(src_aligned_up);
2090   HeapWord* const top_aligned_up = sd.region_align_up(src_space_top);
2091   const RegionData* const top_region_ptr = sd.addr_to_region_ptr(top_aligned_up);
2092 
2093   while (src_region_ptr < top_region_ptr && src_region_ptr->data_size() == 0) {
2094     ++src_region_ptr;
2095   }
2096 
2097   if (src_region_ptr < top_region_ptr) {
2098     // Found the first non-empty region in the same space.
2099     src_region_idx = sd.region(src_region_ptr);
2100     closure.set_source(sd.region_to_addr(src_region_idx));
2101     return src_region_idx;
2102   }
2103 
2104   // Switch to a new source space and find the first non-empty region.
2105   uint space_id = src_space_id + 1;
2106   assert(space_id < last_space_id, "not enough spaces");
2107 
2108   for (/* empty */; space_id < last_space_id; ++space_id) {
2109     HeapWord* bottom = _space_info[space_id].space()->bottom();
2110     HeapWord* top = _space_info[space_id].space()->top();
2111     // Skip empty space
2112     if (bottom == top) {
2113       continue;
2114     }
2115 
2116     // Identify the first region that contains live words in this space
2117     size_t cur_region = sd.addr_to_region_idx(bottom);
2118     size_t end_region = sd.addr_to_region_idx(sd.region_align_up(top));
2119 
2120     for (/* empty */ ; cur_region < end_region; ++cur_region) {
2121       RegionData* cur = sd.region(cur_region);
2122       if (cur->live_obj_size() > 0) {
2123         HeapWord* region_start_addr = sd.region_to_addr(cur_region);
2124 
2125         src_space_id = SpaceId(space_id);
2126         src_space_top = top;
2127         closure.set_source(region_start_addr);
2128         return cur_region;
2129       }
2130     }
2131   }
2132 
2133   ShouldNotReachHere();
2134 }
2135 
2136 HeapWord* PSParallelCompact::partial_obj_end(HeapWord* region_start_addr) {
2137   ParallelCompactData& sd = summary_data();
2138   assert(sd.is_region_aligned(region_start_addr), "precondition");
2139 
2140   // Use per-region partial_obj_size to locate the end of the obj, that extends
2141   // to region_start_addr.
2142   size_t start_region_idx = sd.addr_to_region_idx(region_start_addr);
2143   size_t end_region_idx = sd.region_count();
2144   size_t accumulated_size = 0;
2145   for (size_t region_idx = start_region_idx; region_idx < end_region_idx; ++region_idx) {
2146     size_t cur_partial_obj_size = sd.region(region_idx)->partial_obj_size();
2147     accumulated_size += cur_partial_obj_size;
2148     if (cur_partial_obj_size != ParallelCompactData::RegionSize) {
2149       break;
2150     }
2151   }
2152   return region_start_addr + accumulated_size;
2153 }
2154 
2155 static markWord safe_mark_word_prototype(HeapWord* cur_addr, HeapWord* end_addr) {
2156   // If the original markWord contains bits that cannot be reconstructed because
2157   // the header cannot be safely read, a placeholder is used. In this case,
2158   // the correct markWord is preserved before compaction and restored after
2159   // compaction completes.
2160   size_t remaining_words = pointer_delta(end_addr, cur_addr);
2161 
2162   if (UseCompactObjectHeaders || (Arguments::is_valhalla_enabled() && safe_to_read_header(remaining_words))) {
2163     return cast_to_oop(cur_addr)->klass()->prototype_header();
2164   } else {
2165     return markWord::prototype();
2166   }
2167 }
2168 
2169 // Use region_idx as the destination region, and evacuate all live objs on its
2170 // source regions to this destination region.
2171 void PSParallelCompact::fill_region(ParCompactionManager* cm, MoveAndUpdateClosure& closure, size_t region_idx)
2172 {
2173   ParMarkBitMap* const bitmap = mark_bitmap();
2174   ParallelCompactData& sd = summary_data();
2175   RegionData* const region_ptr = sd.region(region_idx);
2176 
2177   // Get the source region and related info.
2178   size_t src_region_idx = region_ptr->source_region();
2179   SpaceId src_space_id = space_id(sd.region_to_addr(src_region_idx));
2180   HeapWord* src_space_top = _space_info[src_space_id].space()->top();
2181   HeapWord* dest_addr = sd.region_to_addr(region_idx);
2182 
2183   closure.set_source(first_src_addr(dest_addr, src_space_id, src_region_idx));
2184 
2185   // Adjust src_region_idx to prepare for decrementing destination counts (the
2186   // destination count is not decremented when a region is copied to itself).
2187   if (src_region_idx == region_idx) {
2188     src_region_idx += 1;
2189   }
2190 
2191   // source-region:
2192   //
2193   // **********
2194   // |   ~~~  |
2195   // **********
2196   //      ^
2197   //      |-- closure.source() / first_src_addr
2198   //
2199   //
2200   // ~~~ : live words
2201   //
2202   // destination-region:
2203   //
2204   // **********
2205   // |        |
2206   // **********
2207   // ^
2208   // |-- region-start
2209   if (bitmap->is_unmarked(closure.source())) {
2210     // An object overflows the previous destination region, so this
2211     // destination region should copy the remainder of the object or as much as
2212     // will fit.
2213     HeapWord* const old_src_addr = closure.source();
2214     {
2215       HeapWord* region_start = sd.region_align_down(closure.source());
2216       HeapWord* obj_start = bitmap->find_obj_beg_reverse(region_start, closure.source());
2217       HeapWord* obj_end;
2218       if (obj_start != closure.source()) {
2219         assert(bitmap->is_marked(obj_start), "inv");
2220         // Found the actual obj-start, try to find the obj-end using either
2221         // size() if this obj is completely contained in the current region.
2222         HeapWord* next_region_start = region_start + ParallelCompactData::RegionSize;
2223         HeapWord* partial_obj_start = (next_region_start >= src_space_top)
2224                                       ? nullptr
2225                                       : sd.addr_to_region_ptr(next_region_start)->partial_obj_addr();
2226         // This obj extends to next region iff partial_obj_addr of the *next*
2227         // region is the same as obj-start.
2228         if (partial_obj_start == obj_start) {
2229           // This obj extends to next region.
2230           obj_end = partial_obj_end(next_region_start);
2231         } else {
2232           // Completely contained in this region; safe to use size().
2233           obj_end = obj_start + cast_to_oop(obj_start)->size();
2234         }
2235       } else {
2236         // This obj extends to current region.
2237         obj_end = partial_obj_end(region_start);
2238       }
2239       size_t partial_obj_size = pointer_delta(obj_end, closure.source());
2240       closure.copy_partial_obj(partial_obj_size);
2241     }
2242 
2243     if (closure.is_full()) {
2244       decrement_destination_counts(cm, src_space_id, src_region_idx, closure.source());
2245       closure.complete_region(dest_addr, region_ptr);
2246       return;
2247     }
2248 
2249     // Finished copying without using up the current destination-region
2250     HeapWord* const end_addr = sd.region_align_down(closure.source());
2251     if (sd.region_align_down(old_src_addr) != end_addr) {
2252       assert(sd.region_align_up(old_src_addr) == end_addr, "only one region");
2253       // The partial object was copied from more than one source region.
2254       decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr);
2255 
2256       // Move to the next source region, possibly switching spaces as well.  All
2257       // args except end_addr may be modified.
2258       src_region_idx = next_src_region(closure, src_space_id, src_space_top, end_addr);
2259     }
2260   }
2261 
2262   // Handle the rest obj-by-obj, where we know obj-start.
2263   do {
2264     HeapWord* cur_addr = closure.source();
2265     HeapWord* const end_addr = MIN2(sd.region_align_up(cur_addr + 1),
2266                                     src_space_top);
2267     // To handle the case where the final obj in source region extends to next region.
2268     HeapWord* final_obj_start = (end_addr == src_space_top)
2269                                 ? nullptr
2270                                 : sd.addr_to_region_ptr(end_addr)->partial_obj_addr();
2271     // Apply closure on objs inside [cur_addr, end_addr)
2272     do {
2273       cur_addr = bitmap->find_obj_beg(cur_addr, end_addr);
2274       if (cur_addr == end_addr) {
2275         break;
2276       }
2277       size_t obj_size;
2278       if (final_obj_start == cur_addr) {
2279         obj_size = pointer_delta(partial_obj_end(end_addr), cur_addr);
2280       } else {
2281         // This obj doesn't extend into next region; size() is safe to use.
2282         obj_size = cast_to_oop(cur_addr)->size();
2283       }
2284 
2285       markWord mark = safe_mark_word_prototype(cur_addr, end_addr);
2286 
2287       // Perform the move and update of the object
2288       closure.do_addr(cur_addr, obj_size, mark);
2289 
2290       cur_addr += obj_size;
2291     } while (cur_addr < end_addr && !closure.is_full());
2292 
2293     if (closure.is_full()) {
2294       decrement_destination_counts(cm, src_space_id, src_region_idx, closure.source());
2295       closure.complete_region(dest_addr, region_ptr);
2296       return;
2297     }
2298 
2299     decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr);
2300 
2301     // Move to the next source region, possibly switching spaces as well.  All
2302     // args except end_addr may be modified.
2303     src_region_idx = next_src_region(closure, src_space_id, src_space_top, end_addr);
2304   } while (true);
2305 }
2306 
2307 void PSParallelCompact::fill_and_update_region(ParCompactionManager* cm, size_t region_idx)
2308 {
2309   MoveAndUpdateClosure cl(mark_bitmap(), region_idx);
2310   fill_region(cm, cl, region_idx);
2311 }
2312 
2313 void PSParallelCompact::fill_and_update_shadow_region(ParCompactionManager* cm, size_t region_idx)
2314 {
2315   // Get a shadow region first
2316   ParallelCompactData& sd = summary_data();
2317   RegionData* const region_ptr = sd.region(region_idx);
2318   size_t shadow_region = ParCompactionManager::pop_shadow_region_mt_safe(region_ptr);
2319   // The InvalidShadow return value indicates the corresponding heap region is available,
2320   // so use MoveAndUpdateClosure to fill the normal region. Otherwise, use
2321   // MoveAndUpdateShadowClosure to fill the acquired shadow region.
2322   if (shadow_region == ParCompactionManager::InvalidShadow) {
2323     MoveAndUpdateClosure cl(mark_bitmap(), region_idx);
2324     region_ptr->shadow_to_normal();
2325     return fill_region(cm, cl, region_idx);
2326   } else {
2327     MoveAndUpdateShadowClosure cl(mark_bitmap(), region_idx, shadow_region);
2328     return fill_region(cm, cl, region_idx);
2329   }
2330 }
2331 
2332 void PSParallelCompact::copy_back(HeapWord *shadow_addr, HeapWord *region_addr)
2333 {
2334   Copy::aligned_conjoint_words(shadow_addr, region_addr, _summary_data.RegionSize);
2335 }
2336 
2337 bool PSParallelCompact::steal_unavailable_region(ParCompactionManager* cm, size_t &region_idx)
2338 {
2339   size_t next = cm->next_shadow_region();
2340   ParallelCompactData& sd = summary_data();
2341   size_t old_new_top = sd.addr_to_region_idx(_space_info[old_space_id].new_top());
2342   uint active_gc_threads = ParallelScavengeHeap::heap()->workers().active_workers();
2343 
2344   while (next < old_new_top) {
2345     if (sd.region(next)->mark_shadow()) {
2346       region_idx = next;
2347       return true;
2348     }
2349     next = cm->move_next_shadow_region_by(active_gc_threads);
2350   }
2351 
2352   return false;
2353 }
2354 
2355 // The shadow region is an optimization to address region dependencies in full GC. The basic
2356 // idea is making more regions available by temporally storing their live objects in empty
2357 // shadow regions to resolve dependencies between them and the destination regions. Therefore,
2358 // GC threads need not wait destination regions to be available before processing sources.
2359 //
2360 // A typical workflow would be:
2361 // After draining its own stack and failing to steal from others, a GC worker would pick an
2362 // unavailable region (destination count > 0) and get a shadow region. Then the worker fills
2363 // the shadow region by copying live objects from source regions of the unavailable one. Once
2364 // the unavailable region becomes available, the data in the shadow region will be copied back.
2365 // Shadow regions are empty regions in the to-space and regions between top and end of other spaces.
2366 void PSParallelCompact::initialize_shadow_regions(uint parallel_gc_threads)
2367 {
2368   const ParallelCompactData& sd = PSParallelCompact::summary_data();
2369 
2370   for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2371     SpaceInfo* const space_info = _space_info + id;
2372     MutableSpace* const space = space_info->space();
2373 
2374     const size_t beg_region =
2375       sd.addr_to_region_idx(sd.region_align_up(MAX2(space_info->new_top(), space->top())));
2376     const size_t end_region =
2377       sd.addr_to_region_idx(sd.region_align_down(space->end()));
2378 
2379     for (size_t cur = beg_region; cur < end_region; ++cur) {
2380       ParCompactionManager::push_shadow_region(cur);
2381     }
2382   }
2383 
2384   size_t beg_region = sd.addr_to_region_idx(_space_info[old_space_id].dense_prefix());
2385   for (uint i = 0; i < parallel_gc_threads; i++) {
2386     ParCompactionManager *cm = ParCompactionManager::gc_thread_compaction_manager(i);
2387     cm->set_next_shadow_region(beg_region + i);
2388   }
2389 }
2390 
2391 void MoveAndUpdateClosure::copy_partial_obj(size_t partial_obj_size)
2392 {
2393   size_t words = MIN2(partial_obj_size, words_remaining());
2394 
2395   // This test is necessary; if omitted, the pointer updates to a partial object
2396   // that crosses the dense prefix boundary could be overwritten.
2397   if (source() != copy_destination()) {
2398     DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
2399     Copy::aligned_conjoint_words(source(), copy_destination(), words);
2400   }
2401   update_state(words);
2402 }
2403 
2404 void MoveAndUpdateClosure::complete_region(HeapWord* dest_addr, PSParallelCompact::RegionData* region_ptr) {
2405   assert(region_ptr->shadow_state() == ParallelCompactData::RegionData::NormalRegion, "Region should be finished");
2406   region_ptr->set_completed();
2407 }
2408 
2409 void MoveAndUpdateClosure::do_addr(HeapWord* addr, size_t words, markWord mark) {
2410   assert(destination() != nullptr, "sanity");
2411   _source = addr;
2412 
2413   // The start_array must be updated even if the object is not moving.
2414   if (_start_array != nullptr) {
2415     _start_array->update_for_block(destination(), destination() + words);
2416   }
2417 
2418   // Avoid overflow
2419   words = MIN2(words, words_remaining());
2420   assert(words > 0, "inv");
2421 
2422   if (copy_destination() != source()) {
2423     DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
2424     assert(source() != destination(), "inv");
2425     assert(FullGCForwarding::is_forwarded(cast_to_oop(source())), "inv");
2426     assert(FullGCForwarding::forwardee(cast_to_oop(source())) == cast_to_oop(destination()), "inv");
2427     Copy::aligned_conjoint_words(source(), copy_destination(), words);
2428     cast_to_oop(copy_destination())->set_mark(mark);
2429   }
2430 
2431   update_state(words);
2432 }
2433 
2434 void MoveAndUpdateShadowClosure::complete_region(HeapWord* dest_addr, PSParallelCompact::RegionData* region_ptr) {
2435   assert(region_ptr->shadow_state() == ParallelCompactData::RegionData::ShadowRegion, "Region should be shadow");
2436   // Record the shadow region index
2437   region_ptr->set_shadow_region(_shadow);
2438   // Mark the shadow region as filled to indicate the data is ready to be
2439   // copied back
2440   region_ptr->mark_filled();
2441   // Try to copy the content of the shadow region back to its corresponding
2442   // heap region if available; the GC thread that decreases the destination
2443   // count to zero will do the copying otherwise (see
2444   // PSParallelCompact::decrement_destination_counts).
2445   if (((region_ptr->available() && region_ptr->claim()) || region_ptr->claimed()) && region_ptr->mark_copied()) {
2446     region_ptr->set_completed();
2447     PSParallelCompact::copy_back(PSParallelCompact::summary_data().region_to_addr(_shadow), dest_addr);
2448     ParCompactionManager::push_shadow_region_mt_safe(_shadow);
2449   }
2450 }