rev 60421 : [mq]: 8248401-stefank-review

   1 /*
   2  * Copyright (c) 2005, 2020, 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 "precompiled.hpp"
  26 #include "aot/aotLoader.hpp"
  27 #include "classfile/classLoaderDataGraph.hpp"
  28 #include "classfile/javaClasses.inline.hpp"
  29 #include "classfile/stringTable.hpp"
  30 #include "classfile/symbolTable.hpp"
  31 #include "classfile/systemDictionary.hpp"
  32 #include "code/codeCache.hpp"
  33 #include "gc/parallel/parallelArguments.hpp"
  34 #include "gc/parallel/parallelScavengeHeap.inline.hpp"
  35 #include "gc/parallel/parMarkBitMap.inline.hpp"
  36 #include "gc/parallel/psAdaptiveSizePolicy.hpp"
  37 #include "gc/parallel/psCompactionManager.inline.hpp"
  38 #include "gc/parallel/psOldGen.hpp"
  39 #include "gc/parallel/psParallelCompact.inline.hpp"
  40 #include "gc/parallel/psPromotionManager.inline.hpp"
  41 #include "gc/parallel/psRootType.hpp"
  42 #include "gc/parallel/psScavenge.hpp"
  43 #include "gc/parallel/psYoungGen.hpp"
  44 #include "gc/shared/gcCause.hpp"
  45 #include "gc/shared/gcHeapSummary.hpp"
  46 #include "gc/shared/gcId.hpp"
  47 #include "gc/shared/gcLocker.hpp"
  48 #include "gc/shared/gcTimer.hpp"
  49 #include "gc/shared/gcTrace.hpp"
  50 #include "gc/shared/gcTraceTime.inline.hpp"
  51 #include "gc/shared/isGCActiveMark.hpp"
  52 #include "gc/shared/oopStorage.inline.hpp"
  53 #include "gc/shared/oopStorageSet.inline.hpp"
  54 #include "gc/shared/oopStorageSetParState.inline.hpp"
  55 #include "gc/shared/referencePolicy.hpp"
  56 #include "gc/shared/referenceProcessor.hpp"
  57 #include "gc/shared/referenceProcessorPhaseTimes.hpp"
  58 #include "gc/shared/spaceDecorator.inline.hpp"
  59 #include "gc/shared/taskTerminator.hpp"
  60 #include "gc/shared/weakProcessor.hpp"
  61 #include "gc/shared/workerPolicy.hpp"
  62 #include "gc/shared/workgroup.hpp"
  63 #include "logging/log.hpp"
  64 #include "memory/iterator.inline.hpp"
  65 #include "memory/resourceArea.hpp"
  66 #include "memory/universe.hpp"
  67 #include "oops/access.inline.hpp"
  68 #include "oops/instanceClassLoaderKlass.inline.hpp"
  69 #include "oops/instanceKlass.inline.hpp"
  70 #include "oops/instanceMirrorKlass.inline.hpp"
  71 #include "oops/methodData.hpp"
  72 #include "oops/objArrayKlass.inline.hpp"
  73 #include "oops/oop.inline.hpp"
  74 #include "runtime/atomic.hpp"
  75 #include "runtime/handles.inline.hpp"
  76 #include "runtime/safepoint.hpp"
  77 #include "runtime/vmThread.hpp"
  78 #include "services/memTracker.hpp"
  79 #include "services/memoryService.hpp"
  80 #include "utilities/align.hpp"
  81 #include "utilities/debug.hpp"
  82 #include "utilities/events.hpp"
  83 #include "utilities/formatBuffer.hpp"
  84 #include "utilities/macros.hpp"
  85 #include "utilities/stack.inline.hpp"
  86 #if INCLUDE_JVMCI
  87 #include "jvmci/jvmci.hpp"
  88 #endif
  89 
  90 #include <math.h>
  91 
  92 // All sizes are in HeapWords.
  93 const size_t ParallelCompactData::Log2RegionSize  = 16; // 64K words
  94 const size_t ParallelCompactData::RegionSize      = (size_t)1 << Log2RegionSize;
  95 const size_t ParallelCompactData::RegionSizeBytes =
  96   RegionSize << LogHeapWordSize;
  97 const size_t ParallelCompactData::RegionSizeOffsetMask = RegionSize - 1;
  98 const size_t ParallelCompactData::RegionAddrOffsetMask = RegionSizeBytes - 1;
  99 const size_t ParallelCompactData::RegionAddrMask       = ~RegionAddrOffsetMask;
 100 
 101 const size_t ParallelCompactData::Log2BlockSize   = 7; // 128 words
 102 const size_t ParallelCompactData::BlockSize       = (size_t)1 << Log2BlockSize;
 103 const size_t ParallelCompactData::BlockSizeBytes  =
 104   BlockSize << LogHeapWordSize;
 105 const size_t ParallelCompactData::BlockSizeOffsetMask = BlockSize - 1;
 106 const size_t ParallelCompactData::BlockAddrOffsetMask = BlockSizeBytes - 1;
 107 const size_t ParallelCompactData::BlockAddrMask       = ~BlockAddrOffsetMask;
 108 
 109 const size_t ParallelCompactData::BlocksPerRegion = RegionSize / BlockSize;
 110 const size_t ParallelCompactData::Log2BlocksPerRegion =
 111   Log2RegionSize - Log2BlockSize;
 112 
 113 const ParallelCompactData::RegionData::region_sz_t
 114 ParallelCompactData::RegionData::dc_shift = 27;
 115 
 116 const ParallelCompactData::RegionData::region_sz_t
 117 ParallelCompactData::RegionData::dc_mask = ~0U << dc_shift;
 118 
 119 const ParallelCompactData::RegionData::region_sz_t
 120 ParallelCompactData::RegionData::dc_one = 0x1U << dc_shift;
 121 
 122 const ParallelCompactData::RegionData::region_sz_t
 123 ParallelCompactData::RegionData::los_mask = ~dc_mask;
 124 
 125 const ParallelCompactData::RegionData::region_sz_t
 126 ParallelCompactData::RegionData::dc_claimed = 0x8U << dc_shift;
 127 
 128 const ParallelCompactData::RegionData::region_sz_t
 129 ParallelCompactData::RegionData::dc_completed = 0xcU << dc_shift;
 130 
 131 SpaceInfo PSParallelCompact::_space_info[PSParallelCompact::last_space_id];
 132 
 133 SpanSubjectToDiscoveryClosure PSParallelCompact::_span_based_discoverer;
 134 ReferenceProcessor* PSParallelCompact::_ref_processor = NULL;
 135 
 136 double PSParallelCompact::_dwl_mean;
 137 double PSParallelCompact::_dwl_std_dev;
 138 double PSParallelCompact::_dwl_first_term;
 139 double PSParallelCompact::_dwl_adjustment;
 140 #ifdef  ASSERT
 141 bool   PSParallelCompact::_dwl_initialized = false;
 142 #endif  // #ifdef ASSERT
 143 
 144 void SplitInfo::record(size_t src_region_idx, size_t partial_obj_size,
 145                        HeapWord* destination)
 146 {
 147   assert(src_region_idx != 0, "invalid src_region_idx");
 148   assert(partial_obj_size != 0, "invalid partial_obj_size argument");
 149   assert(destination != NULL, "invalid destination argument");
 150 
 151   _src_region_idx = src_region_idx;
 152   _partial_obj_size = partial_obj_size;
 153   _destination = destination;
 154 
 155   // These fields may not be updated below, so make sure they're clear.
 156   assert(_dest_region_addr == NULL, "should have been cleared");
 157   assert(_first_src_addr == NULL, "should have been cleared");
 158 
 159   // Determine the number of destination regions for the partial object.
 160   HeapWord* const last_word = destination + partial_obj_size - 1;
 161   const ParallelCompactData& sd = PSParallelCompact::summary_data();
 162   HeapWord* const beg_region_addr = sd.region_align_down(destination);
 163   HeapWord* const end_region_addr = sd.region_align_down(last_word);
 164 
 165   if (beg_region_addr == end_region_addr) {
 166     // One destination region.
 167     _destination_count = 1;
 168     if (end_region_addr == destination) {
 169       // The destination falls on a region boundary, thus the first word of the
 170       // partial object will be the first word copied to the destination region.
 171       _dest_region_addr = end_region_addr;
 172       _first_src_addr = sd.region_to_addr(src_region_idx);
 173     }
 174   } else {
 175     // Two destination regions.  When copied, the partial object will cross a
 176     // destination region boundary, so a word somewhere within the partial
 177     // object will be the first word copied to the second destination region.
 178     _destination_count = 2;
 179     _dest_region_addr = end_region_addr;
 180     const size_t ofs = pointer_delta(end_region_addr, destination);
 181     assert(ofs < _partial_obj_size, "sanity");
 182     _first_src_addr = sd.region_to_addr(src_region_idx) + ofs;
 183   }
 184 }
 185 
 186 void SplitInfo::clear()
 187 {
 188   _src_region_idx = 0;
 189   _partial_obj_size = 0;
 190   _destination = NULL;
 191   _destination_count = 0;
 192   _dest_region_addr = NULL;
 193   _first_src_addr = NULL;
 194   assert(!is_valid(), "sanity");
 195 }
 196 
 197 #ifdef  ASSERT
 198 void SplitInfo::verify_clear()
 199 {
 200   assert(_src_region_idx == 0, "not clear");
 201   assert(_partial_obj_size == 0, "not clear");
 202   assert(_destination == NULL, "not clear");
 203   assert(_destination_count == 0, "not clear");
 204   assert(_dest_region_addr == NULL, "not clear");
 205   assert(_first_src_addr == NULL, "not clear");
 206 }
 207 #endif  // #ifdef ASSERT
 208 
 209 
 210 void PSParallelCompact::print_on_error(outputStream* st) {
 211   _mark_bitmap.print_on_error(st);
 212 }
 213 
 214 #ifndef PRODUCT
 215 const char* PSParallelCompact::space_names[] = {
 216   "old ", "eden", "from", "to  "
 217 };
 218 
 219 void PSParallelCompact::print_region_ranges() {
 220   if (!log_develop_is_enabled(Trace, gc, compaction)) {
 221     return;
 222   }
 223   Log(gc, compaction) log;
 224   ResourceMark rm;
 225   LogStream ls(log.trace());
 226   Universe::print_on(&ls);
 227   log.trace("space  bottom     top        end        new_top");
 228   log.trace("------ ---------- ---------- ---------- ----------");
 229 
 230   for (unsigned int id = 0; id < last_space_id; ++id) {
 231     const MutableSpace* space = _space_info[id].space();
 232     log.trace("%u %s "
 233               SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " "
 234               SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " ",
 235               id, space_names[id],
 236               summary_data().addr_to_region_idx(space->bottom()),
 237               summary_data().addr_to_region_idx(space->top()),
 238               summary_data().addr_to_region_idx(space->end()),
 239               summary_data().addr_to_region_idx(_space_info[id].new_top()));
 240   }
 241 }
 242 
 243 void
 244 print_generic_summary_region(size_t i, const ParallelCompactData::RegionData* c)
 245 {
 246 #define REGION_IDX_FORMAT        SIZE_FORMAT_W(7)
 247 #define REGION_DATA_FORMAT       SIZE_FORMAT_W(5)
 248 
 249   ParallelCompactData& sd = PSParallelCompact::summary_data();
 250   size_t dci = c->destination() ? sd.addr_to_region_idx(c->destination()) : 0;
 251   log_develop_trace(gc, compaction)(
 252       REGION_IDX_FORMAT " " PTR_FORMAT " "
 253       REGION_IDX_FORMAT " " PTR_FORMAT " "
 254       REGION_DATA_FORMAT " " REGION_DATA_FORMAT " "
 255       REGION_DATA_FORMAT " " REGION_IDX_FORMAT " %d",
 256       i, p2i(c->data_location()), dci, p2i(c->destination()),
 257       c->partial_obj_size(), c->live_obj_size(),
 258       c->data_size(), c->source_region(), c->destination_count());
 259 
 260 #undef  REGION_IDX_FORMAT
 261 #undef  REGION_DATA_FORMAT
 262 }
 263 
 264 void
 265 print_generic_summary_data(ParallelCompactData& summary_data,
 266                            HeapWord* const beg_addr,
 267                            HeapWord* const end_addr)
 268 {
 269   size_t total_words = 0;
 270   size_t i = summary_data.addr_to_region_idx(beg_addr);
 271   const size_t last = summary_data.addr_to_region_idx(end_addr);
 272   HeapWord* pdest = 0;
 273 
 274   while (i < last) {
 275     ParallelCompactData::RegionData* c = summary_data.region(i);
 276     if (c->data_size() != 0 || c->destination() != pdest) {
 277       print_generic_summary_region(i, c);
 278       total_words += c->data_size();
 279       pdest = c->destination();
 280     }
 281     ++i;
 282   }
 283 
 284   log_develop_trace(gc, compaction)("summary_data_bytes=" SIZE_FORMAT, total_words * HeapWordSize);
 285 }
 286 
 287 void
 288 PSParallelCompact::print_generic_summary_data(ParallelCompactData& summary_data,
 289                                               HeapWord* const beg_addr,
 290                                               HeapWord* const end_addr) {
 291   ::print_generic_summary_data(summary_data,beg_addr, end_addr);
 292 }
 293 
 294 void
 295 print_generic_summary_data(ParallelCompactData& summary_data,
 296                            SpaceInfo* space_info)
 297 {
 298   if (!log_develop_is_enabled(Trace, gc, compaction)) {
 299     return;
 300   }
 301 
 302   for (unsigned int id = 0; id < PSParallelCompact::last_space_id; ++id) {
 303     const MutableSpace* space = space_info[id].space();
 304     print_generic_summary_data(summary_data, space->bottom(),
 305                                MAX2(space->top(), space_info[id].new_top()));
 306   }
 307 }
 308 
 309 void
 310 print_initial_summary_data(ParallelCompactData& summary_data,
 311                            const MutableSpace* space) {
 312   if (space->top() == space->bottom()) {
 313     return;
 314   }
 315 
 316   const size_t region_size = ParallelCompactData::RegionSize;
 317   typedef ParallelCompactData::RegionData RegionData;
 318   HeapWord* const top_aligned_up = summary_data.region_align_up(space->top());
 319   const size_t end_region = summary_data.addr_to_region_idx(top_aligned_up);
 320   const RegionData* c = summary_data.region(end_region - 1);
 321   HeapWord* end_addr = c->destination() + c->data_size();
 322   const size_t live_in_space = pointer_delta(end_addr, space->bottom());
 323 
 324   // Print (and count) the full regions at the beginning of the space.
 325   size_t full_region_count = 0;
 326   size_t i = summary_data.addr_to_region_idx(space->bottom());
 327   while (i < end_region && summary_data.region(i)->data_size() == region_size) {
 328     ParallelCompactData::RegionData* c = summary_data.region(i);
 329     log_develop_trace(gc, compaction)(
 330         SIZE_FORMAT_W(5) " " PTR_FORMAT " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d",
 331         i, p2i(c->destination()),
 332         c->partial_obj_size(), c->live_obj_size(),
 333         c->data_size(), c->source_region(), c->destination_count());
 334     ++full_region_count;
 335     ++i;
 336   }
 337 
 338   size_t live_to_right = live_in_space - full_region_count * region_size;
 339 
 340   double max_reclaimed_ratio = 0.0;
 341   size_t max_reclaimed_ratio_region = 0;
 342   size_t max_dead_to_right = 0;
 343   size_t max_live_to_right = 0;
 344 
 345   // Print the 'reclaimed ratio' for regions while there is something live in
 346   // the region or to the right of it.  The remaining regions are empty (and
 347   // uninteresting), and computing the ratio will result in division by 0.
 348   while (i < end_region && live_to_right > 0) {
 349     c = summary_data.region(i);
 350     HeapWord* const region_addr = summary_data.region_to_addr(i);
 351     const size_t used_to_right = pointer_delta(space->top(), region_addr);
 352     const size_t dead_to_right = used_to_right - live_to_right;
 353     const double reclaimed_ratio = double(dead_to_right) / live_to_right;
 354 
 355     if (reclaimed_ratio > max_reclaimed_ratio) {
 356             max_reclaimed_ratio = reclaimed_ratio;
 357             max_reclaimed_ratio_region = i;
 358             max_dead_to_right = dead_to_right;
 359             max_live_to_right = live_to_right;
 360     }
 361 
 362     ParallelCompactData::RegionData* c = summary_data.region(i);
 363     log_develop_trace(gc, compaction)(
 364         SIZE_FORMAT_W(5) " " PTR_FORMAT " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d"
 365         "%12.10f " SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10),
 366         i, p2i(c->destination()),
 367         c->partial_obj_size(), c->live_obj_size(),
 368         c->data_size(), c->source_region(), c->destination_count(),
 369         reclaimed_ratio, dead_to_right, live_to_right);
 370 
 371 
 372     live_to_right -= c->data_size();
 373     ++i;
 374   }
 375 
 376   // Any remaining regions are empty.  Print one more if there is one.
 377   if (i < end_region) {
 378     ParallelCompactData::RegionData* c = summary_data.region(i);
 379     log_develop_trace(gc, compaction)(
 380         SIZE_FORMAT_W(5) " " PTR_FORMAT " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d",
 381          i, p2i(c->destination()),
 382          c->partial_obj_size(), c->live_obj_size(),
 383          c->data_size(), c->source_region(), c->destination_count());
 384   }
 385 
 386   log_develop_trace(gc, compaction)("max:  " SIZE_FORMAT_W(4) " d2r=" SIZE_FORMAT_W(10) " l2r=" SIZE_FORMAT_W(10) " max_ratio=%14.12f",
 387                                     max_reclaimed_ratio_region, max_dead_to_right, max_live_to_right, max_reclaimed_ratio);
 388 }
 389 
 390 void
 391 print_initial_summary_data(ParallelCompactData& summary_data,
 392                            SpaceInfo* space_info) {
 393   if (!log_develop_is_enabled(Trace, gc, compaction)) {
 394     return;
 395   }
 396 
 397   unsigned int id = PSParallelCompact::old_space_id;
 398   const MutableSpace* space;
 399   do {
 400     space = space_info[id].space();
 401     print_initial_summary_data(summary_data, space);
 402   } while (++id < PSParallelCompact::eden_space_id);
 403 
 404   do {
 405     space = space_info[id].space();
 406     print_generic_summary_data(summary_data, space->bottom(), space->top());
 407   } while (++id < PSParallelCompact::last_space_id);
 408 }
 409 #endif  // #ifndef PRODUCT
 410 
 411 #ifdef  ASSERT
 412 size_t add_obj_count;
 413 size_t add_obj_size;
 414 size_t mark_bitmap_count;
 415 size_t mark_bitmap_size;
 416 #endif  // #ifdef ASSERT
 417 
 418 ParallelCompactData::ParallelCompactData() :
 419   _region_start(NULL),
 420   DEBUG_ONLY(_region_end(NULL) COMMA)
 421   _region_vspace(NULL),
 422   _reserved_byte_size(0),
 423   _region_data(NULL),
 424   _region_count(0),
 425   _block_vspace(NULL),
 426   _block_data(NULL),
 427   _block_count(0) {}
 428 
 429 bool ParallelCompactData::initialize(MemRegion covered_region)
 430 {
 431   _region_start = covered_region.start();
 432   const size_t region_size = covered_region.word_size();
 433   DEBUG_ONLY(_region_end = _region_start + region_size;)
 434 
 435   assert(region_align_down(_region_start) == _region_start,
 436          "region start not aligned");
 437   assert((region_size & RegionSizeOffsetMask) == 0,
 438          "region size not a multiple of RegionSize");
 439 
 440   bool result = initialize_region_data(region_size) && initialize_block_data();
 441   return result;
 442 }
 443 
 444 PSVirtualSpace*
 445 ParallelCompactData::create_vspace(size_t count, size_t element_size)
 446 {
 447   const size_t raw_bytes = count * element_size;
 448   const size_t page_sz = os::page_size_for_region_aligned(raw_bytes, 10);
 449   const size_t granularity = os::vm_allocation_granularity();
 450   _reserved_byte_size = align_up(raw_bytes, MAX2(page_sz, granularity));
 451 
 452   const size_t rs_align = page_sz == (size_t) os::vm_page_size() ? 0 :
 453     MAX2(page_sz, granularity);
 454   ReservedSpace rs(_reserved_byte_size, rs_align, rs_align > 0);
 455   os::trace_page_sizes("Parallel Compact Data", raw_bytes, raw_bytes, page_sz, rs.base(),
 456                        rs.size());
 457 
 458   MemTracker::record_virtual_memory_type((address)rs.base(), mtGC);
 459 
 460   PSVirtualSpace* vspace = new PSVirtualSpace(rs, page_sz);
 461   if (vspace != 0) {
 462     if (vspace->expand_by(_reserved_byte_size)) {
 463       return vspace;
 464     }
 465     delete vspace;
 466     // Release memory reserved in the space.
 467     rs.release();
 468   }
 469 
 470   return 0;
 471 }
 472 
 473 bool ParallelCompactData::initialize_region_data(size_t region_size)
 474 {
 475   const size_t count = (region_size + RegionSizeOffsetMask) >> Log2RegionSize;
 476   _region_vspace = create_vspace(count, sizeof(RegionData));
 477   if (_region_vspace != 0) {
 478     _region_data = (RegionData*)_region_vspace->reserved_low_addr();
 479     _region_count = count;
 480     return true;
 481   }
 482   return false;
 483 }
 484 
 485 bool ParallelCompactData::initialize_block_data()
 486 {
 487   assert(_region_count != 0, "region data must be initialized first");
 488   const size_t count = _region_count << Log2BlocksPerRegion;
 489   _block_vspace = create_vspace(count, sizeof(BlockData));
 490   if (_block_vspace != 0) {
 491     _block_data = (BlockData*)_block_vspace->reserved_low_addr();
 492     _block_count = count;
 493     return true;
 494   }
 495   return false;
 496 }
 497 
 498 void ParallelCompactData::clear()
 499 {
 500   memset(_region_data, 0, _region_vspace->committed_size());
 501   memset(_block_data, 0, _block_vspace->committed_size());
 502 }
 503 
 504 void ParallelCompactData::clear_range(size_t beg_region, size_t end_region) {
 505   assert(beg_region <= _region_count, "beg_region out of range");
 506   assert(end_region <= _region_count, "end_region out of range");
 507   assert(RegionSize % BlockSize == 0, "RegionSize not a multiple of BlockSize");
 508 
 509   const size_t region_cnt = end_region - beg_region;
 510   memset(_region_data + beg_region, 0, region_cnt * sizeof(RegionData));
 511 
 512   const size_t beg_block = beg_region * BlocksPerRegion;
 513   const size_t block_cnt = region_cnt * BlocksPerRegion;
 514   memset(_block_data + beg_block, 0, block_cnt * sizeof(BlockData));
 515 }
 516 
 517 HeapWord* ParallelCompactData::partial_obj_end(size_t region_idx) const
 518 {
 519   const RegionData* cur_cp = region(region_idx);
 520   const RegionData* const end_cp = region(region_count() - 1);
 521 
 522   HeapWord* result = region_to_addr(region_idx);
 523   if (cur_cp < end_cp) {
 524     do {
 525       result += cur_cp->partial_obj_size();
 526     } while (cur_cp->partial_obj_size() == RegionSize && ++cur_cp < end_cp);
 527   }
 528   return result;
 529 }
 530 
 531 void ParallelCompactData::add_obj(HeapWord* addr, size_t len)
 532 {
 533   const size_t obj_ofs = pointer_delta(addr, _region_start);
 534   const size_t beg_region = obj_ofs >> Log2RegionSize;
 535   const size_t end_region = (obj_ofs + len - 1) >> Log2RegionSize;
 536 
 537   DEBUG_ONLY(Atomic::inc(&add_obj_count);)
 538   DEBUG_ONLY(Atomic::add(&add_obj_size, len);)
 539 
 540   if (beg_region == end_region) {
 541     // All in one region.
 542     _region_data[beg_region].add_live_obj(len);
 543     return;
 544   }
 545 
 546   // First region.
 547   const size_t beg_ofs = region_offset(addr);
 548   _region_data[beg_region].add_live_obj(RegionSize - beg_ofs);
 549 
 550   Klass* klass = ((oop)addr)->klass();
 551   // Middle regions--completely spanned by this object.
 552   for (size_t region = beg_region + 1; region < end_region; ++region) {
 553     _region_data[region].set_partial_obj_size(RegionSize);
 554     _region_data[region].set_partial_obj_addr(addr);
 555   }
 556 
 557   // Last region.
 558   const size_t end_ofs = region_offset(addr + len - 1);
 559   _region_data[end_region].set_partial_obj_size(end_ofs + 1);
 560   _region_data[end_region].set_partial_obj_addr(addr);
 561 }
 562 
 563 void
 564 ParallelCompactData::summarize_dense_prefix(HeapWord* beg, HeapWord* end)
 565 {
 566   assert(region_offset(beg) == 0, "not RegionSize aligned");
 567   assert(region_offset(end) == 0, "not RegionSize aligned");
 568 
 569   size_t cur_region = addr_to_region_idx(beg);
 570   const size_t end_region = addr_to_region_idx(end);
 571   HeapWord* addr = beg;
 572   while (cur_region < end_region) {
 573     _region_data[cur_region].set_destination(addr);
 574     _region_data[cur_region].set_destination_count(0);
 575     _region_data[cur_region].set_source_region(cur_region);
 576     _region_data[cur_region].set_data_location(addr);
 577 
 578     // Update live_obj_size so the region appears completely full.
 579     size_t live_size = RegionSize - _region_data[cur_region].partial_obj_size();
 580     _region_data[cur_region].set_live_obj_size(live_size);
 581 
 582     ++cur_region;
 583     addr += RegionSize;
 584   }
 585 }
 586 
 587 // Find the point at which a space can be split and, if necessary, record the
 588 // split point.
 589 //
 590 // If the current src region (which overflowed the destination space) doesn't
 591 // have a partial object, the split point is at the beginning of the current src
 592 // region (an "easy" split, no extra bookkeeping required).
 593 //
 594 // If the current src region has a partial object, the split point is in the
 595 // region where that partial object starts (call it the split_region).  If
 596 // split_region has a partial object, then the split point is just after that
 597 // partial object (a "hard" split where we have to record the split data and
 598 // zero the partial_obj_size field).  With a "hard" split, we know that the
 599 // partial_obj ends within split_region because the partial object that caused
 600 // the overflow starts in split_region.  If split_region doesn't have a partial
 601 // obj, then the split is at the beginning of split_region (another "easy"
 602 // split).
 603 HeapWord*
 604 ParallelCompactData::summarize_split_space(size_t src_region,
 605                                            SplitInfo& split_info,
 606                                            HeapWord* destination,
 607                                            HeapWord* target_end,
 608                                            HeapWord** target_next)
 609 {
 610   assert(destination <= target_end, "sanity");
 611   assert(destination + _region_data[src_region].data_size() > target_end,
 612     "region should not fit into target space");
 613   assert(is_region_aligned(target_end), "sanity");
 614 
 615   size_t split_region = src_region;
 616   HeapWord* split_destination = destination;
 617   size_t partial_obj_size = _region_data[src_region].partial_obj_size();
 618 
 619   if (destination + partial_obj_size > target_end) {
 620     // The split point is just after the partial object (if any) in the
 621     // src_region that contains the start of the object that overflowed the
 622     // destination space.
 623     //
 624     // Find the start of the "overflow" object and set split_region to the
 625     // region containing it.
 626     HeapWord* const overflow_obj = _region_data[src_region].partial_obj_addr();
 627     split_region = addr_to_region_idx(overflow_obj);
 628 
 629     // Clear the source_region field of all destination regions whose first word
 630     // came from data after the split point (a non-null source_region field
 631     // implies a region must be filled).
 632     //
 633     // An alternative to the simple loop below:  clear during post_compact(),
 634     // which uses memcpy instead of individual stores, and is easy to
 635     // parallelize.  (The downside is that it clears the entire RegionData
 636     // object as opposed to just one field.)
 637     //
 638     // post_compact() would have to clear the summary data up to the highest
 639     // address that was written during the summary phase, which would be
 640     //
 641     //         max(top, max(new_top, clear_top))
 642     //
 643     // where clear_top is a new field in SpaceInfo.  Would have to set clear_top
 644     // to target_end.
 645     const RegionData* const sr = region(split_region);
 646     const size_t beg_idx =
 647       addr_to_region_idx(region_align_up(sr->destination() +
 648                                          sr->partial_obj_size()));
 649     const size_t end_idx = addr_to_region_idx(target_end);
 650 
 651     log_develop_trace(gc, compaction)("split:  clearing source_region field in [" SIZE_FORMAT ", " SIZE_FORMAT ")", beg_idx, end_idx);
 652     for (size_t idx = beg_idx; idx < end_idx; ++idx) {
 653       _region_data[idx].set_source_region(0);
 654     }
 655 
 656     // Set split_destination and partial_obj_size to reflect the split region.
 657     split_destination = sr->destination();
 658     partial_obj_size = sr->partial_obj_size();
 659   }
 660 
 661   // The split is recorded only if a partial object extends onto the region.
 662   if (partial_obj_size != 0) {
 663     _region_data[split_region].set_partial_obj_size(0);
 664     split_info.record(split_region, partial_obj_size, split_destination);
 665   }
 666 
 667   // Setup the continuation addresses.
 668   *target_next = split_destination + partial_obj_size;
 669   HeapWord* const source_next = region_to_addr(split_region) + partial_obj_size;
 670 
 671   if (log_develop_is_enabled(Trace, gc, compaction)) {
 672     const char * split_type = partial_obj_size == 0 ? "easy" : "hard";
 673     log_develop_trace(gc, compaction)("%s split:  src=" PTR_FORMAT " src_c=" SIZE_FORMAT " pos=" SIZE_FORMAT,
 674                                       split_type, p2i(source_next), split_region, partial_obj_size);
 675     log_develop_trace(gc, compaction)("%s split:  dst=" PTR_FORMAT " dst_c=" SIZE_FORMAT " tn=" PTR_FORMAT,
 676                                       split_type, p2i(split_destination),
 677                                       addr_to_region_idx(split_destination),
 678                                       p2i(*target_next));
 679 
 680     if (partial_obj_size != 0) {
 681       HeapWord* const po_beg = split_info.destination();
 682       HeapWord* const po_end = po_beg + split_info.partial_obj_size();
 683       log_develop_trace(gc, compaction)("%s split:  po_beg=" PTR_FORMAT " " SIZE_FORMAT " po_end=" PTR_FORMAT " " SIZE_FORMAT,
 684                                         split_type,
 685                                         p2i(po_beg), addr_to_region_idx(po_beg),
 686                                         p2i(po_end), addr_to_region_idx(po_end));
 687     }
 688   }
 689 
 690   return source_next;
 691 }
 692 
 693 bool ParallelCompactData::summarize(SplitInfo& split_info,
 694                                     HeapWord* source_beg, HeapWord* source_end,
 695                                     HeapWord** source_next,
 696                                     HeapWord* target_beg, HeapWord* target_end,
 697                                     HeapWord** target_next)
 698 {
 699   HeapWord* const source_next_val = source_next == NULL ? NULL : *source_next;
 700   log_develop_trace(gc, compaction)(
 701       "sb=" PTR_FORMAT " se=" PTR_FORMAT " sn=" PTR_FORMAT
 702       "tb=" PTR_FORMAT " te=" PTR_FORMAT " tn=" PTR_FORMAT,
 703       p2i(source_beg), p2i(source_end), p2i(source_next_val),
 704       p2i(target_beg), p2i(target_end), p2i(*target_next));
 705 
 706   size_t cur_region = addr_to_region_idx(source_beg);
 707   const size_t end_region = addr_to_region_idx(region_align_up(source_end));
 708 
 709   HeapWord *dest_addr = target_beg;
 710   while (cur_region < end_region) {
 711     // The destination must be set even if the region has no data.
 712     _region_data[cur_region].set_destination(dest_addr);
 713 
 714     size_t words = _region_data[cur_region].data_size();
 715     if (words > 0) {
 716       // If cur_region does not fit entirely into the target space, find a point
 717       // at which the source space can be 'split' so that part is copied to the
 718       // target space and the rest is copied elsewhere.
 719       if (dest_addr + words > target_end) {
 720         assert(source_next != NULL, "source_next is NULL when splitting");
 721         *source_next = summarize_split_space(cur_region, split_info, dest_addr,
 722                                              target_end, target_next);
 723         return false;
 724       }
 725 
 726       // Compute the destination_count for cur_region, and if necessary, update
 727       // source_region for a destination region.  The source_region field is
 728       // updated if cur_region is the first (left-most) region to be copied to a
 729       // destination region.
 730       //
 731       // The destination_count calculation is a bit subtle.  A region that has
 732       // data that compacts into itself does not count itself as a destination.
 733       // This maintains the invariant that a zero count means the region is
 734       // available and can be claimed and then filled.
 735       uint destination_count = 0;
 736       if (split_info.is_split(cur_region)) {
 737         // The current region has been split:  the partial object will be copied
 738         // to one destination space and the remaining data will be copied to
 739         // another destination space.  Adjust the initial destination_count and,
 740         // if necessary, set the source_region field if the partial object will
 741         // cross a destination region boundary.
 742         destination_count = split_info.destination_count();
 743         if (destination_count == 2) {
 744           size_t dest_idx = addr_to_region_idx(split_info.dest_region_addr());
 745           _region_data[dest_idx].set_source_region(cur_region);
 746         }
 747       }
 748 
 749       HeapWord* const last_addr = dest_addr + words - 1;
 750       const size_t dest_region_1 = addr_to_region_idx(dest_addr);
 751       const size_t dest_region_2 = addr_to_region_idx(last_addr);
 752 
 753       // Initially assume that the destination regions will be the same and
 754       // adjust the value below if necessary.  Under this assumption, if
 755       // cur_region == dest_region_2, then cur_region will be compacted
 756       // completely into itself.
 757       destination_count += cur_region == dest_region_2 ? 0 : 1;
 758       if (dest_region_1 != dest_region_2) {
 759         // Destination regions differ; adjust destination_count.
 760         destination_count += 1;
 761         // Data from cur_region will be copied to the start of dest_region_2.
 762         _region_data[dest_region_2].set_source_region(cur_region);
 763       } else if (region_offset(dest_addr) == 0) {
 764         // Data from cur_region will be copied to the start of the destination
 765         // region.
 766         _region_data[dest_region_1].set_source_region(cur_region);
 767       }
 768 
 769       _region_data[cur_region].set_destination_count(destination_count);
 770       _region_data[cur_region].set_data_location(region_to_addr(cur_region));
 771       dest_addr += words;
 772     }
 773 
 774     ++cur_region;
 775   }
 776 
 777   *target_next = dest_addr;
 778   return true;
 779 }
 780 
 781 HeapWord* ParallelCompactData::calc_new_pointer(HeapWord* addr, ParCompactionManager* cm) {
 782   assert(addr != NULL, "Should detect NULL oop earlier");
 783   assert(ParallelScavengeHeap::heap()->is_in(addr), "not in heap");
 784   assert(PSParallelCompact::mark_bitmap()->is_marked(addr), "not marked");
 785 
 786   // Region covering the object.
 787   RegionData* const region_ptr = addr_to_region_ptr(addr);
 788   HeapWord* result = region_ptr->destination();
 789 
 790   // If the entire Region is live, the new location is region->destination + the
 791   // offset of the object within in the Region.
 792 
 793   // Run some performance tests to determine if this special case pays off.  It
 794   // is worth it for pointers into the dense prefix.  If the optimization to
 795   // avoid pointer updates in regions that only point to the dense prefix is
 796   // ever implemented, this should be revisited.
 797   if (region_ptr->data_size() == RegionSize) {
 798     result += region_offset(addr);
 799     return result;
 800   }
 801 
 802   // Otherwise, the new location is region->destination + block offset + the
 803   // number of live words in the Block that are (a) to the left of addr and (b)
 804   // due to objects that start in the Block.
 805 
 806   // Fill in the block table if necessary.  This is unsynchronized, so multiple
 807   // threads may fill the block table for a region (harmless, since it is
 808   // idempotent).
 809   if (!region_ptr->blocks_filled()) {
 810     PSParallelCompact::fill_blocks(addr_to_region_idx(addr));
 811     region_ptr->set_blocks_filled();
 812   }
 813 
 814   HeapWord* const search_start = block_align_down(addr);
 815   const size_t block_offset = addr_to_block_ptr(addr)->offset();
 816 
 817   const ParMarkBitMap* bitmap = PSParallelCompact::mark_bitmap();
 818   const size_t live = bitmap->live_words_in_range(cm, search_start, oop(addr));
 819   result += block_offset + live;
 820   DEBUG_ONLY(PSParallelCompact::check_new_location(addr, result));
 821   return result;
 822 }
 823 
 824 #ifdef ASSERT
 825 void ParallelCompactData::verify_clear(const PSVirtualSpace* vspace)
 826 {
 827   const size_t* const beg = (const size_t*)vspace->committed_low_addr();
 828   const size_t* const end = (const size_t*)vspace->committed_high_addr();
 829   for (const size_t* p = beg; p < end; ++p) {
 830     assert(*p == 0, "not zero");
 831   }
 832 }
 833 
 834 void ParallelCompactData::verify_clear()
 835 {
 836   verify_clear(_region_vspace);
 837   verify_clear(_block_vspace);
 838 }
 839 #endif  // #ifdef ASSERT
 840 
 841 STWGCTimer          PSParallelCompact::_gc_timer;
 842 ParallelOldTracer   PSParallelCompact::_gc_tracer;
 843 elapsedTimer        PSParallelCompact::_accumulated_time;
 844 unsigned int        PSParallelCompact::_total_invocations = 0;
 845 unsigned int        PSParallelCompact::_maximum_compaction_gc_num = 0;
 846 CollectorCounters*  PSParallelCompact::_counters = NULL;
 847 ParMarkBitMap       PSParallelCompact::_mark_bitmap;
 848 ParallelCompactData PSParallelCompact::_summary_data;
 849 
 850 PSParallelCompact::IsAliveClosure PSParallelCompact::_is_alive_closure;
 851 
 852 bool PSParallelCompact::IsAliveClosure::do_object_b(oop p) { return mark_bitmap()->is_marked(p); }
 853 
 854 class PCReferenceProcessor: public ReferenceProcessor {
 855 public:
 856   PCReferenceProcessor(
 857     BoolObjectClosure* is_subject_to_discovery,
 858     BoolObjectClosure* is_alive_non_header) :
 859       ReferenceProcessor(is_subject_to_discovery,
 860       ParallelRefProcEnabled && (ParallelGCThreads > 1), // mt processing
 861       ParallelGCThreads,   // mt processing degree
 862       true,                // mt discovery
 863       ParallelGCThreads,   // mt discovery degree
 864       true,                // atomic_discovery
 865       is_alive_non_header) {
 866   }
 867 
 868   template<typename T> bool discover(oop obj, ReferenceType type) {
 869     T* referent_addr = (T*) java_lang_ref_Reference::referent_addr_raw(obj);
 870     T heap_oop = RawAccess<>::oop_load(referent_addr);
 871     oop referent = CompressedOops::decode_not_null(heap_oop);
 872     return PSParallelCompact::mark_bitmap()->is_unmarked(referent)
 873         && ReferenceProcessor::discover_reference(obj, type);
 874   }
 875   virtual bool discover_reference(oop obj, ReferenceType type) {
 876     if (UseCompressedOops) {
 877       return discover<narrowOop>(obj, type);
 878     } else {
 879       return discover<oop>(obj, type);
 880     }
 881   }
 882 };
 883 
 884 void PSParallelCompact::post_initialize() {
 885   ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
 886   _span_based_discoverer.set_span(heap->reserved_region());
 887   _ref_processor =
 888     new PCReferenceProcessor(&_span_based_discoverer,
 889                              &_is_alive_closure); // non-header is alive closure
 890 
 891   _counters = new CollectorCounters("Parallel full collection pauses", 1);
 892 
 893   // Initialize static fields in ParCompactionManager.
 894   ParCompactionManager::initialize(mark_bitmap());
 895 }
 896 
 897 bool PSParallelCompact::initialize() {
 898   ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
 899   MemRegion mr = heap->reserved_region();
 900 
 901   // Was the old gen get allocated successfully?
 902   if (!heap->old_gen()->is_allocated()) {
 903     return false;
 904   }
 905 
 906   initialize_space_info();
 907   initialize_dead_wood_limiter();
 908 
 909   if (!_mark_bitmap.initialize(mr)) {
 910     vm_shutdown_during_initialization(
 911       err_msg("Unable to allocate " SIZE_FORMAT "KB bitmaps for parallel "
 912       "garbage collection for the requested " SIZE_FORMAT "KB heap.",
 913       _mark_bitmap.reserved_byte_size()/K, mr.byte_size()/K));
 914     return false;
 915   }
 916 
 917   if (!_summary_data.initialize(mr)) {
 918     vm_shutdown_during_initialization(
 919       err_msg("Unable to allocate " SIZE_FORMAT "KB card tables for parallel "
 920       "garbage collection for the requested " SIZE_FORMAT "KB heap.",
 921       _summary_data.reserved_byte_size()/K, mr.byte_size()/K));
 922     return false;
 923   }
 924 
 925   return true;
 926 }
 927 
 928 void PSParallelCompact::initialize_space_info()
 929 {
 930   memset(&_space_info, 0, sizeof(_space_info));
 931 
 932   ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
 933   PSYoungGen* young_gen = heap->young_gen();
 934 
 935   _space_info[old_space_id].set_space(heap->old_gen()->object_space());
 936   _space_info[eden_space_id].set_space(young_gen->eden_space());
 937   _space_info[from_space_id].set_space(young_gen->from_space());
 938   _space_info[to_space_id].set_space(young_gen->to_space());
 939 
 940   _space_info[old_space_id].set_start_array(heap->old_gen()->start_array());
 941 }
 942 
 943 void PSParallelCompact::initialize_dead_wood_limiter()
 944 {
 945   const size_t max = 100;
 946   _dwl_mean = double(MIN2(ParallelOldDeadWoodLimiterMean, max)) / 100.0;
 947   _dwl_std_dev = double(MIN2(ParallelOldDeadWoodLimiterStdDev, max)) / 100.0;
 948   _dwl_first_term = 1.0 / (sqrt(2.0 * M_PI) * _dwl_std_dev);
 949   DEBUG_ONLY(_dwl_initialized = true;)
 950   _dwl_adjustment = normal_distribution(1.0);
 951 }
 952 
 953 void
 954 PSParallelCompact::clear_data_covering_space(SpaceId id)
 955 {
 956   // At this point, top is the value before GC, new_top() is the value that will
 957   // be set at the end of GC.  The marking bitmap is cleared to top; nothing
 958   // should be marked above top.  The summary data is cleared to the larger of
 959   // top & new_top.
 960   MutableSpace* const space = _space_info[id].space();
 961   HeapWord* const bot = space->bottom();
 962   HeapWord* const top = space->top();
 963   HeapWord* const max_top = MAX2(top, _space_info[id].new_top());
 964 
 965   const idx_t beg_bit = _mark_bitmap.addr_to_bit(bot);
 966   const idx_t end_bit = _mark_bitmap.align_range_end(_mark_bitmap.addr_to_bit(top));
 967   _mark_bitmap.clear_range(beg_bit, end_bit);
 968 
 969   const size_t beg_region = _summary_data.addr_to_region_idx(bot);
 970   const size_t end_region =
 971     _summary_data.addr_to_region_idx(_summary_data.region_align_up(max_top));
 972   _summary_data.clear_range(beg_region, end_region);
 973 
 974   // Clear the data used to 'split' regions.
 975   SplitInfo& split_info = _space_info[id].split_info();
 976   if (split_info.is_valid()) {
 977     split_info.clear();
 978   }
 979   DEBUG_ONLY(split_info.verify_clear();)
 980 }
 981 
 982 void PSParallelCompact::pre_compact()
 983 {
 984   // Update the from & to space pointers in space_info, since they are swapped
 985   // at each young gen gc.  Do the update unconditionally (even though a
 986   // promotion failure does not swap spaces) because an unknown number of young
 987   // collections will have swapped the spaces an unknown number of times.
 988   GCTraceTime(Debug, gc, phases) tm("Pre Compact", &_gc_timer);
 989   ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
 990   _space_info[from_space_id].set_space(heap->young_gen()->from_space());
 991   _space_info[to_space_id].set_space(heap->young_gen()->to_space());
 992 
 993   DEBUG_ONLY(add_obj_count = add_obj_size = 0;)
 994   DEBUG_ONLY(mark_bitmap_count = mark_bitmap_size = 0;)
 995 
 996   // Increment the invocation count
 997   heap->increment_total_collections(true);
 998 
 999   // We need to track unique mark sweep invocations as well.
1000   _total_invocations++;
1001 
1002   heap->print_heap_before_gc();
1003   heap->trace_heap_before_gc(&_gc_tracer);
1004 
1005   // Fill in TLABs
1006   heap->ensure_parsability(true);  // retire TLABs
1007 
1008   if (VerifyBeforeGC && heap->total_collections() >= VerifyGCStartAt) {
1009     Universe::verify("Before GC");
1010   }
1011 
1012   // Verify object start arrays
1013   if (VerifyObjectStartArray &&
1014       VerifyBeforeGC) {
1015     heap->old_gen()->verify_object_start_array();
1016   }
1017 
1018   DEBUG_ONLY(mark_bitmap()->verify_clear();)
1019   DEBUG_ONLY(summary_data().verify_clear();)
1020 
1021   ParCompactionManager::reset_all_bitmap_query_caches();
1022 }
1023 
1024 void PSParallelCompact::post_compact()
1025 {
1026   GCTraceTime(Info, gc, phases) tm("Post Compact", &_gc_timer);
1027   ParCompactionManager::remove_all_shadow_regions();
1028 
1029   for (unsigned int id = old_space_id; id < last_space_id; ++id) {
1030     // Clear the marking bitmap, summary data and split info.
1031     clear_data_covering_space(SpaceId(id));
1032     // Update top().  Must be done after clearing the bitmap and summary data.
1033     _space_info[id].publish_new_top();
1034   }
1035 
1036   MutableSpace* const eden_space = _space_info[eden_space_id].space();
1037   MutableSpace* const from_space = _space_info[from_space_id].space();
1038   MutableSpace* const to_space   = _space_info[to_space_id].space();
1039 
1040   ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
1041   bool eden_empty = eden_space->is_empty();
1042 
1043   // Update heap occupancy information which is used as input to the soft ref
1044   // clearing policy at the next gc.
1045   Universe::update_heap_info_at_gc();
1046 
1047   Universe::heap()->next_whole_heap_examined();
1048 
1049   bool young_gen_empty = eden_empty && from_space->is_empty() &&
1050     to_space->is_empty();
1051 
1052   PSCardTable* ct = heap->card_table();
1053   MemRegion old_mr = heap->old_gen()->reserved();
1054   if (young_gen_empty) {
1055     ct->clear(MemRegion(old_mr.start(), old_mr.end()));
1056   } else {
1057     ct->invalidate(MemRegion(old_mr.start(), old_mr.end()));
1058   }
1059 
1060   // Delete metaspaces for unloaded class loaders and clean up loader_data graph
1061   ClassLoaderDataGraph::purge();
1062   MetaspaceUtils::verify_metrics();
1063 
1064   heap->prune_scavengable_nmethods();
1065 
1066 #if COMPILER2_OR_JVMCI
1067   DerivedPointerTable::update_pointers();
1068 #endif
1069 
1070   if (ZapUnusedHeapArea) {
1071     heap->gen_mangle_unused_area();
1072   }


1073 }
1074 
1075 HeapWord*
1076 PSParallelCompact::compute_dense_prefix_via_density(const SpaceId id,
1077                                                     bool maximum_compaction)
1078 {
1079   const size_t region_size = ParallelCompactData::RegionSize;
1080   const ParallelCompactData& sd = summary_data();
1081 
1082   const MutableSpace* const space = _space_info[id].space();
1083   HeapWord* const top_aligned_up = sd.region_align_up(space->top());
1084   const RegionData* const beg_cp = sd.addr_to_region_ptr(space->bottom());
1085   const RegionData* const end_cp = sd.addr_to_region_ptr(top_aligned_up);
1086 
1087   // Skip full regions at the beginning of the space--they are necessarily part
1088   // of the dense prefix.
1089   size_t full_count = 0;
1090   const RegionData* cp;
1091   for (cp = beg_cp; cp < end_cp && cp->data_size() == region_size; ++cp) {
1092     ++full_count;
1093   }
1094 
1095   assert(total_invocations() >= _maximum_compaction_gc_num, "sanity");
1096   const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num;
1097   const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval;
1098   if (maximum_compaction || cp == end_cp || interval_ended) {
1099     _maximum_compaction_gc_num = total_invocations();
1100     return sd.region_to_addr(cp);
1101   }
1102 
1103   HeapWord* const new_top = _space_info[id].new_top();
1104   const size_t space_live = pointer_delta(new_top, space->bottom());
1105   const size_t space_used = space->used_in_words();
1106   const size_t space_capacity = space->capacity_in_words();
1107 
1108   const double cur_density = double(space_live) / space_capacity;
1109   const double deadwood_density =
1110     (1.0 - cur_density) * (1.0 - cur_density) * cur_density * cur_density;
1111   const size_t deadwood_goal = size_t(space_capacity * deadwood_density);
1112 
1113   log_develop_debug(gc, compaction)(
1114       "cur_dens=%5.3f dw_dens=%5.3f dw_goal=" SIZE_FORMAT,
1115       cur_density, deadwood_density, deadwood_goal);
1116   log_develop_debug(gc, compaction)(
1117       "space_live=" SIZE_FORMAT " space_used=" SIZE_FORMAT " "
1118       "space_cap=" SIZE_FORMAT,
1119       space_live, space_used,
1120       space_capacity);
1121 
1122   // XXX - Use binary search?
1123   HeapWord* dense_prefix = sd.region_to_addr(cp);
1124   const RegionData* full_cp = cp;
1125   const RegionData* const top_cp = sd.addr_to_region_ptr(space->top() - 1);
1126   while (cp < end_cp) {
1127     HeapWord* region_destination = cp->destination();
1128     const size_t cur_deadwood = pointer_delta(dense_prefix, region_destination);
1129 
1130     log_develop_trace(gc, compaction)(
1131         "c#=" SIZE_FORMAT_W(4) " dst=" PTR_FORMAT " "
1132         "dp=" PTR_FORMAT " cdw=" SIZE_FORMAT_W(8),
1133         sd.region(cp), p2i(region_destination),
1134         p2i(dense_prefix), cur_deadwood);
1135 
1136     if (cur_deadwood >= deadwood_goal) {
1137       // Found the region that has the correct amount of deadwood to the left.
1138       // This typically occurs after crossing a fairly sparse set of regions, so
1139       // iterate backwards over those sparse regions, looking for the region
1140       // that has the lowest density of live objects 'to the right.'
1141       size_t space_to_left = sd.region(cp) * region_size;
1142       size_t live_to_left = space_to_left - cur_deadwood;
1143       size_t space_to_right = space_capacity - space_to_left;
1144       size_t live_to_right = space_live - live_to_left;
1145       double density_to_right = double(live_to_right) / space_to_right;
1146       while (cp > full_cp) {
1147         --cp;
1148         const size_t prev_region_live_to_right = live_to_right -
1149           cp->data_size();
1150         const size_t prev_region_space_to_right = space_to_right + region_size;
1151         double prev_region_density_to_right =
1152           double(prev_region_live_to_right) / prev_region_space_to_right;
1153         if (density_to_right <= prev_region_density_to_right) {
1154           return dense_prefix;
1155         }
1156 
1157         log_develop_trace(gc, compaction)(
1158             "backing up from c=" SIZE_FORMAT_W(4) " d2r=%10.8f "
1159             "pc_d2r=%10.8f",
1160             sd.region(cp), density_to_right,
1161             prev_region_density_to_right);
1162 
1163         dense_prefix -= region_size;
1164         live_to_right = prev_region_live_to_right;
1165         space_to_right = prev_region_space_to_right;
1166         density_to_right = prev_region_density_to_right;
1167       }
1168       return dense_prefix;
1169     }
1170 
1171     dense_prefix += region_size;
1172     ++cp;
1173   }
1174 
1175   return dense_prefix;
1176 }
1177 
1178 #ifndef PRODUCT
1179 void PSParallelCompact::print_dense_prefix_stats(const char* const algorithm,
1180                                                  const SpaceId id,
1181                                                  const bool maximum_compaction,
1182                                                  HeapWord* const addr)
1183 {
1184   const size_t region_idx = summary_data().addr_to_region_idx(addr);
1185   RegionData* const cp = summary_data().region(region_idx);
1186   const MutableSpace* const space = _space_info[id].space();
1187   HeapWord* const new_top = _space_info[id].new_top();
1188 
1189   const size_t space_live = pointer_delta(new_top, space->bottom());
1190   const size_t dead_to_left = pointer_delta(addr, cp->destination());
1191   const size_t space_cap = space->capacity_in_words();
1192   const double dead_to_left_pct = double(dead_to_left) / space_cap;
1193   const size_t live_to_right = new_top - cp->destination();
1194   const size_t dead_to_right = space->top() - addr - live_to_right;
1195 
1196   log_develop_debug(gc, compaction)(
1197       "%s=" PTR_FORMAT " dpc=" SIZE_FORMAT_W(5) " "
1198       "spl=" SIZE_FORMAT " "
1199       "d2l=" SIZE_FORMAT " d2l%%=%6.4f "
1200       "d2r=" SIZE_FORMAT " l2r=" SIZE_FORMAT " "
1201       "ratio=%10.8f",
1202       algorithm, p2i(addr), region_idx,
1203       space_live,
1204       dead_to_left, dead_to_left_pct,
1205       dead_to_right, live_to_right,
1206       double(dead_to_right) / live_to_right);
1207 }
1208 #endif  // #ifndef PRODUCT
1209 
1210 // Return a fraction indicating how much of the generation can be treated as
1211 // "dead wood" (i.e., not reclaimed).  The function uses a normal distribution
1212 // based on the density of live objects in the generation to determine a limit,
1213 // which is then adjusted so the return value is min_percent when the density is
1214 // 1.
1215 //
1216 // The following table shows some return values for a different values of the
1217 // standard deviation (ParallelOldDeadWoodLimiterStdDev); the mean is 0.5 and
1218 // min_percent is 1.
1219 //
1220 //                          fraction allowed as dead wood
1221 //         -----------------------------------------------------------------
1222 // density std_dev=70 std_dev=75 std_dev=80 std_dev=85 std_dev=90 std_dev=95
1223 // ------- ---------- ---------- ---------- ---------- ---------- ----------
1224 // 0.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000
1225 // 0.05000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941
1226 // 0.10000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272
1227 // 0.15000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066
1228 // 0.20000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975
1229 // 0.25000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313
1230 // 0.30000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132
1231 // 0.35000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289
1232 // 0.40000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500
1233 // 0.45000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386
1234 // 0.50000 0.13832410 0.11599237 0.09847664 0.08456518 0.07338887 0.06431510
1235 // 0.55000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386
1236 // 0.60000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500
1237 // 0.65000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289
1238 // 0.70000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132
1239 // 0.75000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313
1240 // 0.80000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975
1241 // 0.85000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066
1242 // 0.90000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272
1243 // 0.95000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941
1244 // 1.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000
1245 
1246 double PSParallelCompact::dead_wood_limiter(double density, size_t min_percent)
1247 {
1248   assert(_dwl_initialized, "uninitialized");
1249 
1250   // The raw limit is the value of the normal distribution at x = density.
1251   const double raw_limit = normal_distribution(density);
1252 
1253   // Adjust the raw limit so it becomes the minimum when the density is 1.
1254   //
1255   // First subtract the adjustment value (which is simply the precomputed value
1256   // normal_distribution(1.0)); this yields a value of 0 when the density is 1.
1257   // Then add the minimum value, so the minimum is returned when the density is
1258   // 1.  Finally, prevent negative values, which occur when the mean is not 0.5.
1259   const double min = double(min_percent) / 100.0;
1260   const double limit = raw_limit - _dwl_adjustment + min;
1261   return MAX2(limit, 0.0);
1262 }
1263 
1264 ParallelCompactData::RegionData*
1265 PSParallelCompact::first_dead_space_region(const RegionData* beg,
1266                                            const RegionData* end)
1267 {
1268   const size_t region_size = ParallelCompactData::RegionSize;
1269   ParallelCompactData& sd = summary_data();
1270   size_t left = sd.region(beg);
1271   size_t right = end > beg ? sd.region(end) - 1 : left;
1272 
1273   // Binary search.
1274   while (left < right) {
1275     // Equivalent to (left + right) / 2, but does not overflow.
1276     const size_t middle = left + (right - left) / 2;
1277     RegionData* const middle_ptr = sd.region(middle);
1278     HeapWord* const dest = middle_ptr->destination();
1279     HeapWord* const addr = sd.region_to_addr(middle);
1280     assert(dest != NULL, "sanity");
1281     assert(dest <= addr, "must move left");
1282 
1283     if (middle > left && dest < addr) {
1284       right = middle - 1;
1285     } else if (middle < right && middle_ptr->data_size() == region_size) {
1286       left = middle + 1;
1287     } else {
1288       return middle_ptr;
1289     }
1290   }
1291   return sd.region(left);
1292 }
1293 
1294 ParallelCompactData::RegionData*
1295 PSParallelCompact::dead_wood_limit_region(const RegionData* beg,
1296                                           const RegionData* end,
1297                                           size_t dead_words)
1298 {
1299   ParallelCompactData& sd = summary_data();
1300   size_t left = sd.region(beg);
1301   size_t right = end > beg ? sd.region(end) - 1 : left;
1302 
1303   // Binary search.
1304   while (left < right) {
1305     // Equivalent to (left + right) / 2, but does not overflow.
1306     const size_t middle = left + (right - left) / 2;
1307     RegionData* const middle_ptr = sd.region(middle);
1308     HeapWord* const dest = middle_ptr->destination();
1309     HeapWord* const addr = sd.region_to_addr(middle);
1310     assert(dest != NULL, "sanity");
1311     assert(dest <= addr, "must move left");
1312 
1313     const size_t dead_to_left = pointer_delta(addr, dest);
1314     if (middle > left && dead_to_left > dead_words) {
1315       right = middle - 1;
1316     } else if (middle < right && dead_to_left < dead_words) {
1317       left = middle + 1;
1318     } else {
1319       return middle_ptr;
1320     }
1321   }
1322   return sd.region(left);
1323 }
1324 
1325 // The result is valid during the summary phase, after the initial summarization
1326 // of each space into itself, and before final summarization.
1327 inline double
1328 PSParallelCompact::reclaimed_ratio(const RegionData* const cp,
1329                                    HeapWord* const bottom,
1330                                    HeapWord* const top,
1331                                    HeapWord* const new_top)
1332 {
1333   ParallelCompactData& sd = summary_data();
1334 
1335   assert(cp != NULL, "sanity");
1336   assert(bottom != NULL, "sanity");
1337   assert(top != NULL, "sanity");
1338   assert(new_top != NULL, "sanity");
1339   assert(top >= new_top, "summary data problem?");
1340   assert(new_top > bottom, "space is empty; should not be here");
1341   assert(new_top >= cp->destination(), "sanity");
1342   assert(top >= sd.region_to_addr(cp), "sanity");
1343 
1344   HeapWord* const destination = cp->destination();
1345   const size_t dense_prefix_live  = pointer_delta(destination, bottom);
1346   const size_t compacted_region_live = pointer_delta(new_top, destination);
1347   const size_t compacted_region_used = pointer_delta(top,
1348                                                      sd.region_to_addr(cp));
1349   const size_t reclaimable = compacted_region_used - compacted_region_live;
1350 
1351   const double divisor = dense_prefix_live + 1.25 * compacted_region_live;
1352   return double(reclaimable) / divisor;
1353 }
1354 
1355 // Return the address of the end of the dense prefix, a.k.a. the start of the
1356 // compacted region.  The address is always on a region boundary.
1357 //
1358 // Completely full regions at the left are skipped, since no compaction can
1359 // occur in those regions.  Then the maximum amount of dead wood to allow is
1360 // computed, based on the density (amount live / capacity) of the generation;
1361 // the region with approximately that amount of dead space to the left is
1362 // identified as the limit region.  Regions between the last completely full
1363 // region and the limit region are scanned and the one that has the best
1364 // (maximum) reclaimed_ratio() is selected.
1365 HeapWord*
1366 PSParallelCompact::compute_dense_prefix(const SpaceId id,
1367                                         bool maximum_compaction)
1368 {
1369   const size_t region_size = ParallelCompactData::RegionSize;
1370   const ParallelCompactData& sd = summary_data();
1371 
1372   const MutableSpace* const space = _space_info[id].space();
1373   HeapWord* const top = space->top();
1374   HeapWord* const top_aligned_up = sd.region_align_up(top);
1375   HeapWord* const new_top = _space_info[id].new_top();
1376   HeapWord* const new_top_aligned_up = sd.region_align_up(new_top);
1377   HeapWord* const bottom = space->bottom();
1378   const RegionData* const beg_cp = sd.addr_to_region_ptr(bottom);
1379   const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up);
1380   const RegionData* const new_top_cp =
1381     sd.addr_to_region_ptr(new_top_aligned_up);
1382 
1383   // Skip full regions at the beginning of the space--they are necessarily part
1384   // of the dense prefix.
1385   const RegionData* const full_cp = first_dead_space_region(beg_cp, new_top_cp);
1386   assert(full_cp->destination() == sd.region_to_addr(full_cp) ||
1387          space->is_empty(), "no dead space allowed to the left");
1388   assert(full_cp->data_size() < region_size || full_cp == new_top_cp - 1,
1389          "region must have dead space");
1390 
1391   // The gc number is saved whenever a maximum compaction is done, and used to
1392   // determine when the maximum compaction interval has expired.  This avoids
1393   // successive max compactions for different reasons.
1394   assert(total_invocations() >= _maximum_compaction_gc_num, "sanity");
1395   const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num;
1396   const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval ||
1397     total_invocations() == HeapFirstMaximumCompactionCount;
1398   if (maximum_compaction || full_cp == top_cp || interval_ended) {
1399     _maximum_compaction_gc_num = total_invocations();
1400     return sd.region_to_addr(full_cp);
1401   }
1402 
1403   const size_t space_live = pointer_delta(new_top, bottom);
1404   const size_t space_used = space->used_in_words();
1405   const size_t space_capacity = space->capacity_in_words();
1406 
1407   const double density = double(space_live) / double(space_capacity);
1408   const size_t min_percent_free = MarkSweepDeadRatio;
1409   const double limiter = dead_wood_limiter(density, min_percent_free);
1410   const size_t dead_wood_max = space_used - space_live;
1411   const size_t dead_wood_limit = MIN2(size_t(space_capacity * limiter),
1412                                       dead_wood_max);
1413 
1414   log_develop_debug(gc, compaction)(
1415       "space_live=" SIZE_FORMAT " space_used=" SIZE_FORMAT " "
1416       "space_cap=" SIZE_FORMAT,
1417       space_live, space_used,
1418       space_capacity);
1419   log_develop_debug(gc, compaction)(
1420       "dead_wood_limiter(%6.4f, " SIZE_FORMAT ")=%6.4f "
1421       "dead_wood_max=" SIZE_FORMAT " dead_wood_limit=" SIZE_FORMAT,
1422       density, min_percent_free, limiter,
1423       dead_wood_max, dead_wood_limit);
1424 
1425   // Locate the region with the desired amount of dead space to the left.
1426   const RegionData* const limit_cp =
1427     dead_wood_limit_region(full_cp, top_cp, dead_wood_limit);
1428 
1429   // Scan from the first region with dead space to the limit region and find the
1430   // one with the best (largest) reclaimed ratio.
1431   double best_ratio = 0.0;
1432   const RegionData* best_cp = full_cp;
1433   for (const RegionData* cp = full_cp; cp < limit_cp; ++cp) {
1434     double tmp_ratio = reclaimed_ratio(cp, bottom, top, new_top);
1435     if (tmp_ratio > best_ratio) {
1436       best_cp = cp;
1437       best_ratio = tmp_ratio;
1438     }
1439   }
1440 
1441   return sd.region_to_addr(best_cp);
1442 }
1443 
1444 void PSParallelCompact::summarize_spaces_quick()
1445 {
1446   for (unsigned int i = 0; i < last_space_id; ++i) {
1447     const MutableSpace* space = _space_info[i].space();
1448     HeapWord** nta = _space_info[i].new_top_addr();
1449     bool result = _summary_data.summarize(_space_info[i].split_info(),
1450                                           space->bottom(), space->top(), NULL,
1451                                           space->bottom(), space->end(), nta);
1452     assert(result, "space must fit into itself");
1453     _space_info[i].set_dense_prefix(space->bottom());
1454   }
1455 }
1456 
1457 void PSParallelCompact::fill_dense_prefix_end(SpaceId id)
1458 {
1459   HeapWord* const dense_prefix_end = dense_prefix(id);
1460   const RegionData* region = _summary_data.addr_to_region_ptr(dense_prefix_end);
1461   const idx_t dense_prefix_bit = _mark_bitmap.addr_to_bit(dense_prefix_end);
1462   if (dead_space_crosses_boundary(region, dense_prefix_bit)) {
1463     // Only enough dead space is filled so that any remaining dead space to the
1464     // left is larger than the minimum filler object.  (The remainder is filled
1465     // during the copy/update phase.)
1466     //
1467     // The size of the dead space to the right of the boundary is not a
1468     // concern, since compaction will be able to use whatever space is
1469     // available.
1470     //
1471     // Here '||' is the boundary, 'x' represents a don't care bit and a box
1472     // surrounds the space to be filled with an object.
1473     //
1474     // In the 32-bit VM, each bit represents two 32-bit words:
1475     //                              +---+
1476     // a) beg_bits:  ...  x   x   x | 0 | ||   0   x  x  ...
1477     //    end_bits:  ...  x   x   x | 0 | ||   0   x  x  ...
1478     //                              +---+
1479     //
1480     // In the 64-bit VM, each bit represents one 64-bit word:
1481     //                              +------------+
1482     // b) beg_bits:  ...  x   x   x | 0   ||   0 | x  x  ...
1483     //    end_bits:  ...  x   x   1 | 0   ||   0 | x  x  ...
1484     //                              +------------+
1485     //                          +-------+
1486     // c) beg_bits:  ...  x   x | 0   0 | ||   0   x  x  ...
1487     //    end_bits:  ...  x   1 | 0   0 | ||   0   x  x  ...
1488     //                          +-------+
1489     //                      +-----------+
1490     // d) beg_bits:  ...  x | 0   0   0 | ||   0   x  x  ...
1491     //    end_bits:  ...  1 | 0   0   0 | ||   0   x  x  ...
1492     //                      +-----------+
1493     //                          +-------+
1494     // e) beg_bits:  ...  0   0 | 0   0 | ||   0   x  x  ...
1495     //    end_bits:  ...  0   0 | 0   0 | ||   0   x  x  ...
1496     //                          +-------+
1497 
1498     // Initially assume case a, c or e will apply.
1499     size_t obj_len = CollectedHeap::min_fill_size();
1500     HeapWord* obj_beg = dense_prefix_end - obj_len;
1501 
1502 #ifdef  _LP64
1503     if (MinObjAlignment > 1) { // object alignment > heap word size
1504       // Cases a, c or e.
1505     } else if (_mark_bitmap.is_obj_end(dense_prefix_bit - 2)) {
1506       // Case b above.
1507       obj_beg = dense_prefix_end - 1;
1508     } else if (!_mark_bitmap.is_obj_end(dense_prefix_bit - 3) &&
1509                _mark_bitmap.is_obj_end(dense_prefix_bit - 4)) {
1510       // Case d above.
1511       obj_beg = dense_prefix_end - 3;
1512       obj_len = 3;
1513     }
1514 #endif  // #ifdef _LP64
1515 
1516     CollectedHeap::fill_with_object(obj_beg, obj_len);
1517     _mark_bitmap.mark_obj(obj_beg, obj_len);
1518     _summary_data.add_obj(obj_beg, obj_len);
1519     assert(start_array(id) != NULL, "sanity");
1520     start_array(id)->allocate_block(obj_beg);
1521   }
1522 }
1523 
1524 void
1525 PSParallelCompact::summarize_space(SpaceId id, bool maximum_compaction)
1526 {
1527   assert(id < last_space_id, "id out of range");
1528   assert(_space_info[id].dense_prefix() == _space_info[id].space()->bottom(),
1529          "should have been reset in summarize_spaces_quick()");
1530 
1531   const MutableSpace* space = _space_info[id].space();
1532   if (_space_info[id].new_top() != space->bottom()) {
1533     HeapWord* dense_prefix_end = compute_dense_prefix(id, maximum_compaction);
1534     _space_info[id].set_dense_prefix(dense_prefix_end);
1535 
1536 #ifndef PRODUCT
1537     if (log_is_enabled(Debug, gc, compaction)) {
1538       print_dense_prefix_stats("ratio", id, maximum_compaction,
1539                                dense_prefix_end);
1540       HeapWord* addr = compute_dense_prefix_via_density(id, maximum_compaction);
1541       print_dense_prefix_stats("density", id, maximum_compaction, addr);
1542     }
1543 #endif  // #ifndef PRODUCT
1544 
1545     // Recompute the summary data, taking into account the dense prefix.  If
1546     // every last byte will be reclaimed, then the existing summary data which
1547     // compacts everything can be left in place.
1548     if (!maximum_compaction && dense_prefix_end != space->bottom()) {
1549       // If dead space crosses the dense prefix boundary, it is (at least
1550       // partially) filled with a dummy object, marked live and added to the
1551       // summary data.  This simplifies the copy/update phase and must be done
1552       // before the final locations of objects are determined, to prevent
1553       // leaving a fragment of dead space that is too small to fill.
1554       fill_dense_prefix_end(id);
1555 
1556       // Compute the destination of each Region, and thus each object.
1557       _summary_data.summarize_dense_prefix(space->bottom(), dense_prefix_end);
1558       _summary_data.summarize(_space_info[id].split_info(),
1559                               dense_prefix_end, space->top(), NULL,
1560                               dense_prefix_end, space->end(),
1561                               _space_info[id].new_top_addr());
1562     }
1563   }
1564 
1565   if (log_develop_is_enabled(Trace, gc, compaction)) {
1566     const size_t region_size = ParallelCompactData::RegionSize;
1567     HeapWord* const dense_prefix_end = _space_info[id].dense_prefix();
1568     const size_t dp_region = _summary_data.addr_to_region_idx(dense_prefix_end);
1569     const size_t dp_words = pointer_delta(dense_prefix_end, space->bottom());
1570     HeapWord* const new_top = _space_info[id].new_top();
1571     const HeapWord* nt_aligned_up = _summary_data.region_align_up(new_top);
1572     const size_t cr_words = pointer_delta(nt_aligned_up, dense_prefix_end);
1573     log_develop_trace(gc, compaction)(
1574         "id=%d cap=" SIZE_FORMAT " dp=" PTR_FORMAT " "
1575         "dp_region=" SIZE_FORMAT " " "dp_count=" SIZE_FORMAT " "
1576         "cr_count=" SIZE_FORMAT " " "nt=" PTR_FORMAT,
1577         id, space->capacity_in_words(), p2i(dense_prefix_end),
1578         dp_region, dp_words / region_size,
1579         cr_words / region_size, p2i(new_top));
1580   }
1581 }
1582 
1583 #ifndef PRODUCT
1584 void PSParallelCompact::summary_phase_msg(SpaceId dst_space_id,
1585                                           HeapWord* dst_beg, HeapWord* dst_end,
1586                                           SpaceId src_space_id,
1587                                           HeapWord* src_beg, HeapWord* src_end)
1588 {
1589   log_develop_trace(gc, compaction)(
1590       "Summarizing %d [%s] into %d [%s]:  "
1591       "src=" PTR_FORMAT "-" PTR_FORMAT " "
1592       SIZE_FORMAT "-" SIZE_FORMAT " "
1593       "dst=" PTR_FORMAT "-" PTR_FORMAT " "
1594       SIZE_FORMAT "-" SIZE_FORMAT,
1595       src_space_id, space_names[src_space_id],
1596       dst_space_id, space_names[dst_space_id],
1597       p2i(src_beg), p2i(src_end),
1598       _summary_data.addr_to_region_idx(src_beg),
1599       _summary_data.addr_to_region_idx(src_end),
1600       p2i(dst_beg), p2i(dst_end),
1601       _summary_data.addr_to_region_idx(dst_beg),
1602       _summary_data.addr_to_region_idx(dst_end));
1603 }
1604 #endif  // #ifndef PRODUCT
1605 
1606 void PSParallelCompact::summary_phase(ParCompactionManager* cm,
1607                                       bool maximum_compaction)
1608 {
1609   GCTraceTime(Info, gc, phases) tm("Summary Phase", &_gc_timer);
1610 
1611   log_develop_debug(gc, marking)(
1612       "add_obj_count=" SIZE_FORMAT " "
1613       "add_obj_bytes=" SIZE_FORMAT,
1614       add_obj_count,
1615       add_obj_size * HeapWordSize);
1616   log_develop_debug(gc, marking)(
1617       "mark_bitmap_count=" SIZE_FORMAT " "
1618       "mark_bitmap_bytes=" SIZE_FORMAT,
1619       mark_bitmap_count,
1620       mark_bitmap_size * HeapWordSize);
1621 
1622   // Quick summarization of each space into itself, to see how much is live.
1623   summarize_spaces_quick();
1624 
1625   log_develop_trace(gc, compaction)("summary phase:  after summarizing each space to self");
1626   NOT_PRODUCT(print_region_ranges());
1627   NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info));
1628 
1629   // The amount of live data that will end up in old space (assuming it fits).
1630   size_t old_space_total_live = 0;
1631   for (unsigned int id = old_space_id; id < last_space_id; ++id) {
1632     old_space_total_live += pointer_delta(_space_info[id].new_top(),
1633                                           _space_info[id].space()->bottom());
1634   }
1635 
1636   MutableSpace* const old_space = _space_info[old_space_id].space();
1637   const size_t old_capacity = old_space->capacity_in_words();
1638   if (old_space_total_live > old_capacity) {
1639     // XXX - should also try to expand
1640     maximum_compaction = true;
1641   }
1642 
1643   // Old generations.
1644   summarize_space(old_space_id, maximum_compaction);
1645 
1646   // Summarize the remaining spaces in the young gen.  The initial target space
1647   // is the old gen.  If a space does not fit entirely into the target, then the
1648   // remainder is compacted into the space itself and that space becomes the new
1649   // target.
1650   SpaceId dst_space_id = old_space_id;
1651   HeapWord* dst_space_end = old_space->end();
1652   HeapWord** new_top_addr = _space_info[dst_space_id].new_top_addr();
1653   for (unsigned int id = eden_space_id; id < last_space_id; ++id) {
1654     const MutableSpace* space = _space_info[id].space();
1655     const size_t live = pointer_delta(_space_info[id].new_top(),
1656                                       space->bottom());
1657     const size_t available = pointer_delta(dst_space_end, *new_top_addr);
1658 
1659     NOT_PRODUCT(summary_phase_msg(dst_space_id, *new_top_addr, dst_space_end,
1660                                   SpaceId(id), space->bottom(), space->top());)
1661     if (live > 0 && live <= available) {
1662       // All the live data will fit.
1663       bool done = _summary_data.summarize(_space_info[id].split_info(),
1664                                           space->bottom(), space->top(),
1665                                           NULL,
1666                                           *new_top_addr, dst_space_end,
1667                                           new_top_addr);
1668       assert(done, "space must fit into old gen");
1669 
1670       // Reset the new_top value for the space.
1671       _space_info[id].set_new_top(space->bottom());
1672     } else if (live > 0) {
1673       // Attempt to fit part of the source space into the target space.
1674       HeapWord* next_src_addr = NULL;
1675       bool done = _summary_data.summarize(_space_info[id].split_info(),
1676                                           space->bottom(), space->top(),
1677                                           &next_src_addr,
1678                                           *new_top_addr, dst_space_end,
1679                                           new_top_addr);
1680       assert(!done, "space should not fit into old gen");
1681       assert(next_src_addr != NULL, "sanity");
1682 
1683       // The source space becomes the new target, so the remainder is compacted
1684       // within the space itself.
1685       dst_space_id = SpaceId(id);
1686       dst_space_end = space->end();
1687       new_top_addr = _space_info[id].new_top_addr();
1688       NOT_PRODUCT(summary_phase_msg(dst_space_id,
1689                                     space->bottom(), dst_space_end,
1690                                     SpaceId(id), next_src_addr, space->top());)
1691       done = _summary_data.summarize(_space_info[id].split_info(),
1692                                      next_src_addr, space->top(),
1693                                      NULL,
1694                                      space->bottom(), dst_space_end,
1695                                      new_top_addr);
1696       assert(done, "space must fit when compacted into itself");
1697       assert(*new_top_addr <= space->top(), "usage should not grow");
1698     }
1699   }
1700 
1701   log_develop_trace(gc, compaction)("Summary_phase:  after final summarization");
1702   NOT_PRODUCT(print_region_ranges());
1703   NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info));
1704 }
1705 
1706 // This method should contain all heap-specific policy for invoking a full
1707 // collection.  invoke_no_policy() will only attempt to compact the heap; it
1708 // will do nothing further.  If we need to bail out for policy reasons, scavenge
1709 // before full gc, or any other specialized behavior, it needs to be added here.
1710 //
1711 // Note that this method should only be called from the vm_thread while at a
1712 // safepoint.
1713 //
1714 // Note that the all_soft_refs_clear flag in the soft ref policy
1715 // may be true because this method can be called without intervening
1716 // activity.  For example when the heap space is tight and full measure
1717 // are being taken to free space.
1718 void PSParallelCompact::invoke(bool maximum_heap_compaction) {
1719   assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
1720   assert(Thread::current() == (Thread*)VMThread::vm_thread(),
1721          "should be in vm thread");
1722 
1723   ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
1724   GCCause::Cause gc_cause = heap->gc_cause();
1725   assert(!heap->is_gc_active(), "not reentrant");
1726 
1727   PSAdaptiveSizePolicy* policy = heap->size_policy();
1728   IsGCActiveMark mark;
1729 
1730   if (ScavengeBeforeFullGC) {
1731     PSScavenge::invoke_no_policy();
1732   }
1733 
1734   const bool clear_all_soft_refs =
1735     heap->soft_ref_policy()->should_clear_all_soft_refs();
1736 
1737   PSParallelCompact::invoke_no_policy(clear_all_soft_refs ||
1738                                       maximum_heap_compaction);
1739 }
1740 
1741 // This method contains no policy. You should probably
1742 // be calling invoke() instead.
1743 bool PSParallelCompact::invoke_no_policy(bool maximum_heap_compaction) {
1744   assert(SafepointSynchronize::is_at_safepoint(), "must be at a safepoint");
1745   assert(ref_processor() != NULL, "Sanity");
1746 
1747   if (GCLocker::check_active_before_gc()) {
1748     return false;
1749   }
1750 
1751   ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
1752 
1753   GCIdMark gc_id_mark;
1754   _gc_timer.register_gc_start();
1755   _gc_tracer.report_gc_start(heap->gc_cause(), _gc_timer.gc_start());
1756 
1757   TimeStamp marking_start;
1758   TimeStamp compaction_start;
1759   TimeStamp collection_exit;
1760 
1761   GCCause::Cause gc_cause = heap->gc_cause();
1762   PSYoungGen* young_gen = heap->young_gen();
1763   PSOldGen* old_gen = heap->old_gen();
1764   PSAdaptiveSizePolicy* size_policy = heap->size_policy();
1765 
1766   // The scope of casr should end after code that can change
1767   // SoftRefPolicy::_should_clear_all_soft_refs.
1768   ClearedAllSoftRefs casr(maximum_heap_compaction,
1769                           heap->soft_ref_policy());
1770 
1771   if (ZapUnusedHeapArea) {
1772     // Save information needed to minimize mangling
1773     heap->record_gen_tops_before_GC();
1774   }
1775 
1776   // Make sure data structures are sane, make the heap parsable, and do other
1777   // miscellaneous bookkeeping.
1778   pre_compact();
1779 
1780   const PreGenGCValues pre_gc_values = heap->get_pre_gc_values();
1781 
1782   // Get the compaction manager reserved for the VM thread.
1783   ParCompactionManager* const vmthread_cm =
1784     ParCompactionManager::manager_array(ParallelScavengeHeap::heap()->workers().total_workers());
1785 
1786   {
1787     ResourceMark rm;
1788 
1789     const uint active_workers =
1790       WorkerPolicy::calc_active_workers(ParallelScavengeHeap::heap()->workers().total_workers(),
1791                                         ParallelScavengeHeap::heap()->workers().active_workers(),
1792                                         Threads::number_of_non_daemon_threads());
1793     ParallelScavengeHeap::heap()->workers().update_active_workers(active_workers);
1794 
1795     GCTraceCPUTime tcpu;
1796     GCTraceTime(Info, gc) tm("Pause Full", NULL, gc_cause, true);
1797 
1798     heap->pre_full_gc_dump(&_gc_timer);
1799 
1800     TraceCollectorStats tcs(counters());
1801     TraceMemoryManagerStats tms(heap->old_gc_manager(), gc_cause);
1802 
1803     if (log_is_enabled(Debug, gc, heap, exit)) {
1804       accumulated_time()->start();
1805     }
1806 
1807     // Let the size policy know we're starting
1808     size_policy->major_collection_begin();
1809 
1810 #if COMPILER2_OR_JVMCI
1811     DerivedPointerTable::clear();
1812 #endif
1813 
1814     ref_processor()->enable_discovery();
1815     ref_processor()->setup_policy(maximum_heap_compaction);
1816 
1817     bool marked_for_unloading = false;
1818 
1819     marking_start.update();
1820     marking_phase(vmthread_cm, maximum_heap_compaction, &_gc_tracer);
1821 
1822     bool max_on_system_gc = UseMaximumCompactionOnSystemGC
1823       && GCCause::is_user_requested_gc(gc_cause);
1824     summary_phase(vmthread_cm, maximum_heap_compaction || max_on_system_gc);
1825 
1826 #if COMPILER2_OR_JVMCI
1827     assert(DerivedPointerTable::is_active(), "Sanity");
1828     DerivedPointerTable::set_active(false);
1829 #endif
1830 
1831     // adjust_roots() updates Universe::_intArrayKlassObj which is
1832     // needed by the compaction for filling holes in the dense prefix.
1833     adjust_roots(vmthread_cm);
1834 
1835     compaction_start.update();
1836     compact();
1837 
1838     // Reset the mark bitmap, summary data, and do other bookkeeping.  Must be
1839     // done before resizing.
1840     post_compact();
1841 
1842     // Let the size policy know we're done
1843     size_policy->major_collection_end(old_gen->used_in_bytes(), gc_cause);
1844 
1845     if (UseAdaptiveSizePolicy) {
1846       log_debug(gc, ergo)("AdaptiveSizeStart: collection: %d ", heap->total_collections());
1847       log_trace(gc, ergo)("old_gen_capacity: " SIZE_FORMAT " young_gen_capacity: " SIZE_FORMAT,
1848                           old_gen->capacity_in_bytes(), young_gen->capacity_in_bytes());
1849 
1850       // Don't check if the size_policy is ready here.  Let
1851       // the size_policy check that internally.
1852       if (UseAdaptiveGenerationSizePolicyAtMajorCollection &&
1853           AdaptiveSizePolicy::should_update_promo_stats(gc_cause)) {
1854         // Swap the survivor spaces if from_space is empty. The
1855         // resize_young_gen() called below is normally used after
1856         // a successful young GC and swapping of survivor spaces;
1857         // otherwise, it will fail to resize the young gen with
1858         // the current implementation.
1859         if (young_gen->from_space()->is_empty()) {
1860           young_gen->from_space()->clear(SpaceDecorator::Mangle);
1861           young_gen->swap_spaces();
1862         }
1863 
1864         // Calculate optimal free space amounts
1865         assert(young_gen->max_gen_size() >
1866           young_gen->from_space()->capacity_in_bytes() +
1867           young_gen->to_space()->capacity_in_bytes(),
1868           "Sizes of space in young gen are out-of-bounds");
1869 
1870         size_t young_live = young_gen->used_in_bytes();
1871         size_t eden_live = young_gen->eden_space()->used_in_bytes();
1872         size_t old_live = old_gen->used_in_bytes();
1873         size_t cur_eden = young_gen->eden_space()->capacity_in_bytes();
1874         size_t max_old_gen_size = old_gen->max_gen_size();
1875         size_t max_eden_size = young_gen->max_gen_size() -
1876           young_gen->from_space()->capacity_in_bytes() -
1877           young_gen->to_space()->capacity_in_bytes();
1878 
1879         // Used for diagnostics
1880         size_policy->clear_generation_free_space_flags();
1881 
1882         size_policy->compute_generations_free_space(young_live,
1883                                                     eden_live,
1884                                                     old_live,
1885                                                     cur_eden,
1886                                                     max_old_gen_size,
1887                                                     max_eden_size,
1888                                                     true /* full gc*/);
1889 
1890         size_policy->check_gc_overhead_limit(eden_live,
1891                                              max_old_gen_size,
1892                                              max_eden_size,
1893                                              true /* full gc*/,
1894                                              gc_cause,
1895                                              heap->soft_ref_policy());
1896 
1897         size_policy->decay_supplemental_growth(true /* full gc*/);
1898 
1899         heap->resize_old_gen(
1900           size_policy->calculated_old_free_size_in_bytes());
1901 
1902         heap->resize_young_gen(size_policy->calculated_eden_size_in_bytes(),
1903                                size_policy->calculated_survivor_size_in_bytes());
1904       }
1905 
1906       log_debug(gc, ergo)("AdaptiveSizeStop: collection: %d ", heap->total_collections());
1907     }
1908 
1909     if (UsePerfData) {
1910       PSGCAdaptivePolicyCounters* const counters = heap->gc_policy_counters();
1911       counters->update_counters();
1912       counters->update_old_capacity(old_gen->capacity_in_bytes());
1913       counters->update_young_capacity(young_gen->capacity_in_bytes());
1914     }
1915 
1916     heap->resize_all_tlabs();
1917 
1918     // Resize the metaspace capacity after a collection
1919     MetaspaceGC::compute_new_size();
1920 
1921     if (log_is_enabled(Debug, gc, heap, exit)) {
1922       accumulated_time()->stop();
1923     }
1924 
1925     heap->print_heap_change(pre_gc_values);
1926 
1927     // Track memory usage and detect low memory
1928     MemoryService::track_memory_usage();
1929     heap->update_counters();
1930 
1931     heap->post_full_gc_dump(&_gc_timer);
1932   }
1933 
1934 #ifdef ASSERT
1935   for (size_t i = 0; i < ParallelGCThreads + 1; ++i) {
1936     ParCompactionManager* const cm =
1937       ParCompactionManager::manager_array(int(i));
1938     assert(cm->marking_stack()->is_empty(),       "should be empty");
1939     assert(cm->region_stack()->is_empty(), "Region stack " SIZE_FORMAT " is not empty", i);
1940   }
1941 #endif // ASSERT
1942 
1943   if (VerifyAfterGC && heap->total_collections() >= VerifyGCStartAt) {
1944     Universe::verify("After GC");
1945   }
1946 
1947   // Re-verify object start arrays
1948   if (VerifyObjectStartArray &&
1949       VerifyAfterGC) {
1950     old_gen->verify_object_start_array();
1951   }
1952 
1953   if (ZapUnusedHeapArea) {
1954     old_gen->object_space()->check_mangled_unused_area_complete();
1955   }
1956 
1957   NOT_PRODUCT(ref_processor()->verify_no_references_recorded());
1958 
1959   collection_exit.update();
1960 
1961   heap->print_heap_after_gc();
1962   heap->trace_heap_after_gc(&_gc_tracer);
1963 
1964   log_debug(gc, task, time)("VM-Thread " JLONG_FORMAT " " JLONG_FORMAT " " JLONG_FORMAT,
1965                          marking_start.ticks(), compaction_start.ticks(),
1966                          collection_exit.ticks());
1967 
1968   AdaptiveSizePolicyOutput::print(size_policy, heap->total_collections());
1969 
1970   _gc_timer.register_gc_end();
1971 
1972   _gc_tracer.report_dense_prefix(dense_prefix(old_space_id));
1973   _gc_tracer.report_gc_end(_gc_timer.gc_end(), _gc_timer.time_partitions());
1974 
1975   return true;
1976 }
1977 
1978 class PCAddThreadRootsMarkingTaskClosure : public ThreadClosure {
1979 private:
1980   uint _worker_id;
1981 
1982 public:
1983   PCAddThreadRootsMarkingTaskClosure(uint worker_id) : _worker_id(worker_id) { }
1984   void do_thread(Thread* thread) {
1985     assert(ParallelScavengeHeap::heap()->is_gc_active(), "called outside gc");
1986 
1987     ResourceMark rm;
1988 
1989     ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(_worker_id);
1990 
1991     PCMarkAndPushClosure mark_and_push_closure(cm);
1992     MarkingCodeBlobClosure mark_and_push_in_blobs(&mark_and_push_closure, !CodeBlobToOopClosure::FixRelocations);
1993 
1994     thread->oops_do(&mark_and_push_closure, &mark_and_push_in_blobs);
1995 
1996     // Do the real work
1997     cm->follow_marking_stacks();
1998   }
1999 };
2000 
2001 static void mark_from_roots_work(ParallelRootType::Value root_type, uint worker_id) {
2002   assert(ParallelScavengeHeap::heap()->is_gc_active(), "called outside gc");
2003 
2004   ParCompactionManager* cm =
2005     ParCompactionManager::gc_thread_compaction_manager(worker_id);
2006   PCMarkAndPushClosure mark_and_push_closure(cm);
2007 
2008   switch (root_type) {
2009     case ParallelRootType::universe:
2010       Universe::oops_do(&mark_and_push_closure);
2011       break;
2012 
2013     case ParallelRootType::object_synchronizer:
2014       ObjectSynchronizer::oops_do(&mark_and_push_closure);
2015       break;
2016 
2017     case ParallelRootType::class_loader_data:
2018       {
2019         CLDToOopClosure cld_closure(&mark_and_push_closure, ClassLoaderData::_claim_strong);
2020         ClassLoaderDataGraph::always_strong_cld_do(&cld_closure);
2021       }
2022       break;
2023 
2024     case ParallelRootType::code_cache:
2025       // Do not treat nmethods as strong roots for mark/sweep, since we can unload them.
2026       //ScavengableNMethods::scavengable_nmethods_do(CodeBlobToOopClosure(&mark_and_push_closure));
2027       AOTLoader::oops_do(&mark_and_push_closure);
2028       break;
2029 
2030     case ParallelRootType::sentinel:
2031     DEBUG_ONLY(default:) // DEBUG_ONLY hack will create compile error on release builds (-Wswitch) and runtime check on debug builds
2032       fatal("Bad enumeration value: %u", root_type);
2033       break;
2034   }
2035 
2036   // Do the real work
2037   cm->follow_marking_stacks();
2038 }
2039 
2040 static void steal_marking_work(TaskTerminator& terminator, uint worker_id) {
2041   assert(ParallelScavengeHeap::heap()->is_gc_active(), "called outside gc");
2042 
2043   ParCompactionManager* cm =
2044     ParCompactionManager::gc_thread_compaction_manager(worker_id);
2045 
2046   oop obj = NULL;
2047   ObjArrayTask task;
2048   do {
2049     while (ParCompactionManager::steal_objarray(worker_id,  task)) {
2050       cm->follow_array((objArrayOop)task.obj(), task.index());
2051       cm->follow_marking_stacks();
2052     }
2053     while (ParCompactionManager::steal(worker_id, obj)) {
2054       cm->follow_contents(obj);
2055       cm->follow_marking_stacks();
2056     }
2057   } while (!terminator.offer_termination());
2058 }
2059 
2060 class MarkFromRootsTask : public AbstractGangTask {
2061   typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
2062   StrongRootsScope _strong_roots_scope; // needed for Threads::possibly_parallel_threads_do
2063   OopStorageSetStrongParState<false /* concurrent */, false /* is_const */> _oop_storage_set_par_state;
2064   SequentialSubTasksDone _subtasks;
2065   TaskTerminator _terminator;
2066   uint _active_workers;
2067 
2068 public:
2069   MarkFromRootsTask(uint active_workers) :
2070       AbstractGangTask("MarkFromRootsTask"),
2071       _strong_roots_scope(active_workers),
2072       _subtasks(),
2073       _terminator(active_workers, ParCompactionManager::oop_task_queues()),
2074       _active_workers(active_workers) {
2075     _subtasks.set_n_threads(active_workers);
2076     _subtasks.set_n_tasks(ParallelRootType::sentinel);
2077   }
2078 
2079   virtual void work(uint worker_id) {
2080     for (uint task = 0; _subtasks.try_claim_task(task); /*empty*/ ) {
2081       mark_from_roots_work(static_cast<ParallelRootType::Value>(task), worker_id);
2082     }
2083     _subtasks.all_tasks_completed();
2084 
2085     PCAddThreadRootsMarkingTaskClosure closure(worker_id);
2086     Threads::possibly_parallel_threads_do(true /*parallel */, &closure);
2087 
2088     // Mark from OopStorages
2089     {
2090       ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(worker_id);
2091       PCMarkAndPushClosure closure(cm);
2092       _oop_storage_set_par_state.oops_do(&closure);
2093       // Do the real work
2094       cm->follow_marking_stacks();
2095     }
2096 
2097     if (_active_workers > 1) {
2098       steal_marking_work(_terminator, worker_id);
2099     }
2100   }
2101 };
2102 
2103 class PCRefProcTask : public AbstractGangTask {
2104   typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
2105   ProcessTask& _task;
2106   uint _ergo_workers;
2107   TaskTerminator _terminator;
2108 
2109 public:
2110   PCRefProcTask(ProcessTask& task, uint ergo_workers) :
2111       AbstractGangTask("PCRefProcTask"),
2112       _task(task),
2113       _ergo_workers(ergo_workers),
2114       _terminator(_ergo_workers, ParCompactionManager::oop_task_queues()) {
2115   }
2116 
2117   virtual void work(uint worker_id) {
2118     ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
2119     assert(ParallelScavengeHeap::heap()->is_gc_active(), "called outside gc");
2120 
2121     ParCompactionManager* cm =
2122       ParCompactionManager::gc_thread_compaction_manager(worker_id);
2123     PCMarkAndPushClosure mark_and_push_closure(cm);
2124     ParCompactionManager::FollowStackClosure follow_stack_closure(cm);
2125     _task.work(worker_id, *PSParallelCompact::is_alive_closure(),
2126                mark_and_push_closure, follow_stack_closure);
2127 
2128     steal_marking_work(_terminator, worker_id);
2129   }
2130 };
2131 
2132 class RefProcTaskExecutor: public AbstractRefProcTaskExecutor {
2133   void execute(ProcessTask& process_task, uint ergo_workers) {
2134     assert(ParallelScavengeHeap::heap()->workers().active_workers() == ergo_workers,
2135            "Ergonomically chosen workers (%u) must be equal to active workers (%u)",
2136            ergo_workers, ParallelScavengeHeap::heap()->workers().active_workers());
2137 
2138     PCRefProcTask task(process_task, ergo_workers);
2139     ParallelScavengeHeap::heap()->workers().run_task(&task);
2140   }
2141 };
2142 
2143 void PSParallelCompact::marking_phase(ParCompactionManager* cm,
2144                                       bool maximum_heap_compaction,
2145                                       ParallelOldTracer *gc_tracer) {
2146   // Recursively traverse all live objects and mark them
2147   GCTraceTime(Info, gc, phases) tm("Marking Phase", &_gc_timer);
2148 
2149   ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
2150   uint active_gc_threads = ParallelScavengeHeap::heap()->workers().active_workers();
2151 
2152   PCMarkAndPushClosure mark_and_push_closure(cm);
2153   ParCompactionManager::FollowStackClosure follow_stack_closure(cm);
2154 
2155   // Need new claim bits before marking starts.
2156   ClassLoaderDataGraph::clear_claimed_marks();
2157 
2158   {
2159     GCTraceTime(Debug, gc, phases) tm("Par Mark", &_gc_timer);
2160 
2161     MarkFromRootsTask task(active_gc_threads);
2162     ParallelScavengeHeap::heap()->workers().run_task(&task);
2163   }
2164 
2165   // Process reference objects found during marking
2166   {
2167     GCTraceTime(Debug, gc, phases) tm("Reference Processing", &_gc_timer);
2168 
2169     ReferenceProcessorStats stats;
2170     ReferenceProcessorPhaseTimes pt(&_gc_timer, ref_processor()->max_num_queues());
2171 
2172     if (ref_processor()->processing_is_mt()) {
2173       ref_processor()->set_active_mt_degree(active_gc_threads);
2174 
2175       RefProcTaskExecutor task_executor;
2176       stats = ref_processor()->process_discovered_references(
2177         is_alive_closure(), &mark_and_push_closure, &follow_stack_closure,
2178         &task_executor, &pt);
2179     } else {
2180       stats = ref_processor()->process_discovered_references(
2181         is_alive_closure(), &mark_and_push_closure, &follow_stack_closure, NULL,
2182         &pt);
2183     }
2184 
2185     gc_tracer->report_gc_reference_stats(stats);
2186     pt.print_all_references();
2187   }
2188 
2189   // This is the point where the entire marking should have completed.
2190   assert(cm->marking_stacks_empty(), "Marking should have completed");
2191 
2192   {
2193     GCTraceTime(Debug, gc, phases) tm("Weak Processing", &_gc_timer);
2194     WeakProcessor::weak_oops_do(is_alive_closure(), &do_nothing_cl);
2195   }
2196 
2197   {
2198     GCTraceTime(Debug, gc, phases) tm_m("Class Unloading", &_gc_timer);
2199 
2200     // Follow system dictionary roots and unload classes.
2201     bool purged_class = SystemDictionary::do_unloading(&_gc_timer);
2202 
2203     // Unload nmethods.
2204     CodeCache::do_unloading(is_alive_closure(), purged_class);
2205 
2206     // Prune dead klasses from subklass/sibling/implementor lists.
2207     Klass::clean_weak_klass_links(purged_class);
2208 
2209     // Clean JVMCI metadata handles.
2210     JVMCI_ONLY(JVMCI::do_unloading(purged_class));
2211   }
2212 
2213   _gc_tracer.report_object_count_after_gc(is_alive_closure());
2214 }
2215 
2216 void PSParallelCompact::adjust_roots(ParCompactionManager* cm) {
2217   // Adjust the pointers to reflect the new locations
2218   GCTraceTime(Info, gc, phases) tm("Adjust Roots", &_gc_timer);
2219 
2220   // Need new claim bits when tracing through and adjusting pointers.
2221   ClassLoaderDataGraph::clear_claimed_marks();
2222 
2223   PCAdjustPointerClosure oop_closure(cm);
2224 
2225   // General strong roots.
2226   Universe::oops_do(&oop_closure);
2227   Threads::oops_do(&oop_closure, NULL);
2228   ObjectSynchronizer::oops_do(&oop_closure);
2229   OopStorageSet::strong_oops_do(&oop_closure);
2230   CLDToOopClosure cld_closure(&oop_closure, ClassLoaderData::_claim_strong);
2231   ClassLoaderDataGraph::cld_do(&cld_closure);
2232 
2233   // Now adjust pointers in remaining weak roots.  (All of which should
2234   // have been cleared if they pointed to non-surviving objects.)
2235   WeakProcessor::oops_do(&oop_closure);
2236 
2237   CodeBlobToOopClosure adjust_from_blobs(&oop_closure, CodeBlobToOopClosure::FixRelocations);
2238   CodeCache::blobs_do(&adjust_from_blobs);
2239   AOT_ONLY(AOTLoader::oops_do(&oop_closure);)
2240 
2241   ref_processor()->weak_oops_do(&oop_closure);
2242   // Roots were visited so references into the young gen in roots
2243   // may have been scanned.  Process them also.
2244   // Should the reference processor have a span that excludes
2245   // young gen objects?
2246   PSScavenge::reference_processor()->weak_oops_do(&oop_closure);
2247 }
2248 
2249 // Helper class to print 8 region numbers per line and then print the total at the end.
2250 class FillableRegionLogger : public StackObj {
2251 private:
2252   Log(gc, compaction) log;
2253   static const int LineLength = 8;
2254   size_t _regions[LineLength];
2255   int _next_index;
2256   bool _enabled;
2257   size_t _total_regions;
2258 public:
2259   FillableRegionLogger() : _next_index(0), _enabled(log_develop_is_enabled(Trace, gc, compaction)), _total_regions(0) { }
2260   ~FillableRegionLogger() {
2261     log.trace(SIZE_FORMAT " initially fillable regions", _total_regions);
2262   }
2263 
2264   void print_line() {
2265     if (!_enabled || _next_index == 0) {
2266       return;
2267     }
2268     FormatBuffer<> line("Fillable: ");
2269     for (int i = 0; i < _next_index; i++) {
2270       line.append(" " SIZE_FORMAT_W(7), _regions[i]);
2271     }
2272     log.trace("%s", line.buffer());
2273     _next_index = 0;
2274   }
2275 
2276   void handle(size_t region) {
2277     if (!_enabled) {
2278       return;
2279     }
2280     _regions[_next_index++] = region;
2281     if (_next_index == LineLength) {
2282       print_line();
2283     }
2284     _total_regions++;
2285   }
2286 };
2287 
2288 void PSParallelCompact::prepare_region_draining_tasks(uint parallel_gc_threads)
2289 {
2290   GCTraceTime(Trace, gc, phases) tm("Drain Task Setup", &_gc_timer);
2291 
2292   // Find the threads that are active
2293   uint worker_id = 0;
2294 
2295   // Find all regions that are available (can be filled immediately) and
2296   // distribute them to the thread stacks.  The iteration is done in reverse
2297   // order (high to low) so the regions will be removed in ascending order.
2298 
2299   const ParallelCompactData& sd = PSParallelCompact::summary_data();
2300 
2301   // id + 1 is used to test termination so unsigned  can
2302   // be used with an old_space_id == 0.
2303   FillableRegionLogger region_logger;
2304   for (unsigned int id = to_space_id; id + 1 > old_space_id; --id) {
2305     SpaceInfo* const space_info = _space_info + id;
2306     MutableSpace* const space = space_info->space();
2307     HeapWord* const new_top = space_info->new_top();
2308 
2309     const size_t beg_region = sd.addr_to_region_idx(space_info->dense_prefix());
2310     const size_t end_region =
2311       sd.addr_to_region_idx(sd.region_align_up(new_top));
2312 
2313     for (size_t cur = end_region - 1; cur + 1 > beg_region; --cur) {
2314       if (sd.region(cur)->claim_unsafe()) {
2315         ParCompactionManager* cm = ParCompactionManager::manager_array(worker_id);
2316         bool result = sd.region(cur)->mark_normal();
2317         assert(result, "Must succeed at this point.");
2318         cm->region_stack()->push(cur);
2319         region_logger.handle(cur);
2320         // Assign regions to tasks in round-robin fashion.
2321         if (++worker_id == parallel_gc_threads) {
2322           worker_id = 0;
2323         }
2324       }
2325     }
2326     region_logger.print_line();
2327   }
2328 }
2329 
2330 class TaskQueue : StackObj {
2331   volatile uint _counter;
2332   uint _size;
2333   uint _insert_index;
2334   PSParallelCompact::UpdateDensePrefixTask* _backing_array;
2335 public:
2336   explicit TaskQueue(uint size) : _counter(0), _size(size), _insert_index(0), _backing_array(NULL) {
2337     _backing_array = NEW_C_HEAP_ARRAY(PSParallelCompact::UpdateDensePrefixTask, _size, mtGC);
2338   }
2339   ~TaskQueue() {
2340     assert(_counter >= _insert_index, "not all queue elements were claimed");
2341     FREE_C_HEAP_ARRAY(T, _backing_array);
2342   }
2343 
2344   void push(const PSParallelCompact::UpdateDensePrefixTask& value) {
2345     assert(_insert_index < _size, "too small backing array");
2346     _backing_array[_insert_index++] = value;
2347   }
2348 
2349   bool try_claim(PSParallelCompact::UpdateDensePrefixTask& reference) {
2350     uint claimed = Atomic::fetch_and_add(&_counter, 1u);
2351     if (claimed < _insert_index) {
2352       reference = _backing_array[claimed];
2353       return true;
2354     } else {
2355       return false;
2356     }
2357   }
2358 };
2359 
2360 #define PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING 4
2361 
2362 void PSParallelCompact::enqueue_dense_prefix_tasks(TaskQueue& task_queue,
2363                                                    uint parallel_gc_threads) {
2364   GCTraceTime(Trace, gc, phases) tm("Dense Prefix Task Setup", &_gc_timer);
2365 
2366   ParallelCompactData& sd = PSParallelCompact::summary_data();
2367 
2368   // Iterate over all the spaces adding tasks for updating
2369   // regions in the dense prefix.  Assume that 1 gc thread
2370   // will work on opening the gaps and the remaining gc threads
2371   // will work on the dense prefix.
2372   unsigned int space_id;
2373   for (space_id = old_space_id; space_id < last_space_id; ++ space_id) {
2374     HeapWord* const dense_prefix_end = _space_info[space_id].dense_prefix();
2375     const MutableSpace* const space = _space_info[space_id].space();
2376 
2377     if (dense_prefix_end == space->bottom()) {
2378       // There is no dense prefix for this space.
2379       continue;
2380     }
2381 
2382     // The dense prefix is before this region.
2383     size_t region_index_end_dense_prefix =
2384         sd.addr_to_region_idx(dense_prefix_end);
2385     RegionData* const dense_prefix_cp =
2386       sd.region(region_index_end_dense_prefix);
2387     assert(dense_prefix_end == space->end() ||
2388            dense_prefix_cp->available() ||
2389            dense_prefix_cp->claimed(),
2390            "The region after the dense prefix should always be ready to fill");
2391 
2392     size_t region_index_start = sd.addr_to_region_idx(space->bottom());
2393 
2394     // Is there dense prefix work?
2395     size_t total_dense_prefix_regions =
2396       region_index_end_dense_prefix - region_index_start;
2397     // How many regions of the dense prefix should be given to
2398     // each thread?
2399     if (total_dense_prefix_regions > 0) {
2400       uint tasks_for_dense_prefix = 1;
2401       if (total_dense_prefix_regions <=
2402           (parallel_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING)) {
2403         // Don't over partition.  This assumes that
2404         // PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING is a small integer value
2405         // so there are not many regions to process.
2406         tasks_for_dense_prefix = parallel_gc_threads;
2407       } else {
2408         // Over partition
2409         tasks_for_dense_prefix = parallel_gc_threads *
2410           PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING;
2411       }
2412       size_t regions_per_thread = total_dense_prefix_regions /
2413         tasks_for_dense_prefix;
2414       // Give each thread at least 1 region.
2415       if (regions_per_thread == 0) {
2416         regions_per_thread = 1;
2417       }
2418 
2419       for (uint k = 0; k < tasks_for_dense_prefix; k++) {
2420         if (region_index_start >= region_index_end_dense_prefix) {
2421           break;
2422         }
2423         // region_index_end is not processed
2424         size_t region_index_end = MIN2(region_index_start + regions_per_thread,
2425                                        region_index_end_dense_prefix);
2426         task_queue.push(UpdateDensePrefixTask(SpaceId(space_id),
2427                                               region_index_start,
2428                                               region_index_end));
2429         region_index_start = region_index_end;
2430       }
2431     }
2432     // This gets any part of the dense prefix that did not
2433     // fit evenly.
2434     if (region_index_start < region_index_end_dense_prefix) {
2435       task_queue.push(UpdateDensePrefixTask(SpaceId(space_id),
2436                                             region_index_start,
2437                                             region_index_end_dense_prefix));
2438     }
2439   }
2440 }
2441 
2442 #ifdef ASSERT
2443 // Write a histogram of the number of times the block table was filled for a
2444 // region.
2445 void PSParallelCompact::write_block_fill_histogram()
2446 {
2447   if (!log_develop_is_enabled(Trace, gc, compaction)) {
2448     return;
2449   }
2450 
2451   Log(gc, compaction) log;
2452   ResourceMark rm;
2453   LogStream ls(log.trace());
2454   outputStream* out = &ls;
2455 
2456   typedef ParallelCompactData::RegionData rd_t;
2457   ParallelCompactData& sd = summary_data();
2458 
2459   for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2460     MutableSpace* const spc = _space_info[id].space();
2461     if (spc->bottom() != spc->top()) {
2462       const rd_t* const beg = sd.addr_to_region_ptr(spc->bottom());
2463       HeapWord* const top_aligned_up = sd.region_align_up(spc->top());
2464       const rd_t* const end = sd.addr_to_region_ptr(top_aligned_up);
2465 
2466       size_t histo[5] = { 0, 0, 0, 0, 0 };
2467       const size_t histo_len = sizeof(histo) / sizeof(size_t);
2468       const size_t region_cnt = pointer_delta(end, beg, sizeof(rd_t));
2469 
2470       for (const rd_t* cur = beg; cur < end; ++cur) {
2471         ++histo[MIN2(cur->blocks_filled_count(), histo_len - 1)];
2472       }
2473       out->print("Block fill histogram: %u %-4s" SIZE_FORMAT_W(5), id, space_names[id], region_cnt);
2474       for (size_t i = 0; i < histo_len; ++i) {
2475         out->print(" " SIZE_FORMAT_W(5) " %5.1f%%",
2476                    histo[i], 100.0 * histo[i] / region_cnt);
2477       }
2478       out->cr();
2479     }
2480   }
2481 }
2482 #endif // #ifdef ASSERT
2483 
2484 static void compaction_with_stealing_work(TaskTerminator* terminator, uint worker_id) {
2485   assert(ParallelScavengeHeap::heap()->is_gc_active(), "called outside gc");
2486 
2487   ParCompactionManager* cm =
2488     ParCompactionManager::gc_thread_compaction_manager(worker_id);
2489 
2490   // Drain the stacks that have been preloaded with regions
2491   // that are ready to fill.
2492 
2493   cm->drain_region_stacks();
2494 
2495   guarantee(cm->region_stack()->is_empty(), "Not empty");
2496 
2497   size_t region_index = 0;
2498 
2499   while (true) {
2500     if (ParCompactionManager::steal(worker_id, region_index)) {
2501       PSParallelCompact::fill_and_update_region(cm, region_index);
2502       cm->drain_region_stacks();
2503     } else if (PSParallelCompact::steal_unavailable_region(cm, region_index)) {
2504       // Fill and update an unavailable region with the help of a shadow region
2505       PSParallelCompact::fill_and_update_shadow_region(cm, region_index);
2506       cm->drain_region_stacks();
2507     } else {
2508       if (terminator->offer_termination()) {
2509         break;
2510       }
2511       // Go around again.
2512     }
2513   }
2514   return;
2515 }
2516 
2517 class UpdateDensePrefixAndCompactionTask: public AbstractGangTask {
2518   typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
2519   TaskQueue& _tq;
2520   TaskTerminator _terminator;
2521   uint _active_workers;
2522 
2523 public:
2524   UpdateDensePrefixAndCompactionTask(TaskQueue& tq, uint active_workers) :
2525       AbstractGangTask("UpdateDensePrefixAndCompactionTask"),
2526       _tq(tq),
2527       _terminator(active_workers, ParCompactionManager::region_task_queues()),
2528       _active_workers(active_workers) {
2529   }
2530   virtual void work(uint worker_id) {
2531     ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(worker_id);
2532 
2533     for (PSParallelCompact::UpdateDensePrefixTask task; _tq.try_claim(task); /* empty */) {
2534       PSParallelCompact::update_and_deadwood_in_dense_prefix(cm,
2535                                                              task._space_id,
2536                                                              task._region_index_start,
2537                                                              task._region_index_end);
2538     }
2539 
2540     // Once a thread has drained it's stack, it should try to steal regions from
2541     // other threads.
2542     compaction_with_stealing_work(&_terminator, worker_id);
2543   }
2544 };
2545 
2546 void PSParallelCompact::compact() {
2547   GCTraceTime(Info, gc, phases) tm("Compaction Phase", &_gc_timer);
2548 
2549   ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
2550   PSOldGen* old_gen = heap->old_gen();
2551   old_gen->start_array()->reset();
2552   uint active_gc_threads = ParallelScavengeHeap::heap()->workers().active_workers();
2553 
2554   // for [0..last_space_id)
2555   //     for [0..active_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING)
2556   //         push
2557   //     push
2558   //
2559   // max push count is thus: last_space_id * (active_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING + 1)
2560   TaskQueue task_queue(last_space_id * (active_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING + 1));
2561   initialize_shadow_regions(active_gc_threads);
2562   prepare_region_draining_tasks(active_gc_threads);
2563   enqueue_dense_prefix_tasks(task_queue, active_gc_threads);
2564 
2565   {
2566     GCTraceTime(Trace, gc, phases) tm("Par Compact", &_gc_timer);
2567 
2568     UpdateDensePrefixAndCompactionTask task(task_queue, active_gc_threads);
2569     ParallelScavengeHeap::heap()->workers().run_task(&task);
2570 
2571 #ifdef  ASSERT
2572     // Verify that all regions have been processed before the deferred updates.
2573     for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2574       verify_complete(SpaceId(id));
2575     }
2576 #endif
2577   }
2578 
2579   {
2580     // Update the deferred objects, if any.  Any compaction manager can be used.
2581     GCTraceTime(Trace, gc, phases) tm("Deferred Updates", &_gc_timer);
2582     ParCompactionManager* cm = ParCompactionManager::manager_array(0);
2583     for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2584       update_deferred_objects(cm, SpaceId(id));
2585     }
2586   }
2587 
2588   DEBUG_ONLY(write_block_fill_histogram());
2589 }
2590 
2591 #ifdef  ASSERT
2592 void PSParallelCompact::verify_complete(SpaceId space_id) {
2593   // All Regions between space bottom() to new_top() should be marked as filled
2594   // and all Regions between new_top() and top() should be available (i.e.,
2595   // should have been emptied).
2596   ParallelCompactData& sd = summary_data();
2597   SpaceInfo si = _space_info[space_id];
2598   HeapWord* new_top_addr = sd.region_align_up(si.new_top());
2599   HeapWord* old_top_addr = sd.region_align_up(si.space()->top());
2600   const size_t beg_region = sd.addr_to_region_idx(si.space()->bottom());
2601   const size_t new_top_region = sd.addr_to_region_idx(new_top_addr);
2602   const size_t old_top_region = sd.addr_to_region_idx(old_top_addr);
2603 
2604   bool issued_a_warning = false;
2605 
2606   size_t cur_region;
2607   for (cur_region = beg_region; cur_region < new_top_region; ++cur_region) {
2608     const RegionData* const c = sd.region(cur_region);
2609     if (!c->completed()) {
2610       log_warning(gc)("region " SIZE_FORMAT " not filled: destination_count=%u",
2611                       cur_region, c->destination_count());
2612       issued_a_warning = true;
2613     }
2614   }
2615 
2616   for (cur_region = new_top_region; cur_region < old_top_region; ++cur_region) {
2617     const RegionData* const c = sd.region(cur_region);
2618     if (!c->available()) {
2619       log_warning(gc)("region " SIZE_FORMAT " not empty: destination_count=%u",
2620                       cur_region, c->destination_count());
2621       issued_a_warning = true;
2622     }
2623   }
2624 
2625   if (issued_a_warning) {
2626     print_region_ranges();
2627   }
2628 }
2629 #endif  // #ifdef ASSERT
2630 
2631 inline void UpdateOnlyClosure::do_addr(HeapWord* addr) {
2632   _start_array->allocate_block(addr);
2633   compaction_manager()->update_contents(oop(addr));
2634 }
2635 
2636 // Update interior oops in the ranges of regions [beg_region, end_region).
2637 void
2638 PSParallelCompact::update_and_deadwood_in_dense_prefix(ParCompactionManager* cm,
2639                                                        SpaceId space_id,
2640                                                        size_t beg_region,
2641                                                        size_t end_region) {
2642   ParallelCompactData& sd = summary_data();
2643   ParMarkBitMap* const mbm = mark_bitmap();
2644 
2645   HeapWord* beg_addr = sd.region_to_addr(beg_region);
2646   HeapWord* const end_addr = sd.region_to_addr(end_region);
2647   assert(beg_region <= end_region, "bad region range");
2648   assert(end_addr <= dense_prefix(space_id), "not in the dense prefix");
2649 
2650 #ifdef  ASSERT
2651   // Claim the regions to avoid triggering an assert when they are marked as
2652   // filled.
2653   for (size_t claim_region = beg_region; claim_region < end_region; ++claim_region) {
2654     assert(sd.region(claim_region)->claim_unsafe(), "claim() failed");
2655   }
2656 #endif  // #ifdef ASSERT
2657 
2658   if (beg_addr != space(space_id)->bottom()) {
2659     // Find the first live object or block of dead space that *starts* in this
2660     // range of regions.  If a partial object crosses onto the region, skip it;
2661     // it will be marked for 'deferred update' when the object head is
2662     // processed.  If dead space crosses onto the region, it is also skipped; it
2663     // will be filled when the prior region is processed.  If neither of those
2664     // apply, the first word in the region is the start of a live object or dead
2665     // space.
2666     assert(beg_addr > space(space_id)->bottom(), "sanity");
2667     const RegionData* const cp = sd.region(beg_region);
2668     if (cp->partial_obj_size() != 0) {
2669       beg_addr = sd.partial_obj_end(beg_region);
2670     } else if (dead_space_crosses_boundary(cp, mbm->addr_to_bit(beg_addr))) {
2671       beg_addr = mbm->find_obj_beg(beg_addr, end_addr);
2672     }
2673   }
2674 
2675   if (beg_addr < end_addr) {
2676     // A live object or block of dead space starts in this range of Regions.
2677      HeapWord* const dense_prefix_end = dense_prefix(space_id);
2678 
2679     // Create closures and iterate.
2680     UpdateOnlyClosure update_closure(mbm, cm, space_id);
2681     FillClosure fill_closure(cm, space_id);
2682     ParMarkBitMap::IterationStatus status;
2683     status = mbm->iterate(&update_closure, &fill_closure, beg_addr, end_addr,
2684                           dense_prefix_end);
2685     if (status == ParMarkBitMap::incomplete) {
2686       update_closure.do_addr(update_closure.source());
2687     }
2688   }
2689 
2690   // Mark the regions as filled.
2691   RegionData* const beg_cp = sd.region(beg_region);
2692   RegionData* const end_cp = sd.region(end_region);
2693   for (RegionData* cp = beg_cp; cp < end_cp; ++cp) {
2694     cp->set_completed();
2695   }
2696 }
2697 
2698 // Return the SpaceId for the space containing addr.  If addr is not in the
2699 // heap, last_space_id is returned.  In debug mode it expects the address to be
2700 // in the heap and asserts such.
2701 PSParallelCompact::SpaceId PSParallelCompact::space_id(HeapWord* addr) {
2702   assert(ParallelScavengeHeap::heap()->is_in_reserved(addr), "addr not in the heap");
2703 
2704   for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2705     if (_space_info[id].space()->contains(addr)) {
2706       return SpaceId(id);
2707     }
2708   }
2709 
2710   assert(false, "no space contains the addr");
2711   return last_space_id;
2712 }
2713 
2714 void PSParallelCompact::update_deferred_objects(ParCompactionManager* cm,
2715                                                 SpaceId id) {
2716   assert(id < last_space_id, "bad space id");
2717 
2718   ParallelCompactData& sd = summary_data();
2719   const SpaceInfo* const space_info = _space_info + id;
2720   ObjectStartArray* const start_array = space_info->start_array();
2721 
2722   const MutableSpace* const space = space_info->space();
2723   assert(space_info->dense_prefix() >= space->bottom(), "dense_prefix not set");
2724   HeapWord* const beg_addr = space_info->dense_prefix();
2725   HeapWord* const end_addr = sd.region_align_up(space_info->new_top());
2726 
2727   const RegionData* const beg_region = sd.addr_to_region_ptr(beg_addr);
2728   const RegionData* const end_region = sd.addr_to_region_ptr(end_addr);
2729   const RegionData* cur_region;
2730   for (cur_region = beg_region; cur_region < end_region; ++cur_region) {
2731     HeapWord* const addr = cur_region->deferred_obj_addr();
2732     if (addr != NULL) {
2733       if (start_array != NULL) {
2734         start_array->allocate_block(addr);
2735       }
2736       cm->update_contents(oop(addr));
2737       assert(oopDesc::is_oop_or_null(oop(addr)), "Expected an oop or NULL at " PTR_FORMAT, p2i(oop(addr)));
2738     }
2739   }
2740 }
2741 
2742 // Skip over count live words starting from beg, and return the address of the
2743 // next live word.  Unless marked, the word corresponding to beg is assumed to
2744 // be dead.  Callers must either ensure beg does not correspond to the middle of
2745 // an object, or account for those live words in some other way.  Callers must
2746 // also ensure that there are enough live words in the range [beg, end) to skip.
2747 HeapWord*
2748 PSParallelCompact::skip_live_words(HeapWord* beg, HeapWord* end, size_t count)
2749 {
2750   assert(count > 0, "sanity");
2751 
2752   ParMarkBitMap* m = mark_bitmap();
2753   idx_t bits_to_skip = m->words_to_bits(count);
2754   idx_t cur_beg = m->addr_to_bit(beg);
2755   const idx_t search_end = m->align_range_end(m->addr_to_bit(end));
2756 
2757   do {
2758     cur_beg = m->find_obj_beg(cur_beg, search_end);
2759     idx_t cur_end = m->find_obj_end(cur_beg, search_end);
2760     const size_t obj_bits = cur_end - cur_beg + 1;
2761     if (obj_bits > bits_to_skip) {
2762       return m->bit_to_addr(cur_beg + bits_to_skip);
2763     }
2764     bits_to_skip -= obj_bits;
2765     cur_beg = cur_end + 1;
2766   } while (bits_to_skip > 0);
2767 
2768   // Skipping the desired number of words landed just past the end of an object.
2769   // Find the start of the next object.
2770   cur_beg = m->find_obj_beg(cur_beg, search_end);
2771   assert(cur_beg < m->addr_to_bit(end), "not enough live words to skip");
2772   return m->bit_to_addr(cur_beg);
2773 }
2774 
2775 HeapWord* PSParallelCompact::first_src_addr(HeapWord* const dest_addr,
2776                                             SpaceId src_space_id,
2777                                             size_t src_region_idx)
2778 {
2779   assert(summary_data().is_region_aligned(dest_addr), "not aligned");
2780 
2781   const SplitInfo& split_info = _space_info[src_space_id].split_info();
2782   if (split_info.dest_region_addr() == dest_addr) {
2783     // The partial object ending at the split point contains the first word to
2784     // be copied to dest_addr.
2785     return split_info.first_src_addr();
2786   }
2787 
2788   const ParallelCompactData& sd = summary_data();
2789   ParMarkBitMap* const bitmap = mark_bitmap();
2790   const size_t RegionSize = ParallelCompactData::RegionSize;
2791 
2792   assert(sd.is_region_aligned(dest_addr), "not aligned");
2793   const RegionData* const src_region_ptr = sd.region(src_region_idx);
2794   const size_t partial_obj_size = src_region_ptr->partial_obj_size();
2795   HeapWord* const src_region_destination = src_region_ptr->destination();
2796 
2797   assert(dest_addr >= src_region_destination, "wrong src region");
2798   assert(src_region_ptr->data_size() > 0, "src region cannot be empty");
2799 
2800   HeapWord* const src_region_beg = sd.region_to_addr(src_region_idx);
2801   HeapWord* const src_region_end = src_region_beg + RegionSize;
2802 
2803   HeapWord* addr = src_region_beg;
2804   if (dest_addr == src_region_destination) {
2805     // Return the first live word in the source region.
2806     if (partial_obj_size == 0) {
2807       addr = bitmap->find_obj_beg(addr, src_region_end);
2808       assert(addr < src_region_end, "no objects start in src region");
2809     }
2810     return addr;
2811   }
2812 
2813   // Must skip some live data.
2814   size_t words_to_skip = dest_addr - src_region_destination;
2815   assert(src_region_ptr->data_size() > words_to_skip, "wrong src region");
2816 
2817   if (partial_obj_size >= words_to_skip) {
2818     // All the live words to skip are part of the partial object.
2819     addr += words_to_skip;
2820     if (partial_obj_size == words_to_skip) {
2821       // Find the first live word past the partial object.
2822       addr = bitmap->find_obj_beg(addr, src_region_end);
2823       assert(addr < src_region_end, "wrong src region");
2824     }
2825     return addr;
2826   }
2827 
2828   // Skip over the partial object (if any).
2829   if (partial_obj_size != 0) {
2830     words_to_skip -= partial_obj_size;
2831     addr += partial_obj_size;
2832   }
2833 
2834   // Skip over live words due to objects that start in the region.
2835   addr = skip_live_words(addr, src_region_end, words_to_skip);
2836   assert(addr < src_region_end, "wrong src region");
2837   return addr;
2838 }
2839 
2840 void PSParallelCompact::decrement_destination_counts(ParCompactionManager* cm,
2841                                                      SpaceId src_space_id,
2842                                                      size_t beg_region,
2843                                                      HeapWord* end_addr)
2844 {
2845   ParallelCompactData& sd = summary_data();
2846 
2847 #ifdef ASSERT
2848   MutableSpace* const src_space = _space_info[src_space_id].space();
2849   HeapWord* const beg_addr = sd.region_to_addr(beg_region);
2850   assert(src_space->contains(beg_addr) || beg_addr == src_space->end(),
2851          "src_space_id does not match beg_addr");
2852   assert(src_space->contains(end_addr) || end_addr == src_space->end(),
2853          "src_space_id does not match end_addr");
2854 #endif // #ifdef ASSERT
2855 
2856   RegionData* const beg = sd.region(beg_region);
2857   RegionData* const end = sd.addr_to_region_ptr(sd.region_align_up(end_addr));
2858 
2859   // Regions up to new_top() are enqueued if they become available.
2860   HeapWord* const new_top = _space_info[src_space_id].new_top();
2861   RegionData* const enqueue_end =
2862     sd.addr_to_region_ptr(sd.region_align_up(new_top));
2863 
2864   for (RegionData* cur = beg; cur < end; ++cur) {
2865     assert(cur->data_size() > 0, "region must have live data");
2866     cur->decrement_destination_count();
2867     if (cur < enqueue_end && cur->available() && cur->claim()) {
2868       if (cur->mark_normal()) {
2869         cm->push_region(sd.region(cur));
2870       } else if (cur->mark_copied()) {
2871         // Try to copy the content of the shadow region back to its corresponding
2872         // heap region if the shadow region is filled. Otherwise, the GC thread
2873         // fills the shadow region will copy the data back (see
2874         // MoveAndUpdateShadowClosure::complete_region).
2875         copy_back(sd.region_to_addr(cur->shadow_region()), sd.region_to_addr(cur));
2876         ParCompactionManager::push_shadow_region_mt_safe(cur->shadow_region());
2877         cur->set_completed();
2878       }
2879     }
2880   }
2881 }
2882 
2883 size_t PSParallelCompact::next_src_region(MoveAndUpdateClosure& closure,
2884                                           SpaceId& src_space_id,
2885                                           HeapWord*& src_space_top,
2886                                           HeapWord* end_addr)
2887 {
2888   typedef ParallelCompactData::RegionData RegionData;
2889 
2890   ParallelCompactData& sd = PSParallelCompact::summary_data();
2891   const size_t region_size = ParallelCompactData::RegionSize;
2892 
2893   size_t src_region_idx = 0;
2894 
2895   // Skip empty regions (if any) up to the top of the space.
2896   HeapWord* const src_aligned_up = sd.region_align_up(end_addr);
2897   RegionData* src_region_ptr = sd.addr_to_region_ptr(src_aligned_up);
2898   HeapWord* const top_aligned_up = sd.region_align_up(src_space_top);
2899   const RegionData* const top_region_ptr =
2900     sd.addr_to_region_ptr(top_aligned_up);
2901   while (src_region_ptr < top_region_ptr && src_region_ptr->data_size() == 0) {
2902     ++src_region_ptr;
2903   }
2904 
2905   if (src_region_ptr < top_region_ptr) {
2906     // The next source region is in the current space.  Update src_region_idx
2907     // and the source address to match src_region_ptr.
2908     src_region_idx = sd.region(src_region_ptr);
2909     HeapWord* const src_region_addr = sd.region_to_addr(src_region_idx);
2910     if (src_region_addr > closure.source()) {
2911       closure.set_source(src_region_addr);
2912     }
2913     return src_region_idx;
2914   }
2915 
2916   // Switch to a new source space and find the first non-empty region.
2917   unsigned int space_id = src_space_id + 1;
2918   assert(space_id < last_space_id, "not enough spaces");
2919 
2920   HeapWord* const destination = closure.destination();
2921 
2922   do {
2923     MutableSpace* space = _space_info[space_id].space();
2924     HeapWord* const bottom = space->bottom();
2925     const RegionData* const bottom_cp = sd.addr_to_region_ptr(bottom);
2926 
2927     // Iterate over the spaces that do not compact into themselves.
2928     if (bottom_cp->destination() != bottom) {
2929       HeapWord* const top_aligned_up = sd.region_align_up(space->top());
2930       const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up);
2931 
2932       for (const RegionData* src_cp = bottom_cp; src_cp < top_cp; ++src_cp) {
2933         if (src_cp->live_obj_size() > 0) {
2934           // Found it.
2935           assert(src_cp->destination() == destination,
2936                  "first live obj in the space must match the destination");
2937           assert(src_cp->partial_obj_size() == 0,
2938                  "a space cannot begin with a partial obj");
2939 
2940           src_space_id = SpaceId(space_id);
2941           src_space_top = space->top();
2942           const size_t src_region_idx = sd.region(src_cp);
2943           closure.set_source(sd.region_to_addr(src_region_idx));
2944           return src_region_idx;
2945         } else {
2946           assert(src_cp->data_size() == 0, "sanity");
2947         }
2948       }
2949     }
2950   } while (++space_id < last_space_id);
2951 
2952   assert(false, "no source region was found");
2953   return 0;
2954 }
2955 
2956 void PSParallelCompact::fill_region(ParCompactionManager* cm, MoveAndUpdateClosure& closure, size_t region_idx)
2957 {
2958   typedef ParMarkBitMap::IterationStatus IterationStatus;
2959   ParMarkBitMap* const bitmap = mark_bitmap();
2960   ParallelCompactData& sd = summary_data();
2961   RegionData* const region_ptr = sd.region(region_idx);
2962 
2963   // Get the source region and related info.
2964   size_t src_region_idx = region_ptr->source_region();
2965   SpaceId src_space_id = space_id(sd.region_to_addr(src_region_idx));
2966   HeapWord* src_space_top = _space_info[src_space_id].space()->top();
2967   HeapWord* dest_addr = sd.region_to_addr(region_idx);
2968 
2969   closure.set_source(first_src_addr(dest_addr, src_space_id, src_region_idx));
2970 
2971   // Adjust src_region_idx to prepare for decrementing destination counts (the
2972   // destination count is not decremented when a region is copied to itself).
2973   if (src_region_idx == region_idx) {
2974     src_region_idx += 1;
2975   }
2976 
2977   if (bitmap->is_unmarked(closure.source())) {
2978     // The first source word is in the middle of an object; copy the remainder
2979     // of the object or as much as will fit.  The fact that pointer updates were
2980     // deferred will be noted when the object header is processed.
2981     HeapWord* const old_src_addr = closure.source();
2982     closure.copy_partial_obj();
2983     if (closure.is_full()) {
2984       decrement_destination_counts(cm, src_space_id, src_region_idx,
2985                                    closure.source());
2986       region_ptr->set_deferred_obj_addr(NULL);
2987       closure.complete_region(cm, dest_addr, region_ptr);
2988       return;
2989     }
2990 
2991     HeapWord* const end_addr = sd.region_align_down(closure.source());
2992     if (sd.region_align_down(old_src_addr) != end_addr) {
2993       // The partial object was copied from more than one source region.
2994       decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr);
2995 
2996       // Move to the next source region, possibly switching spaces as well.  All
2997       // args except end_addr may be modified.
2998       src_region_idx = next_src_region(closure, src_space_id, src_space_top,
2999                                        end_addr);
3000     }
3001   }
3002 
3003   do {
3004     HeapWord* const cur_addr = closure.source();
3005     HeapWord* const end_addr = MIN2(sd.region_align_up(cur_addr + 1),
3006                                     src_space_top);
3007     IterationStatus status = bitmap->iterate(&closure, cur_addr, end_addr);
3008 
3009     if (status == ParMarkBitMap::incomplete) {
3010       // The last obj that starts in the source region does not end in the
3011       // region.
3012       assert(closure.source() < end_addr, "sanity");
3013       HeapWord* const obj_beg = closure.source();
3014       HeapWord* const range_end = MIN2(obj_beg + closure.words_remaining(),
3015                                        src_space_top);
3016       HeapWord* const obj_end = bitmap->find_obj_end(obj_beg, range_end);
3017       if (obj_end < range_end) {
3018         // The end was found; the entire object will fit.
3019         status = closure.do_addr(obj_beg, bitmap->obj_size(obj_beg, obj_end));
3020         assert(status != ParMarkBitMap::would_overflow, "sanity");
3021       } else {
3022         // The end was not found; the object will not fit.
3023         assert(range_end < src_space_top, "obj cannot cross space boundary");
3024         status = ParMarkBitMap::would_overflow;
3025       }
3026     }
3027 
3028     if (status == ParMarkBitMap::would_overflow) {
3029       // The last object did not fit.  Note that interior oop updates were
3030       // deferred, then copy enough of the object to fill the region.
3031       region_ptr->set_deferred_obj_addr(closure.destination());
3032       status = closure.copy_until_full(); // copies from closure.source()
3033 
3034       decrement_destination_counts(cm, src_space_id, src_region_idx,
3035                                    closure.source());
3036       closure.complete_region(cm, dest_addr, region_ptr);
3037       return;
3038     }
3039 
3040     if (status == ParMarkBitMap::full) {
3041       decrement_destination_counts(cm, src_space_id, src_region_idx,
3042                                    closure.source());
3043       region_ptr->set_deferred_obj_addr(NULL);
3044       closure.complete_region(cm, dest_addr, region_ptr);
3045       return;
3046     }
3047 
3048     decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr);
3049 
3050     // Move to the next source region, possibly switching spaces as well.  All
3051     // args except end_addr may be modified.
3052     src_region_idx = next_src_region(closure, src_space_id, src_space_top,
3053                                      end_addr);
3054   } while (true);
3055 }
3056 
3057 void PSParallelCompact::fill_and_update_region(ParCompactionManager* cm, size_t region_idx)
3058 {
3059   MoveAndUpdateClosure cl(mark_bitmap(), cm, region_idx);
3060   fill_region(cm, cl, region_idx);
3061 }
3062 
3063 void PSParallelCompact::fill_and_update_shadow_region(ParCompactionManager* cm, size_t region_idx)
3064 {
3065   // Get a shadow region first
3066   ParallelCompactData& sd = summary_data();
3067   RegionData* const region_ptr = sd.region(region_idx);
3068   size_t shadow_region = ParCompactionManager::pop_shadow_region_mt_safe(region_ptr);
3069   // The InvalidShadow return value indicates the corresponding heap region is available,
3070   // so use MoveAndUpdateClosure to fill the normal region. Otherwise, use
3071   // MoveAndUpdateShadowClosure to fill the acquired shadow region.
3072   if (shadow_region == ParCompactionManager::InvalidShadow) {
3073     MoveAndUpdateClosure cl(mark_bitmap(), cm, region_idx);
3074     region_ptr->shadow_to_normal();
3075     return fill_region(cm, cl, region_idx);
3076   } else {
3077     MoveAndUpdateShadowClosure cl(mark_bitmap(), cm, region_idx, shadow_region);
3078     return fill_region(cm, cl, region_idx);
3079   }
3080 }
3081 
3082 void PSParallelCompact::copy_back(HeapWord *shadow_addr, HeapWord *region_addr)
3083 {
3084   Copy::aligned_conjoint_words(shadow_addr, region_addr, _summary_data.RegionSize);
3085 }
3086 
3087 bool PSParallelCompact::steal_unavailable_region(ParCompactionManager* cm, size_t &region_idx)
3088 {
3089   size_t next = cm->next_shadow_region();
3090   ParallelCompactData& sd = summary_data();
3091   size_t old_new_top = sd.addr_to_region_idx(_space_info[old_space_id].new_top());
3092   uint active_gc_threads = ParallelScavengeHeap::heap()->workers().active_workers();
3093 
3094   while (next < old_new_top) {
3095     if (sd.region(next)->mark_shadow()) {
3096       region_idx = next;
3097       return true;
3098     }
3099     next = cm->move_next_shadow_region_by(active_gc_threads);
3100   }
3101 
3102   return false;
3103 }
3104 
3105 // The shadow region is an optimization to address region dependencies in full GC. The basic
3106 // idea is making more regions available by temporally storing their live objects in empty
3107 // shadow regions to resolve dependencies between them and the destination regions. Therefore,
3108 // GC threads need not wait destination regions to be available before processing sources.
3109 //
3110 // A typical workflow would be:
3111 // After draining its own stack and failing to steal from others, a GC worker would pick an
3112 // unavailable region (destination count > 0) and get a shadow region. Then the worker fills
3113 // the shadow region by copying live objects from source regions of the unavailable one. Once
3114 // the unavailable region becomes available, the data in the shadow region will be copied back.
3115 // Shadow regions are empty regions in the to-space and regions between top and end of other spaces.
3116 //
3117 // For more details, please refer to ยง4.2 of the VEE'19 paper:
3118 // Haoyu Li, Mingyu Wu, Binyu Zang, and Haibo Chen. 2019. ScissorGC: scalable and efficient
3119 // compaction for Java full garbage collection. In Proceedings of the 15th ACM SIGPLAN/SIGOPS
3120 // International Conference on Virtual Execution Environments (VEE 2019). ACM, New York, NY, USA,
3121 // 108-121. DOI: https://doi.org/10.1145/3313808.3313820
3122 void PSParallelCompact::initialize_shadow_regions(uint parallel_gc_threads)
3123 {
3124   const ParallelCompactData& sd = PSParallelCompact::summary_data();
3125 
3126   for (unsigned int id = old_space_id; id < last_space_id; ++id) {
3127     SpaceInfo* const space_info = _space_info + id;
3128     MutableSpace* const space = space_info->space();
3129 
3130     const size_t beg_region =
3131       sd.addr_to_region_idx(sd.region_align_up(MAX2(space_info->new_top(), space->top())));
3132     const size_t end_region =
3133       sd.addr_to_region_idx(sd.region_align_down(space->end()));
3134 
3135     for (size_t cur = beg_region; cur < end_region; ++cur) {
3136       ParCompactionManager::push_shadow_region(cur);
3137     }
3138   }
3139 
3140   size_t beg_region = sd.addr_to_region_idx(_space_info[old_space_id].dense_prefix());
3141   for (uint i = 0; i < parallel_gc_threads; i++) {
3142     ParCompactionManager *cm = ParCompactionManager::manager_array(i);
3143     cm->set_next_shadow_region(beg_region + i);
3144   }
3145 }
3146 
3147 void PSParallelCompact::fill_blocks(size_t region_idx)
3148 {
3149   // Fill in the block table elements for the specified region.  Each block
3150   // table element holds the number of live words in the region that are to the
3151   // left of the first object that starts in the block.  Thus only blocks in
3152   // which an object starts need to be filled.
3153   //
3154   // The algorithm scans the section of the bitmap that corresponds to the
3155   // region, keeping a running total of the live words.  When an object start is
3156   // found, if it's the first to start in the block that contains it, the
3157   // current total is written to the block table element.
3158   const size_t Log2BlockSize = ParallelCompactData::Log2BlockSize;
3159   const size_t Log2RegionSize = ParallelCompactData::Log2RegionSize;
3160   const size_t RegionSize = ParallelCompactData::RegionSize;
3161 
3162   ParallelCompactData& sd = summary_data();
3163   const size_t partial_obj_size = sd.region(region_idx)->partial_obj_size();
3164   if (partial_obj_size >= RegionSize) {
3165     return; // No objects start in this region.
3166   }
3167 
3168   // Ensure the first loop iteration decides that the block has changed.
3169   size_t cur_block = sd.block_count();
3170 
3171   const ParMarkBitMap* const bitmap = mark_bitmap();
3172 
3173   const size_t Log2BitsPerBlock = Log2BlockSize - LogMinObjAlignment;
3174   assert((size_t)1 << Log2BitsPerBlock ==
3175          bitmap->words_to_bits(ParallelCompactData::BlockSize), "sanity");
3176 
3177   size_t beg_bit = bitmap->words_to_bits(region_idx << Log2RegionSize);
3178   const size_t range_end = beg_bit + bitmap->words_to_bits(RegionSize);
3179   size_t live_bits = bitmap->words_to_bits(partial_obj_size);
3180   beg_bit = bitmap->find_obj_beg(beg_bit + live_bits, range_end);
3181   while (beg_bit < range_end) {
3182     const size_t new_block = beg_bit >> Log2BitsPerBlock;
3183     if (new_block != cur_block) {
3184       cur_block = new_block;
3185       sd.block(cur_block)->set_offset(bitmap->bits_to_words(live_bits));
3186     }
3187 
3188     const size_t end_bit = bitmap->find_obj_end(beg_bit, range_end);
3189     if (end_bit < range_end - 1) {
3190       live_bits += end_bit - beg_bit + 1;
3191       beg_bit = bitmap->find_obj_beg(end_bit + 1, range_end);
3192     } else {
3193       return;
3194     }
3195   }
3196 }
3197 
3198 ParMarkBitMap::IterationStatus MoveAndUpdateClosure::copy_until_full()
3199 {
3200   if (source() != copy_destination()) {
3201     DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3202     Copy::aligned_conjoint_words(source(), copy_destination(), words_remaining());
3203   }
3204   update_state(words_remaining());
3205   assert(is_full(), "sanity");
3206   return ParMarkBitMap::full;
3207 }
3208 
3209 void MoveAndUpdateClosure::copy_partial_obj()
3210 {
3211   size_t words = words_remaining();
3212 
3213   HeapWord* const range_end = MIN2(source() + words, bitmap()->region_end());
3214   HeapWord* const end_addr = bitmap()->find_obj_end(source(), range_end);
3215   if (end_addr < range_end) {
3216     words = bitmap()->obj_size(source(), end_addr);
3217   }
3218 
3219   // This test is necessary; if omitted, the pointer updates to a partial object
3220   // that crosses the dense prefix boundary could be overwritten.
3221   if (source() != copy_destination()) {
3222     DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3223     Copy::aligned_conjoint_words(source(), copy_destination(), words);
3224   }
3225   update_state(words);
3226 }
3227 
3228 void MoveAndUpdateClosure::complete_region(ParCompactionManager *cm, HeapWord *dest_addr,
3229                                            PSParallelCompact::RegionData *region_ptr) {
3230   assert(region_ptr->shadow_state() == ParallelCompactData::RegionData::NormalRegion, "Region should be finished");
3231   region_ptr->set_completed();
3232 }
3233 
3234 ParMarkBitMapClosure::IterationStatus
3235 MoveAndUpdateClosure::do_addr(HeapWord* addr, size_t words) {
3236   assert(destination() != NULL, "sanity");
3237   assert(bitmap()->obj_size(addr) == words, "bad size");
3238 
3239   _source = addr;
3240   assert(PSParallelCompact::summary_data().calc_new_pointer(source(), compaction_manager()) ==
3241          destination(), "wrong destination");
3242 
3243   if (words > words_remaining()) {
3244     return ParMarkBitMap::would_overflow;
3245   }
3246 
3247   // The start_array must be updated even if the object is not moving.
3248   if (_start_array != NULL) {
3249     _start_array->allocate_block(destination());
3250   }
3251 
3252   if (copy_destination() != source()) {
3253     DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3254     Copy::aligned_conjoint_words(source(), copy_destination(), words);
3255   }
3256 
3257   oop moved_oop = (oop) copy_destination();
3258   compaction_manager()->update_contents(moved_oop);
3259   assert(oopDesc::is_oop_or_null(moved_oop), "Expected an oop or NULL at " PTR_FORMAT, p2i(moved_oop));
3260 
3261   update_state(words);
3262   assert(copy_destination() == cast_from_oop<HeapWord*>(moved_oop) + moved_oop->size(), "sanity");
3263   return is_full() ? ParMarkBitMap::full : ParMarkBitMap::incomplete;
3264 }
3265 
3266 void MoveAndUpdateShadowClosure::complete_region(ParCompactionManager *cm, HeapWord *dest_addr,
3267                                                  PSParallelCompact::RegionData *region_ptr) {
3268   assert(region_ptr->shadow_state() == ParallelCompactData::RegionData::ShadowRegion, "Region should be shadow");
3269   // Record the shadow region index
3270   region_ptr->set_shadow_region(_shadow);
3271   // Mark the shadow region as filled to indicate the data is ready to be
3272   // copied back
3273   region_ptr->mark_filled();
3274   // Try to copy the content of the shadow region back to its corresponding
3275   // heap region if available; the GC thread that decreases the destination
3276   // count to zero will do the copying otherwise (see
3277   // PSParallelCompact::decrement_destination_counts).
3278   if (((region_ptr->available() && region_ptr->claim()) || region_ptr->claimed()) && region_ptr->mark_copied()) {
3279     region_ptr->set_completed();
3280     PSParallelCompact::copy_back(PSParallelCompact::summary_data().region_to_addr(_shadow), dest_addr);
3281     ParCompactionManager::push_shadow_region_mt_safe(_shadow);
3282   }
3283 }
3284 
3285 UpdateOnlyClosure::UpdateOnlyClosure(ParMarkBitMap* mbm,
3286                                      ParCompactionManager* cm,
3287                                      PSParallelCompact::SpaceId space_id) :
3288   ParMarkBitMapClosure(mbm, cm),
3289   _space_id(space_id),
3290   _start_array(PSParallelCompact::start_array(space_id))
3291 {
3292 }
3293 
3294 // Updates the references in the object to their new values.
3295 ParMarkBitMapClosure::IterationStatus
3296 UpdateOnlyClosure::do_addr(HeapWord* addr, size_t words) {
3297   do_addr(addr);
3298   return ParMarkBitMap::incomplete;
3299 }
3300 
3301 FillClosure::FillClosure(ParCompactionManager* cm, PSParallelCompact::SpaceId space_id) :
3302   ParMarkBitMapClosure(PSParallelCompact::mark_bitmap(), cm),
3303   _start_array(PSParallelCompact::start_array(space_id))
3304 {
3305   assert(space_id == PSParallelCompact::old_space_id,
3306          "cannot use FillClosure in the young gen");
3307 }
3308 
3309 ParMarkBitMapClosure::IterationStatus
3310 FillClosure::do_addr(HeapWord* addr, size_t size) {
3311   CollectedHeap::fill_with_objects(addr, size);
3312   HeapWord* const end = addr + size;
3313   do {
3314     _start_array->allocate_block(addr);
3315     addr += oop(addr)->size();
3316   } while (addr < end);
3317   return ParMarkBitMap::incomplete;
3318 }
--- EOF ---