1 /*
   2  * Copyright (c) 2005, 2021, 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 "classfile/classLoaderDataGraph.hpp"
  27 #include "classfile/javaClasses.inline.hpp"
  28 #include "classfile/stringTable.hpp"
  29 #include "classfile/symbolTable.hpp"
  30 #include "classfile/systemDictionary.hpp"
  31 #include "code/codeCache.hpp"
  32 #include "compiler/oopMap.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/psStringDedup.hpp"
  44 #include "gc/parallel/psYoungGen.hpp"
  45 #include "gc/shared/gcCause.hpp"
  46 #include "gc/shared/gcHeapSummary.hpp"
  47 #include "gc/shared/gcId.hpp"
  48 #include "gc/shared/gcLocker.hpp"
  49 #include "gc/shared/gcTimer.hpp"
  50 #include "gc/shared/gcTrace.hpp"
  51 #include "gc/shared/gcTraceTime.inline.hpp"
  52 #include "gc/shared/isGCActiveMark.hpp"
  53 #include "gc/shared/oopStorage.inline.hpp"
  54 #include "gc/shared/oopStorageSet.inline.hpp"
  55 #include "gc/shared/oopStorageSetParState.inline.hpp"
  56 #include "gc/shared/referencePolicy.hpp"
  57 #include "gc/shared/referenceProcessor.hpp"
  58 #include "gc/shared/referenceProcessorPhaseTimes.hpp"
  59 #include "gc/shared/spaceDecorator.inline.hpp"
  60 #include "gc/shared/taskTerminator.hpp"
  61 #include "gc/shared/weakProcessor.inline.hpp"
  62 #include "gc/shared/workerPolicy.hpp"
  63 #include "gc/shared/workerThread.hpp"
  64 #include "gc/shared/workerUtils.hpp"
  65 #include "logging/log.hpp"
  66 #include "memory/iterator.inline.hpp"
  67 #include "memory/metaspaceUtils.hpp"
  68 #include "memory/resourceArea.hpp"
  69 #include "memory/universe.hpp"
  70 #include "oops/access.inline.hpp"
  71 #include "oops/instanceClassLoaderKlass.inline.hpp"
  72 #include "oops/instanceKlass.inline.hpp"
  73 #include "oops/instanceMirrorKlass.inline.hpp"
  74 #include "oops/methodData.hpp"
  75 #include "oops/objArrayKlass.inline.hpp"
  76 #include "oops/oop.inline.hpp"
  77 #include "runtime/atomic.hpp"
  78 #include "runtime/handles.inline.hpp"
  79 #include "runtime/java.hpp"
  80 #include "runtime/safepoint.hpp"
  81 #include "runtime/vmThread.hpp"
  82 #include "services/memTracker.hpp"
  83 #include "services/memoryService.hpp"
  84 #include "utilities/align.hpp"
  85 #include "utilities/debug.hpp"
  86 #include "utilities/events.hpp"
  87 #include "utilities/formatBuffer.hpp"
  88 #include "utilities/macros.hpp"
  89 #include "utilities/stack.inline.hpp"
  90 #if INCLUDE_JVMCI
  91 #include "jvmci/jvmci.hpp"
  92 #endif
  93 
  94 #include <math.h>
  95 
  96 // All sizes are in HeapWords.
  97 const size_t ParallelCompactData::Log2RegionSize  = 16; // 64K words
  98 const size_t ParallelCompactData::RegionSize      = (size_t)1 << Log2RegionSize;
  99 const size_t ParallelCompactData::RegionSizeBytes =
 100   RegionSize << LogHeapWordSize;
 101 const size_t ParallelCompactData::RegionSizeOffsetMask = RegionSize - 1;
 102 const size_t ParallelCompactData::RegionAddrOffsetMask = RegionSizeBytes - 1;
 103 const size_t ParallelCompactData::RegionAddrMask       = ~RegionAddrOffsetMask;
 104 
 105 const size_t ParallelCompactData::Log2BlockSize   = 7; // 128 words
 106 const size_t ParallelCompactData::BlockSize       = (size_t)1 << Log2BlockSize;
 107 const size_t ParallelCompactData::BlockSizeBytes  =
 108   BlockSize << LogHeapWordSize;
 109 const size_t ParallelCompactData::BlockSizeOffsetMask = BlockSize - 1;
 110 const size_t ParallelCompactData::BlockAddrOffsetMask = BlockSizeBytes - 1;
 111 const size_t ParallelCompactData::BlockAddrMask       = ~BlockAddrOffsetMask;
 112 
 113 const size_t ParallelCompactData::BlocksPerRegion = RegionSize / BlockSize;
 114 const size_t ParallelCompactData::Log2BlocksPerRegion =
 115   Log2RegionSize - Log2BlockSize;
 116 
 117 const ParallelCompactData::RegionData::region_sz_t
 118 ParallelCompactData::RegionData::dc_shift = 27;
 119 
 120 const ParallelCompactData::RegionData::region_sz_t
 121 ParallelCompactData::RegionData::dc_mask = ~0U << dc_shift;
 122 
 123 const ParallelCompactData::RegionData::region_sz_t
 124 ParallelCompactData::RegionData::dc_one = 0x1U << dc_shift;
 125 
 126 const ParallelCompactData::RegionData::region_sz_t
 127 ParallelCompactData::RegionData::los_mask = ~dc_mask;
 128 
 129 const ParallelCompactData::RegionData::region_sz_t
 130 ParallelCompactData::RegionData::dc_claimed = 0x8U << dc_shift;
 131 
 132 const ParallelCompactData::RegionData::region_sz_t
 133 ParallelCompactData::RegionData::dc_completed = 0xcU << dc_shift;
 134 
 135 SpaceInfo PSParallelCompact::_space_info[PSParallelCompact::last_space_id];
 136 
 137 SpanSubjectToDiscoveryClosure PSParallelCompact::_span_based_discoverer;
 138 ReferenceProcessor* PSParallelCompact::_ref_processor = NULL;
 139 
 140 double PSParallelCompact::_dwl_mean;
 141 double PSParallelCompact::_dwl_std_dev;
 142 double PSParallelCompact::_dwl_first_term;
 143 double PSParallelCompact::_dwl_adjustment;
 144 #ifdef  ASSERT
 145 bool   PSParallelCompact::_dwl_initialized = false;
 146 #endif  // #ifdef ASSERT
 147 
 148 void SplitInfo::record(size_t src_region_idx, size_t partial_obj_size,
 149                        HeapWord* destination)
 150 {
 151   assert(src_region_idx != 0, "invalid src_region_idx");
 152   assert(partial_obj_size != 0, "invalid partial_obj_size argument");
 153   assert(destination != NULL, "invalid destination argument");
 154 
 155   _src_region_idx = src_region_idx;
 156   _partial_obj_size = partial_obj_size;
 157   _destination = destination;
 158 
 159   // These fields may not be updated below, so make sure they're clear.
 160   assert(_dest_region_addr == NULL, "should have been cleared");
 161   assert(_first_src_addr == NULL, "should have been cleared");
 162 
 163   // Determine the number of destination regions for the partial object.
 164   HeapWord* const last_word = destination + partial_obj_size - 1;
 165   const ParallelCompactData& sd = PSParallelCompact::summary_data();
 166   HeapWord* const beg_region_addr = sd.region_align_down(destination);
 167   HeapWord* const end_region_addr = sd.region_align_down(last_word);
 168 
 169   if (beg_region_addr == end_region_addr) {
 170     // One destination region.
 171     _destination_count = 1;
 172     if (end_region_addr == destination) {
 173       // The destination falls on a region boundary, thus the first word of the
 174       // partial object will be the first word copied to the destination region.
 175       _dest_region_addr = end_region_addr;
 176       _first_src_addr = sd.region_to_addr(src_region_idx);
 177     }
 178   } else {
 179     // Two destination regions.  When copied, the partial object will cross a
 180     // destination region boundary, so a word somewhere within the partial
 181     // object will be the first word copied to the second destination region.
 182     _destination_count = 2;
 183     _dest_region_addr = end_region_addr;
 184     const size_t ofs = pointer_delta(end_region_addr, destination);
 185     assert(ofs < _partial_obj_size, "sanity");
 186     _first_src_addr = sd.region_to_addr(src_region_idx) + ofs;
 187   }
 188 }
 189 
 190 void SplitInfo::clear()
 191 {
 192   _src_region_idx = 0;
 193   _partial_obj_size = 0;
 194   _destination = NULL;
 195   _destination_count = 0;
 196   _dest_region_addr = NULL;
 197   _first_src_addr = NULL;
 198   assert(!is_valid(), "sanity");
 199 }
 200 
 201 #ifdef  ASSERT
 202 void SplitInfo::verify_clear()
 203 {
 204   assert(_src_region_idx == 0, "not clear");
 205   assert(_partial_obj_size == 0, "not clear");
 206   assert(_destination == NULL, "not clear");
 207   assert(_destination_count == 0, "not clear");
 208   assert(_dest_region_addr == NULL, "not clear");
 209   assert(_first_src_addr == NULL, "not clear");
 210 }
 211 #endif  // #ifdef ASSERT
 212 
 213 
 214 void PSParallelCompact::print_on_error(outputStream* st) {
 215   _mark_bitmap.print_on_error(st);
 216 }
 217 
 218 #ifndef PRODUCT
 219 const char* PSParallelCompact::space_names[] = {
 220   "old ", "eden", "from", "to  "
 221 };
 222 
 223 void PSParallelCompact::print_region_ranges() {
 224   if (!log_develop_is_enabled(Trace, gc, compaction)) {
 225     return;
 226   }
 227   Log(gc, compaction) log;
 228   ResourceMark rm;
 229   LogStream ls(log.trace());
 230   Universe::print_on(&ls);
 231   log.trace("space  bottom     top        end        new_top");
 232   log.trace("------ ---------- ---------- ---------- ----------");
 233 
 234   for (unsigned int id = 0; id < last_space_id; ++id) {
 235     const MutableSpace* space = _space_info[id].space();
 236     log.trace("%u %s "
 237               SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " "
 238               SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " ",
 239               id, space_names[id],
 240               summary_data().addr_to_region_idx(space->bottom()),
 241               summary_data().addr_to_region_idx(space->top()),
 242               summary_data().addr_to_region_idx(space->end()),
 243               summary_data().addr_to_region_idx(_space_info[id].new_top()));
 244   }
 245 }
 246 
 247 void
 248 print_generic_summary_region(size_t i, const ParallelCompactData::RegionData* c)
 249 {
 250 #define REGION_IDX_FORMAT        SIZE_FORMAT_W(7)
 251 #define REGION_DATA_FORMAT       SIZE_FORMAT_W(5)
 252 
 253   ParallelCompactData& sd = PSParallelCompact::summary_data();
 254   size_t dci = c->destination() ? sd.addr_to_region_idx(c->destination()) : 0;
 255   log_develop_trace(gc, compaction)(
 256       REGION_IDX_FORMAT " " PTR_FORMAT " "
 257       REGION_IDX_FORMAT " " PTR_FORMAT " "
 258       REGION_DATA_FORMAT " " REGION_DATA_FORMAT " "
 259       REGION_DATA_FORMAT " " REGION_IDX_FORMAT " %d",
 260       i, p2i(c->data_location()), dci, p2i(c->destination()),
 261       c->partial_obj_size(), c->live_obj_size(),
 262       c->data_size(), c->source_region(), c->destination_count());
 263 
 264 #undef  REGION_IDX_FORMAT
 265 #undef  REGION_DATA_FORMAT
 266 }
 267 
 268 void
 269 print_generic_summary_data(ParallelCompactData& summary_data,
 270                            HeapWord* const beg_addr,
 271                            HeapWord* const end_addr)
 272 {
 273   size_t total_words = 0;
 274   size_t i = summary_data.addr_to_region_idx(beg_addr);
 275   const size_t last = summary_data.addr_to_region_idx(end_addr);
 276   HeapWord* pdest = 0;
 277 
 278   while (i < last) {
 279     ParallelCompactData::RegionData* c = summary_data.region(i);
 280     if (c->data_size() != 0 || c->destination() != pdest) {
 281       print_generic_summary_region(i, c);
 282       total_words += c->data_size();
 283       pdest = c->destination();
 284     }
 285     ++i;
 286   }
 287 
 288   log_develop_trace(gc, compaction)("summary_data_bytes=" SIZE_FORMAT, total_words * HeapWordSize);
 289 }
 290 
 291 void
 292 PSParallelCompact::print_generic_summary_data(ParallelCompactData& summary_data,
 293                                               HeapWord* const beg_addr,
 294                                               HeapWord* const end_addr) {
 295   ::print_generic_summary_data(summary_data,beg_addr, end_addr);
 296 }
 297 
 298 void
 299 print_generic_summary_data(ParallelCompactData& summary_data,
 300                            SpaceInfo* space_info)
 301 {
 302   if (!log_develop_is_enabled(Trace, gc, compaction)) {
 303     return;
 304   }
 305 
 306   for (unsigned int id = 0; id < PSParallelCompact::last_space_id; ++id) {
 307     const MutableSpace* space = space_info[id].space();
 308     print_generic_summary_data(summary_data, space->bottom(),
 309                                MAX2(space->top(), space_info[id].new_top()));
 310   }
 311 }
 312 
 313 void
 314 print_initial_summary_data(ParallelCompactData& summary_data,
 315                            const MutableSpace* space) {
 316   if (space->top() == space->bottom()) {
 317     return;
 318   }
 319 
 320   const size_t region_size = ParallelCompactData::RegionSize;
 321   typedef ParallelCompactData::RegionData RegionData;
 322   HeapWord* const top_aligned_up = summary_data.region_align_up(space->top());
 323   const size_t end_region = summary_data.addr_to_region_idx(top_aligned_up);
 324   const RegionData* c = summary_data.region(end_region - 1);
 325   HeapWord* end_addr = c->destination() + c->data_size();
 326   const size_t live_in_space = pointer_delta(end_addr, space->bottom());
 327 
 328   // Print (and count) the full regions at the beginning of the space.
 329   size_t full_region_count = 0;
 330   size_t i = summary_data.addr_to_region_idx(space->bottom());
 331   while (i < end_region && summary_data.region(i)->data_size() == region_size) {
 332     ParallelCompactData::RegionData* c = summary_data.region(i);
 333     log_develop_trace(gc, compaction)(
 334         SIZE_FORMAT_W(5) " " PTR_FORMAT " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d",
 335         i, p2i(c->destination()),
 336         c->partial_obj_size(), c->live_obj_size(),
 337         c->data_size(), c->source_region(), c->destination_count());
 338     ++full_region_count;
 339     ++i;
 340   }
 341 
 342   size_t live_to_right = live_in_space - full_region_count * region_size;
 343 
 344   double max_reclaimed_ratio = 0.0;
 345   size_t max_reclaimed_ratio_region = 0;
 346   size_t max_dead_to_right = 0;
 347   size_t max_live_to_right = 0;
 348 
 349   // Print the 'reclaimed ratio' for regions while there is something live in
 350   // the region or to the right of it.  The remaining regions are empty (and
 351   // uninteresting), and computing the ratio will result in division by 0.
 352   while (i < end_region && live_to_right > 0) {
 353     c = summary_data.region(i);
 354     HeapWord* const region_addr = summary_data.region_to_addr(i);
 355     const size_t used_to_right = pointer_delta(space->top(), region_addr);
 356     const size_t dead_to_right = used_to_right - live_to_right;
 357     const double reclaimed_ratio = double(dead_to_right) / live_to_right;
 358 
 359     if (reclaimed_ratio > max_reclaimed_ratio) {
 360             max_reclaimed_ratio = reclaimed_ratio;
 361             max_reclaimed_ratio_region = i;
 362             max_dead_to_right = dead_to_right;
 363             max_live_to_right = live_to_right;
 364     }
 365 
 366     ParallelCompactData::RegionData* c = summary_data.region(i);
 367     log_develop_trace(gc, compaction)(
 368         SIZE_FORMAT_W(5) " " PTR_FORMAT " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d"
 369         "%12.10f " SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10),
 370         i, p2i(c->destination()),
 371         c->partial_obj_size(), c->live_obj_size(),
 372         c->data_size(), c->source_region(), c->destination_count(),
 373         reclaimed_ratio, dead_to_right, live_to_right);
 374 
 375 
 376     live_to_right -= c->data_size();
 377     ++i;
 378   }
 379 
 380   // Any remaining regions are empty.  Print one more if there is one.
 381   if (i < end_region) {
 382     ParallelCompactData::RegionData* c = summary_data.region(i);
 383     log_develop_trace(gc, compaction)(
 384         SIZE_FORMAT_W(5) " " PTR_FORMAT " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d",
 385          i, p2i(c->destination()),
 386          c->partial_obj_size(), c->live_obj_size(),
 387          c->data_size(), c->source_region(), c->destination_count());
 388   }
 389 
 390   log_develop_trace(gc, compaction)("max:  " SIZE_FORMAT_W(4) " d2r=" SIZE_FORMAT_W(10) " l2r=" SIZE_FORMAT_W(10) " max_ratio=%14.12f",
 391                                     max_reclaimed_ratio_region, max_dead_to_right, max_live_to_right, max_reclaimed_ratio);
 392 }
 393 
 394 void
 395 print_initial_summary_data(ParallelCompactData& summary_data,
 396                            SpaceInfo* space_info) {
 397   if (!log_develop_is_enabled(Trace, gc, compaction)) {
 398     return;
 399   }
 400 
 401   unsigned int id = PSParallelCompact::old_space_id;
 402   const MutableSpace* space;
 403   do {
 404     space = space_info[id].space();
 405     print_initial_summary_data(summary_data, space);
 406   } while (++id < PSParallelCompact::eden_space_id);
 407 
 408   do {
 409     space = space_info[id].space();
 410     print_generic_summary_data(summary_data, space->bottom(), space->top());
 411   } while (++id < PSParallelCompact::last_space_id);
 412 }
 413 #endif  // #ifndef PRODUCT
 414 
 415 ParallelCompactData::ParallelCompactData() :
 416   _region_start(NULL),
 417   DEBUG_ONLY(_region_end(NULL) COMMA)
 418   _region_vspace(NULL),
 419   _reserved_byte_size(0),
 420   _region_data(NULL),
 421   _region_count(0),
 422   _block_vspace(NULL),
 423   _block_data(NULL),
 424   _block_count(0) {}
 425 
 426 bool ParallelCompactData::initialize(MemRegion covered_region)
 427 {
 428   _region_start = covered_region.start();
 429   const size_t region_size = covered_region.word_size();
 430   DEBUG_ONLY(_region_end = _region_start + region_size;)
 431 
 432   assert(region_align_down(_region_start) == _region_start,
 433          "region start not aligned");
 434   assert((region_size & RegionSizeOffsetMask) == 0,
 435          "region size not a multiple of RegionSize");
 436 
 437   bool result = initialize_region_data(region_size) && initialize_block_data();
 438   return result;
 439 }
 440 
 441 PSVirtualSpace*
 442 ParallelCompactData::create_vspace(size_t count, size_t element_size)
 443 {
 444   const size_t raw_bytes = count * element_size;
 445   const size_t page_sz = os::page_size_for_region_aligned(raw_bytes, 10);
 446   const size_t granularity = os::vm_allocation_granularity();
 447   _reserved_byte_size = align_up(raw_bytes, MAX2(page_sz, granularity));
 448 
 449   const size_t rs_align = page_sz == (size_t) os::vm_page_size() ? 0 :
 450     MAX2(page_sz, granularity);
 451   ReservedSpace rs(_reserved_byte_size, rs_align, page_sz);
 452   os::trace_page_sizes("Parallel Compact Data", raw_bytes, raw_bytes, page_sz, rs.base(),
 453                        rs.size());
 454 
 455   MemTracker::record_virtual_memory_type((address)rs.base(), mtGC);
 456 
 457   PSVirtualSpace* vspace = new PSVirtualSpace(rs, page_sz);
 458   if (vspace != 0) {
 459     if (vspace->expand_by(_reserved_byte_size)) {
 460       return vspace;
 461     }
 462     delete vspace;
 463     // Release memory reserved in the space.
 464     rs.release();
 465   }
 466 
 467   return 0;
 468 }
 469 
 470 bool ParallelCompactData::initialize_region_data(size_t region_size)
 471 {
 472   const size_t count = (region_size + RegionSizeOffsetMask) >> Log2RegionSize;
 473   _region_vspace = create_vspace(count, sizeof(RegionData));
 474   if (_region_vspace != 0) {
 475     _region_data = (RegionData*)_region_vspace->reserved_low_addr();
 476     _region_count = count;
 477     return true;
 478   }
 479   return false;
 480 }
 481 
 482 bool ParallelCompactData::initialize_block_data()
 483 {
 484   assert(_region_count != 0, "region data must be initialized first");
 485   const size_t count = _region_count << Log2BlocksPerRegion;
 486   _block_vspace = create_vspace(count, sizeof(BlockData));
 487   if (_block_vspace != 0) {
 488     _block_data = (BlockData*)_block_vspace->reserved_low_addr();
 489     _block_count = count;
 490     return true;
 491   }
 492   return false;
 493 }
 494 
 495 void ParallelCompactData::clear()
 496 {
 497   memset(_region_data, 0, _region_vspace->committed_size());
 498   memset(_block_data, 0, _block_vspace->committed_size());
 499 }
 500 
 501 void ParallelCompactData::clear_range(size_t beg_region, size_t end_region) {
 502   assert(beg_region <= _region_count, "beg_region out of range");
 503   assert(end_region <= _region_count, "end_region out of range");
 504   assert(RegionSize % BlockSize == 0, "RegionSize not a multiple of BlockSize");
 505 
 506   const size_t region_cnt = end_region - beg_region;
 507   memset(_region_data + beg_region, 0, region_cnt * sizeof(RegionData));
 508 
 509   const size_t beg_block = beg_region * BlocksPerRegion;
 510   const size_t block_cnt = region_cnt * BlocksPerRegion;
 511   memset(_block_data + beg_block, 0, block_cnt * sizeof(BlockData));
 512 }
 513 
 514 HeapWord* ParallelCompactData::partial_obj_end(size_t region_idx) const
 515 {
 516   const RegionData* cur_cp = region(region_idx);
 517   const RegionData* const end_cp = region(region_count() - 1);
 518 
 519   HeapWord* result = region_to_addr(region_idx);
 520   if (cur_cp < end_cp) {
 521     do {
 522       result += cur_cp->partial_obj_size();
 523     } while (cur_cp->partial_obj_size() == RegionSize && ++cur_cp < end_cp);
 524   }
 525   return result;
 526 }
 527 
 528 void ParallelCompactData::add_obj(HeapWord* addr, size_t len)
 529 {
 530   const size_t obj_ofs = pointer_delta(addr, _region_start);
 531   const size_t beg_region = obj_ofs >> Log2RegionSize;
 532   // end_region is inclusive
 533   const size_t end_region = (obj_ofs + len - 1) >> Log2RegionSize;
 534 
 535   if (beg_region == end_region) {
 536     // All in one region.
 537     _region_data[beg_region].add_live_obj(len);
 538     return;
 539   }
 540 
 541   // First region.
 542   const size_t beg_ofs = region_offset(addr);
 543   _region_data[beg_region].add_live_obj(RegionSize - beg_ofs);
 544 
 545   // Middle regions--completely spanned by this object.
 546   for (size_t region = beg_region + 1; region < end_region; ++region) {
 547     _region_data[region].set_partial_obj_size(RegionSize);
 548     _region_data[region].set_partial_obj_addr(addr);
 549   }
 550 
 551   // Last region.
 552   const size_t end_ofs = region_offset(addr + len - 1);
 553   _region_data[end_region].set_partial_obj_size(end_ofs + 1);
 554   _region_data[end_region].set_partial_obj_addr(addr);
 555 }
 556 
 557 void
 558 ParallelCompactData::summarize_dense_prefix(HeapWord* beg, HeapWord* end)
 559 {
 560   assert(is_region_aligned(beg), "not RegionSize aligned");
 561   assert(is_region_aligned(end), "not RegionSize aligned");
 562 
 563   size_t cur_region = addr_to_region_idx(beg);
 564   const size_t end_region = addr_to_region_idx(end);
 565   HeapWord* addr = beg;
 566   while (cur_region < end_region) {
 567     _region_data[cur_region].set_destination(addr);
 568     _region_data[cur_region].set_destination_count(0);
 569     _region_data[cur_region].set_source_region(cur_region);
 570     _region_data[cur_region].set_data_location(addr);
 571 
 572     // Update live_obj_size so the region appears completely full.
 573     size_t live_size = RegionSize - _region_data[cur_region].partial_obj_size();
 574     _region_data[cur_region].set_live_obj_size(live_size);
 575 
 576     ++cur_region;
 577     addr += RegionSize;
 578   }
 579 }
 580 
 581 // Find the point at which a space can be split and, if necessary, record the
 582 // split point.
 583 //
 584 // If the current src region (which overflowed the destination space) doesn't
 585 // have a partial object, the split point is at the beginning of the current src
 586 // region (an "easy" split, no extra bookkeeping required).
 587 //
 588 // If the current src region has a partial object, the split point is in the
 589 // region where that partial object starts (call it the split_region).  If
 590 // split_region has a partial object, then the split point is just after that
 591 // partial object (a "hard" split where we have to record the split data and
 592 // zero the partial_obj_size field).  With a "hard" split, we know that the
 593 // partial_obj ends within split_region because the partial object that caused
 594 // the overflow starts in split_region.  If split_region doesn't have a partial
 595 // obj, then the split is at the beginning of split_region (another "easy"
 596 // split).
 597 HeapWord*
 598 ParallelCompactData::summarize_split_space(size_t src_region,
 599                                            SplitInfo& split_info,
 600                                            HeapWord* destination,
 601                                            HeapWord* target_end,
 602                                            HeapWord** target_next)
 603 {
 604   assert(destination <= target_end, "sanity");
 605   assert(destination + _region_data[src_region].data_size() > target_end,
 606     "region should not fit into target space");
 607   assert(is_region_aligned(target_end), "sanity");
 608 
 609   size_t split_region = src_region;
 610   HeapWord* split_destination = destination;
 611   size_t partial_obj_size = _region_data[src_region].partial_obj_size();
 612 
 613   if (destination + partial_obj_size > target_end) {
 614     // The split point is just after the partial object (if any) in the
 615     // src_region that contains the start of the object that overflowed the
 616     // destination space.
 617     //
 618     // Find the start of the "overflow" object and set split_region to the
 619     // region containing it.
 620     HeapWord* const overflow_obj = _region_data[src_region].partial_obj_addr();
 621     split_region = addr_to_region_idx(overflow_obj);
 622 
 623     // Clear the source_region field of all destination regions whose first word
 624     // came from data after the split point (a non-null source_region field
 625     // implies a region must be filled).
 626     //
 627     // An alternative to the simple loop below:  clear during post_compact(),
 628     // which uses memcpy instead of individual stores, and is easy to
 629     // parallelize.  (The downside is that it clears the entire RegionData
 630     // object as opposed to just one field.)
 631     //
 632     // post_compact() would have to clear the summary data up to the highest
 633     // address that was written during the summary phase, which would be
 634     //
 635     //         max(top, max(new_top, clear_top))
 636     //
 637     // where clear_top is a new field in SpaceInfo.  Would have to set clear_top
 638     // to target_end.
 639     const RegionData* const sr = region(split_region);
 640     const size_t beg_idx =
 641       addr_to_region_idx(region_align_up(sr->destination() +
 642                                          sr->partial_obj_size()));
 643     const size_t end_idx = addr_to_region_idx(target_end);
 644 
 645     log_develop_trace(gc, compaction)("split:  clearing source_region field in [" SIZE_FORMAT ", " SIZE_FORMAT ")", beg_idx, end_idx);
 646     for (size_t idx = beg_idx; idx < end_idx; ++idx) {
 647       _region_data[idx].set_source_region(0);
 648     }
 649 
 650     // Set split_destination and partial_obj_size to reflect the split region.
 651     split_destination = sr->destination();
 652     partial_obj_size = sr->partial_obj_size();
 653   }
 654 
 655   // The split is recorded only if a partial object extends onto the region.
 656   if (partial_obj_size != 0) {
 657     _region_data[split_region].set_partial_obj_size(0);
 658     split_info.record(split_region, partial_obj_size, split_destination);
 659   }
 660 
 661   // Setup the continuation addresses.
 662   *target_next = split_destination + partial_obj_size;
 663   HeapWord* const source_next = region_to_addr(split_region) + partial_obj_size;
 664 
 665   if (log_develop_is_enabled(Trace, gc, compaction)) {
 666     const char * split_type = partial_obj_size == 0 ? "easy" : "hard";
 667     log_develop_trace(gc, compaction)("%s split:  src=" PTR_FORMAT " src_c=" SIZE_FORMAT " pos=" SIZE_FORMAT,
 668                                       split_type, p2i(source_next), split_region, partial_obj_size);
 669     log_develop_trace(gc, compaction)("%s split:  dst=" PTR_FORMAT " dst_c=" SIZE_FORMAT " tn=" PTR_FORMAT,
 670                                       split_type, p2i(split_destination),
 671                                       addr_to_region_idx(split_destination),
 672                                       p2i(*target_next));
 673 
 674     if (partial_obj_size != 0) {
 675       HeapWord* const po_beg = split_info.destination();
 676       HeapWord* const po_end = po_beg + split_info.partial_obj_size();
 677       log_develop_trace(gc, compaction)("%s split:  po_beg=" PTR_FORMAT " " SIZE_FORMAT " po_end=" PTR_FORMAT " " SIZE_FORMAT,
 678                                         split_type,
 679                                         p2i(po_beg), addr_to_region_idx(po_beg),
 680                                         p2i(po_end), addr_to_region_idx(po_end));
 681     }
 682   }
 683 
 684   return source_next;
 685 }
 686 
 687 bool ParallelCompactData::summarize(SplitInfo& split_info,
 688                                     HeapWord* source_beg, HeapWord* source_end,
 689                                     HeapWord** source_next,
 690                                     HeapWord* target_beg, HeapWord* target_end,
 691                                     HeapWord** target_next)
 692 {
 693   HeapWord* const source_next_val = source_next == NULL ? NULL : *source_next;
 694   log_develop_trace(gc, compaction)(
 695       "sb=" PTR_FORMAT " se=" PTR_FORMAT " sn=" PTR_FORMAT
 696       "tb=" PTR_FORMAT " te=" PTR_FORMAT " tn=" PTR_FORMAT,
 697       p2i(source_beg), p2i(source_end), p2i(source_next_val),
 698       p2i(target_beg), p2i(target_end), p2i(*target_next));
 699 
 700   size_t cur_region = addr_to_region_idx(source_beg);
 701   const size_t end_region = addr_to_region_idx(region_align_up(source_end));
 702 
 703   HeapWord *dest_addr = target_beg;
 704   while (cur_region < end_region) {
 705     // The destination must be set even if the region has no data.
 706     _region_data[cur_region].set_destination(dest_addr);
 707 
 708     size_t words = _region_data[cur_region].data_size();
 709     if (words > 0) {
 710       // If cur_region does not fit entirely into the target space, find a point
 711       // at which the source space can be 'split' so that part is copied to the
 712       // target space and the rest is copied elsewhere.
 713       if (dest_addr + words > target_end) {
 714         assert(source_next != NULL, "source_next is NULL when splitting");
 715         *source_next = summarize_split_space(cur_region, split_info, dest_addr,
 716                                              target_end, target_next);
 717         return false;
 718       }
 719 
 720       // Compute the destination_count for cur_region, and if necessary, update
 721       // source_region for a destination region.  The source_region field is
 722       // updated if cur_region is the first (left-most) region to be copied to a
 723       // destination region.
 724       //
 725       // The destination_count calculation is a bit subtle.  A region that has
 726       // data that compacts into itself does not count itself as a destination.
 727       // This maintains the invariant that a zero count means the region is
 728       // available and can be claimed and then filled.
 729       uint destination_count = 0;
 730       if (split_info.is_split(cur_region)) {
 731         // The current region has been split:  the partial object will be copied
 732         // to one destination space and the remaining data will be copied to
 733         // another destination space.  Adjust the initial destination_count and,
 734         // if necessary, set the source_region field if the partial object will
 735         // cross a destination region boundary.
 736         destination_count = split_info.destination_count();
 737         if (destination_count == 2) {
 738           size_t dest_idx = addr_to_region_idx(split_info.dest_region_addr());
 739           _region_data[dest_idx].set_source_region(cur_region);
 740         }
 741       }
 742 
 743       HeapWord* const last_addr = dest_addr + words - 1;
 744       const size_t dest_region_1 = addr_to_region_idx(dest_addr);
 745       const size_t dest_region_2 = addr_to_region_idx(last_addr);
 746 
 747       // Initially assume that the destination regions will be the same and
 748       // adjust the value below if necessary.  Under this assumption, if
 749       // cur_region == dest_region_2, then cur_region will be compacted
 750       // completely into itself.
 751       destination_count += cur_region == dest_region_2 ? 0 : 1;
 752       if (dest_region_1 != dest_region_2) {
 753         // Destination regions differ; adjust destination_count.
 754         destination_count += 1;
 755         // Data from cur_region will be copied to the start of dest_region_2.
 756         _region_data[dest_region_2].set_source_region(cur_region);
 757       } else if (is_region_aligned(dest_addr)) {
 758         // Data from cur_region will be copied to the start of the destination
 759         // region.
 760         _region_data[dest_region_1].set_source_region(cur_region);
 761       }
 762 
 763       _region_data[cur_region].set_destination_count(destination_count);
 764       _region_data[cur_region].set_data_location(region_to_addr(cur_region));
 765       dest_addr += words;
 766     }
 767 
 768     ++cur_region;
 769   }
 770 
 771   *target_next = dest_addr;
 772   return true;
 773 }
 774 
 775 HeapWord* ParallelCompactData::calc_new_pointer(HeapWord* addr, ParCompactionManager* cm) const {
 776   assert(addr != NULL, "Should detect NULL oop earlier");
 777   assert(ParallelScavengeHeap::heap()->is_in(addr), "not in heap");
 778   assert(PSParallelCompact::mark_bitmap()->is_marked(addr), "not marked");
 779 
 780   // Region covering the object.
 781   RegionData* const region_ptr = addr_to_region_ptr(addr);
 782   HeapWord* result = region_ptr->destination();
 783 
 784   // If the entire Region is live, the new location is region->destination + the
 785   // offset of the object within in the Region.
 786 
 787   // Run some performance tests to determine if this special case pays off.  It
 788   // is worth it for pointers into the dense prefix.  If the optimization to
 789   // avoid pointer updates in regions that only point to the dense prefix is
 790   // ever implemented, this should be revisited.
 791   if (region_ptr->data_size() == RegionSize) {
 792     result += region_offset(addr);
 793     return result;
 794   }
 795 
 796   // Otherwise, the new location is region->destination + block offset + the
 797   // number of live words in the Block that are (a) to the left of addr and (b)
 798   // due to objects that start in the Block.
 799 
 800   // Fill in the block table if necessary.  This is unsynchronized, so multiple
 801   // threads may fill the block table for a region (harmless, since it is
 802   // idempotent).
 803   if (!region_ptr->blocks_filled()) {
 804     PSParallelCompact::fill_blocks(addr_to_region_idx(addr));
 805     region_ptr->set_blocks_filled();
 806   }
 807 
 808   HeapWord* const search_start = block_align_down(addr);
 809   const size_t block_offset = addr_to_block_ptr(addr)->offset();
 810 
 811   const ParMarkBitMap* bitmap = PSParallelCompact::mark_bitmap();
 812   const size_t live = bitmap->live_words_in_range(cm, search_start, cast_to_oop(addr));
 813   result += block_offset + live;
 814   DEBUG_ONLY(PSParallelCompact::check_new_location(addr, result));
 815   return result;
 816 }
 817 
 818 #ifdef ASSERT
 819 void ParallelCompactData::verify_clear(const PSVirtualSpace* vspace)
 820 {
 821   const size_t* const beg = (const size_t*)vspace->committed_low_addr();
 822   const size_t* const end = (const size_t*)vspace->committed_high_addr();
 823   for (const size_t* p = beg; p < end; ++p) {
 824     assert(*p == 0, "not zero");
 825   }
 826 }
 827 
 828 void ParallelCompactData::verify_clear()
 829 {
 830   verify_clear(_region_vspace);
 831   verify_clear(_block_vspace);
 832 }
 833 #endif  // #ifdef ASSERT
 834 
 835 STWGCTimer          PSParallelCompact::_gc_timer;
 836 ParallelOldTracer   PSParallelCompact::_gc_tracer;
 837 elapsedTimer        PSParallelCompact::_accumulated_time;
 838 unsigned int        PSParallelCompact::_total_invocations = 0;
 839 unsigned int        PSParallelCompact::_maximum_compaction_gc_num = 0;
 840 CollectorCounters*  PSParallelCompact::_counters = NULL;
 841 ParMarkBitMap       PSParallelCompact::_mark_bitmap;
 842 ParallelCompactData PSParallelCompact::_summary_data;
 843 
 844 PSParallelCompact::IsAliveClosure PSParallelCompact::_is_alive_closure;
 845 
 846 bool PSParallelCompact::IsAliveClosure::do_object_b(oop p) { return mark_bitmap()->is_marked(p); }
 847 
 848 class PCReferenceProcessor: public ReferenceProcessor {
 849 public:
 850   PCReferenceProcessor(
 851     BoolObjectClosure* is_subject_to_discovery,
 852     BoolObjectClosure* is_alive_non_header) :
 853       ReferenceProcessor(is_subject_to_discovery,
 854       ParallelGCThreads,   // mt processing degree
 855       ParallelGCThreads,   // mt discovery degree
 856       true,                // atomic_discovery
 857       is_alive_non_header) {
 858   }
 859 
 860   template<typename T> bool discover(oop obj, ReferenceType type) {
 861     T* referent_addr = (T*) java_lang_ref_Reference::referent_addr_raw(obj);
 862     T heap_oop = RawAccess<>::oop_load(referent_addr);
 863     oop referent = CompressedOops::decode_not_null(heap_oop);
 864     return PSParallelCompact::mark_bitmap()->is_unmarked(referent)
 865         && ReferenceProcessor::discover_reference(obj, type);
 866   }
 867   virtual bool discover_reference(oop obj, ReferenceType type) {
 868     if (UseCompressedOops) {
 869       return discover<narrowOop>(obj, type);
 870     } else {
 871       return discover<oop>(obj, type);
 872     }
 873   }
 874 };
 875 
 876 void PSParallelCompact::post_initialize() {
 877   ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
 878   _span_based_discoverer.set_span(heap->reserved_region());
 879   _ref_processor =
 880     new PCReferenceProcessor(&_span_based_discoverer,
 881                              &_is_alive_closure); // non-header is alive closure
 882 
 883   _counters = new CollectorCounters("Parallel full collection pauses", 1);
 884 
 885   // Initialize static fields in ParCompactionManager.
 886   ParCompactionManager::initialize(mark_bitmap());
 887 }
 888 
 889 bool PSParallelCompact::initialize() {
 890   ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
 891   MemRegion mr = heap->reserved_region();
 892 
 893   // Was the old gen get allocated successfully?
 894   if (!heap->old_gen()->is_allocated()) {
 895     return false;
 896   }
 897 
 898   initialize_space_info();
 899   initialize_dead_wood_limiter();
 900 
 901   if (!_mark_bitmap.initialize(mr)) {
 902     vm_shutdown_during_initialization(
 903       err_msg("Unable to allocate " SIZE_FORMAT "KB bitmaps for parallel "
 904       "garbage collection for the requested " SIZE_FORMAT "KB heap.",
 905       _mark_bitmap.reserved_byte_size()/K, mr.byte_size()/K));
 906     return false;
 907   }
 908 
 909   if (!_summary_data.initialize(mr)) {
 910     vm_shutdown_during_initialization(
 911       err_msg("Unable to allocate " SIZE_FORMAT "KB card tables for parallel "
 912       "garbage collection for the requested " SIZE_FORMAT "KB heap.",
 913       _summary_data.reserved_byte_size()/K, mr.byte_size()/K));
 914     return false;
 915   }
 916 
 917   return true;
 918 }
 919 
 920 void PSParallelCompact::initialize_space_info()
 921 {
 922   memset(&_space_info, 0, sizeof(_space_info));
 923 
 924   ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
 925   PSYoungGen* young_gen = heap->young_gen();
 926 
 927   _space_info[old_space_id].set_space(heap->old_gen()->object_space());
 928   _space_info[eden_space_id].set_space(young_gen->eden_space());
 929   _space_info[from_space_id].set_space(young_gen->from_space());
 930   _space_info[to_space_id].set_space(young_gen->to_space());
 931 
 932   _space_info[old_space_id].set_start_array(heap->old_gen()->start_array());
 933 }
 934 
 935 void PSParallelCompact::initialize_dead_wood_limiter()
 936 {
 937   const size_t max = 100;
 938   _dwl_mean = double(MIN2(ParallelOldDeadWoodLimiterMean, max)) / 100.0;
 939   _dwl_std_dev = double(MIN2(ParallelOldDeadWoodLimiterStdDev, max)) / 100.0;
 940   _dwl_first_term = 1.0 / (sqrt(2.0 * M_PI) * _dwl_std_dev);
 941   DEBUG_ONLY(_dwl_initialized = true;)
 942   _dwl_adjustment = normal_distribution(1.0);
 943 }
 944 
 945 void
 946 PSParallelCompact::clear_data_covering_space(SpaceId id)
 947 {
 948   // At this point, top is the value before GC, new_top() is the value that will
 949   // be set at the end of GC.  The marking bitmap is cleared to top; nothing
 950   // should be marked above top.  The summary data is cleared to the larger of
 951   // top & new_top.
 952   MutableSpace* const space = _space_info[id].space();
 953   HeapWord* const bot = space->bottom();
 954   HeapWord* const top = space->top();
 955   HeapWord* const max_top = MAX2(top, _space_info[id].new_top());
 956 
 957   const idx_t beg_bit = _mark_bitmap.addr_to_bit(bot);
 958   const idx_t end_bit = _mark_bitmap.align_range_end(_mark_bitmap.addr_to_bit(top));
 959   _mark_bitmap.clear_range(beg_bit, end_bit);
 960 
 961   const size_t beg_region = _summary_data.addr_to_region_idx(bot);
 962   const size_t end_region =
 963     _summary_data.addr_to_region_idx(_summary_data.region_align_up(max_top));
 964   _summary_data.clear_range(beg_region, end_region);
 965 
 966   // Clear the data used to 'split' regions.
 967   SplitInfo& split_info = _space_info[id].split_info();
 968   if (split_info.is_valid()) {
 969     split_info.clear();
 970   }
 971   DEBUG_ONLY(split_info.verify_clear();)
 972 }
 973 
 974 void PSParallelCompact::pre_compact()
 975 {
 976   // Update the from & to space pointers in space_info, since they are swapped
 977   // at each young gen gc.  Do the update unconditionally (even though a
 978   // promotion failure does not swap spaces) because an unknown number of young
 979   // collections will have swapped the spaces an unknown number of times.
 980   GCTraceTime(Debug, gc, phases) tm("Pre Compact", &_gc_timer);
 981   ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
 982   _space_info[from_space_id].set_space(heap->young_gen()->from_space());
 983   _space_info[to_space_id].set_space(heap->young_gen()->to_space());
 984 
 985   // Increment the invocation count
 986   heap->increment_total_collections(true);
 987 


 988   // We need to track unique mark sweep invocations as well.
 989   _total_invocations++;
 990 
 991   heap->print_heap_before_gc();
 992   heap->trace_heap_before_gc(&_gc_tracer);
 993 
 994   // Fill in TLABs
 995   heap->ensure_parsability(true);  // retire TLABs
 996 
 997   if (VerifyBeforeGC && heap->total_collections() >= VerifyGCStartAt) {
 998     Universe::verify("Before GC");
 999   }
1000 
1001   // Verify object start arrays
1002   if (VerifyObjectStartArray &&
1003       VerifyBeforeGC) {
1004     heap->old_gen()->verify_object_start_array();
1005   }
1006 
1007   DEBUG_ONLY(mark_bitmap()->verify_clear();)
1008   DEBUG_ONLY(summary_data().verify_clear();)
1009 
1010   ParCompactionManager::reset_all_bitmap_query_caches();
1011 }
1012 
1013 void PSParallelCompact::post_compact()
1014 {
1015   GCTraceTime(Info, gc, phases) tm("Post Compact", &_gc_timer);
1016   ParCompactionManager::remove_all_shadow_regions();
1017 


1018   for (unsigned int id = old_space_id; id < last_space_id; ++id) {
1019     // Clear the marking bitmap, summary data and split info.
1020     clear_data_covering_space(SpaceId(id));
1021     // Update top().  Must be done after clearing the bitmap and summary data.
1022     _space_info[id].publish_new_top();
1023   }
1024 
1025   ParCompactionManager::flush_all_string_dedup_requests();
1026 
1027   MutableSpace* const eden_space = _space_info[eden_space_id].space();
1028   MutableSpace* const from_space = _space_info[from_space_id].space();
1029   MutableSpace* const to_space   = _space_info[to_space_id].space();
1030 
1031   ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
1032   bool eden_empty = eden_space->is_empty();
1033 
1034   // Update heap occupancy information which is used as input to the soft ref
1035   // clearing policy at the next gc.
1036   Universe::heap()->update_capacity_and_used_at_gc();
1037 
1038   bool young_gen_empty = eden_empty && from_space->is_empty() &&
1039     to_space->is_empty();
1040 
1041   PSCardTable* ct = heap->card_table();
1042   MemRegion old_mr = heap->old_gen()->reserved();
1043   if (young_gen_empty) {
1044     ct->clear(MemRegion(old_mr.start(), old_mr.end()));
1045   } else {
1046     ct->invalidate(MemRegion(old_mr.start(), old_mr.end()));
1047   }
1048 
1049   // Delete metaspaces for unloaded class loaders and clean up loader_data graph
1050   ClassLoaderDataGraph::purge(/*at_safepoint*/true);
1051   DEBUG_ONLY(MetaspaceUtils::verify();)
1052 
1053   heap->prune_scavengable_nmethods();
1054 
1055 #if COMPILER2_OR_JVMCI
1056   DerivedPointerTable::update_pointers();
1057 #endif
1058 
1059   if (ZapUnusedHeapArea) {
1060     heap->gen_mangle_unused_area();
1061   }
1062 
1063   // Signal that we have completed a visit to all live objects.
1064   Universe::heap()->record_whole_heap_examined_timestamp();
1065 }
1066 
1067 HeapWord*
1068 PSParallelCompact::compute_dense_prefix_via_density(const SpaceId id,
1069                                                     bool maximum_compaction)
1070 {
1071   const size_t region_size = ParallelCompactData::RegionSize;
1072   const ParallelCompactData& sd = summary_data();
1073 
1074   const MutableSpace* const space = _space_info[id].space();
1075   HeapWord* const top_aligned_up = sd.region_align_up(space->top());
1076   const RegionData* const beg_cp = sd.addr_to_region_ptr(space->bottom());
1077   const RegionData* const end_cp = sd.addr_to_region_ptr(top_aligned_up);
1078 
1079   // Skip full regions at the beginning of the space--they are necessarily part
1080   // of the dense prefix.
1081   size_t full_count = 0;
1082   const RegionData* cp;
1083   for (cp = beg_cp; cp < end_cp && cp->data_size() == region_size; ++cp) {
1084     ++full_count;
1085   }
1086 
1087   assert(total_invocations() >= _maximum_compaction_gc_num, "sanity");
1088   const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num;
1089   const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval;
1090   if (maximum_compaction || cp == end_cp || interval_ended) {
1091     _maximum_compaction_gc_num = total_invocations();
1092     return sd.region_to_addr(cp);
1093   }
1094 
1095   HeapWord* const new_top = _space_info[id].new_top();
1096   const size_t space_live = pointer_delta(new_top, space->bottom());
1097   const size_t space_used = space->used_in_words();
1098   const size_t space_capacity = space->capacity_in_words();
1099 
1100   const double cur_density = double(space_live) / space_capacity;
1101   const double deadwood_density =
1102     (1.0 - cur_density) * (1.0 - cur_density) * cur_density * cur_density;
1103   const size_t deadwood_goal = size_t(space_capacity * deadwood_density);
1104 
1105   log_develop_debug(gc, compaction)(
1106       "cur_dens=%5.3f dw_dens=%5.3f dw_goal=" SIZE_FORMAT,
1107       cur_density, deadwood_density, deadwood_goal);
1108   log_develop_debug(gc, compaction)(
1109       "space_live=" SIZE_FORMAT " space_used=" SIZE_FORMAT " "
1110       "space_cap=" SIZE_FORMAT,
1111       space_live, space_used,
1112       space_capacity);
1113 
1114   // XXX - Use binary search?
1115   HeapWord* dense_prefix = sd.region_to_addr(cp);
1116   const RegionData* full_cp = cp;
1117   const RegionData* const top_cp = sd.addr_to_region_ptr(space->top() - 1);
1118   while (cp < end_cp) {
1119     HeapWord* region_destination = cp->destination();
1120     const size_t cur_deadwood = pointer_delta(dense_prefix, region_destination);
1121 
1122     log_develop_trace(gc, compaction)(
1123         "c#=" SIZE_FORMAT_W(4) " dst=" PTR_FORMAT " "
1124         "dp=" PTR_FORMAT " cdw=" SIZE_FORMAT_W(8),
1125         sd.region(cp), p2i(region_destination),
1126         p2i(dense_prefix), cur_deadwood);
1127 
1128     if (cur_deadwood >= deadwood_goal) {
1129       // Found the region that has the correct amount of deadwood to the left.
1130       // This typically occurs after crossing a fairly sparse set of regions, so
1131       // iterate backwards over those sparse regions, looking for the region
1132       // that has the lowest density of live objects 'to the right.'
1133       size_t space_to_left = sd.region(cp) * region_size;
1134       size_t live_to_left = space_to_left - cur_deadwood;
1135       size_t space_to_right = space_capacity - space_to_left;
1136       size_t live_to_right = space_live - live_to_left;
1137       double density_to_right = double(live_to_right) / space_to_right;
1138       while (cp > full_cp) {
1139         --cp;
1140         const size_t prev_region_live_to_right = live_to_right -
1141           cp->data_size();
1142         const size_t prev_region_space_to_right = space_to_right + region_size;
1143         double prev_region_density_to_right =
1144           double(prev_region_live_to_right) / prev_region_space_to_right;
1145         if (density_to_right <= prev_region_density_to_right) {
1146           return dense_prefix;
1147         }
1148 
1149         log_develop_trace(gc, compaction)(
1150             "backing up from c=" SIZE_FORMAT_W(4) " d2r=%10.8f "
1151             "pc_d2r=%10.8f",
1152             sd.region(cp), density_to_right,
1153             prev_region_density_to_right);
1154 
1155         dense_prefix -= region_size;
1156         live_to_right = prev_region_live_to_right;
1157         space_to_right = prev_region_space_to_right;
1158         density_to_right = prev_region_density_to_right;
1159       }
1160       return dense_prefix;
1161     }
1162 
1163     dense_prefix += region_size;
1164     ++cp;
1165   }
1166 
1167   return dense_prefix;
1168 }
1169 
1170 #ifndef PRODUCT
1171 void PSParallelCompact::print_dense_prefix_stats(const char* const algorithm,
1172                                                  const SpaceId id,
1173                                                  const bool maximum_compaction,
1174                                                  HeapWord* const addr)
1175 {
1176   const size_t region_idx = summary_data().addr_to_region_idx(addr);
1177   RegionData* const cp = summary_data().region(region_idx);
1178   const MutableSpace* const space = _space_info[id].space();
1179   HeapWord* const new_top = _space_info[id].new_top();
1180 
1181   const size_t space_live = pointer_delta(new_top, space->bottom());
1182   const size_t dead_to_left = pointer_delta(addr, cp->destination());
1183   const size_t space_cap = space->capacity_in_words();
1184   const double dead_to_left_pct = double(dead_to_left) / space_cap;
1185   const size_t live_to_right = new_top - cp->destination();
1186   const size_t dead_to_right = space->top() - addr - live_to_right;
1187 
1188   log_develop_debug(gc, compaction)(
1189       "%s=" PTR_FORMAT " dpc=" SIZE_FORMAT_W(5) " "
1190       "spl=" SIZE_FORMAT " "
1191       "d2l=" SIZE_FORMAT " d2l%%=%6.4f "
1192       "d2r=" SIZE_FORMAT " l2r=" SIZE_FORMAT " "
1193       "ratio=%10.8f",
1194       algorithm, p2i(addr), region_idx,
1195       space_live,
1196       dead_to_left, dead_to_left_pct,
1197       dead_to_right, live_to_right,
1198       double(dead_to_right) / live_to_right);
1199 }
1200 #endif  // #ifndef PRODUCT
1201 
1202 // Return a fraction indicating how much of the generation can be treated as
1203 // "dead wood" (i.e., not reclaimed).  The function uses a normal distribution
1204 // based on the density of live objects in the generation to determine a limit,
1205 // which is then adjusted so the return value is min_percent when the density is
1206 // 1.
1207 //
1208 // The following table shows some return values for a different values of the
1209 // standard deviation (ParallelOldDeadWoodLimiterStdDev); the mean is 0.5 and
1210 // min_percent is 1.
1211 //
1212 //                          fraction allowed as dead wood
1213 //         -----------------------------------------------------------------
1214 // density std_dev=70 std_dev=75 std_dev=80 std_dev=85 std_dev=90 std_dev=95
1215 // ------- ---------- ---------- ---------- ---------- ---------- ----------
1216 // 0.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000
1217 // 0.05000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941
1218 // 0.10000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272
1219 // 0.15000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066
1220 // 0.20000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975
1221 // 0.25000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313
1222 // 0.30000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132
1223 // 0.35000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289
1224 // 0.40000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500
1225 // 0.45000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386
1226 // 0.50000 0.13832410 0.11599237 0.09847664 0.08456518 0.07338887 0.06431510
1227 // 0.55000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386
1228 // 0.60000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500
1229 // 0.65000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289
1230 // 0.70000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132
1231 // 0.75000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313
1232 // 0.80000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975
1233 // 0.85000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066
1234 // 0.90000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272
1235 // 0.95000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941
1236 // 1.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000
1237 
1238 double PSParallelCompact::dead_wood_limiter(double density, size_t min_percent)
1239 {
1240   assert(_dwl_initialized, "uninitialized");
1241 
1242   // The raw limit is the value of the normal distribution at x = density.
1243   const double raw_limit = normal_distribution(density);
1244 
1245   // Adjust the raw limit so it becomes the minimum when the density is 1.
1246   //
1247   // First subtract the adjustment value (which is simply the precomputed value
1248   // normal_distribution(1.0)); this yields a value of 0 when the density is 1.
1249   // Then add the minimum value, so the minimum is returned when the density is
1250   // 1.  Finally, prevent negative values, which occur when the mean is not 0.5.
1251   const double min = double(min_percent) / 100.0;
1252   const double limit = raw_limit - _dwl_adjustment + min;
1253   return MAX2(limit, 0.0);
1254 }
1255 
1256 ParallelCompactData::RegionData*
1257 PSParallelCompact::first_dead_space_region(const RegionData* beg,
1258                                            const RegionData* end)
1259 {
1260   const size_t region_size = ParallelCompactData::RegionSize;
1261   ParallelCompactData& sd = summary_data();
1262   size_t left = sd.region(beg);
1263   size_t right = end > beg ? sd.region(end) - 1 : left;
1264 
1265   // Binary search.
1266   while (left < right) {
1267     // Equivalent to (left + right) / 2, but does not overflow.
1268     const size_t middle = left + (right - left) / 2;
1269     RegionData* const middle_ptr = sd.region(middle);
1270     HeapWord* const dest = middle_ptr->destination();
1271     HeapWord* const addr = sd.region_to_addr(middle);
1272     assert(dest != NULL, "sanity");
1273     assert(dest <= addr, "must move left");
1274 
1275     if (middle > left && dest < addr) {
1276       right = middle - 1;
1277     } else if (middle < right && middle_ptr->data_size() == region_size) {
1278       left = middle + 1;
1279     } else {
1280       return middle_ptr;
1281     }
1282   }
1283   return sd.region(left);
1284 }
1285 
1286 ParallelCompactData::RegionData*
1287 PSParallelCompact::dead_wood_limit_region(const RegionData* beg,
1288                                           const RegionData* end,
1289                                           size_t dead_words)
1290 {
1291   ParallelCompactData& sd = summary_data();
1292   size_t left = sd.region(beg);
1293   size_t right = end > beg ? sd.region(end) - 1 : left;
1294 
1295   // Binary search.
1296   while (left < right) {
1297     // Equivalent to (left + right) / 2, but does not overflow.
1298     const size_t middle = left + (right - left) / 2;
1299     RegionData* const middle_ptr = sd.region(middle);
1300     HeapWord* const dest = middle_ptr->destination();
1301     HeapWord* const addr = sd.region_to_addr(middle);
1302     assert(dest != NULL, "sanity");
1303     assert(dest <= addr, "must move left");
1304 
1305     const size_t dead_to_left = pointer_delta(addr, dest);
1306     if (middle > left && dead_to_left > dead_words) {
1307       right = middle - 1;
1308     } else if (middle < right && dead_to_left < dead_words) {
1309       left = middle + 1;
1310     } else {
1311       return middle_ptr;
1312     }
1313   }
1314   return sd.region(left);
1315 }
1316 
1317 // The result is valid during the summary phase, after the initial summarization
1318 // of each space into itself, and before final summarization.
1319 inline double
1320 PSParallelCompact::reclaimed_ratio(const RegionData* const cp,
1321                                    HeapWord* const bottom,
1322                                    HeapWord* const top,
1323                                    HeapWord* const new_top)
1324 {
1325   ParallelCompactData& sd = summary_data();
1326 
1327   assert(cp != NULL, "sanity");
1328   assert(bottom != NULL, "sanity");
1329   assert(top != NULL, "sanity");
1330   assert(new_top != NULL, "sanity");
1331   assert(top >= new_top, "summary data problem?");
1332   assert(new_top > bottom, "space is empty; should not be here");
1333   assert(new_top >= cp->destination(), "sanity");
1334   assert(top >= sd.region_to_addr(cp), "sanity");
1335 
1336   HeapWord* const destination = cp->destination();
1337   const size_t dense_prefix_live  = pointer_delta(destination, bottom);
1338   const size_t compacted_region_live = pointer_delta(new_top, destination);
1339   const size_t compacted_region_used = pointer_delta(top,
1340                                                      sd.region_to_addr(cp));
1341   const size_t reclaimable = compacted_region_used - compacted_region_live;
1342 
1343   const double divisor = dense_prefix_live + 1.25 * compacted_region_live;
1344   return double(reclaimable) / divisor;
1345 }
1346 
1347 // Return the address of the end of the dense prefix, a.k.a. the start of the
1348 // compacted region.  The address is always on a region boundary.
1349 //
1350 // Completely full regions at the left are skipped, since no compaction can
1351 // occur in those regions.  Then the maximum amount of dead wood to allow is
1352 // computed, based on the density (amount live / capacity) of the generation;
1353 // the region with approximately that amount of dead space to the left is
1354 // identified as the limit region.  Regions between the last completely full
1355 // region and the limit region are scanned and the one that has the best
1356 // (maximum) reclaimed_ratio() is selected.
1357 HeapWord*
1358 PSParallelCompact::compute_dense_prefix(const SpaceId id,
1359                                         bool maximum_compaction)
1360 {
1361   const size_t region_size = ParallelCompactData::RegionSize;
1362   const ParallelCompactData& sd = summary_data();
1363 
1364   const MutableSpace* const space = _space_info[id].space();
1365   HeapWord* const top = space->top();
1366   HeapWord* const top_aligned_up = sd.region_align_up(top);
1367   HeapWord* const new_top = _space_info[id].new_top();
1368   HeapWord* const new_top_aligned_up = sd.region_align_up(new_top);
1369   HeapWord* const bottom = space->bottom();
1370   const RegionData* const beg_cp = sd.addr_to_region_ptr(bottom);
1371   const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up);
1372   const RegionData* const new_top_cp =
1373     sd.addr_to_region_ptr(new_top_aligned_up);
1374 
1375   // Skip full regions at the beginning of the space--they are necessarily part
1376   // of the dense prefix.
1377   const RegionData* const full_cp = first_dead_space_region(beg_cp, new_top_cp);
1378   assert(full_cp->destination() == sd.region_to_addr(full_cp) ||
1379          space->is_empty(), "no dead space allowed to the left");
1380   assert(full_cp->data_size() < region_size || full_cp == new_top_cp - 1,
1381          "region must have dead space");
1382 
1383   // The gc number is saved whenever a maximum compaction is done, and used to
1384   // determine when the maximum compaction interval has expired.  This avoids
1385   // successive max compactions for different reasons.
1386   assert(total_invocations() >= _maximum_compaction_gc_num, "sanity");
1387   const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num;
1388   const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval ||
1389     total_invocations() == HeapFirstMaximumCompactionCount;
1390   if (maximum_compaction || full_cp == top_cp || interval_ended) {
1391     _maximum_compaction_gc_num = total_invocations();
1392     return sd.region_to_addr(full_cp);
1393   }
1394 
1395   const size_t space_live = pointer_delta(new_top, bottom);
1396   const size_t space_used = space->used_in_words();
1397   const size_t space_capacity = space->capacity_in_words();
1398 
1399   const double density = double(space_live) / double(space_capacity);
1400   const size_t min_percent_free = MarkSweepDeadRatio;
1401   const double limiter = dead_wood_limiter(density, min_percent_free);
1402   const size_t dead_wood_max = space_used - space_live;
1403   const size_t dead_wood_limit = MIN2(size_t(space_capacity * limiter),
1404                                       dead_wood_max);
1405 
1406   log_develop_debug(gc, compaction)(
1407       "space_live=" SIZE_FORMAT " space_used=" SIZE_FORMAT " "
1408       "space_cap=" SIZE_FORMAT,
1409       space_live, space_used,
1410       space_capacity);
1411   log_develop_debug(gc, compaction)(
1412       "dead_wood_limiter(%6.4f, " SIZE_FORMAT ")=%6.4f "
1413       "dead_wood_max=" SIZE_FORMAT " dead_wood_limit=" SIZE_FORMAT,
1414       density, min_percent_free, limiter,
1415       dead_wood_max, dead_wood_limit);
1416 
1417   // Locate the region with the desired amount of dead space to the left.
1418   const RegionData* const limit_cp =
1419     dead_wood_limit_region(full_cp, top_cp, dead_wood_limit);
1420 
1421   // Scan from the first region with dead space to the limit region and find the
1422   // one with the best (largest) reclaimed ratio.
1423   double best_ratio = 0.0;
1424   const RegionData* best_cp = full_cp;
1425   for (const RegionData* cp = full_cp; cp < limit_cp; ++cp) {
1426     double tmp_ratio = reclaimed_ratio(cp, bottom, top, new_top);
1427     if (tmp_ratio > best_ratio) {
1428       best_cp = cp;
1429       best_ratio = tmp_ratio;
1430     }
1431   }
1432 
1433   return sd.region_to_addr(best_cp);
1434 }
1435 
1436 void PSParallelCompact::summarize_spaces_quick()
1437 {
1438   for (unsigned int i = 0; i < last_space_id; ++i) {
1439     const MutableSpace* space = _space_info[i].space();
1440     HeapWord** nta = _space_info[i].new_top_addr();
1441     bool result = _summary_data.summarize(_space_info[i].split_info(),
1442                                           space->bottom(), space->top(), NULL,
1443                                           space->bottom(), space->end(), nta);
1444     assert(result, "space must fit into itself");
1445     _space_info[i].set_dense_prefix(space->bottom());
1446   }
1447 }
1448 
1449 void PSParallelCompact::fill_dense_prefix_end(SpaceId id)
1450 {
1451   HeapWord* const dense_prefix_end = dense_prefix(id);
1452   const RegionData* region = _summary_data.addr_to_region_ptr(dense_prefix_end);
1453   const idx_t dense_prefix_bit = _mark_bitmap.addr_to_bit(dense_prefix_end);
1454   if (dead_space_crosses_boundary(region, dense_prefix_bit)) {
1455     // Only enough dead space is filled so that any remaining dead space to the
1456     // left is larger than the minimum filler object.  (The remainder is filled
1457     // during the copy/update phase.)
1458     //
1459     // The size of the dead space to the right of the boundary is not a
1460     // concern, since compaction will be able to use whatever space is
1461     // available.
1462     //
1463     // Here '||' is the boundary, 'x' represents a don't care bit and a box
1464     // surrounds the space to be filled with an object.
1465     //
1466     // In the 32-bit VM, each bit represents two 32-bit words:
1467     //                              +---+
1468     // a) beg_bits:  ...  x   x   x | 0 | ||   0   x  x  ...
1469     //    end_bits:  ...  x   x   x | 0 | ||   0   x  x  ...
1470     //                              +---+
1471     //
1472     // In the 64-bit VM, each bit represents one 64-bit word:
1473     //                              +------------+
1474     // b) beg_bits:  ...  x   x   x | 0   ||   0 | x  x  ...
1475     //    end_bits:  ...  x   x   1 | 0   ||   0 | x  x  ...
1476     //                              +------------+
1477     //                          +-------+
1478     // c) beg_bits:  ...  x   x | 0   0 | ||   0   x  x  ...
1479     //    end_bits:  ...  x   1 | 0   0 | ||   0   x  x  ...
1480     //                          +-------+
1481     //                      +-----------+
1482     // d) beg_bits:  ...  x | 0   0   0 | ||   0   x  x  ...
1483     //    end_bits:  ...  1 | 0   0   0 | ||   0   x  x  ...
1484     //                      +-----------+
1485     //                          +-------+
1486     // e) beg_bits:  ...  0   0 | 0   0 | ||   0   x  x  ...
1487     //    end_bits:  ...  0   0 | 0   0 | ||   0   x  x  ...
1488     //                          +-------+
1489 
1490     // Initially assume case a, c or e will apply.
1491     size_t obj_len = CollectedHeap::min_fill_size();
1492     HeapWord* obj_beg = dense_prefix_end - obj_len;
1493 
1494 #ifdef  _LP64
1495     if (MinObjAlignment > 1) { // object alignment > heap word size
1496       // Cases a, c or e.
1497     } else if (_mark_bitmap.is_obj_end(dense_prefix_bit - 2)) {
1498       // Case b above.
1499       obj_beg = dense_prefix_end - 1;
1500     } else if (!_mark_bitmap.is_obj_end(dense_prefix_bit - 3) &&
1501                _mark_bitmap.is_obj_end(dense_prefix_bit - 4)) {
1502       // Case d above.
1503       obj_beg = dense_prefix_end - 3;
1504       obj_len = 3;
1505     }
1506 #endif  // #ifdef _LP64
1507 
1508     CollectedHeap::fill_with_object(obj_beg, obj_len);
1509     _mark_bitmap.mark_obj(obj_beg, obj_len);
1510     _summary_data.add_obj(obj_beg, obj_len);
1511     assert(start_array(id) != NULL, "sanity");
1512     start_array(id)->allocate_block(obj_beg);
1513   }
1514 }
1515 
1516 void
1517 PSParallelCompact::summarize_space(SpaceId id, bool maximum_compaction)
1518 {
1519   assert(id < last_space_id, "id out of range");
1520   assert(_space_info[id].dense_prefix() == _space_info[id].space()->bottom(),
1521          "should have been reset in summarize_spaces_quick()");
1522 
1523   const MutableSpace* space = _space_info[id].space();
1524   if (_space_info[id].new_top() != space->bottom()) {
1525     HeapWord* dense_prefix_end = compute_dense_prefix(id, maximum_compaction);
1526     _space_info[id].set_dense_prefix(dense_prefix_end);
1527 
1528 #ifndef PRODUCT
1529     if (log_is_enabled(Debug, gc, compaction)) {
1530       print_dense_prefix_stats("ratio", id, maximum_compaction,
1531                                dense_prefix_end);
1532       HeapWord* addr = compute_dense_prefix_via_density(id, maximum_compaction);
1533       print_dense_prefix_stats("density", id, maximum_compaction, addr);
1534     }
1535 #endif  // #ifndef PRODUCT
1536 
1537     // Recompute the summary data, taking into account the dense prefix.  If
1538     // every last byte will be reclaimed, then the existing summary data which
1539     // compacts everything can be left in place.
1540     if (!maximum_compaction && dense_prefix_end != space->bottom()) {
1541       // If dead space crosses the dense prefix boundary, it is (at least
1542       // partially) filled with a dummy object, marked live and added to the
1543       // summary data.  This simplifies the copy/update phase and must be done
1544       // before the final locations of objects are determined, to prevent
1545       // leaving a fragment of dead space that is too small to fill.
1546       fill_dense_prefix_end(id);
1547 
1548       // Compute the destination of each Region, and thus each object.
1549       _summary_data.summarize_dense_prefix(space->bottom(), dense_prefix_end);
1550       _summary_data.summarize(_space_info[id].split_info(),
1551                               dense_prefix_end, space->top(), NULL,
1552                               dense_prefix_end, space->end(),
1553                               _space_info[id].new_top_addr());
1554     }
1555   }
1556 
1557   if (log_develop_is_enabled(Trace, gc, compaction)) {
1558     const size_t region_size = ParallelCompactData::RegionSize;
1559     HeapWord* const dense_prefix_end = _space_info[id].dense_prefix();
1560     const size_t dp_region = _summary_data.addr_to_region_idx(dense_prefix_end);
1561     const size_t dp_words = pointer_delta(dense_prefix_end, space->bottom());
1562     HeapWord* const new_top = _space_info[id].new_top();
1563     const HeapWord* nt_aligned_up = _summary_data.region_align_up(new_top);
1564     const size_t cr_words = pointer_delta(nt_aligned_up, dense_prefix_end);
1565     log_develop_trace(gc, compaction)(
1566         "id=%d cap=" SIZE_FORMAT " dp=" PTR_FORMAT " "
1567         "dp_region=" SIZE_FORMAT " " "dp_count=" SIZE_FORMAT " "
1568         "cr_count=" SIZE_FORMAT " " "nt=" PTR_FORMAT,
1569         id, space->capacity_in_words(), p2i(dense_prefix_end),
1570         dp_region, dp_words / region_size,
1571         cr_words / region_size, p2i(new_top));
1572   }
1573 }
1574 
1575 #ifndef PRODUCT
1576 void PSParallelCompact::summary_phase_msg(SpaceId dst_space_id,
1577                                           HeapWord* dst_beg, HeapWord* dst_end,
1578                                           SpaceId src_space_id,
1579                                           HeapWord* src_beg, HeapWord* src_end)
1580 {
1581   log_develop_trace(gc, compaction)(
1582       "Summarizing %d [%s] into %d [%s]:  "
1583       "src=" PTR_FORMAT "-" PTR_FORMAT " "
1584       SIZE_FORMAT "-" SIZE_FORMAT " "
1585       "dst=" PTR_FORMAT "-" PTR_FORMAT " "
1586       SIZE_FORMAT "-" SIZE_FORMAT,
1587       src_space_id, space_names[src_space_id],
1588       dst_space_id, space_names[dst_space_id],
1589       p2i(src_beg), p2i(src_end),
1590       _summary_data.addr_to_region_idx(src_beg),
1591       _summary_data.addr_to_region_idx(src_end),
1592       p2i(dst_beg), p2i(dst_end),
1593       _summary_data.addr_to_region_idx(dst_beg),
1594       _summary_data.addr_to_region_idx(dst_end));
1595 }
1596 #endif  // #ifndef PRODUCT
1597 
1598 void PSParallelCompact::summary_phase(ParCompactionManager* cm,
1599                                       bool maximum_compaction)
1600 {
1601   GCTraceTime(Info, gc, phases) tm("Summary Phase", &_gc_timer);
1602 
1603   // Quick summarization of each space into itself, to see how much is live.
1604   summarize_spaces_quick();
1605 
1606   log_develop_trace(gc, compaction)("summary phase:  after summarizing each space to self");
1607   NOT_PRODUCT(print_region_ranges());
1608   NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info));
1609 
1610   // The amount of live data that will end up in old space (assuming it fits).
1611   size_t old_space_total_live = 0;
1612   for (unsigned int id = old_space_id; id < last_space_id; ++id) {
1613     old_space_total_live += pointer_delta(_space_info[id].new_top(),
1614                                           _space_info[id].space()->bottom());
1615   }
1616 
1617   MutableSpace* const old_space = _space_info[old_space_id].space();
1618   const size_t old_capacity = old_space->capacity_in_words();
1619   if (old_space_total_live > old_capacity) {
1620     // XXX - should also try to expand
1621     maximum_compaction = true;
1622   }
1623 
1624   // Old generations.
1625   summarize_space(old_space_id, maximum_compaction);
1626 
1627   // Summarize the remaining spaces in the young gen.  The initial target space
1628   // is the old gen.  If a space does not fit entirely into the target, then the
1629   // remainder is compacted into the space itself and that space becomes the new
1630   // target.
1631   SpaceId dst_space_id = old_space_id;
1632   HeapWord* dst_space_end = old_space->end();
1633   HeapWord** new_top_addr = _space_info[dst_space_id].new_top_addr();
1634   for (unsigned int id = eden_space_id; id < last_space_id; ++id) {
1635     const MutableSpace* space = _space_info[id].space();
1636     const size_t live = pointer_delta(_space_info[id].new_top(),
1637                                       space->bottom());
1638     const size_t available = pointer_delta(dst_space_end, *new_top_addr);
1639 
1640     NOT_PRODUCT(summary_phase_msg(dst_space_id, *new_top_addr, dst_space_end,
1641                                   SpaceId(id), space->bottom(), space->top());)
1642     if (live > 0 && live <= available) {
1643       // All the live data will fit.
1644       bool done = _summary_data.summarize(_space_info[id].split_info(),
1645                                           space->bottom(), space->top(),
1646                                           NULL,
1647                                           *new_top_addr, dst_space_end,
1648                                           new_top_addr);
1649       assert(done, "space must fit into old gen");
1650 
1651       // Reset the new_top value for the space.
1652       _space_info[id].set_new_top(space->bottom());
1653     } else if (live > 0) {
1654       // Attempt to fit part of the source space into the target space.
1655       HeapWord* next_src_addr = NULL;
1656       bool done = _summary_data.summarize(_space_info[id].split_info(),
1657                                           space->bottom(), space->top(),
1658                                           &next_src_addr,
1659                                           *new_top_addr, dst_space_end,
1660                                           new_top_addr);
1661       assert(!done, "space should not fit into old gen");
1662       assert(next_src_addr != NULL, "sanity");
1663 
1664       // The source space becomes the new target, so the remainder is compacted
1665       // within the space itself.
1666       dst_space_id = SpaceId(id);
1667       dst_space_end = space->end();
1668       new_top_addr = _space_info[id].new_top_addr();
1669       NOT_PRODUCT(summary_phase_msg(dst_space_id,
1670                                     space->bottom(), dst_space_end,
1671                                     SpaceId(id), next_src_addr, space->top());)
1672       done = _summary_data.summarize(_space_info[id].split_info(),
1673                                      next_src_addr, space->top(),
1674                                      NULL,
1675                                      space->bottom(), dst_space_end,
1676                                      new_top_addr);
1677       assert(done, "space must fit when compacted into itself");
1678       assert(*new_top_addr <= space->top(), "usage should not grow");
1679     }
1680   }
1681 
1682   log_develop_trace(gc, compaction)("Summary_phase:  after final summarization");
1683   NOT_PRODUCT(print_region_ranges());
1684   NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info));
1685 }
1686 
1687 // This method should contain all heap-specific policy for invoking a full
1688 // collection.  invoke_no_policy() will only attempt to compact the heap; it
1689 // will do nothing further.  If we need to bail out for policy reasons, scavenge
1690 // before full gc, or any other specialized behavior, it needs to be added here.
1691 //
1692 // Note that this method should only be called from the vm_thread while at a
1693 // safepoint.
1694 //
1695 // Note that the all_soft_refs_clear flag in the soft ref policy
1696 // may be true because this method can be called without intervening
1697 // activity.  For example when the heap space is tight and full measure
1698 // are being taken to free space.
1699 void PSParallelCompact::invoke(bool maximum_heap_compaction) {
1700   assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
1701   assert(Thread::current() == (Thread*)VMThread::vm_thread(),
1702          "should be in vm thread");
1703 
1704   ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
1705   GCCause::Cause gc_cause = heap->gc_cause();
1706   assert(!heap->is_gc_active(), "not reentrant");
1707 
1708   PSAdaptiveSizePolicy* policy = heap->size_policy();
1709   IsGCActiveMark mark;
1710 
1711   if (ScavengeBeforeFullGC) {
1712     PSScavenge::invoke_no_policy();
1713   }
1714 
1715   const bool clear_all_soft_refs =
1716     heap->soft_ref_policy()->should_clear_all_soft_refs();
1717 
1718   PSParallelCompact::invoke_no_policy(clear_all_soft_refs ||
1719                                       maximum_heap_compaction);
1720 }
1721 
1722 // This method contains no policy. You should probably
1723 // be calling invoke() instead.
1724 bool PSParallelCompact::invoke_no_policy(bool maximum_heap_compaction) {
1725   assert(SafepointSynchronize::is_at_safepoint(), "must be at a safepoint");
1726   assert(ref_processor() != NULL, "Sanity");
1727 
1728   if (GCLocker::check_active_before_gc()) {
1729     return false;
1730   }
1731 
1732   ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
1733 
1734   GCIdMark gc_id_mark;
1735   _gc_timer.register_gc_start();
1736   _gc_tracer.report_gc_start(heap->gc_cause(), _gc_timer.gc_start());
1737 
1738   TimeStamp marking_start;
1739   TimeStamp compaction_start;
1740   TimeStamp collection_exit;
1741 
1742   GCCause::Cause gc_cause = heap->gc_cause();
1743   PSYoungGen* young_gen = heap->young_gen();
1744   PSOldGen* old_gen = heap->old_gen();
1745   PSAdaptiveSizePolicy* size_policy = heap->size_policy();
1746 
1747   // The scope of casr should end after code that can change
1748   // SoftRefPolicy::_should_clear_all_soft_refs.
1749   ClearedAllSoftRefs casr(maximum_heap_compaction,
1750                           heap->soft_ref_policy());
1751 
1752   if (ZapUnusedHeapArea) {
1753     // Save information needed to minimize mangling
1754     heap->record_gen_tops_before_GC();
1755   }
1756 
1757   // Make sure data structures are sane, make the heap parsable, and do other
1758   // miscellaneous bookkeeping.
1759   pre_compact();
1760 
1761   const PreGenGCValues pre_gc_values = heap->get_pre_gc_values();
1762 
1763   // Get the compaction manager reserved for the VM thread.
1764   ParCompactionManager* const vmthread_cm = ParCompactionManager::get_vmthread_cm();
1765 
1766   {
1767     const uint active_workers =
1768       WorkerPolicy::calc_active_workers(ParallelScavengeHeap::heap()->workers().max_workers(),
1769                                         ParallelScavengeHeap::heap()->workers().active_workers(),
1770                                         Threads::number_of_non_daemon_threads());
1771     ParallelScavengeHeap::heap()->workers().set_active_workers(active_workers);
1772 
1773     GCTraceCPUTime tcpu;
1774     GCTraceTime(Info, gc) tm("Pause Full", NULL, gc_cause, true);
1775 
1776     heap->pre_full_gc_dump(&_gc_timer);
1777 
1778     TraceCollectorStats tcs(counters());
1779     TraceMemoryManagerStats tms(heap->old_gc_manager(), gc_cause);
1780 
1781     if (log_is_enabled(Debug, gc, heap, exit)) {
1782       accumulated_time()->start();
1783     }
1784 
1785     // Let the size policy know we're starting
1786     size_policy->major_collection_begin();
1787 
1788 #if COMPILER2_OR_JVMCI
1789     DerivedPointerTable::clear();
1790 #endif
1791 
1792     ref_processor()->start_discovery(maximum_heap_compaction);
1793 
1794     marking_start.update();
1795     marking_phase(vmthread_cm, &_gc_tracer);
1796 
1797     bool max_on_system_gc = UseMaximumCompactionOnSystemGC
1798       && GCCause::is_user_requested_gc(gc_cause);
1799     summary_phase(vmthread_cm, maximum_heap_compaction || max_on_system_gc);
1800 
1801 #if COMPILER2_OR_JVMCI
1802     assert(DerivedPointerTable::is_active(), "Sanity");
1803     DerivedPointerTable::set_active(false);
1804 #endif
1805 
1806     // adjust_roots() updates Universe::_intArrayKlassObj which is
1807     // needed by the compaction for filling holes in the dense prefix.
1808     adjust_roots();
1809 
1810     compaction_start.update();
1811     compact();
1812 
1813     ParCompactionManager::verify_all_region_stack_empty();
1814 
1815     // Reset the mark bitmap, summary data, and do other bookkeeping.  Must be
1816     // done before resizing.
1817     post_compact();
1818 
1819     // Let the size policy know we're done
1820     size_policy->major_collection_end(old_gen->used_in_bytes(), gc_cause);
1821 
1822     if (UseAdaptiveSizePolicy) {
1823       log_debug(gc, ergo)("AdaptiveSizeStart: collection: %d ", heap->total_collections());
1824       log_trace(gc, ergo)("old_gen_capacity: " SIZE_FORMAT " young_gen_capacity: " SIZE_FORMAT,
1825                           old_gen->capacity_in_bytes(), young_gen->capacity_in_bytes());
1826 
1827       // Don't check if the size_policy is ready here.  Let
1828       // the size_policy check that internally.
1829       if (UseAdaptiveGenerationSizePolicyAtMajorCollection &&
1830           AdaptiveSizePolicy::should_update_promo_stats(gc_cause)) {
1831         // Swap the survivor spaces if from_space is empty. The
1832         // resize_young_gen() called below is normally used after
1833         // a successful young GC and swapping of survivor spaces;
1834         // otherwise, it will fail to resize the young gen with
1835         // the current implementation.
1836         if (young_gen->from_space()->is_empty()) {
1837           young_gen->from_space()->clear(SpaceDecorator::Mangle);
1838           young_gen->swap_spaces();
1839         }
1840 
1841         // Calculate optimal free space amounts
1842         assert(young_gen->max_gen_size() >
1843           young_gen->from_space()->capacity_in_bytes() +
1844           young_gen->to_space()->capacity_in_bytes(),
1845           "Sizes of space in young gen are out-of-bounds");
1846 
1847         size_t young_live = young_gen->used_in_bytes();
1848         size_t eden_live = young_gen->eden_space()->used_in_bytes();
1849         size_t old_live = old_gen->used_in_bytes();
1850         size_t cur_eden = young_gen->eden_space()->capacity_in_bytes();
1851         size_t max_old_gen_size = old_gen->max_gen_size();
1852         size_t max_eden_size = young_gen->max_gen_size() -
1853           young_gen->from_space()->capacity_in_bytes() -
1854           young_gen->to_space()->capacity_in_bytes();
1855 
1856         // Used for diagnostics
1857         size_policy->clear_generation_free_space_flags();
1858 
1859         size_policy->compute_generations_free_space(young_live,
1860                                                     eden_live,
1861                                                     old_live,
1862                                                     cur_eden,
1863                                                     max_old_gen_size,
1864                                                     max_eden_size,
1865                                                     true /* full gc*/);
1866 
1867         size_policy->check_gc_overhead_limit(eden_live,
1868                                              max_old_gen_size,
1869                                              max_eden_size,
1870                                              true /* full gc*/,
1871                                              gc_cause,
1872                                              heap->soft_ref_policy());
1873 
1874         size_policy->decay_supplemental_growth(true /* full gc*/);
1875 
1876         heap->resize_old_gen(
1877           size_policy->calculated_old_free_size_in_bytes());
1878 
1879         heap->resize_young_gen(size_policy->calculated_eden_size_in_bytes(),
1880                                size_policy->calculated_survivor_size_in_bytes());
1881       }
1882 
1883       log_debug(gc, ergo)("AdaptiveSizeStop: collection: %d ", heap->total_collections());
1884     }
1885 
1886     if (UsePerfData) {
1887       PSGCAdaptivePolicyCounters* const counters = heap->gc_policy_counters();
1888       counters->update_counters();
1889       counters->update_old_capacity(old_gen->capacity_in_bytes());
1890       counters->update_young_capacity(young_gen->capacity_in_bytes());
1891     }
1892 
1893     heap->resize_all_tlabs();
1894 
1895     // Resize the metaspace capacity after a collection
1896     MetaspaceGC::compute_new_size();
1897 
1898     if (log_is_enabled(Debug, gc, heap, exit)) {
1899       accumulated_time()->stop();
1900     }
1901 
1902     heap->print_heap_change(pre_gc_values);
1903 
1904     // Track memory usage and detect low memory
1905     MemoryService::track_memory_usage();
1906     heap->update_counters();
1907 
1908     heap->post_full_gc_dump(&_gc_timer);
1909   }
1910 
1911   if (VerifyAfterGC && heap->total_collections() >= VerifyGCStartAt) {
1912     Universe::verify("After GC");
1913   }
1914 
1915   // Re-verify object start arrays
1916   if (VerifyObjectStartArray &&
1917       VerifyAfterGC) {
1918     old_gen->verify_object_start_array();
1919   }
1920 
1921   if (ZapUnusedHeapArea) {
1922     old_gen->object_space()->check_mangled_unused_area_complete();
1923   }
1924 
1925   collection_exit.update();
1926 
1927   heap->print_heap_after_gc();
1928   heap->trace_heap_after_gc(&_gc_tracer);
1929 
1930   log_debug(gc, task, time)("VM-Thread " JLONG_FORMAT " " JLONG_FORMAT " " JLONG_FORMAT,
1931                          marking_start.ticks(), compaction_start.ticks(),
1932                          collection_exit.ticks());
1933 
1934   AdaptiveSizePolicyOutput::print(size_policy, heap->total_collections());
1935 
1936   _gc_timer.register_gc_end();
1937 
1938   _gc_tracer.report_dense_prefix(dense_prefix(old_space_id));
1939   _gc_tracer.report_gc_end(_gc_timer.gc_end(), _gc_timer.time_partitions());
1940 
1941   return true;
1942 }
1943 
1944 class PCAddThreadRootsMarkingTaskClosure : public ThreadClosure {
1945 private:
1946   uint _worker_id;
1947 
1948 public:
1949   PCAddThreadRootsMarkingTaskClosure(uint worker_id) : _worker_id(worker_id) { }
1950   void do_thread(Thread* thread) {
1951     assert(ParallelScavengeHeap::heap()->is_gc_active(), "called outside gc");
1952 
1953     ResourceMark rm;
1954 
1955     ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(_worker_id);
1956 
1957     PCMarkAndPushClosure mark_and_push_closure(cm);
1958     MarkingCodeBlobClosure mark_and_push_in_blobs(&mark_and_push_closure, !CodeBlobToOopClosure::FixRelocations);
1959 
1960     thread->oops_do(&mark_and_push_closure, &mark_and_push_in_blobs);
1961 
1962     // Do the real work
1963     cm->follow_marking_stacks();
1964   }
1965 };
1966 
1967 static void mark_from_roots_work(ParallelRootType::Value root_type, uint worker_id) {
1968   assert(ParallelScavengeHeap::heap()->is_gc_active(), "called outside gc");
1969 
1970   ParCompactionManager* cm =
1971     ParCompactionManager::gc_thread_compaction_manager(worker_id);
1972   PCMarkAndPushClosure mark_and_push_closure(cm);
1973 
1974   switch (root_type) {
1975     case ParallelRootType::class_loader_data:
1976       {
1977         CLDToOopClosure cld_closure(&mark_and_push_closure, ClassLoaderData::_claim_strong);
1978         ClassLoaderDataGraph::always_strong_cld_do(&cld_closure);
1979       }
1980       break;
1981 
1982     case ParallelRootType::code_cache:
1983       // Do not treat nmethods as strong roots for mark/sweep, since we can unload them.
1984       //ScavengableNMethods::scavengable_nmethods_do(CodeBlobToOopClosure(&mark_and_push_closure));
1985       break;
1986 
1987     case ParallelRootType::sentinel:
1988     DEBUG_ONLY(default:) // DEBUG_ONLY hack will create compile error on release builds (-Wswitch) and runtime check on debug builds
1989       fatal("Bad enumeration value: %u", root_type);
1990       break;
1991   }
1992 
1993   // Do the real work
1994   cm->follow_marking_stacks();
1995 }
1996 
1997 void steal_marking_work(TaskTerminator& terminator, uint worker_id) {
1998   assert(ParallelScavengeHeap::heap()->is_gc_active(), "called outside gc");
1999 
2000   ParCompactionManager* cm =
2001     ParCompactionManager::gc_thread_compaction_manager(worker_id);
2002 
2003   oop obj = NULL;
2004   ObjArrayTask task;
2005   do {
2006     while (ParCompactionManager::steal_objarray(worker_id,  task)) {
2007       cm->follow_array((objArrayOop)task.obj(), task.index());
2008       cm->follow_marking_stacks();
2009     }
2010     while (ParCompactionManager::steal(worker_id, obj)) {
2011       cm->follow_contents(obj);
2012       cm->follow_marking_stacks();
2013     }
2014   } while (!terminator.offer_termination());
2015 }
2016 
2017 class MarkFromRootsTask : public WorkerTask {
2018   StrongRootsScope _strong_roots_scope; // needed for Threads::possibly_parallel_threads_do
2019   OopStorageSetStrongParState<false /* concurrent */, false /* is_const */> _oop_storage_set_par_state;
2020   SequentialSubTasksDone _subtasks;
2021   TaskTerminator _terminator;
2022   uint _active_workers;
2023 
2024 public:
2025   MarkFromRootsTask(uint active_workers) :
2026       WorkerTask("MarkFromRootsTask"),
2027       _strong_roots_scope(active_workers),
2028       _subtasks(ParallelRootType::sentinel),
2029       _terminator(active_workers, ParCompactionManager::oop_task_queues()),
2030       _active_workers(active_workers) {
2031   }
2032 
2033   virtual void work(uint worker_id) {
2034     for (uint task = 0; _subtasks.try_claim_task(task); /*empty*/ ) {
2035       mark_from_roots_work(static_cast<ParallelRootType::Value>(task), worker_id);
2036     }
2037 
2038     PCAddThreadRootsMarkingTaskClosure closure(worker_id);
2039     Threads::possibly_parallel_threads_do(true /*parallel */, &closure);
2040 
2041     // Mark from OopStorages
2042     {
2043       ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(worker_id);
2044       PCMarkAndPushClosure closure(cm);
2045       _oop_storage_set_par_state.oops_do(&closure);
2046       // Do the real work
2047       cm->follow_marking_stacks();
2048     }
2049 
2050     if (_active_workers > 1) {
2051       steal_marking_work(_terminator, worker_id);
2052     }
2053   }
2054 };
2055 
2056 class ParallelCompactRefProcProxyTask : public RefProcProxyTask {
2057   TaskTerminator _terminator;
2058 
2059 public:
2060   ParallelCompactRefProcProxyTask(uint max_workers)
2061     : RefProcProxyTask("ParallelCompactRefProcProxyTask", max_workers),
2062       _terminator(_max_workers, ParCompactionManager::oop_task_queues()) {}
2063 
2064   void work(uint worker_id) override {
2065     assert(worker_id < _max_workers, "sanity");
2066     ParCompactionManager* cm = (_tm == RefProcThreadModel::Single) ? ParCompactionManager::get_vmthread_cm() : ParCompactionManager::gc_thread_compaction_manager(worker_id);
2067     PCMarkAndPushClosure keep_alive(cm);
2068     BarrierEnqueueDiscoveredFieldClosure enqueue;
2069     ParCompactionManager::FollowStackClosure complete_gc(cm, (_tm == RefProcThreadModel::Single) ? nullptr : &_terminator, worker_id);
2070     _rp_task->rp_work(worker_id, PSParallelCompact::is_alive_closure(), &keep_alive, &enqueue, &complete_gc);
2071   }
2072 
2073   void prepare_run_task_hook() override {
2074     _terminator.reset_for_reuse(_queue_count);
2075   }
2076 };
2077 
2078 void PSParallelCompact::marking_phase(ParCompactionManager* cm,
2079                                       ParallelOldTracer *gc_tracer) {
2080   // Recursively traverse all live objects and mark them
2081   GCTraceTime(Info, gc, phases) tm("Marking Phase", &_gc_timer);
2082 
2083   uint active_gc_threads = ParallelScavengeHeap::heap()->workers().active_workers();
2084 
2085   // Need new claim bits before marking starts.
2086   ClassLoaderDataGraph::clear_claimed_marks();
2087 
2088   {
2089     GCTraceTime(Debug, gc, phases) tm("Par Mark", &_gc_timer);
2090 
2091     MarkFromRootsTask task(active_gc_threads);
2092     ParallelScavengeHeap::heap()->workers().run_task(&task);
2093   }
2094 
2095   // Process reference objects found during marking
2096   {
2097     GCTraceTime(Debug, gc, phases) tm("Reference Processing", &_gc_timer);
2098 
2099     ReferenceProcessorStats stats;
2100     ReferenceProcessorPhaseTimes pt(&_gc_timer, ref_processor()->max_num_queues());
2101 
2102     ref_processor()->set_active_mt_degree(active_gc_threads);
2103     ParallelCompactRefProcProxyTask task(ref_processor()->max_num_queues());
2104     stats = ref_processor()->process_discovered_references(task, pt);
2105 
2106     gc_tracer->report_gc_reference_stats(stats);
2107     pt.print_all_references();
2108   }
2109 
2110   // This is the point where the entire marking should have completed.
2111   ParCompactionManager::verify_all_marking_stack_empty();
2112 
2113   {
2114     GCTraceTime(Debug, gc, phases) tm("Weak Processing", &_gc_timer);
2115     WeakProcessor::weak_oops_do(&ParallelScavengeHeap::heap()->workers(),
2116                                 is_alive_closure(),
2117                                 &do_nothing_cl,
2118                                 1);
2119   }
2120 
2121   {
2122     GCTraceTime(Debug, gc, phases) tm_m("Class Unloading", &_gc_timer);
2123 
2124     // Follow system dictionary roots and unload classes.
2125     bool purged_class = SystemDictionary::do_unloading(&_gc_timer);
2126 
2127     // Unload nmethods.
2128     CodeCache::do_unloading(is_alive_closure(), purged_class);
2129 
2130     // Prune dead klasses from subklass/sibling/implementor lists.
2131     Klass::clean_weak_klass_links(purged_class);
2132 
2133     // Clean JVMCI metadata handles.
2134     JVMCI_ONLY(JVMCI::do_unloading(purged_class));
2135   }
2136 
2137   _gc_tracer.report_object_count_after_gc(is_alive_closure());
2138 }
2139 
2140 #ifdef ASSERT
2141 void PCAdjustPointerClosure::verify_cm(ParCompactionManager* cm) {
2142   assert(cm != NULL, "associate ParCompactionManage should not be NULL");
2143   auto vmthread_cm = ParCompactionManager::get_vmthread_cm();
2144   if (Thread::current()->is_VM_thread()) {
2145     assert(cm == vmthread_cm, "VM threads should use ParCompactionManager from get_vmthread_cm()");
2146   } else {
2147     assert(Thread::current()->is_Worker_thread(), "Must be a GC thread");
2148     assert(cm != vmthread_cm, "GC threads should use ParCompactionManager from gc_thread_compaction_manager()");
2149   }
2150 }
2151 #endif
2152 
2153 class PSAdjustTask final : public WorkerTask {
2154   SubTasksDone                               _sub_tasks;
2155   WeakProcessor::Task                        _weak_proc_task;
2156   OopStorageSetStrongParState<false, false>  _oop_storage_iter;
2157   uint                                       _nworkers;
2158 
2159   enum PSAdjustSubTask {
2160     PSAdjustSubTask_code_cache,
2161 
2162     PSAdjustSubTask_num_elements
2163   };
2164 
2165 public:
2166   PSAdjustTask(uint nworkers) :
2167     WorkerTask("PSAdjust task"),
2168     _sub_tasks(PSAdjustSubTask_num_elements),
2169     _weak_proc_task(nworkers),
2170     _nworkers(nworkers) {
2171     // Need new claim bits when tracing through and adjusting pointers.
2172     ClassLoaderDataGraph::clear_claimed_marks();
2173     if (nworkers > 1) {
2174       Threads::change_thread_claim_token();
2175     }
2176   }
2177 
2178   ~PSAdjustTask() {
2179     Threads::assert_all_threads_claimed();
2180   }
2181 
2182   void work(uint worker_id) {
2183     ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(worker_id);
2184     PCAdjustPointerClosure adjust(cm);
2185     {
2186       ResourceMark rm;
2187       Threads::possibly_parallel_oops_do(_nworkers > 1, &adjust, nullptr);
2188     }
2189     _oop_storage_iter.oops_do(&adjust);
2190     {
2191       CLDToOopClosure cld_closure(&adjust, ClassLoaderData::_claim_strong);
2192       ClassLoaderDataGraph::cld_do(&cld_closure);
2193     }
2194     {
2195       AlwaysTrueClosure always_alive;
2196       _weak_proc_task.work(worker_id, &always_alive, &adjust);
2197     }
2198     if (_sub_tasks.try_claim_task(PSAdjustSubTask_code_cache)) {
2199       CodeBlobToOopClosure adjust_code(&adjust, CodeBlobToOopClosure::FixRelocations);
2200       CodeCache::blobs_do(&adjust_code);
2201     }
2202     _sub_tasks.all_tasks_claimed();
2203   }
2204 };
2205 
2206 void PSParallelCompact::adjust_roots() {
2207   // Adjust the pointers to reflect the new locations
2208   GCTraceTime(Info, gc, phases) tm("Adjust Roots", &_gc_timer);
2209   uint nworkers = ParallelScavengeHeap::heap()->workers().active_workers();
2210   PSAdjustTask task(nworkers);
2211   ParallelScavengeHeap::heap()->workers().run_task(&task);
2212 }
2213 
2214 // Helper class to print 8 region numbers per line and then print the total at the end.
2215 class FillableRegionLogger : public StackObj {
2216 private:
2217   Log(gc, compaction) log;
2218   static const int LineLength = 8;
2219   size_t _regions[LineLength];
2220   int _next_index;
2221   bool _enabled;
2222   size_t _total_regions;
2223 public:
2224   FillableRegionLogger() : _next_index(0), _enabled(log_develop_is_enabled(Trace, gc, compaction)), _total_regions(0) { }
2225   ~FillableRegionLogger() {
2226     log.trace(SIZE_FORMAT " initially fillable regions", _total_regions);
2227   }
2228 
2229   void print_line() {
2230     if (!_enabled || _next_index == 0) {
2231       return;
2232     }
2233     FormatBuffer<> line("Fillable: ");
2234     for (int i = 0; i < _next_index; i++) {
2235       line.append(" " SIZE_FORMAT_W(7), _regions[i]);
2236     }
2237     log.trace("%s", line.buffer());
2238     _next_index = 0;
2239   }
2240 
2241   void handle(size_t region) {
2242     if (!_enabled) {
2243       return;
2244     }
2245     _regions[_next_index++] = region;
2246     if (_next_index == LineLength) {
2247       print_line();
2248     }
2249     _total_regions++;
2250   }
2251 };
2252 
2253 void PSParallelCompact::prepare_region_draining_tasks(uint parallel_gc_threads)
2254 {
2255   GCTraceTime(Trace, gc, phases) tm("Drain Task Setup", &_gc_timer);
2256 
2257   // Find the threads that are active
2258   uint worker_id = 0;
2259 
2260   // Find all regions that are available (can be filled immediately) and
2261   // distribute them to the thread stacks.  The iteration is done in reverse
2262   // order (high to low) so the regions will be removed in ascending order.
2263 
2264   const ParallelCompactData& sd = PSParallelCompact::summary_data();
2265 
2266   // id + 1 is used to test termination so unsigned  can
2267   // be used with an old_space_id == 0.
2268   FillableRegionLogger region_logger;
2269   for (unsigned int id = to_space_id; id + 1 > old_space_id; --id) {
2270     SpaceInfo* const space_info = _space_info + id;
2271     HeapWord* const new_top = space_info->new_top();
2272 
2273     const size_t beg_region = sd.addr_to_region_idx(space_info->dense_prefix());
2274     const size_t end_region =
2275       sd.addr_to_region_idx(sd.region_align_up(new_top));
2276 
2277     for (size_t cur = end_region - 1; cur + 1 > beg_region; --cur) {
2278       if (sd.region(cur)->claim_unsafe()) {
2279         ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(worker_id);
2280         bool result = sd.region(cur)->mark_normal();
2281         assert(result, "Must succeed at this point.");
2282         cm->region_stack()->push(cur);
2283         region_logger.handle(cur);
2284         // Assign regions to tasks in round-robin fashion.
2285         if (++worker_id == parallel_gc_threads) {
2286           worker_id = 0;
2287         }
2288       }
2289     }
2290     region_logger.print_line();
2291   }
2292 }
2293 
2294 class TaskQueue : StackObj {
2295   volatile uint _counter;
2296   uint _size;
2297   uint _insert_index;
2298   PSParallelCompact::UpdateDensePrefixTask* _backing_array;
2299 public:
2300   explicit TaskQueue(uint size) : _counter(0), _size(size), _insert_index(0), _backing_array(NULL) {
2301     _backing_array = NEW_C_HEAP_ARRAY(PSParallelCompact::UpdateDensePrefixTask, _size, mtGC);
2302   }
2303   ~TaskQueue() {
2304     assert(_counter >= _insert_index, "not all queue elements were claimed");
2305     FREE_C_HEAP_ARRAY(T, _backing_array);
2306   }
2307 
2308   void push(const PSParallelCompact::UpdateDensePrefixTask& value) {
2309     assert(_insert_index < _size, "too small backing array");
2310     _backing_array[_insert_index++] = value;
2311   }
2312 
2313   bool try_claim(PSParallelCompact::UpdateDensePrefixTask& reference) {
2314     uint claimed = Atomic::fetch_and_add(&_counter, 1u);
2315     if (claimed < _insert_index) {
2316       reference = _backing_array[claimed];
2317       return true;
2318     } else {
2319       return false;
2320     }
2321   }
2322 };
2323 
2324 #define PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING 4
2325 
2326 void PSParallelCompact::enqueue_dense_prefix_tasks(TaskQueue& task_queue,
2327                                                    uint parallel_gc_threads) {
2328   GCTraceTime(Trace, gc, phases) tm("Dense Prefix Task Setup", &_gc_timer);
2329 
2330   ParallelCompactData& sd = PSParallelCompact::summary_data();
2331 
2332   // Iterate over all the spaces adding tasks for updating
2333   // regions in the dense prefix.  Assume that 1 gc thread
2334   // will work on opening the gaps and the remaining gc threads
2335   // will work on the dense prefix.
2336   unsigned int space_id;
2337   for (space_id = old_space_id; space_id < last_space_id; ++ space_id) {
2338     HeapWord* const dense_prefix_end = _space_info[space_id].dense_prefix();
2339     const MutableSpace* const space = _space_info[space_id].space();
2340 
2341     if (dense_prefix_end == space->bottom()) {
2342       // There is no dense prefix for this space.
2343       continue;
2344     }
2345 
2346     // The dense prefix is before this region.
2347     size_t region_index_end_dense_prefix =
2348         sd.addr_to_region_idx(dense_prefix_end);
2349     RegionData* const dense_prefix_cp =
2350       sd.region(region_index_end_dense_prefix);
2351     assert(dense_prefix_end == space->end() ||
2352            dense_prefix_cp->available() ||
2353            dense_prefix_cp->claimed(),
2354            "The region after the dense prefix should always be ready to fill");
2355 
2356     size_t region_index_start = sd.addr_to_region_idx(space->bottom());
2357 
2358     // Is there dense prefix work?
2359     size_t total_dense_prefix_regions =
2360       region_index_end_dense_prefix - region_index_start;
2361     // How many regions of the dense prefix should be given to
2362     // each thread?
2363     if (total_dense_prefix_regions > 0) {
2364       uint tasks_for_dense_prefix = 1;
2365       if (total_dense_prefix_regions <=
2366           (parallel_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING)) {
2367         // Don't over partition.  This assumes that
2368         // PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING is a small integer value
2369         // so there are not many regions to process.
2370         tasks_for_dense_prefix = parallel_gc_threads;
2371       } else {
2372         // Over partition
2373         tasks_for_dense_prefix = parallel_gc_threads *
2374           PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING;
2375       }
2376       size_t regions_per_thread = total_dense_prefix_regions /
2377         tasks_for_dense_prefix;
2378       // Give each thread at least 1 region.
2379       if (regions_per_thread == 0) {
2380         regions_per_thread = 1;
2381       }
2382 
2383       for (uint k = 0; k < tasks_for_dense_prefix; k++) {
2384         if (region_index_start >= region_index_end_dense_prefix) {
2385           break;
2386         }
2387         // region_index_end is not processed
2388         size_t region_index_end = MIN2(region_index_start + regions_per_thread,
2389                                        region_index_end_dense_prefix);
2390         task_queue.push(UpdateDensePrefixTask(SpaceId(space_id),
2391                                               region_index_start,
2392                                               region_index_end));
2393         region_index_start = region_index_end;
2394       }
2395     }
2396     // This gets any part of the dense prefix that did not
2397     // fit evenly.
2398     if (region_index_start < region_index_end_dense_prefix) {
2399       task_queue.push(UpdateDensePrefixTask(SpaceId(space_id),
2400                                             region_index_start,
2401                                             region_index_end_dense_prefix));
2402     }
2403   }
2404 }
2405 
2406 #ifdef ASSERT
2407 // Write a histogram of the number of times the block table was filled for a
2408 // region.
2409 void PSParallelCompact::write_block_fill_histogram()
2410 {
2411   if (!log_develop_is_enabled(Trace, gc, compaction)) {
2412     return;
2413   }
2414 
2415   Log(gc, compaction) log;
2416   ResourceMark rm;
2417   LogStream ls(log.trace());
2418   outputStream* out = &ls;
2419 
2420   typedef ParallelCompactData::RegionData rd_t;
2421   ParallelCompactData& sd = summary_data();
2422 
2423   for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2424     MutableSpace* const spc = _space_info[id].space();
2425     if (spc->bottom() != spc->top()) {
2426       const rd_t* const beg = sd.addr_to_region_ptr(spc->bottom());
2427       HeapWord* const top_aligned_up = sd.region_align_up(spc->top());
2428       const rd_t* const end = sd.addr_to_region_ptr(top_aligned_up);
2429 
2430       size_t histo[5] = { 0, 0, 0, 0, 0 };
2431       const size_t histo_len = sizeof(histo) / sizeof(size_t);
2432       const size_t region_cnt = pointer_delta(end, beg, sizeof(rd_t));
2433 
2434       for (const rd_t* cur = beg; cur < end; ++cur) {
2435         ++histo[MIN2(cur->blocks_filled_count(), histo_len - 1)];
2436       }
2437       out->print("Block fill histogram: %u %-4s" SIZE_FORMAT_W(5), id, space_names[id], region_cnt);
2438       for (size_t i = 0; i < histo_len; ++i) {
2439         out->print(" " SIZE_FORMAT_W(5) " %5.1f%%",
2440                    histo[i], 100.0 * histo[i] / region_cnt);
2441       }
2442       out->cr();
2443     }
2444   }
2445 }
2446 #endif // #ifdef ASSERT
2447 
2448 static void compaction_with_stealing_work(TaskTerminator* terminator, uint worker_id) {
2449   assert(ParallelScavengeHeap::heap()->is_gc_active(), "called outside gc");
2450 
2451   ParCompactionManager* cm =
2452     ParCompactionManager::gc_thread_compaction_manager(worker_id);
2453 
2454   // Drain the stacks that have been preloaded with regions
2455   // that are ready to fill.
2456 
2457   cm->drain_region_stacks();
2458 
2459   guarantee(cm->region_stack()->is_empty(), "Not empty");
2460 
2461   size_t region_index = 0;
2462 
2463   while (true) {
2464     if (ParCompactionManager::steal(worker_id, region_index)) {
2465       PSParallelCompact::fill_and_update_region(cm, region_index);
2466       cm->drain_region_stacks();
2467     } else if (PSParallelCompact::steal_unavailable_region(cm, region_index)) {
2468       // Fill and update an unavailable region with the help of a shadow region
2469       PSParallelCompact::fill_and_update_shadow_region(cm, region_index);
2470       cm->drain_region_stacks();
2471     } else {
2472       if (terminator->offer_termination()) {
2473         break;
2474       }
2475       // Go around again.
2476     }
2477   }
2478 }
2479 
2480 class UpdateDensePrefixAndCompactionTask: public WorkerTask {
2481   TaskQueue& _tq;
2482   TaskTerminator _terminator;
2483   uint _active_workers;
2484 
2485 public:
2486   UpdateDensePrefixAndCompactionTask(TaskQueue& tq, uint active_workers) :
2487       WorkerTask("UpdateDensePrefixAndCompactionTask"),
2488       _tq(tq),
2489       _terminator(active_workers, ParCompactionManager::region_task_queues()),
2490       _active_workers(active_workers) {
2491   }
2492   virtual void work(uint worker_id) {
2493     ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(worker_id);
2494 
2495     for (PSParallelCompact::UpdateDensePrefixTask task; _tq.try_claim(task); /* empty */) {
2496       PSParallelCompact::update_and_deadwood_in_dense_prefix(cm,
2497                                                              task._space_id,
2498                                                              task._region_index_start,
2499                                                              task._region_index_end);
2500     }
2501 
2502     // Once a thread has drained it's stack, it should try to steal regions from
2503     // other threads.
2504     compaction_with_stealing_work(&_terminator, worker_id);
2505   }
2506 };
2507 
2508 void PSParallelCompact::compact() {
2509   GCTraceTime(Info, gc, phases) tm("Compaction Phase", &_gc_timer);
2510 
2511   ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
2512   PSOldGen* old_gen = heap->old_gen();
2513   old_gen->start_array()->reset();
2514   uint active_gc_threads = ParallelScavengeHeap::heap()->workers().active_workers();
2515 
2516   // for [0..last_space_id)
2517   //     for [0..active_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING)
2518   //         push
2519   //     push
2520   //
2521   // max push count is thus: last_space_id * (active_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING + 1)
2522   TaskQueue task_queue(last_space_id * (active_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING + 1));
2523   initialize_shadow_regions(active_gc_threads);
2524   prepare_region_draining_tasks(active_gc_threads);
2525   enqueue_dense_prefix_tasks(task_queue, active_gc_threads);
2526 
2527   {
2528     GCTraceTime(Trace, gc, phases) tm("Par Compact", &_gc_timer);
2529 
2530     UpdateDensePrefixAndCompactionTask task(task_queue, active_gc_threads);
2531     ParallelScavengeHeap::heap()->workers().run_task(&task);
2532 
2533 #ifdef  ASSERT
2534     // Verify that all regions have been processed before the deferred updates.
2535     for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2536       verify_complete(SpaceId(id));
2537     }
2538 #endif
2539   }
2540 
2541   {
2542     GCTraceTime(Trace, gc, phases) tm("Deferred Updates", &_gc_timer);
2543     // Update the deferred objects, if any. In principle, any compaction
2544     // manager can be used. However, since the current thread is VM thread, we
2545     // use the rightful one to keep the verification logic happy.
2546     ParCompactionManager* cm = ParCompactionManager::get_vmthread_cm();
2547     for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2548       update_deferred_objects(cm, SpaceId(id));
2549     }
2550   }
2551 
2552   DEBUG_ONLY(write_block_fill_histogram());
2553 }
2554 
2555 #ifdef  ASSERT
2556 void PSParallelCompact::verify_complete(SpaceId space_id) {
2557   // All Regions between space bottom() to new_top() should be marked as filled
2558   // and all Regions between new_top() and top() should be available (i.e.,
2559   // should have been emptied).
2560   ParallelCompactData& sd = summary_data();
2561   SpaceInfo si = _space_info[space_id];
2562   HeapWord* new_top_addr = sd.region_align_up(si.new_top());
2563   HeapWord* old_top_addr = sd.region_align_up(si.space()->top());
2564   const size_t beg_region = sd.addr_to_region_idx(si.space()->bottom());
2565   const size_t new_top_region = sd.addr_to_region_idx(new_top_addr);
2566   const size_t old_top_region = sd.addr_to_region_idx(old_top_addr);
2567 
2568   bool issued_a_warning = false;
2569 
2570   size_t cur_region;
2571   for (cur_region = beg_region; cur_region < new_top_region; ++cur_region) {
2572     const RegionData* const c = sd.region(cur_region);
2573     if (!c->completed()) {
2574       log_warning(gc)("region " SIZE_FORMAT " not filled: destination_count=%u",
2575                       cur_region, c->destination_count());
2576       issued_a_warning = true;
2577     }
2578   }
2579 
2580   for (cur_region = new_top_region; cur_region < old_top_region; ++cur_region) {
2581     const RegionData* const c = sd.region(cur_region);
2582     if (!c->available()) {
2583       log_warning(gc)("region " SIZE_FORMAT " not empty: destination_count=%u",
2584                       cur_region, c->destination_count());
2585       issued_a_warning = true;
2586     }
2587   }
2588 
2589   if (issued_a_warning) {
2590     print_region_ranges();
2591   }
2592 }
2593 #endif  // #ifdef ASSERT
2594 
2595 inline void UpdateOnlyClosure::do_addr(HeapWord* addr) {
2596   _start_array->allocate_block(addr);
2597   compaction_manager()->update_contents(cast_to_oop(addr));
2598 }
2599 
2600 // Update interior oops in the ranges of regions [beg_region, end_region).
2601 void
2602 PSParallelCompact::update_and_deadwood_in_dense_prefix(ParCompactionManager* cm,
2603                                                        SpaceId space_id,
2604                                                        size_t beg_region,
2605                                                        size_t end_region) {
2606   ParallelCompactData& sd = summary_data();
2607   ParMarkBitMap* const mbm = mark_bitmap();
2608 
2609   HeapWord* beg_addr = sd.region_to_addr(beg_region);
2610   HeapWord* const end_addr = sd.region_to_addr(end_region);
2611   assert(beg_region <= end_region, "bad region range");
2612   assert(end_addr <= dense_prefix(space_id), "not in the dense prefix");
2613 
2614 #ifdef  ASSERT
2615   // Claim the regions to avoid triggering an assert when they are marked as
2616   // filled.
2617   for (size_t claim_region = beg_region; claim_region < end_region; ++claim_region) {
2618     assert(sd.region(claim_region)->claim_unsafe(), "claim() failed");
2619   }
2620 #endif  // #ifdef ASSERT
2621 
2622   if (beg_addr != space(space_id)->bottom()) {
2623     // Find the first live object or block of dead space that *starts* in this
2624     // range of regions.  If a partial object crosses onto the region, skip it;
2625     // it will be marked for 'deferred update' when the object head is
2626     // processed.  If dead space crosses onto the region, it is also skipped; it
2627     // will be filled when the prior region is processed.  If neither of those
2628     // apply, the first word in the region is the start of a live object or dead
2629     // space.
2630     assert(beg_addr > space(space_id)->bottom(), "sanity");
2631     const RegionData* const cp = sd.region(beg_region);
2632     if (cp->partial_obj_size() != 0) {
2633       beg_addr = sd.partial_obj_end(beg_region);
2634     } else if (dead_space_crosses_boundary(cp, mbm->addr_to_bit(beg_addr))) {
2635       beg_addr = mbm->find_obj_beg(beg_addr, end_addr);
2636     }
2637   }
2638 
2639   if (beg_addr < end_addr) {
2640     // A live object or block of dead space starts in this range of Regions.
2641      HeapWord* const dense_prefix_end = dense_prefix(space_id);
2642 
2643     // Create closures and iterate.
2644     UpdateOnlyClosure update_closure(mbm, cm, space_id);
2645     FillClosure fill_closure(cm, space_id);
2646     ParMarkBitMap::IterationStatus status;
2647     status = mbm->iterate(&update_closure, &fill_closure, beg_addr, end_addr,
2648                           dense_prefix_end);
2649     if (status == ParMarkBitMap::incomplete) {
2650       update_closure.do_addr(update_closure.source());
2651     }
2652   }
2653 
2654   // Mark the regions as filled.
2655   RegionData* const beg_cp = sd.region(beg_region);
2656   RegionData* const end_cp = sd.region(end_region);
2657   for (RegionData* cp = beg_cp; cp < end_cp; ++cp) {
2658     cp->set_completed();
2659   }
2660 }
2661 
2662 // Return the SpaceId for the space containing addr.  If addr is not in the
2663 // heap, last_space_id is returned.  In debug mode it expects the address to be
2664 // in the heap and asserts such.
2665 PSParallelCompact::SpaceId PSParallelCompact::space_id(HeapWord* addr) {
2666   assert(ParallelScavengeHeap::heap()->is_in_reserved(addr), "addr not in the heap");
2667 
2668   for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2669     if (_space_info[id].space()->contains(addr)) {
2670       return SpaceId(id);
2671     }
2672   }
2673 
2674   assert(false, "no space contains the addr");
2675   return last_space_id;
2676 }
2677 
2678 void PSParallelCompact::update_deferred_objects(ParCompactionManager* cm,
2679                                                 SpaceId id) {
2680   assert(id < last_space_id, "bad space id");
2681 
2682   ParallelCompactData& sd = summary_data();
2683   const SpaceInfo* const space_info = _space_info + id;
2684   ObjectStartArray* const start_array = space_info->start_array();
2685 
2686   const MutableSpace* const space = space_info->space();
2687   assert(space_info->dense_prefix() >= space->bottom(), "dense_prefix not set");
2688   HeapWord* const beg_addr = space_info->dense_prefix();
2689   HeapWord* const end_addr = sd.region_align_up(space_info->new_top());
2690 
2691   const RegionData* const beg_region = sd.addr_to_region_ptr(beg_addr);
2692   const RegionData* const end_region = sd.addr_to_region_ptr(end_addr);
2693   const RegionData* cur_region;
2694   for (cur_region = beg_region; cur_region < end_region; ++cur_region) {
2695     HeapWord* const addr = cur_region->deferred_obj_addr();
2696     if (addr != NULL) {
2697       if (start_array != NULL) {
2698         start_array->allocate_block(addr);
2699       }
2700       cm->update_contents(cast_to_oop(addr));
2701       assert(oopDesc::is_oop_or_null(cast_to_oop(addr)), "Expected an oop or NULL at " PTR_FORMAT, p2i(cast_to_oop(addr)));
2702     }
2703   }
2704 }
2705 
2706 // Skip over count live words starting from beg, and return the address of the
2707 // next live word.  Unless marked, the word corresponding to beg is assumed to
2708 // be dead.  Callers must either ensure beg does not correspond to the middle of
2709 // an object, or account for those live words in some other way.  Callers must
2710 // also ensure that there are enough live words in the range [beg, end) to skip.
2711 HeapWord*
2712 PSParallelCompact::skip_live_words(HeapWord* beg, HeapWord* end, size_t count)
2713 {
2714   assert(count > 0, "sanity");
2715 
2716   ParMarkBitMap* m = mark_bitmap();
2717   idx_t bits_to_skip = m->words_to_bits(count);
2718   idx_t cur_beg = m->addr_to_bit(beg);
2719   const idx_t search_end = m->align_range_end(m->addr_to_bit(end));
2720 
2721   do {
2722     cur_beg = m->find_obj_beg(cur_beg, search_end);
2723     idx_t cur_end = m->find_obj_end(cur_beg, search_end);
2724     const size_t obj_bits = cur_end - cur_beg + 1;
2725     if (obj_bits > bits_to_skip) {
2726       return m->bit_to_addr(cur_beg + bits_to_skip);
2727     }
2728     bits_to_skip -= obj_bits;
2729     cur_beg = cur_end + 1;
2730   } while (bits_to_skip > 0);
2731 
2732   // Skipping the desired number of words landed just past the end of an object.
2733   // Find the start of the next object.
2734   cur_beg = m->find_obj_beg(cur_beg, search_end);
2735   assert(cur_beg < m->addr_to_bit(end), "not enough live words to skip");
2736   return m->bit_to_addr(cur_beg);
2737 }
2738 
2739 HeapWord* PSParallelCompact::first_src_addr(HeapWord* const dest_addr,
2740                                             SpaceId src_space_id,
2741                                             size_t src_region_idx)
2742 {
2743   assert(summary_data().is_region_aligned(dest_addr), "not aligned");
2744 
2745   const SplitInfo& split_info = _space_info[src_space_id].split_info();
2746   if (split_info.dest_region_addr() == dest_addr) {
2747     // The partial object ending at the split point contains the first word to
2748     // be copied to dest_addr.
2749     return split_info.first_src_addr();
2750   }
2751 
2752   const ParallelCompactData& sd = summary_data();
2753   ParMarkBitMap* const bitmap = mark_bitmap();
2754   const size_t RegionSize = ParallelCompactData::RegionSize;
2755 
2756   assert(sd.is_region_aligned(dest_addr), "not aligned");
2757   const RegionData* const src_region_ptr = sd.region(src_region_idx);
2758   const size_t partial_obj_size = src_region_ptr->partial_obj_size();
2759   HeapWord* const src_region_destination = src_region_ptr->destination();
2760 
2761   assert(dest_addr >= src_region_destination, "wrong src region");
2762   assert(src_region_ptr->data_size() > 0, "src region cannot be empty");
2763 
2764   HeapWord* const src_region_beg = sd.region_to_addr(src_region_idx);
2765   HeapWord* const src_region_end = src_region_beg + RegionSize;
2766 
2767   HeapWord* addr = src_region_beg;
2768   if (dest_addr == src_region_destination) {
2769     // Return the first live word in the source region.
2770     if (partial_obj_size == 0) {
2771       addr = bitmap->find_obj_beg(addr, src_region_end);
2772       assert(addr < src_region_end, "no objects start in src region");
2773     }
2774     return addr;
2775   }
2776 
2777   // Must skip some live data.
2778   size_t words_to_skip = dest_addr - src_region_destination;
2779   assert(src_region_ptr->data_size() > words_to_skip, "wrong src region");
2780 
2781   if (partial_obj_size >= words_to_skip) {
2782     // All the live words to skip are part of the partial object.
2783     addr += words_to_skip;
2784     if (partial_obj_size == words_to_skip) {
2785       // Find the first live word past the partial object.
2786       addr = bitmap->find_obj_beg(addr, src_region_end);
2787       assert(addr < src_region_end, "wrong src region");
2788     }
2789     return addr;
2790   }
2791 
2792   // Skip over the partial object (if any).
2793   if (partial_obj_size != 0) {
2794     words_to_skip -= partial_obj_size;
2795     addr += partial_obj_size;
2796   }
2797 
2798   // Skip over live words due to objects that start in the region.
2799   addr = skip_live_words(addr, src_region_end, words_to_skip);
2800   assert(addr < src_region_end, "wrong src region");
2801   return addr;
2802 }
2803 
2804 void PSParallelCompact::decrement_destination_counts(ParCompactionManager* cm,
2805                                                      SpaceId src_space_id,
2806                                                      size_t beg_region,
2807                                                      HeapWord* end_addr)
2808 {
2809   ParallelCompactData& sd = summary_data();
2810 
2811 #ifdef ASSERT
2812   MutableSpace* const src_space = _space_info[src_space_id].space();
2813   HeapWord* const beg_addr = sd.region_to_addr(beg_region);
2814   assert(src_space->contains(beg_addr) || beg_addr == src_space->end(),
2815          "src_space_id does not match beg_addr");
2816   assert(src_space->contains(end_addr) || end_addr == src_space->end(),
2817          "src_space_id does not match end_addr");
2818 #endif // #ifdef ASSERT
2819 
2820   RegionData* const beg = sd.region(beg_region);
2821   RegionData* const end = sd.addr_to_region_ptr(sd.region_align_up(end_addr));
2822 
2823   // Regions up to new_top() are enqueued if they become available.
2824   HeapWord* const new_top = _space_info[src_space_id].new_top();
2825   RegionData* const enqueue_end =
2826     sd.addr_to_region_ptr(sd.region_align_up(new_top));
2827 
2828   for (RegionData* cur = beg; cur < end; ++cur) {
2829     assert(cur->data_size() > 0, "region must have live data");
2830     cur->decrement_destination_count();
2831     if (cur < enqueue_end && cur->available() && cur->claim()) {
2832       if (cur->mark_normal()) {
2833         cm->push_region(sd.region(cur));
2834       } else if (cur->mark_copied()) {
2835         // Try to copy the content of the shadow region back to its corresponding
2836         // heap region if the shadow region is filled. Otherwise, the GC thread
2837         // fills the shadow region will copy the data back (see
2838         // MoveAndUpdateShadowClosure::complete_region).
2839         copy_back(sd.region_to_addr(cur->shadow_region()), sd.region_to_addr(cur));
2840         ParCompactionManager::push_shadow_region_mt_safe(cur->shadow_region());
2841         cur->set_completed();
2842       }
2843     }
2844   }
2845 }
2846 
2847 size_t PSParallelCompact::next_src_region(MoveAndUpdateClosure& closure,
2848                                           SpaceId& src_space_id,
2849                                           HeapWord*& src_space_top,
2850                                           HeapWord* end_addr)
2851 {
2852   typedef ParallelCompactData::RegionData RegionData;
2853 
2854   ParallelCompactData& sd = PSParallelCompact::summary_data();
2855   const size_t region_size = ParallelCompactData::RegionSize;
2856 
2857   size_t src_region_idx = 0;
2858 
2859   // Skip empty regions (if any) up to the top of the space.
2860   HeapWord* const src_aligned_up = sd.region_align_up(end_addr);
2861   RegionData* src_region_ptr = sd.addr_to_region_ptr(src_aligned_up);
2862   HeapWord* const top_aligned_up = sd.region_align_up(src_space_top);
2863   const RegionData* const top_region_ptr =
2864     sd.addr_to_region_ptr(top_aligned_up);
2865   while (src_region_ptr < top_region_ptr && src_region_ptr->data_size() == 0) {
2866     ++src_region_ptr;
2867   }
2868 
2869   if (src_region_ptr < top_region_ptr) {
2870     // The next source region is in the current space.  Update src_region_idx
2871     // and the source address to match src_region_ptr.
2872     src_region_idx = sd.region(src_region_ptr);
2873     HeapWord* const src_region_addr = sd.region_to_addr(src_region_idx);
2874     if (src_region_addr > closure.source()) {
2875       closure.set_source(src_region_addr);
2876     }
2877     return src_region_idx;
2878   }
2879 
2880   // Switch to a new source space and find the first non-empty region.
2881   unsigned int space_id = src_space_id + 1;
2882   assert(space_id < last_space_id, "not enough spaces");
2883 
2884   HeapWord* const destination = closure.destination();
2885 
2886   do {
2887     MutableSpace* space = _space_info[space_id].space();
2888     HeapWord* const bottom = space->bottom();
2889     const RegionData* const bottom_cp = sd.addr_to_region_ptr(bottom);
2890 
2891     // Iterate over the spaces that do not compact into themselves.
2892     if (bottom_cp->destination() != bottom) {
2893       HeapWord* const top_aligned_up = sd.region_align_up(space->top());
2894       const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up);
2895 
2896       for (const RegionData* src_cp = bottom_cp; src_cp < top_cp; ++src_cp) {
2897         if (src_cp->live_obj_size() > 0) {
2898           // Found it.
2899           assert(src_cp->destination() == destination,
2900                  "first live obj in the space must match the destination");
2901           assert(src_cp->partial_obj_size() == 0,
2902                  "a space cannot begin with a partial obj");
2903 
2904           src_space_id = SpaceId(space_id);
2905           src_space_top = space->top();
2906           const size_t src_region_idx = sd.region(src_cp);
2907           closure.set_source(sd.region_to_addr(src_region_idx));
2908           return src_region_idx;
2909         } else {
2910           assert(src_cp->data_size() == 0, "sanity");
2911         }
2912       }
2913     }
2914   } while (++space_id < last_space_id);
2915 
2916   assert(false, "no source region was found");
2917   return 0;
2918 }
2919 
2920 void PSParallelCompact::fill_region(ParCompactionManager* cm, MoveAndUpdateClosure& closure, size_t region_idx)
2921 {
2922   typedef ParMarkBitMap::IterationStatus IterationStatus;
2923   ParMarkBitMap* const bitmap = mark_bitmap();
2924   ParallelCompactData& sd = summary_data();
2925   RegionData* const region_ptr = sd.region(region_idx);
2926 
2927   // Get the source region and related info.
2928   size_t src_region_idx = region_ptr->source_region();
2929   SpaceId src_space_id = space_id(sd.region_to_addr(src_region_idx));
2930   HeapWord* src_space_top = _space_info[src_space_id].space()->top();
2931   HeapWord* dest_addr = sd.region_to_addr(region_idx);
2932 
2933   closure.set_source(first_src_addr(dest_addr, src_space_id, src_region_idx));
2934 
2935   // Adjust src_region_idx to prepare for decrementing destination counts (the
2936   // destination count is not decremented when a region is copied to itself).
2937   if (src_region_idx == region_idx) {
2938     src_region_idx += 1;
2939   }
2940 
2941   if (bitmap->is_unmarked(closure.source())) {
2942     // The first source word is in the middle of an object; copy the remainder
2943     // of the object or as much as will fit.  The fact that pointer updates were
2944     // deferred will be noted when the object header is processed.
2945     HeapWord* const old_src_addr = closure.source();
2946     closure.copy_partial_obj();
2947     if (closure.is_full()) {
2948       decrement_destination_counts(cm, src_space_id, src_region_idx,
2949                                    closure.source());
2950       region_ptr->set_deferred_obj_addr(NULL);
2951       closure.complete_region(cm, dest_addr, region_ptr);
2952       return;
2953     }
2954 
2955     HeapWord* const end_addr = sd.region_align_down(closure.source());
2956     if (sd.region_align_down(old_src_addr) != end_addr) {
2957       // The partial object was copied from more than one source region.
2958       decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr);
2959 
2960       // Move to the next source region, possibly switching spaces as well.  All
2961       // args except end_addr may be modified.
2962       src_region_idx = next_src_region(closure, src_space_id, src_space_top,
2963                                        end_addr);
2964     }
2965   }
2966 
2967   do {
2968     HeapWord* const cur_addr = closure.source();
2969     HeapWord* const end_addr = MIN2(sd.region_align_up(cur_addr + 1),
2970                                     src_space_top);
2971     IterationStatus status = bitmap->iterate(&closure, cur_addr, end_addr);
2972 
2973     if (status == ParMarkBitMap::incomplete) {
2974       // The last obj that starts in the source region does not end in the
2975       // region.
2976       assert(closure.source() < end_addr, "sanity");
2977       HeapWord* const obj_beg = closure.source();
2978       HeapWord* const range_end = MIN2(obj_beg + closure.words_remaining(),
2979                                        src_space_top);
2980       HeapWord* const obj_end = bitmap->find_obj_end(obj_beg, range_end);
2981       if (obj_end < range_end) {
2982         // The end was found; the entire object will fit.
2983         status = closure.do_addr(obj_beg, bitmap->obj_size(obj_beg, obj_end));
2984         assert(status != ParMarkBitMap::would_overflow, "sanity");
2985       } else {
2986         // The end was not found; the object will not fit.
2987         assert(range_end < src_space_top, "obj cannot cross space boundary");
2988         status = ParMarkBitMap::would_overflow;
2989       }
2990     }
2991 
2992     if (status == ParMarkBitMap::would_overflow) {
2993       // The last object did not fit.  Note that interior oop updates were
2994       // deferred, then copy enough of the object to fill the region.
2995       region_ptr->set_deferred_obj_addr(closure.destination());
2996       status = closure.copy_until_full(); // copies from closure.source()
2997 
2998       decrement_destination_counts(cm, src_space_id, src_region_idx,
2999                                    closure.source());
3000       closure.complete_region(cm, dest_addr, region_ptr);
3001       return;
3002     }
3003 
3004     if (status == ParMarkBitMap::full) {
3005       decrement_destination_counts(cm, src_space_id, src_region_idx,
3006                                    closure.source());
3007       region_ptr->set_deferred_obj_addr(NULL);
3008       closure.complete_region(cm, dest_addr, region_ptr);
3009       return;
3010     }
3011 
3012     decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr);
3013 
3014     // Move to the next source region, possibly switching spaces as well.  All
3015     // args except end_addr may be modified.
3016     src_region_idx = next_src_region(closure, src_space_id, src_space_top,
3017                                      end_addr);
3018   } while (true);
3019 }
3020 
3021 void PSParallelCompact::fill_and_update_region(ParCompactionManager* cm, size_t region_idx)
3022 {
3023   MoveAndUpdateClosure cl(mark_bitmap(), cm, region_idx);
3024   fill_region(cm, cl, region_idx);
3025 }
3026 
3027 void PSParallelCompact::fill_and_update_shadow_region(ParCompactionManager* cm, size_t region_idx)
3028 {
3029   // Get a shadow region first
3030   ParallelCompactData& sd = summary_data();
3031   RegionData* const region_ptr = sd.region(region_idx);
3032   size_t shadow_region = ParCompactionManager::pop_shadow_region_mt_safe(region_ptr);
3033   // The InvalidShadow return value indicates the corresponding heap region is available,
3034   // so use MoveAndUpdateClosure to fill the normal region. Otherwise, use
3035   // MoveAndUpdateShadowClosure to fill the acquired shadow region.
3036   if (shadow_region == ParCompactionManager::InvalidShadow) {
3037     MoveAndUpdateClosure cl(mark_bitmap(), cm, region_idx);
3038     region_ptr->shadow_to_normal();
3039     return fill_region(cm, cl, region_idx);
3040   } else {
3041     MoveAndUpdateShadowClosure cl(mark_bitmap(), cm, region_idx, shadow_region);
3042     return fill_region(cm, cl, region_idx);
3043   }
3044 }
3045 
3046 void PSParallelCompact::copy_back(HeapWord *shadow_addr, HeapWord *region_addr)
3047 {
3048   Copy::aligned_conjoint_words(shadow_addr, region_addr, _summary_data.RegionSize);
3049 }
3050 
3051 bool PSParallelCompact::steal_unavailable_region(ParCompactionManager* cm, size_t &region_idx)
3052 {
3053   size_t next = cm->next_shadow_region();
3054   ParallelCompactData& sd = summary_data();
3055   size_t old_new_top = sd.addr_to_region_idx(_space_info[old_space_id].new_top());
3056   uint active_gc_threads = ParallelScavengeHeap::heap()->workers().active_workers();
3057 
3058   while (next < old_new_top) {
3059     if (sd.region(next)->mark_shadow()) {
3060       region_idx = next;
3061       return true;
3062     }
3063     next = cm->move_next_shadow_region_by(active_gc_threads);
3064   }
3065 
3066   return false;
3067 }
3068 
3069 // The shadow region is an optimization to address region dependencies in full GC. The basic
3070 // idea is making more regions available by temporally storing their live objects in empty
3071 // shadow regions to resolve dependencies between them and the destination regions. Therefore,
3072 // GC threads need not wait destination regions to be available before processing sources.
3073 //
3074 // A typical workflow would be:
3075 // After draining its own stack and failing to steal from others, a GC worker would pick an
3076 // unavailable region (destination count > 0) and get a shadow region. Then the worker fills
3077 // the shadow region by copying live objects from source regions of the unavailable one. Once
3078 // the unavailable region becomes available, the data in the shadow region will be copied back.
3079 // Shadow regions are empty regions in the to-space and regions between top and end of other spaces.
3080 void PSParallelCompact::initialize_shadow_regions(uint parallel_gc_threads)
3081 {
3082   const ParallelCompactData& sd = PSParallelCompact::summary_data();
3083 
3084   for (unsigned int id = old_space_id; id < last_space_id; ++id) {
3085     SpaceInfo* const space_info = _space_info + id;
3086     MutableSpace* const space = space_info->space();
3087 
3088     const size_t beg_region =
3089       sd.addr_to_region_idx(sd.region_align_up(MAX2(space_info->new_top(), space->top())));
3090     const size_t end_region =
3091       sd.addr_to_region_idx(sd.region_align_down(space->end()));
3092 
3093     for (size_t cur = beg_region; cur < end_region; ++cur) {
3094       ParCompactionManager::push_shadow_region(cur);
3095     }
3096   }
3097 
3098   size_t beg_region = sd.addr_to_region_idx(_space_info[old_space_id].dense_prefix());
3099   for (uint i = 0; i < parallel_gc_threads; i++) {
3100     ParCompactionManager *cm = ParCompactionManager::gc_thread_compaction_manager(i);
3101     cm->set_next_shadow_region(beg_region + i);
3102   }
3103 }
3104 
3105 void PSParallelCompact::fill_blocks(size_t region_idx)
3106 {
3107   // Fill in the block table elements for the specified region.  Each block
3108   // table element holds the number of live words in the region that are to the
3109   // left of the first object that starts in the block.  Thus only blocks in
3110   // which an object starts need to be filled.
3111   //
3112   // The algorithm scans the section of the bitmap that corresponds to the
3113   // region, keeping a running total of the live words.  When an object start is
3114   // found, if it's the first to start in the block that contains it, the
3115   // current total is written to the block table element.
3116   const size_t Log2BlockSize = ParallelCompactData::Log2BlockSize;
3117   const size_t Log2RegionSize = ParallelCompactData::Log2RegionSize;
3118   const size_t RegionSize = ParallelCompactData::RegionSize;
3119 
3120   ParallelCompactData& sd = summary_data();
3121   const size_t partial_obj_size = sd.region(region_idx)->partial_obj_size();
3122   if (partial_obj_size >= RegionSize) {
3123     return; // No objects start in this region.
3124   }
3125 
3126   // Ensure the first loop iteration decides that the block has changed.
3127   size_t cur_block = sd.block_count();
3128 
3129   const ParMarkBitMap* const bitmap = mark_bitmap();
3130 
3131   const size_t Log2BitsPerBlock = Log2BlockSize - LogMinObjAlignment;
3132   assert((size_t)1 << Log2BitsPerBlock ==
3133          bitmap->words_to_bits(ParallelCompactData::BlockSize), "sanity");
3134 
3135   size_t beg_bit = bitmap->words_to_bits(region_idx << Log2RegionSize);
3136   const size_t range_end = beg_bit + bitmap->words_to_bits(RegionSize);
3137   size_t live_bits = bitmap->words_to_bits(partial_obj_size);
3138   beg_bit = bitmap->find_obj_beg(beg_bit + live_bits, range_end);
3139   while (beg_bit < range_end) {
3140     const size_t new_block = beg_bit >> Log2BitsPerBlock;
3141     if (new_block != cur_block) {
3142       cur_block = new_block;
3143       sd.block(cur_block)->set_offset(bitmap->bits_to_words(live_bits));
3144     }
3145 
3146     const size_t end_bit = bitmap->find_obj_end(beg_bit, range_end);
3147     if (end_bit < range_end - 1) {
3148       live_bits += end_bit - beg_bit + 1;
3149       beg_bit = bitmap->find_obj_beg(end_bit + 1, range_end);
3150     } else {
3151       return;
3152     }
3153   }
3154 }
3155 
3156 ParMarkBitMap::IterationStatus MoveAndUpdateClosure::copy_until_full()
3157 {
3158   if (source() != copy_destination()) {
3159     DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3160     Copy::aligned_conjoint_words(source(), copy_destination(), words_remaining());
3161   }
3162   update_state(words_remaining());
3163   assert(is_full(), "sanity");
3164   return ParMarkBitMap::full;
3165 }
3166 
3167 void MoveAndUpdateClosure::copy_partial_obj()
3168 {
3169   size_t words = words_remaining();
3170 
3171   HeapWord* const range_end = MIN2(source() + words, bitmap()->region_end());
3172   HeapWord* const end_addr = bitmap()->find_obj_end(source(), range_end);
3173   if (end_addr < range_end) {
3174     words = bitmap()->obj_size(source(), end_addr);
3175   }
3176 
3177   // This test is necessary; if omitted, the pointer updates to a partial object
3178   // that crosses the dense prefix boundary could be overwritten.
3179   if (source() != copy_destination()) {
3180     DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3181     Copy::aligned_conjoint_words(source(), copy_destination(), words);
3182   }
3183   update_state(words);
3184 }
3185 
3186 void MoveAndUpdateClosure::complete_region(ParCompactionManager *cm, HeapWord *dest_addr,
3187                                            PSParallelCompact::RegionData *region_ptr) {
3188   assert(region_ptr->shadow_state() == ParallelCompactData::RegionData::NormalRegion, "Region should be finished");
3189   region_ptr->set_completed();
3190 }
3191 
3192 ParMarkBitMapClosure::IterationStatus
3193 MoveAndUpdateClosure::do_addr(HeapWord* addr, size_t words) {
3194   assert(destination() != NULL, "sanity");
3195   assert(bitmap()->obj_size(addr) == words, "bad size");
3196 
3197   _source = addr;
3198   assert(PSParallelCompact::summary_data().calc_new_pointer(source(), compaction_manager()) ==
3199          destination(), "wrong destination");
3200 
3201   if (words > words_remaining()) {
3202     return ParMarkBitMap::would_overflow;
3203   }
3204 
3205   // The start_array must be updated even if the object is not moving.
3206   if (_start_array != NULL) {
3207     _start_array->allocate_block(destination());
3208   }
3209 
3210   if (copy_destination() != source()) {
3211     DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3212     Copy::aligned_conjoint_words(source(), copy_destination(), words);
3213   }
3214 
3215   oop moved_oop = cast_to_oop(copy_destination());
3216   compaction_manager()->update_contents(moved_oop);
3217   assert(oopDesc::is_oop_or_null(moved_oop), "Expected an oop or NULL at " PTR_FORMAT, p2i(moved_oop));
3218 
3219   update_state(words);
3220   assert(copy_destination() == cast_from_oop<HeapWord*>(moved_oop) + moved_oop->size(), "sanity");
3221   return is_full() ? ParMarkBitMap::full : ParMarkBitMap::incomplete;
3222 }
3223 
3224 void MoveAndUpdateShadowClosure::complete_region(ParCompactionManager *cm, HeapWord *dest_addr,
3225                                                  PSParallelCompact::RegionData *region_ptr) {
3226   assert(region_ptr->shadow_state() == ParallelCompactData::RegionData::ShadowRegion, "Region should be shadow");
3227   // Record the shadow region index
3228   region_ptr->set_shadow_region(_shadow);
3229   // Mark the shadow region as filled to indicate the data is ready to be
3230   // copied back
3231   region_ptr->mark_filled();
3232   // Try to copy the content of the shadow region back to its corresponding
3233   // heap region if available; the GC thread that decreases the destination
3234   // count to zero will do the copying otherwise (see
3235   // PSParallelCompact::decrement_destination_counts).
3236   if (((region_ptr->available() && region_ptr->claim()) || region_ptr->claimed()) && region_ptr->mark_copied()) {
3237     region_ptr->set_completed();
3238     PSParallelCompact::copy_back(PSParallelCompact::summary_data().region_to_addr(_shadow), dest_addr);
3239     ParCompactionManager::push_shadow_region_mt_safe(_shadow);
3240   }
3241 }
3242 
3243 UpdateOnlyClosure::UpdateOnlyClosure(ParMarkBitMap* mbm,
3244                                      ParCompactionManager* cm,
3245                                      PSParallelCompact::SpaceId space_id) :
3246   ParMarkBitMapClosure(mbm, cm),
3247   _space_id(space_id),
3248   _start_array(PSParallelCompact::start_array(space_id))
3249 {
3250 }
3251 
3252 // Updates the references in the object to their new values.
3253 ParMarkBitMapClosure::IterationStatus
3254 UpdateOnlyClosure::do_addr(HeapWord* addr, size_t words) {
3255   do_addr(addr);
3256   return ParMarkBitMap::incomplete;
3257 }
3258 
3259 FillClosure::FillClosure(ParCompactionManager* cm, PSParallelCompact::SpaceId space_id) :
3260   ParMarkBitMapClosure(PSParallelCompact::mark_bitmap(), cm),
3261   _start_array(PSParallelCompact::start_array(space_id))
3262 {
3263   assert(space_id == PSParallelCompact::old_space_id,
3264          "cannot use FillClosure in the young gen");
3265 }
3266 
3267 ParMarkBitMapClosure::IterationStatus
3268 FillClosure::do_addr(HeapWord* addr, size_t size) {
3269   CollectedHeap::fill_with_objects(addr, size);
3270   HeapWord* const end = addr + size;
3271   do {
3272     _start_array->allocate_block(addr);
3273     addr += cast_to_oop(addr)->size();
3274   } while (addr < end);
3275   return ParMarkBitMap::incomplete;
3276 }
--- EOF ---