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   CodeCache::increment_marking_cycle();
 989 
 990   // We need to track unique mark sweep invocations as well.
 991   _total_invocations++;
 992 
 993   heap->print_heap_before_gc();
 994   heap->trace_heap_before_gc(&_gc_tracer);
 995 
 996   // Fill in TLABs
 997   heap->ensure_parsability(true);  // retire TLABs
 998 
 999   if (VerifyBeforeGC && heap->total_collections() >= VerifyGCStartAt) {
1000     Universe::verify("Before GC");
1001   }
1002 
1003   // Verify object start arrays
1004   if (VerifyObjectStartArray &&
1005       VerifyBeforeGC) {
1006     heap->old_gen()->verify_object_start_array();
1007   }
1008 
1009   DEBUG_ONLY(mark_bitmap()->verify_clear();)
1010   DEBUG_ONLY(summary_data().verify_clear();)
1011 
1012   ParCompactionManager::reset_all_bitmap_query_caches();
1013 }
1014 
1015 void PSParallelCompact::post_compact()
1016 {
1017   GCTraceTime(Info, gc, phases) tm("Post Compact", &_gc_timer);
1018   ParCompactionManager::remove_all_shadow_regions();
1019 
1020   CodeCache::increment_marking_cycle();
1021 
1022   for (unsigned int id = old_space_id; id < last_space_id; ++id) {
1023     // Clear the marking bitmap, summary data and split info.
1024     clear_data_covering_space(SpaceId(id));
1025     // Update top().  Must be done after clearing the bitmap and summary data.
1026     _space_info[id].publish_new_top();
1027   }
1028 
1029   ParCompactionManager::flush_all_string_dedup_requests();
1030 
1031   MutableSpace* const eden_space = _space_info[eden_space_id].space();
1032   MutableSpace* const from_space = _space_info[from_space_id].space();
1033   MutableSpace* const to_space   = _space_info[to_space_id].space();
1034 
1035   ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
1036   bool eden_empty = eden_space->is_empty();
1037 
1038   // Update heap occupancy information which is used as input to the soft ref
1039   // clearing policy at the next gc.
1040   Universe::heap()->update_capacity_and_used_at_gc();
1041 
1042   bool young_gen_empty = eden_empty && from_space->is_empty() &&
1043     to_space->is_empty();
1044 
1045   PSCardTable* ct = heap->card_table();
1046   MemRegion old_mr = heap->old_gen()->reserved();
1047   if (young_gen_empty) {
1048     ct->clear(MemRegion(old_mr.start(), old_mr.end()));
1049   } else {
1050     ct->invalidate(MemRegion(old_mr.start(), old_mr.end()));
1051   }
1052 
1053   // Delete metaspaces for unloaded class loaders and clean up loader_data graph
1054   ClassLoaderDataGraph::purge(/*at_safepoint*/true);
1055   DEBUG_ONLY(MetaspaceUtils::verify();)
1056 
1057   heap->prune_scavengable_nmethods();
1058 
1059 #if COMPILER2_OR_JVMCI
1060   DerivedPointerTable::update_pointers();
1061 #endif
1062 
1063   if (ZapUnusedHeapArea) {
1064     heap->gen_mangle_unused_area();
1065   }
1066 
1067   // Signal that we have completed a visit to all live objects.
1068   Universe::heap()->record_whole_heap_examined_timestamp();
1069 }
1070 
1071 HeapWord*
1072 PSParallelCompact::compute_dense_prefix_via_density(const SpaceId id,
1073                                                     bool maximum_compaction)
1074 {
1075   const size_t region_size = ParallelCompactData::RegionSize;
1076   const ParallelCompactData& sd = summary_data();
1077 
1078   const MutableSpace* const space = _space_info[id].space();
1079   HeapWord* const top_aligned_up = sd.region_align_up(space->top());
1080   const RegionData* const beg_cp = sd.addr_to_region_ptr(space->bottom());
1081   const RegionData* const end_cp = sd.addr_to_region_ptr(top_aligned_up);
1082 
1083   // Skip full regions at the beginning of the space--they are necessarily part
1084   // of the dense prefix.
1085   size_t full_count = 0;
1086   const RegionData* cp;
1087   for (cp = beg_cp; cp < end_cp && cp->data_size() == region_size; ++cp) {
1088     ++full_count;
1089   }
1090 
1091   assert(total_invocations() >= _maximum_compaction_gc_num, "sanity");
1092   const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num;
1093   const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval;
1094   if (maximum_compaction || cp == end_cp || interval_ended) {
1095     _maximum_compaction_gc_num = total_invocations();
1096     return sd.region_to_addr(cp);
1097   }
1098 
1099   HeapWord* const new_top = _space_info[id].new_top();
1100   const size_t space_live = pointer_delta(new_top, space->bottom());
1101   const size_t space_used = space->used_in_words();
1102   const size_t space_capacity = space->capacity_in_words();
1103 
1104   const double cur_density = double(space_live) / space_capacity;
1105   const double deadwood_density =
1106     (1.0 - cur_density) * (1.0 - cur_density) * cur_density * cur_density;
1107   const size_t deadwood_goal = size_t(space_capacity * deadwood_density);
1108 
1109   log_develop_debug(gc, compaction)(
1110       "cur_dens=%5.3f dw_dens=%5.3f dw_goal=" SIZE_FORMAT,
1111       cur_density, deadwood_density, deadwood_goal);
1112   log_develop_debug(gc, compaction)(
1113       "space_live=" SIZE_FORMAT " space_used=" SIZE_FORMAT " "
1114       "space_cap=" SIZE_FORMAT,
1115       space_live, space_used,
1116       space_capacity);
1117 
1118   // XXX - Use binary search?
1119   HeapWord* dense_prefix = sd.region_to_addr(cp);
1120   const RegionData* full_cp = cp;
1121   const RegionData* const top_cp = sd.addr_to_region_ptr(space->top() - 1);
1122   while (cp < end_cp) {
1123     HeapWord* region_destination = cp->destination();
1124     const size_t cur_deadwood = pointer_delta(dense_prefix, region_destination);
1125 
1126     log_develop_trace(gc, compaction)(
1127         "c#=" SIZE_FORMAT_W(4) " dst=" PTR_FORMAT " "
1128         "dp=" PTR_FORMAT " cdw=" SIZE_FORMAT_W(8),
1129         sd.region(cp), p2i(region_destination),
1130         p2i(dense_prefix), cur_deadwood);
1131 
1132     if (cur_deadwood >= deadwood_goal) {
1133       // Found the region that has the correct amount of deadwood to the left.
1134       // This typically occurs after crossing a fairly sparse set of regions, so
1135       // iterate backwards over those sparse regions, looking for the region
1136       // that has the lowest density of live objects 'to the right.'
1137       size_t space_to_left = sd.region(cp) * region_size;
1138       size_t live_to_left = space_to_left - cur_deadwood;
1139       size_t space_to_right = space_capacity - space_to_left;
1140       size_t live_to_right = space_live - live_to_left;
1141       double density_to_right = double(live_to_right) / space_to_right;
1142       while (cp > full_cp) {
1143         --cp;
1144         const size_t prev_region_live_to_right = live_to_right -
1145           cp->data_size();
1146         const size_t prev_region_space_to_right = space_to_right + region_size;
1147         double prev_region_density_to_right =
1148           double(prev_region_live_to_right) / prev_region_space_to_right;
1149         if (density_to_right <= prev_region_density_to_right) {
1150           return dense_prefix;
1151         }
1152 
1153         log_develop_trace(gc, compaction)(
1154             "backing up from c=" SIZE_FORMAT_W(4) " d2r=%10.8f "
1155             "pc_d2r=%10.8f",
1156             sd.region(cp), density_to_right,
1157             prev_region_density_to_right);
1158 
1159         dense_prefix -= region_size;
1160         live_to_right = prev_region_live_to_right;
1161         space_to_right = prev_region_space_to_right;
1162         density_to_right = prev_region_density_to_right;
1163       }
1164       return dense_prefix;
1165     }
1166 
1167     dense_prefix += region_size;
1168     ++cp;
1169   }
1170 
1171   return dense_prefix;
1172 }
1173 
1174 #ifndef PRODUCT
1175 void PSParallelCompact::print_dense_prefix_stats(const char* const algorithm,
1176                                                  const SpaceId id,
1177                                                  const bool maximum_compaction,
1178                                                  HeapWord* const addr)
1179 {
1180   const size_t region_idx = summary_data().addr_to_region_idx(addr);
1181   RegionData* const cp = summary_data().region(region_idx);
1182   const MutableSpace* const space = _space_info[id].space();
1183   HeapWord* const new_top = _space_info[id].new_top();
1184 
1185   const size_t space_live = pointer_delta(new_top, space->bottom());
1186   const size_t dead_to_left = pointer_delta(addr, cp->destination());
1187   const size_t space_cap = space->capacity_in_words();
1188   const double dead_to_left_pct = double(dead_to_left) / space_cap;
1189   const size_t live_to_right = new_top - cp->destination();
1190   const size_t dead_to_right = space->top() - addr - live_to_right;
1191 
1192   log_develop_debug(gc, compaction)(
1193       "%s=" PTR_FORMAT " dpc=" SIZE_FORMAT_W(5) " "
1194       "spl=" SIZE_FORMAT " "
1195       "d2l=" SIZE_FORMAT " d2l%%=%6.4f "
1196       "d2r=" SIZE_FORMAT " l2r=" SIZE_FORMAT " "
1197       "ratio=%10.8f",
1198       algorithm, p2i(addr), region_idx,
1199       space_live,
1200       dead_to_left, dead_to_left_pct,
1201       dead_to_right, live_to_right,
1202       double(dead_to_right) / live_to_right);
1203 }
1204 #endif  // #ifndef PRODUCT
1205 
1206 // Return a fraction indicating how much of the generation can be treated as
1207 // "dead wood" (i.e., not reclaimed).  The function uses a normal distribution
1208 // based on the density of live objects in the generation to determine a limit,
1209 // which is then adjusted so the return value is min_percent when the density is
1210 // 1.
1211 //
1212 // The following table shows some return values for a different values of the
1213 // standard deviation (ParallelOldDeadWoodLimiterStdDev); the mean is 0.5 and
1214 // min_percent is 1.
1215 //
1216 //                          fraction allowed as dead wood
1217 //         -----------------------------------------------------------------
1218 // density std_dev=70 std_dev=75 std_dev=80 std_dev=85 std_dev=90 std_dev=95
1219 // ------- ---------- ---------- ---------- ---------- ---------- ----------
1220 // 0.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000
1221 // 0.05000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941
1222 // 0.10000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272
1223 // 0.15000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066
1224 // 0.20000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975
1225 // 0.25000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313
1226 // 0.30000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132
1227 // 0.35000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289
1228 // 0.40000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500
1229 // 0.45000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386
1230 // 0.50000 0.13832410 0.11599237 0.09847664 0.08456518 0.07338887 0.06431510
1231 // 0.55000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386
1232 // 0.60000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500
1233 // 0.65000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289
1234 // 0.70000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132
1235 // 0.75000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313
1236 // 0.80000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975
1237 // 0.85000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066
1238 // 0.90000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272
1239 // 0.95000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941
1240 // 1.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000
1241 
1242 double PSParallelCompact::dead_wood_limiter(double density, size_t min_percent)
1243 {
1244   assert(_dwl_initialized, "uninitialized");
1245 
1246   // The raw limit is the value of the normal distribution at x = density.
1247   const double raw_limit = normal_distribution(density);
1248 
1249   // Adjust the raw limit so it becomes the minimum when the density is 1.
1250   //
1251   // First subtract the adjustment value (which is simply the precomputed value
1252   // normal_distribution(1.0)); this yields a value of 0 when the density is 1.
1253   // Then add the minimum value, so the minimum is returned when the density is
1254   // 1.  Finally, prevent negative values, which occur when the mean is not 0.5.
1255   const double min = double(min_percent) / 100.0;
1256   const double limit = raw_limit - _dwl_adjustment + min;
1257   return MAX2(limit, 0.0);
1258 }
1259 
1260 ParallelCompactData::RegionData*
1261 PSParallelCompact::first_dead_space_region(const RegionData* beg,
1262                                            const RegionData* end)
1263 {
1264   const size_t region_size = ParallelCompactData::RegionSize;
1265   ParallelCompactData& sd = summary_data();
1266   size_t left = sd.region(beg);
1267   size_t right = end > beg ? sd.region(end) - 1 : left;
1268 
1269   // Binary search.
1270   while (left < right) {
1271     // Equivalent to (left + right) / 2, but does not overflow.
1272     const size_t middle = left + (right - left) / 2;
1273     RegionData* const middle_ptr = sd.region(middle);
1274     HeapWord* const dest = middle_ptr->destination();
1275     HeapWord* const addr = sd.region_to_addr(middle);
1276     assert(dest != NULL, "sanity");
1277     assert(dest <= addr, "must move left");
1278 
1279     if (middle > left && dest < addr) {
1280       right = middle - 1;
1281     } else if (middle < right && middle_ptr->data_size() == region_size) {
1282       left = middle + 1;
1283     } else {
1284       return middle_ptr;
1285     }
1286   }
1287   return sd.region(left);
1288 }
1289 
1290 ParallelCompactData::RegionData*
1291 PSParallelCompact::dead_wood_limit_region(const RegionData* beg,
1292                                           const RegionData* end,
1293                                           size_t dead_words)
1294 {
1295   ParallelCompactData& sd = summary_data();
1296   size_t left = sd.region(beg);
1297   size_t right = end > beg ? sd.region(end) - 1 : left;
1298 
1299   // Binary search.
1300   while (left < right) {
1301     // Equivalent to (left + right) / 2, but does not overflow.
1302     const size_t middle = left + (right - left) / 2;
1303     RegionData* const middle_ptr = sd.region(middle);
1304     HeapWord* const dest = middle_ptr->destination();
1305     HeapWord* const addr = sd.region_to_addr(middle);
1306     assert(dest != NULL, "sanity");
1307     assert(dest <= addr, "must move left");
1308 
1309     const size_t dead_to_left = pointer_delta(addr, dest);
1310     if (middle > left && dead_to_left > dead_words) {
1311       right = middle - 1;
1312     } else if (middle < right && dead_to_left < dead_words) {
1313       left = middle + 1;
1314     } else {
1315       return middle_ptr;
1316     }
1317   }
1318   return sd.region(left);
1319 }
1320 
1321 // The result is valid during the summary phase, after the initial summarization
1322 // of each space into itself, and before final summarization.
1323 inline double
1324 PSParallelCompact::reclaimed_ratio(const RegionData* const cp,
1325                                    HeapWord* const bottom,
1326                                    HeapWord* const top,
1327                                    HeapWord* const new_top)
1328 {
1329   ParallelCompactData& sd = summary_data();
1330 
1331   assert(cp != NULL, "sanity");
1332   assert(bottom != NULL, "sanity");
1333   assert(top != NULL, "sanity");
1334   assert(new_top != NULL, "sanity");
1335   assert(top >= new_top, "summary data problem?");
1336   assert(new_top > bottom, "space is empty; should not be here");
1337   assert(new_top >= cp->destination(), "sanity");
1338   assert(top >= sd.region_to_addr(cp), "sanity");
1339 
1340   HeapWord* const destination = cp->destination();
1341   const size_t dense_prefix_live  = pointer_delta(destination, bottom);
1342   const size_t compacted_region_live = pointer_delta(new_top, destination);
1343   const size_t compacted_region_used = pointer_delta(top,
1344                                                      sd.region_to_addr(cp));
1345   const size_t reclaimable = compacted_region_used - compacted_region_live;
1346 
1347   const double divisor = dense_prefix_live + 1.25 * compacted_region_live;
1348   return double(reclaimable) / divisor;
1349 }
1350 
1351 // Return the address of the end of the dense prefix, a.k.a. the start of the
1352 // compacted region.  The address is always on a region boundary.
1353 //
1354 // Completely full regions at the left are skipped, since no compaction can
1355 // occur in those regions.  Then the maximum amount of dead wood to allow is
1356 // computed, based on the density (amount live / capacity) of the generation;
1357 // the region with approximately that amount of dead space to the left is
1358 // identified as the limit region.  Regions between the last completely full
1359 // region and the limit region are scanned and the one that has the best
1360 // (maximum) reclaimed_ratio() is selected.
1361 HeapWord*
1362 PSParallelCompact::compute_dense_prefix(const SpaceId id,
1363                                         bool maximum_compaction)
1364 {
1365   const size_t region_size = ParallelCompactData::RegionSize;
1366   const ParallelCompactData& sd = summary_data();
1367 
1368   const MutableSpace* const space = _space_info[id].space();
1369   HeapWord* const top = space->top();
1370   HeapWord* const top_aligned_up = sd.region_align_up(top);
1371   HeapWord* const new_top = _space_info[id].new_top();
1372   HeapWord* const new_top_aligned_up = sd.region_align_up(new_top);
1373   HeapWord* const bottom = space->bottom();
1374   const RegionData* const beg_cp = sd.addr_to_region_ptr(bottom);
1375   const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up);
1376   const RegionData* const new_top_cp =
1377     sd.addr_to_region_ptr(new_top_aligned_up);
1378 
1379   // Skip full regions at the beginning of the space--they are necessarily part
1380   // of the dense prefix.
1381   const RegionData* const full_cp = first_dead_space_region(beg_cp, new_top_cp);
1382   assert(full_cp->destination() == sd.region_to_addr(full_cp) ||
1383          space->is_empty(), "no dead space allowed to the left");
1384   assert(full_cp->data_size() < region_size || full_cp == new_top_cp - 1,
1385          "region must have dead space");
1386 
1387   // The gc number is saved whenever a maximum compaction is done, and used to
1388   // determine when the maximum compaction interval has expired.  This avoids
1389   // successive max compactions for different reasons.
1390   assert(total_invocations() >= _maximum_compaction_gc_num, "sanity");
1391   const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num;
1392   const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval ||
1393     total_invocations() == HeapFirstMaximumCompactionCount;
1394   if (maximum_compaction || full_cp == top_cp || interval_ended) {
1395     _maximum_compaction_gc_num = total_invocations();
1396     return sd.region_to_addr(full_cp);
1397   }
1398 
1399   const size_t space_live = pointer_delta(new_top, bottom);
1400   const size_t space_used = space->used_in_words();
1401   const size_t space_capacity = space->capacity_in_words();
1402 
1403   const double density = double(space_live) / double(space_capacity);
1404   const size_t min_percent_free = MarkSweepDeadRatio;
1405   const double limiter = dead_wood_limiter(density, min_percent_free);
1406   const size_t dead_wood_max = space_used - space_live;
1407   const size_t dead_wood_limit = MIN2(size_t(space_capacity * limiter),
1408                                       dead_wood_max);
1409 
1410   log_develop_debug(gc, compaction)(
1411       "space_live=" SIZE_FORMAT " space_used=" SIZE_FORMAT " "
1412       "space_cap=" SIZE_FORMAT,
1413       space_live, space_used,
1414       space_capacity);
1415   log_develop_debug(gc, compaction)(
1416       "dead_wood_limiter(%6.4f, " SIZE_FORMAT ")=%6.4f "
1417       "dead_wood_max=" SIZE_FORMAT " dead_wood_limit=" SIZE_FORMAT,
1418       density, min_percent_free, limiter,
1419       dead_wood_max, dead_wood_limit);
1420 
1421   // Locate the region with the desired amount of dead space to the left.
1422   const RegionData* const limit_cp =
1423     dead_wood_limit_region(full_cp, top_cp, dead_wood_limit);
1424 
1425   // Scan from the first region with dead space to the limit region and find the
1426   // one with the best (largest) reclaimed ratio.
1427   double best_ratio = 0.0;
1428   const RegionData* best_cp = full_cp;
1429   for (const RegionData* cp = full_cp; cp < limit_cp; ++cp) {
1430     double tmp_ratio = reclaimed_ratio(cp, bottom, top, new_top);
1431     if (tmp_ratio > best_ratio) {
1432       best_cp = cp;
1433       best_ratio = tmp_ratio;
1434     }
1435   }
1436 
1437   return sd.region_to_addr(best_cp);
1438 }
1439 
1440 void PSParallelCompact::summarize_spaces_quick()
1441 {
1442   for (unsigned int i = 0; i < last_space_id; ++i) {
1443     const MutableSpace* space = _space_info[i].space();
1444     HeapWord** nta = _space_info[i].new_top_addr();
1445     bool result = _summary_data.summarize(_space_info[i].split_info(),
1446                                           space->bottom(), space->top(), NULL,
1447                                           space->bottom(), space->end(), nta);
1448     assert(result, "space must fit into itself");
1449     _space_info[i].set_dense_prefix(space->bottom());
1450   }
1451 }
1452 
1453 void PSParallelCompact::fill_dense_prefix_end(SpaceId id)
1454 {
1455   HeapWord* const dense_prefix_end = dense_prefix(id);
1456   const RegionData* region = _summary_data.addr_to_region_ptr(dense_prefix_end);
1457   const idx_t dense_prefix_bit = _mark_bitmap.addr_to_bit(dense_prefix_end);
1458   if (dead_space_crosses_boundary(region, dense_prefix_bit)) {
1459     // Only enough dead space is filled so that any remaining dead space to the
1460     // left is larger than the minimum filler object.  (The remainder is filled
1461     // during the copy/update phase.)
1462     //
1463     // The size of the dead space to the right of the boundary is not a
1464     // concern, since compaction will be able to use whatever space is
1465     // available.
1466     //
1467     // Here '||' is the boundary, 'x' represents a don't care bit and a box
1468     // surrounds the space to be filled with an object.
1469     //
1470     // In the 32-bit VM, each bit represents two 32-bit words:
1471     //                              +---+
1472     // a) beg_bits:  ...  x   x   x | 0 | ||   0   x  x  ...
1473     //    end_bits:  ...  x   x   x | 0 | ||   0   x  x  ...
1474     //                              +---+
1475     //
1476     // In the 64-bit VM, each bit represents one 64-bit word:
1477     //                              +------------+
1478     // b) beg_bits:  ...  x   x   x | 0   ||   0 | x  x  ...
1479     //    end_bits:  ...  x   x   1 | 0   ||   0 | x  x  ...
1480     //                              +------------+
1481     //                          +-------+
1482     // c) beg_bits:  ...  x   x | 0   0 | ||   0   x  x  ...
1483     //    end_bits:  ...  x   1 | 0   0 | ||   0   x  x  ...
1484     //                          +-------+
1485     //                      +-----------+
1486     // d) beg_bits:  ...  x | 0   0   0 | ||   0   x  x  ...
1487     //    end_bits:  ...  1 | 0   0   0 | ||   0   x  x  ...
1488     //                      +-----------+
1489     //                          +-------+
1490     // e) beg_bits:  ...  0   0 | 0   0 | ||   0   x  x  ...
1491     //    end_bits:  ...  0   0 | 0   0 | ||   0   x  x  ...
1492     //                          +-------+
1493 
1494     // Initially assume case a, c or e will apply.
1495     size_t obj_len = CollectedHeap::min_fill_size();
1496     HeapWord* obj_beg = dense_prefix_end - obj_len;
1497 
1498 #ifdef  _LP64
1499     if (MinObjAlignment > 1) { // object alignment > heap word size
1500       // Cases a, c or e.
1501     } else if (_mark_bitmap.is_obj_end(dense_prefix_bit - 2)) {
1502       // Case b above.
1503       obj_beg = dense_prefix_end - 1;
1504     } else if (!_mark_bitmap.is_obj_end(dense_prefix_bit - 3) &&
1505                _mark_bitmap.is_obj_end(dense_prefix_bit - 4)) {
1506       // Case d above.
1507       obj_beg = dense_prefix_end - 3;
1508       obj_len = 3;
1509     }
1510 #endif  // #ifdef _LP64
1511 
1512     CollectedHeap::fill_with_object(obj_beg, obj_len);
1513     _mark_bitmap.mark_obj(obj_beg, obj_len);
1514     _summary_data.add_obj(obj_beg, obj_len);
1515     assert(start_array(id) != NULL, "sanity");
1516     start_array(id)->allocate_block(obj_beg);
1517   }
1518 }
1519 
1520 void
1521 PSParallelCompact::summarize_space(SpaceId id, bool maximum_compaction)
1522 {
1523   assert(id < last_space_id, "id out of range");
1524   assert(_space_info[id].dense_prefix() == _space_info[id].space()->bottom(),
1525          "should have been reset in summarize_spaces_quick()");
1526 
1527   const MutableSpace* space = _space_info[id].space();
1528   if (_space_info[id].new_top() != space->bottom()) {
1529     HeapWord* dense_prefix_end = compute_dense_prefix(id, maximum_compaction);
1530     _space_info[id].set_dense_prefix(dense_prefix_end);
1531 
1532 #ifndef PRODUCT
1533     if (log_is_enabled(Debug, gc, compaction)) {
1534       print_dense_prefix_stats("ratio", id, maximum_compaction,
1535                                dense_prefix_end);
1536       HeapWord* addr = compute_dense_prefix_via_density(id, maximum_compaction);
1537       print_dense_prefix_stats("density", id, maximum_compaction, addr);
1538     }
1539 #endif  // #ifndef PRODUCT
1540 
1541     // Recompute the summary data, taking into account the dense prefix.  If
1542     // every last byte will be reclaimed, then the existing summary data which
1543     // compacts everything can be left in place.
1544     if (!maximum_compaction && dense_prefix_end != space->bottom()) {
1545       // If dead space crosses the dense prefix boundary, it is (at least
1546       // partially) filled with a dummy object, marked live and added to the
1547       // summary data.  This simplifies the copy/update phase and must be done
1548       // before the final locations of objects are determined, to prevent
1549       // leaving a fragment of dead space that is too small to fill.
1550       fill_dense_prefix_end(id);
1551 
1552       // Compute the destination of each Region, and thus each object.
1553       _summary_data.summarize_dense_prefix(space->bottom(), dense_prefix_end);
1554       _summary_data.summarize(_space_info[id].split_info(),
1555                               dense_prefix_end, space->top(), NULL,
1556                               dense_prefix_end, space->end(),
1557                               _space_info[id].new_top_addr());
1558     }
1559   }
1560 
1561   if (log_develop_is_enabled(Trace, gc, compaction)) {
1562     const size_t region_size = ParallelCompactData::RegionSize;
1563     HeapWord* const dense_prefix_end = _space_info[id].dense_prefix();
1564     const size_t dp_region = _summary_data.addr_to_region_idx(dense_prefix_end);
1565     const size_t dp_words = pointer_delta(dense_prefix_end, space->bottom());
1566     HeapWord* const new_top = _space_info[id].new_top();
1567     const HeapWord* nt_aligned_up = _summary_data.region_align_up(new_top);
1568     const size_t cr_words = pointer_delta(nt_aligned_up, dense_prefix_end);
1569     log_develop_trace(gc, compaction)(
1570         "id=%d cap=" SIZE_FORMAT " dp=" PTR_FORMAT " "
1571         "dp_region=" SIZE_FORMAT " " "dp_count=" SIZE_FORMAT " "
1572         "cr_count=" SIZE_FORMAT " " "nt=" PTR_FORMAT,
1573         id, space->capacity_in_words(), p2i(dense_prefix_end),
1574         dp_region, dp_words / region_size,
1575         cr_words / region_size, p2i(new_top));
1576   }
1577 }
1578 
1579 #ifndef PRODUCT
1580 void PSParallelCompact::summary_phase_msg(SpaceId dst_space_id,
1581                                           HeapWord* dst_beg, HeapWord* dst_end,
1582                                           SpaceId src_space_id,
1583                                           HeapWord* src_beg, HeapWord* src_end)
1584 {
1585   log_develop_trace(gc, compaction)(
1586       "Summarizing %d [%s] into %d [%s]:  "
1587       "src=" PTR_FORMAT "-" PTR_FORMAT " "
1588       SIZE_FORMAT "-" SIZE_FORMAT " "
1589       "dst=" PTR_FORMAT "-" PTR_FORMAT " "
1590       SIZE_FORMAT "-" SIZE_FORMAT,
1591       src_space_id, space_names[src_space_id],
1592       dst_space_id, space_names[dst_space_id],
1593       p2i(src_beg), p2i(src_end),
1594       _summary_data.addr_to_region_idx(src_beg),
1595       _summary_data.addr_to_region_idx(src_end),
1596       p2i(dst_beg), p2i(dst_end),
1597       _summary_data.addr_to_region_idx(dst_beg),
1598       _summary_data.addr_to_region_idx(dst_end));
1599 }
1600 #endif  // #ifndef PRODUCT
1601 
1602 void PSParallelCompact::summary_phase(ParCompactionManager* cm,
1603                                       bool maximum_compaction)
1604 {
1605   GCTraceTime(Info, gc, phases) tm("Summary Phase", &_gc_timer);
1606 
1607   // Quick summarization of each space into itself, to see how much is live.
1608   summarize_spaces_quick();
1609 
1610   log_develop_trace(gc, compaction)("summary phase:  after summarizing each space to self");
1611   NOT_PRODUCT(print_region_ranges());
1612   NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info));
1613 
1614   // The amount of live data that will end up in old space (assuming it fits).
1615   size_t old_space_total_live = 0;
1616   for (unsigned int id = old_space_id; id < last_space_id; ++id) {
1617     old_space_total_live += pointer_delta(_space_info[id].new_top(),
1618                                           _space_info[id].space()->bottom());
1619   }
1620 
1621   MutableSpace* const old_space = _space_info[old_space_id].space();
1622   const size_t old_capacity = old_space->capacity_in_words();
1623   if (old_space_total_live > old_capacity) {
1624     // XXX - should also try to expand
1625     maximum_compaction = true;
1626   }
1627 
1628   // Old generations.
1629   summarize_space(old_space_id, maximum_compaction);
1630 
1631   // Summarize the remaining spaces in the young gen.  The initial target space
1632   // is the old gen.  If a space does not fit entirely into the target, then the
1633   // remainder is compacted into the space itself and that space becomes the new
1634   // target.
1635   SpaceId dst_space_id = old_space_id;
1636   HeapWord* dst_space_end = old_space->end();
1637   HeapWord** new_top_addr = _space_info[dst_space_id].new_top_addr();
1638   for (unsigned int id = eden_space_id; id < last_space_id; ++id) {
1639     const MutableSpace* space = _space_info[id].space();
1640     const size_t live = pointer_delta(_space_info[id].new_top(),
1641                                       space->bottom());
1642     const size_t available = pointer_delta(dst_space_end, *new_top_addr);
1643 
1644     NOT_PRODUCT(summary_phase_msg(dst_space_id, *new_top_addr, dst_space_end,
1645                                   SpaceId(id), space->bottom(), space->top());)
1646     if (live > 0 && live <= available) {
1647       // All the live data will fit.
1648       bool done = _summary_data.summarize(_space_info[id].split_info(),
1649                                           space->bottom(), space->top(),
1650                                           NULL,
1651                                           *new_top_addr, dst_space_end,
1652                                           new_top_addr);
1653       assert(done, "space must fit into old gen");
1654 
1655       // Reset the new_top value for the space.
1656       _space_info[id].set_new_top(space->bottom());
1657     } else if (live > 0) {
1658       // Attempt to fit part of the source space into the target space.
1659       HeapWord* next_src_addr = NULL;
1660       bool done = _summary_data.summarize(_space_info[id].split_info(),
1661                                           space->bottom(), space->top(),
1662                                           &next_src_addr,
1663                                           *new_top_addr, dst_space_end,
1664                                           new_top_addr);
1665       assert(!done, "space should not fit into old gen");
1666       assert(next_src_addr != NULL, "sanity");
1667 
1668       // The source space becomes the new target, so the remainder is compacted
1669       // within the space itself.
1670       dst_space_id = SpaceId(id);
1671       dst_space_end = space->end();
1672       new_top_addr = _space_info[id].new_top_addr();
1673       NOT_PRODUCT(summary_phase_msg(dst_space_id,
1674                                     space->bottom(), dst_space_end,
1675                                     SpaceId(id), next_src_addr, space->top());)
1676       done = _summary_data.summarize(_space_info[id].split_info(),
1677                                      next_src_addr, space->top(),
1678                                      NULL,
1679                                      space->bottom(), dst_space_end,
1680                                      new_top_addr);
1681       assert(done, "space must fit when compacted into itself");
1682       assert(*new_top_addr <= space->top(), "usage should not grow");
1683     }
1684   }
1685 
1686   log_develop_trace(gc, compaction)("Summary_phase:  after final summarization");
1687   NOT_PRODUCT(print_region_ranges());
1688   NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info));
1689 }
1690 
1691 // This method should contain all heap-specific policy for invoking a full
1692 // collection.  invoke_no_policy() will only attempt to compact the heap; it
1693 // will do nothing further.  If we need to bail out for policy reasons, scavenge
1694 // before full gc, or any other specialized behavior, it needs to be added here.
1695 //
1696 // Note that this method should only be called from the vm_thread while at a
1697 // safepoint.
1698 //
1699 // Note that the all_soft_refs_clear flag in the soft ref policy
1700 // may be true because this method can be called without intervening
1701 // activity.  For example when the heap space is tight and full measure
1702 // are being taken to free space.
1703 void PSParallelCompact::invoke(bool maximum_heap_compaction) {
1704   assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
1705   assert(Thread::current() == (Thread*)VMThread::vm_thread(),
1706          "should be in vm thread");
1707 
1708   ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
1709   GCCause::Cause gc_cause = heap->gc_cause();
1710   assert(!heap->is_gc_active(), "not reentrant");
1711 
1712   PSAdaptiveSizePolicy* policy = heap->size_policy();
1713   IsGCActiveMark mark;
1714 
1715   if (ScavengeBeforeFullGC) {
1716     PSScavenge::invoke_no_policy();
1717   }
1718 
1719   const bool clear_all_soft_refs =
1720     heap->soft_ref_policy()->should_clear_all_soft_refs();
1721 
1722   PSParallelCompact::invoke_no_policy(clear_all_soft_refs ||
1723                                       maximum_heap_compaction);
1724 }
1725 
1726 // This method contains no policy. You should probably
1727 // be calling invoke() instead.
1728 bool PSParallelCompact::invoke_no_policy(bool maximum_heap_compaction) {
1729   assert(SafepointSynchronize::is_at_safepoint(), "must be at a safepoint");
1730   assert(ref_processor() != NULL, "Sanity");
1731 
1732   if (GCLocker::check_active_before_gc()) {
1733     return false;
1734   }
1735 
1736   ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
1737 
1738   GCIdMark gc_id_mark;
1739   _gc_timer.register_gc_start();
1740   _gc_tracer.report_gc_start(heap->gc_cause(), _gc_timer.gc_start());
1741 
1742   TimeStamp marking_start;
1743   TimeStamp compaction_start;
1744   TimeStamp collection_exit;
1745 
1746   GCCause::Cause gc_cause = heap->gc_cause();
1747   PSYoungGen* young_gen = heap->young_gen();
1748   PSOldGen* old_gen = heap->old_gen();
1749   PSAdaptiveSizePolicy* size_policy = heap->size_policy();
1750 
1751   // The scope of casr should end after code that can change
1752   // SoftRefPolicy::_should_clear_all_soft_refs.
1753   ClearedAllSoftRefs casr(maximum_heap_compaction,
1754                           heap->soft_ref_policy());
1755 
1756   if (ZapUnusedHeapArea) {
1757     // Save information needed to minimize mangling
1758     heap->record_gen_tops_before_GC();
1759   }
1760 
1761   // Make sure data structures are sane, make the heap parsable, and do other
1762   // miscellaneous bookkeeping.
1763   pre_compact();
1764 
1765   const PreGenGCValues pre_gc_values = heap->get_pre_gc_values();
1766 
1767   // Get the compaction manager reserved for the VM thread.
1768   ParCompactionManager* const vmthread_cm = ParCompactionManager::get_vmthread_cm();
1769 
1770   {
1771     const uint active_workers =
1772       WorkerPolicy::calc_active_workers(ParallelScavengeHeap::heap()->workers().max_workers(),
1773                                         ParallelScavengeHeap::heap()->workers().active_workers(),
1774                                         Threads::number_of_non_daemon_threads());
1775     ParallelScavengeHeap::heap()->workers().set_active_workers(active_workers);
1776 
1777     GCTraceCPUTime tcpu;
1778     GCTraceTime(Info, gc) tm("Pause Full", NULL, gc_cause, true);
1779 
1780     heap->pre_full_gc_dump(&_gc_timer);
1781 
1782     TraceCollectorStats tcs(counters());
1783     TraceMemoryManagerStats tms(heap->old_gc_manager(), gc_cause);
1784 
1785     if (log_is_enabled(Debug, gc, heap, exit)) {
1786       accumulated_time()->start();
1787     }
1788 
1789     // Let the size policy know we're starting
1790     size_policy->major_collection_begin();
1791 
1792 #if COMPILER2_OR_JVMCI
1793     DerivedPointerTable::clear();
1794 #endif
1795 
1796     ref_processor()->start_discovery(maximum_heap_compaction);
1797 
1798     marking_start.update();
1799     marking_phase(vmthread_cm, &_gc_tracer);
1800 
1801     bool max_on_system_gc = UseMaximumCompactionOnSystemGC
1802       && GCCause::is_user_requested_gc(gc_cause);
1803     summary_phase(vmthread_cm, maximum_heap_compaction || max_on_system_gc);
1804 
1805 #if COMPILER2_OR_JVMCI
1806     assert(DerivedPointerTable::is_active(), "Sanity");
1807     DerivedPointerTable::set_active(false);
1808 #endif
1809 
1810     // adjust_roots() updates Universe::_intArrayKlassObj which is
1811     // needed by the compaction for filling holes in the dense prefix.
1812     adjust_roots();
1813 
1814     compaction_start.update();
1815     compact();
1816 
1817     ParCompactionManager::verify_all_region_stack_empty();
1818 
1819     // Reset the mark bitmap, summary data, and do other bookkeeping.  Must be
1820     // done before resizing.
1821     post_compact();
1822 
1823     // Let the size policy know we're done
1824     size_policy->major_collection_end(old_gen->used_in_bytes(), gc_cause);
1825 
1826     if (UseAdaptiveSizePolicy) {
1827       log_debug(gc, ergo)("AdaptiveSizeStart: collection: %d ", heap->total_collections());
1828       log_trace(gc, ergo)("old_gen_capacity: " SIZE_FORMAT " young_gen_capacity: " SIZE_FORMAT,
1829                           old_gen->capacity_in_bytes(), young_gen->capacity_in_bytes());
1830 
1831       // Don't check if the size_policy is ready here.  Let
1832       // the size_policy check that internally.
1833       if (UseAdaptiveGenerationSizePolicyAtMajorCollection &&
1834           AdaptiveSizePolicy::should_update_promo_stats(gc_cause)) {
1835         // Swap the survivor spaces if from_space is empty. The
1836         // resize_young_gen() called below is normally used after
1837         // a successful young GC and swapping of survivor spaces;
1838         // otherwise, it will fail to resize the young gen with
1839         // the current implementation.
1840         if (young_gen->from_space()->is_empty()) {
1841           young_gen->from_space()->clear(SpaceDecorator::Mangle);
1842           young_gen->swap_spaces();
1843         }
1844 
1845         // Calculate optimal free space amounts
1846         assert(young_gen->max_gen_size() >
1847           young_gen->from_space()->capacity_in_bytes() +
1848           young_gen->to_space()->capacity_in_bytes(),
1849           "Sizes of space in young gen are out-of-bounds");
1850 
1851         size_t young_live = young_gen->used_in_bytes();
1852         size_t eden_live = young_gen->eden_space()->used_in_bytes();
1853         size_t old_live = old_gen->used_in_bytes();
1854         size_t cur_eden = young_gen->eden_space()->capacity_in_bytes();
1855         size_t max_old_gen_size = old_gen->max_gen_size();
1856         size_t max_eden_size = young_gen->max_gen_size() -
1857           young_gen->from_space()->capacity_in_bytes() -
1858           young_gen->to_space()->capacity_in_bytes();
1859 
1860         // Used for diagnostics
1861         size_policy->clear_generation_free_space_flags();
1862 
1863         size_policy->compute_generations_free_space(young_live,
1864                                                     eden_live,
1865                                                     old_live,
1866                                                     cur_eden,
1867                                                     max_old_gen_size,
1868                                                     max_eden_size,
1869                                                     true /* full gc*/);
1870 
1871         size_policy->check_gc_overhead_limit(eden_live,
1872                                              max_old_gen_size,
1873                                              max_eden_size,
1874                                              true /* full gc*/,
1875                                              gc_cause,
1876                                              heap->soft_ref_policy());
1877 
1878         size_policy->decay_supplemental_growth(true /* full gc*/);
1879 
1880         heap->resize_old_gen(
1881           size_policy->calculated_old_free_size_in_bytes());
1882 
1883         heap->resize_young_gen(size_policy->calculated_eden_size_in_bytes(),
1884                                size_policy->calculated_survivor_size_in_bytes());
1885       }
1886 
1887       log_debug(gc, ergo)("AdaptiveSizeStop: collection: %d ", heap->total_collections());
1888     }
1889 
1890     if (UsePerfData) {
1891       PSGCAdaptivePolicyCounters* const counters = heap->gc_policy_counters();
1892       counters->update_counters();
1893       counters->update_old_capacity(old_gen->capacity_in_bytes());
1894       counters->update_young_capacity(young_gen->capacity_in_bytes());
1895     }
1896 
1897     heap->resize_all_tlabs();
1898 
1899     // Resize the metaspace capacity after a collection
1900     MetaspaceGC::compute_new_size();
1901 
1902     if (log_is_enabled(Debug, gc, heap, exit)) {
1903       accumulated_time()->stop();
1904     }
1905 
1906     heap->print_heap_change(pre_gc_values);
1907 
1908     // Track memory usage and detect low memory
1909     MemoryService::track_memory_usage();
1910     heap->update_counters();
1911 
1912     heap->post_full_gc_dump(&_gc_timer);
1913   }
1914 
1915   if (VerifyAfterGC && heap->total_collections() >= VerifyGCStartAt) {
1916     Universe::verify("After GC");
1917   }
1918 
1919   // Re-verify object start arrays
1920   if (VerifyObjectStartArray &&
1921       VerifyAfterGC) {
1922     old_gen->verify_object_start_array();
1923   }
1924 
1925   if (ZapUnusedHeapArea) {
1926     old_gen->object_space()->check_mangled_unused_area_complete();
1927   }
1928 
1929   collection_exit.update();
1930 
1931   heap->print_heap_after_gc();
1932   heap->trace_heap_after_gc(&_gc_tracer);
1933 
1934   log_debug(gc, task, time)("VM-Thread " JLONG_FORMAT " " JLONG_FORMAT " " JLONG_FORMAT,
1935                          marking_start.ticks(), compaction_start.ticks(),
1936                          collection_exit.ticks());
1937 
1938   AdaptiveSizePolicyOutput::print(size_policy, heap->total_collections());
1939 
1940   _gc_timer.register_gc_end();
1941 
1942   _gc_tracer.report_dense_prefix(dense_prefix(old_space_id));
1943   _gc_tracer.report_gc_end(_gc_timer.gc_end(), _gc_timer.time_partitions());
1944 
1945   return true;
1946 }
1947 
1948 class PCAddThreadRootsMarkingTaskClosure : public ThreadClosure {
1949 private:
1950   uint _worker_id;
1951 
1952 public:
1953   PCAddThreadRootsMarkingTaskClosure(uint worker_id) : _worker_id(worker_id) { }
1954   void do_thread(Thread* thread) {
1955     assert(ParallelScavengeHeap::heap()->is_gc_active(), "called outside gc");
1956 
1957     ResourceMark rm;
1958 
1959     ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(_worker_id);
1960 
1961     PCMarkAndPushClosure mark_and_push_closure(cm);
1962     MarkingCodeBlobClosure mark_and_push_in_blobs(&mark_and_push_closure, !CodeBlobToOopClosure::FixRelocations, true /* keepalive nmethods */);
1963 
1964     thread->oops_do(&mark_and_push_closure, &mark_and_push_in_blobs);
1965 
1966     // Do the real work
1967     cm->follow_marking_stacks();
1968   }
1969 };
1970 
1971 static void mark_from_roots_work(ParallelRootType::Value root_type, uint worker_id) {
1972   assert(ParallelScavengeHeap::heap()->is_gc_active(), "called outside gc");
1973 
1974   ParCompactionManager* cm =
1975     ParCompactionManager::gc_thread_compaction_manager(worker_id);
1976   PCMarkAndPushClosure mark_and_push_closure(cm);
1977 
1978   switch (root_type) {
1979     case ParallelRootType::class_loader_data:
1980       {
1981         CLDToOopClosure cld_closure(&mark_and_push_closure, ClassLoaderData::_claim_strong);
1982         ClassLoaderDataGraph::always_strong_cld_do(&cld_closure);
1983       }
1984       break;
1985 
1986     case ParallelRootType::code_cache:
1987       // Do not treat nmethods as strong roots for mark/sweep, since we can unload them.
1988       //ScavengableNMethods::scavengable_nmethods_do(CodeBlobToOopClosure(&mark_and_push_closure));
1989       break;
1990 
1991     case ParallelRootType::sentinel:
1992     DEBUG_ONLY(default:) // DEBUG_ONLY hack will create compile error on release builds (-Wswitch) and runtime check on debug builds
1993       fatal("Bad enumeration value: %u", root_type);
1994       break;
1995   }
1996 
1997   // Do the real work
1998   cm->follow_marking_stacks();
1999 }
2000 
2001 void steal_marking_work(TaskTerminator& terminator, uint worker_id) {
2002   assert(ParallelScavengeHeap::heap()->is_gc_active(), "called outside gc");
2003 
2004   ParCompactionManager* cm =
2005     ParCompactionManager::gc_thread_compaction_manager(worker_id);
2006 
2007   oop obj = NULL;
2008   ObjArrayTask task;
2009   do {
2010     while (ParCompactionManager::steal_objarray(worker_id,  task)) {
2011       cm->follow_array((objArrayOop)task.obj(), task.index());
2012       cm->follow_marking_stacks();
2013     }
2014     while (ParCompactionManager::steal(worker_id, obj)) {
2015       cm->follow_contents(obj);
2016       cm->follow_marking_stacks();
2017     }
2018   } while (!terminator.offer_termination());
2019 }
2020 
2021 class MarkFromRootsTask : public WorkerTask {
2022   StrongRootsScope _strong_roots_scope; // needed for Threads::possibly_parallel_threads_do
2023   OopStorageSetStrongParState<false /* concurrent */, false /* is_const */> _oop_storage_set_par_state;
2024   SequentialSubTasksDone _subtasks;
2025   TaskTerminator _terminator;
2026   uint _active_workers;
2027 
2028 public:
2029   MarkFromRootsTask(uint active_workers) :
2030       WorkerTask("MarkFromRootsTask"),
2031       _strong_roots_scope(active_workers),
2032       _subtasks(ParallelRootType::sentinel),
2033       _terminator(active_workers, ParCompactionManager::oop_task_queues()),
2034       _active_workers(active_workers) {
2035   }
2036 
2037   virtual void work(uint worker_id) {
2038     for (uint task = 0; _subtasks.try_claim_task(task); /*empty*/ ) {
2039       mark_from_roots_work(static_cast<ParallelRootType::Value>(task), worker_id);
2040     }
2041 
2042     PCAddThreadRootsMarkingTaskClosure closure(worker_id);
2043     Threads::possibly_parallel_threads_do(true /*parallel */, &closure);
2044 
2045     // Mark from OopStorages
2046     {
2047       ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(worker_id);
2048       PCMarkAndPushClosure closure(cm);
2049       _oop_storage_set_par_state.oops_do(&closure);
2050       // Do the real work
2051       cm->follow_marking_stacks();
2052     }
2053 
2054     if (_active_workers > 1) {
2055       steal_marking_work(_terminator, worker_id);
2056     }
2057   }
2058 };
2059 
2060 class ParallelCompactRefProcProxyTask : public RefProcProxyTask {
2061   TaskTerminator _terminator;
2062 
2063 public:
2064   ParallelCompactRefProcProxyTask(uint max_workers)
2065     : RefProcProxyTask("ParallelCompactRefProcProxyTask", max_workers),
2066       _terminator(_max_workers, ParCompactionManager::oop_task_queues()) {}
2067 
2068   void work(uint worker_id) override {
2069     assert(worker_id < _max_workers, "sanity");
2070     ParCompactionManager* cm = (_tm == RefProcThreadModel::Single) ? ParCompactionManager::get_vmthread_cm() : ParCompactionManager::gc_thread_compaction_manager(worker_id);
2071     PCMarkAndPushClosure keep_alive(cm);
2072     BarrierEnqueueDiscoveredFieldClosure enqueue;
2073     ParCompactionManager::FollowStackClosure complete_gc(cm, (_tm == RefProcThreadModel::Single) ? nullptr : &_terminator, worker_id);
2074     _rp_task->rp_work(worker_id, PSParallelCompact::is_alive_closure(), &keep_alive, &enqueue, &complete_gc);
2075   }
2076 
2077   void prepare_run_task_hook() override {
2078     _terminator.reset_for_reuse(_queue_count);
2079   }
2080 };
2081 
2082 void PSParallelCompact::marking_phase(ParCompactionManager* cm,
2083                                       ParallelOldTracer *gc_tracer) {
2084   // Recursively traverse all live objects and mark them
2085   GCTraceTime(Info, gc, phases) tm("Marking Phase", &_gc_timer);
2086 
2087   uint active_gc_threads = ParallelScavengeHeap::heap()->workers().active_workers();
2088 
2089   // Need new claim bits before marking starts.
2090   ClassLoaderDataGraph::clear_claimed_marks();
2091 
2092   {
2093     GCTraceTime(Debug, gc, phases) tm("Par Mark", &_gc_timer);
2094 
2095     MarkFromRootsTask task(active_gc_threads);
2096     ParallelScavengeHeap::heap()->workers().run_task(&task);
2097   }
2098 
2099   // Process reference objects found during marking
2100   {
2101     GCTraceTime(Debug, gc, phases) tm("Reference Processing", &_gc_timer);
2102 
2103     ReferenceProcessorStats stats;
2104     ReferenceProcessorPhaseTimes pt(&_gc_timer, ref_processor()->max_num_queues());
2105 
2106     ref_processor()->set_active_mt_degree(active_gc_threads);
2107     ParallelCompactRefProcProxyTask task(ref_processor()->max_num_queues());
2108     stats = ref_processor()->process_discovered_references(task, pt);
2109 
2110     gc_tracer->report_gc_reference_stats(stats);
2111     pt.print_all_references();
2112   }
2113 
2114   // This is the point where the entire marking should have completed.
2115   ParCompactionManager::verify_all_marking_stack_empty();
2116 
2117   {
2118     GCTraceTime(Debug, gc, phases) tm("Weak Processing", &_gc_timer);
2119     WeakProcessor::weak_oops_do(&ParallelScavengeHeap::heap()->workers(),
2120                                 is_alive_closure(),
2121                                 &do_nothing_cl,
2122                                 1);
2123   }
2124 
2125   {
2126     GCTraceTime(Debug, gc, phases) tm_m("Class Unloading", &_gc_timer);
2127 
2128     // Follow system dictionary roots and unload classes.
2129     bool purged_class = SystemDictionary::do_unloading(&_gc_timer);
2130 
2131     // Unload nmethods.
2132     CodeCache::do_unloading(is_alive_closure(), purged_class);
2133 
2134     // Prune dead klasses from subklass/sibling/implementor lists.
2135     Klass::clean_weak_klass_links(purged_class);
2136 
2137     // Clean JVMCI metadata handles.
2138     JVMCI_ONLY(JVMCI::do_unloading(purged_class));
2139   }
2140 
2141   _gc_tracer.report_object_count_after_gc(is_alive_closure());
2142 }
2143 
2144 #ifdef ASSERT
2145 void PCAdjustPointerClosure::verify_cm(ParCompactionManager* cm) {
2146   assert(cm != NULL, "associate ParCompactionManage should not be NULL");
2147   auto vmthread_cm = ParCompactionManager::get_vmthread_cm();
2148   if (Thread::current()->is_VM_thread()) {
2149     assert(cm == vmthread_cm, "VM threads should use ParCompactionManager from get_vmthread_cm()");
2150   } else {
2151     assert(Thread::current()->is_Worker_thread(), "Must be a GC thread");
2152     assert(cm != vmthread_cm, "GC threads should use ParCompactionManager from gc_thread_compaction_manager()");
2153   }
2154 }
2155 #endif
2156 
2157 class PSAdjustTask final : public WorkerTask {
2158   SubTasksDone                               _sub_tasks;
2159   WeakProcessor::Task                        _weak_proc_task;
2160   OopStorageSetStrongParState<false, false>  _oop_storage_iter;
2161   uint                                       _nworkers;
2162 
2163   enum PSAdjustSubTask {
2164     PSAdjustSubTask_code_cache,
2165 
2166     PSAdjustSubTask_num_elements
2167   };
2168 
2169 public:
2170   PSAdjustTask(uint nworkers) :
2171     WorkerTask("PSAdjust task"),
2172     _sub_tasks(PSAdjustSubTask_num_elements),
2173     _weak_proc_task(nworkers),
2174     _nworkers(nworkers) {
2175     // Need new claim bits when tracing through and adjusting pointers.
2176     ClassLoaderDataGraph::clear_claimed_marks();
2177     if (nworkers > 1) {
2178       Threads::change_thread_claim_token();
2179     }
2180   }
2181 
2182   ~PSAdjustTask() {
2183     Threads::assert_all_threads_claimed();
2184   }
2185 
2186   void work(uint worker_id) {
2187     ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(worker_id);
2188     PCAdjustPointerClosure adjust(cm);
2189     {
2190       ResourceMark rm;
2191       Threads::possibly_parallel_oops_do(_nworkers > 1, &adjust, nullptr);
2192     }
2193     _oop_storage_iter.oops_do(&adjust);
2194     {
2195       CLDToOopClosure cld_closure(&adjust, ClassLoaderData::_claim_strong);
2196       ClassLoaderDataGraph::cld_do(&cld_closure);
2197     }
2198     {
2199       AlwaysTrueClosure always_alive;
2200       _weak_proc_task.work(worker_id, &always_alive, &adjust);
2201     }
2202     if (_sub_tasks.try_claim_task(PSAdjustSubTask_code_cache)) {
2203       CodeBlobToOopClosure adjust_code(&adjust, CodeBlobToOopClosure::FixRelocations);
2204       CodeCache::blobs_do(&adjust_code);
2205     }
2206     _sub_tasks.all_tasks_claimed();
2207   }
2208 };
2209 
2210 void PSParallelCompact::adjust_roots() {
2211   // Adjust the pointers to reflect the new locations
2212   GCTraceTime(Info, gc, phases) tm("Adjust Roots", &_gc_timer);
2213   uint nworkers = ParallelScavengeHeap::heap()->workers().active_workers();
2214   PSAdjustTask task(nworkers);
2215   ParallelScavengeHeap::heap()->workers().run_task(&task);
2216 }
2217 
2218 // Helper class to print 8 region numbers per line and then print the total at the end.
2219 class FillableRegionLogger : public StackObj {
2220 private:
2221   Log(gc, compaction) log;
2222   static const int LineLength = 8;
2223   size_t _regions[LineLength];
2224   int _next_index;
2225   bool _enabled;
2226   size_t _total_regions;
2227 public:
2228   FillableRegionLogger() : _next_index(0), _enabled(log_develop_is_enabled(Trace, gc, compaction)), _total_regions(0) { }
2229   ~FillableRegionLogger() {
2230     log.trace(SIZE_FORMAT " initially fillable regions", _total_regions);
2231   }
2232 
2233   void print_line() {
2234     if (!_enabled || _next_index == 0) {
2235       return;
2236     }
2237     FormatBuffer<> line("Fillable: ");
2238     for (int i = 0; i < _next_index; i++) {
2239       line.append(" " SIZE_FORMAT_W(7), _regions[i]);
2240     }
2241     log.trace("%s", line.buffer());
2242     _next_index = 0;
2243   }
2244 
2245   void handle(size_t region) {
2246     if (!_enabled) {
2247       return;
2248     }
2249     _regions[_next_index++] = region;
2250     if (_next_index == LineLength) {
2251       print_line();
2252     }
2253     _total_regions++;
2254   }
2255 };
2256 
2257 void PSParallelCompact::prepare_region_draining_tasks(uint parallel_gc_threads)
2258 {
2259   GCTraceTime(Trace, gc, phases) tm("Drain Task Setup", &_gc_timer);
2260 
2261   // Find the threads that are active
2262   uint worker_id = 0;
2263 
2264   // Find all regions that are available (can be filled immediately) and
2265   // distribute them to the thread stacks.  The iteration is done in reverse
2266   // order (high to low) so the regions will be removed in ascending order.
2267 
2268   const ParallelCompactData& sd = PSParallelCompact::summary_data();
2269 
2270   // id + 1 is used to test termination so unsigned  can
2271   // be used with an old_space_id == 0.
2272   FillableRegionLogger region_logger;
2273   for (unsigned int id = to_space_id; id + 1 > old_space_id; --id) {
2274     SpaceInfo* const space_info = _space_info + id;
2275     HeapWord* const new_top = space_info->new_top();
2276 
2277     const size_t beg_region = sd.addr_to_region_idx(space_info->dense_prefix());
2278     const size_t end_region =
2279       sd.addr_to_region_idx(sd.region_align_up(new_top));
2280 
2281     for (size_t cur = end_region - 1; cur + 1 > beg_region; --cur) {
2282       if (sd.region(cur)->claim_unsafe()) {
2283         ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(worker_id);
2284         bool result = sd.region(cur)->mark_normal();
2285         assert(result, "Must succeed at this point.");
2286         cm->region_stack()->push(cur);
2287         region_logger.handle(cur);
2288         // Assign regions to tasks in round-robin fashion.
2289         if (++worker_id == parallel_gc_threads) {
2290           worker_id = 0;
2291         }
2292       }
2293     }
2294     region_logger.print_line();
2295   }
2296 }
2297 
2298 class TaskQueue : StackObj {
2299   volatile uint _counter;
2300   uint _size;
2301   uint _insert_index;
2302   PSParallelCompact::UpdateDensePrefixTask* _backing_array;
2303 public:
2304   explicit TaskQueue(uint size) : _counter(0), _size(size), _insert_index(0), _backing_array(NULL) {
2305     _backing_array = NEW_C_HEAP_ARRAY(PSParallelCompact::UpdateDensePrefixTask, _size, mtGC);
2306   }
2307   ~TaskQueue() {
2308     assert(_counter >= _insert_index, "not all queue elements were claimed");
2309     FREE_C_HEAP_ARRAY(T, _backing_array);
2310   }
2311 
2312   void push(const PSParallelCompact::UpdateDensePrefixTask& value) {
2313     assert(_insert_index < _size, "too small backing array");
2314     _backing_array[_insert_index++] = value;
2315   }
2316 
2317   bool try_claim(PSParallelCompact::UpdateDensePrefixTask& reference) {
2318     uint claimed = Atomic::fetch_and_add(&_counter, 1u);
2319     if (claimed < _insert_index) {
2320       reference = _backing_array[claimed];
2321       return true;
2322     } else {
2323       return false;
2324     }
2325   }
2326 };
2327 
2328 #define PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING 4
2329 
2330 void PSParallelCompact::enqueue_dense_prefix_tasks(TaskQueue& task_queue,
2331                                                    uint parallel_gc_threads) {
2332   GCTraceTime(Trace, gc, phases) tm("Dense Prefix Task Setup", &_gc_timer);
2333 
2334   ParallelCompactData& sd = PSParallelCompact::summary_data();
2335 
2336   // Iterate over all the spaces adding tasks for updating
2337   // regions in the dense prefix.  Assume that 1 gc thread
2338   // will work on opening the gaps and the remaining gc threads
2339   // will work on the dense prefix.
2340   unsigned int space_id;
2341   for (space_id = old_space_id; space_id < last_space_id; ++ space_id) {
2342     HeapWord* const dense_prefix_end = _space_info[space_id].dense_prefix();
2343     const MutableSpace* const space = _space_info[space_id].space();
2344 
2345     if (dense_prefix_end == space->bottom()) {
2346       // There is no dense prefix for this space.
2347       continue;
2348     }
2349 
2350     // The dense prefix is before this region.
2351     size_t region_index_end_dense_prefix =
2352         sd.addr_to_region_idx(dense_prefix_end);
2353     RegionData* const dense_prefix_cp =
2354       sd.region(region_index_end_dense_prefix);
2355     assert(dense_prefix_end == space->end() ||
2356            dense_prefix_cp->available() ||
2357            dense_prefix_cp->claimed(),
2358            "The region after the dense prefix should always be ready to fill");
2359 
2360     size_t region_index_start = sd.addr_to_region_idx(space->bottom());
2361 
2362     // Is there dense prefix work?
2363     size_t total_dense_prefix_regions =
2364       region_index_end_dense_prefix - region_index_start;
2365     // How many regions of the dense prefix should be given to
2366     // each thread?
2367     if (total_dense_prefix_regions > 0) {
2368       uint tasks_for_dense_prefix = 1;
2369       if (total_dense_prefix_regions <=
2370           (parallel_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING)) {
2371         // Don't over partition.  This assumes that
2372         // PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING is a small integer value
2373         // so there are not many regions to process.
2374         tasks_for_dense_prefix = parallel_gc_threads;
2375       } else {
2376         // Over partition
2377         tasks_for_dense_prefix = parallel_gc_threads *
2378           PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING;
2379       }
2380       size_t regions_per_thread = total_dense_prefix_regions /
2381         tasks_for_dense_prefix;
2382       // Give each thread at least 1 region.
2383       if (regions_per_thread == 0) {
2384         regions_per_thread = 1;
2385       }
2386 
2387       for (uint k = 0; k < tasks_for_dense_prefix; k++) {
2388         if (region_index_start >= region_index_end_dense_prefix) {
2389           break;
2390         }
2391         // region_index_end is not processed
2392         size_t region_index_end = MIN2(region_index_start + regions_per_thread,
2393                                        region_index_end_dense_prefix);
2394         task_queue.push(UpdateDensePrefixTask(SpaceId(space_id),
2395                                               region_index_start,
2396                                               region_index_end));
2397         region_index_start = region_index_end;
2398       }
2399     }
2400     // This gets any part of the dense prefix that did not
2401     // fit evenly.
2402     if (region_index_start < region_index_end_dense_prefix) {
2403       task_queue.push(UpdateDensePrefixTask(SpaceId(space_id),
2404                                             region_index_start,
2405                                             region_index_end_dense_prefix));
2406     }
2407   }
2408 }
2409 
2410 #ifdef ASSERT
2411 // Write a histogram of the number of times the block table was filled for a
2412 // region.
2413 void PSParallelCompact::write_block_fill_histogram()
2414 {
2415   if (!log_develop_is_enabled(Trace, gc, compaction)) {
2416     return;
2417   }
2418 
2419   Log(gc, compaction) log;
2420   ResourceMark rm;
2421   LogStream ls(log.trace());
2422   outputStream* out = &ls;
2423 
2424   typedef ParallelCompactData::RegionData rd_t;
2425   ParallelCompactData& sd = summary_data();
2426 
2427   for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2428     MutableSpace* const spc = _space_info[id].space();
2429     if (spc->bottom() != spc->top()) {
2430       const rd_t* const beg = sd.addr_to_region_ptr(spc->bottom());
2431       HeapWord* const top_aligned_up = sd.region_align_up(spc->top());
2432       const rd_t* const end = sd.addr_to_region_ptr(top_aligned_up);
2433 
2434       size_t histo[5] = { 0, 0, 0, 0, 0 };
2435       const size_t histo_len = sizeof(histo) / sizeof(size_t);
2436       const size_t region_cnt = pointer_delta(end, beg, sizeof(rd_t));
2437 
2438       for (const rd_t* cur = beg; cur < end; ++cur) {
2439         ++histo[MIN2(cur->blocks_filled_count(), histo_len - 1)];
2440       }
2441       out->print("Block fill histogram: %u %-4s" SIZE_FORMAT_W(5), id, space_names[id], region_cnt);
2442       for (size_t i = 0; i < histo_len; ++i) {
2443         out->print(" " SIZE_FORMAT_W(5) " %5.1f%%",
2444                    histo[i], 100.0 * histo[i] / region_cnt);
2445       }
2446       out->cr();
2447     }
2448   }
2449 }
2450 #endif // #ifdef ASSERT
2451 
2452 static void compaction_with_stealing_work(TaskTerminator* terminator, uint worker_id) {
2453   assert(ParallelScavengeHeap::heap()->is_gc_active(), "called outside gc");
2454 
2455   ParCompactionManager* cm =
2456     ParCompactionManager::gc_thread_compaction_manager(worker_id);
2457 
2458   // Drain the stacks that have been preloaded with regions
2459   // that are ready to fill.
2460 
2461   cm->drain_region_stacks();
2462 
2463   guarantee(cm->region_stack()->is_empty(), "Not empty");
2464 
2465   size_t region_index = 0;
2466 
2467   while (true) {
2468     if (ParCompactionManager::steal(worker_id, region_index)) {
2469       PSParallelCompact::fill_and_update_region(cm, region_index);
2470       cm->drain_region_stacks();
2471     } else if (PSParallelCompact::steal_unavailable_region(cm, region_index)) {
2472       // Fill and update an unavailable region with the help of a shadow region
2473       PSParallelCompact::fill_and_update_shadow_region(cm, region_index);
2474       cm->drain_region_stacks();
2475     } else {
2476       if (terminator->offer_termination()) {
2477         break;
2478       }
2479       // Go around again.
2480     }
2481   }
2482 }
2483 
2484 class UpdateDensePrefixAndCompactionTask: public WorkerTask {
2485   TaskQueue& _tq;
2486   TaskTerminator _terminator;
2487   uint _active_workers;
2488 
2489 public:
2490   UpdateDensePrefixAndCompactionTask(TaskQueue& tq, uint active_workers) :
2491       WorkerTask("UpdateDensePrefixAndCompactionTask"),
2492       _tq(tq),
2493       _terminator(active_workers, ParCompactionManager::region_task_queues()),
2494       _active_workers(active_workers) {
2495   }
2496   virtual void work(uint worker_id) {
2497     ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(worker_id);
2498 
2499     for (PSParallelCompact::UpdateDensePrefixTask task; _tq.try_claim(task); /* empty */) {
2500       PSParallelCompact::update_and_deadwood_in_dense_prefix(cm,
2501                                                              task._space_id,
2502                                                              task._region_index_start,
2503                                                              task._region_index_end);
2504     }
2505 
2506     // Once a thread has drained it's stack, it should try to steal regions from
2507     // other threads.
2508     compaction_with_stealing_work(&_terminator, worker_id);
2509   }
2510 };
2511 
2512 void PSParallelCompact::compact() {
2513   GCTraceTime(Info, gc, phases) tm("Compaction Phase", &_gc_timer);
2514 
2515   ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
2516   PSOldGen* old_gen = heap->old_gen();
2517   old_gen->start_array()->reset();
2518   uint active_gc_threads = ParallelScavengeHeap::heap()->workers().active_workers();
2519 
2520   // for [0..last_space_id)
2521   //     for [0..active_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING)
2522   //         push
2523   //     push
2524   //
2525   // max push count is thus: last_space_id * (active_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING + 1)
2526   TaskQueue task_queue(last_space_id * (active_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING + 1));
2527   initialize_shadow_regions(active_gc_threads);
2528   prepare_region_draining_tasks(active_gc_threads);
2529   enqueue_dense_prefix_tasks(task_queue, active_gc_threads);
2530 
2531   {
2532     GCTraceTime(Trace, gc, phases) tm("Par Compact", &_gc_timer);
2533 
2534     UpdateDensePrefixAndCompactionTask task(task_queue, active_gc_threads);
2535     ParallelScavengeHeap::heap()->workers().run_task(&task);
2536 
2537 #ifdef  ASSERT
2538     // Verify that all regions have been processed before the deferred updates.
2539     for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2540       verify_complete(SpaceId(id));
2541     }
2542 #endif
2543   }
2544 
2545   {
2546     GCTraceTime(Trace, gc, phases) tm("Deferred Updates", &_gc_timer);
2547     // Update the deferred objects, if any. In principle, any compaction
2548     // manager can be used. However, since the current thread is VM thread, we
2549     // use the rightful one to keep the verification logic happy.
2550     ParCompactionManager* cm = ParCompactionManager::get_vmthread_cm();
2551     for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2552       update_deferred_objects(cm, SpaceId(id));
2553     }
2554   }
2555 
2556   DEBUG_ONLY(write_block_fill_histogram());
2557 }
2558 
2559 #ifdef  ASSERT
2560 void PSParallelCompact::verify_complete(SpaceId space_id) {
2561   // All Regions between space bottom() to new_top() should be marked as filled
2562   // and all Regions between new_top() and top() should be available (i.e.,
2563   // should have been emptied).
2564   ParallelCompactData& sd = summary_data();
2565   SpaceInfo si = _space_info[space_id];
2566   HeapWord* new_top_addr = sd.region_align_up(si.new_top());
2567   HeapWord* old_top_addr = sd.region_align_up(si.space()->top());
2568   const size_t beg_region = sd.addr_to_region_idx(si.space()->bottom());
2569   const size_t new_top_region = sd.addr_to_region_idx(new_top_addr);
2570   const size_t old_top_region = sd.addr_to_region_idx(old_top_addr);
2571 
2572   bool issued_a_warning = false;
2573 
2574   size_t cur_region;
2575   for (cur_region = beg_region; cur_region < new_top_region; ++cur_region) {
2576     const RegionData* const c = sd.region(cur_region);
2577     if (!c->completed()) {
2578       log_warning(gc)("region " SIZE_FORMAT " not filled: destination_count=%u",
2579                       cur_region, c->destination_count());
2580       issued_a_warning = true;
2581     }
2582   }
2583 
2584   for (cur_region = new_top_region; cur_region < old_top_region; ++cur_region) {
2585     const RegionData* const c = sd.region(cur_region);
2586     if (!c->available()) {
2587       log_warning(gc)("region " SIZE_FORMAT " not empty: destination_count=%u",
2588                       cur_region, c->destination_count());
2589       issued_a_warning = true;
2590     }
2591   }
2592 
2593   if (issued_a_warning) {
2594     print_region_ranges();
2595   }
2596 }
2597 #endif  // #ifdef ASSERT
2598 
2599 inline void UpdateOnlyClosure::do_addr(HeapWord* addr) {
2600   _start_array->allocate_block(addr);
2601   compaction_manager()->update_contents(cast_to_oop(addr));
2602 }
2603 
2604 // Update interior oops in the ranges of regions [beg_region, end_region).
2605 void
2606 PSParallelCompact::update_and_deadwood_in_dense_prefix(ParCompactionManager* cm,
2607                                                        SpaceId space_id,
2608                                                        size_t beg_region,
2609                                                        size_t end_region) {
2610   ParallelCompactData& sd = summary_data();
2611   ParMarkBitMap* const mbm = mark_bitmap();
2612 
2613   HeapWord* beg_addr = sd.region_to_addr(beg_region);
2614   HeapWord* const end_addr = sd.region_to_addr(end_region);
2615   assert(beg_region <= end_region, "bad region range");
2616   assert(end_addr <= dense_prefix(space_id), "not in the dense prefix");
2617 
2618 #ifdef  ASSERT
2619   // Claim the regions to avoid triggering an assert when they are marked as
2620   // filled.
2621   for (size_t claim_region = beg_region; claim_region < end_region; ++claim_region) {
2622     assert(sd.region(claim_region)->claim_unsafe(), "claim() failed");
2623   }
2624 #endif  // #ifdef ASSERT
2625 
2626   if (beg_addr != space(space_id)->bottom()) {
2627     // Find the first live object or block of dead space that *starts* in this
2628     // range of regions.  If a partial object crosses onto the region, skip it;
2629     // it will be marked for 'deferred update' when the object head is
2630     // processed.  If dead space crosses onto the region, it is also skipped; it
2631     // will be filled when the prior region is processed.  If neither of those
2632     // apply, the first word in the region is the start of a live object or dead
2633     // space.
2634     assert(beg_addr > space(space_id)->bottom(), "sanity");
2635     const RegionData* const cp = sd.region(beg_region);
2636     if (cp->partial_obj_size() != 0) {
2637       beg_addr = sd.partial_obj_end(beg_region);
2638     } else if (dead_space_crosses_boundary(cp, mbm->addr_to_bit(beg_addr))) {
2639       beg_addr = mbm->find_obj_beg(beg_addr, end_addr);
2640     }
2641   }
2642 
2643   if (beg_addr < end_addr) {
2644     // A live object or block of dead space starts in this range of Regions.
2645      HeapWord* const dense_prefix_end = dense_prefix(space_id);
2646 
2647     // Create closures and iterate.
2648     UpdateOnlyClosure update_closure(mbm, cm, space_id);
2649     FillClosure fill_closure(cm, space_id);
2650     ParMarkBitMap::IterationStatus status;
2651     status = mbm->iterate(&update_closure, &fill_closure, beg_addr, end_addr,
2652                           dense_prefix_end);
2653     if (status == ParMarkBitMap::incomplete) {
2654       update_closure.do_addr(update_closure.source());
2655     }
2656   }
2657 
2658   // Mark the regions as filled.
2659   RegionData* const beg_cp = sd.region(beg_region);
2660   RegionData* const end_cp = sd.region(end_region);
2661   for (RegionData* cp = beg_cp; cp < end_cp; ++cp) {
2662     cp->set_completed();
2663   }
2664 }
2665 
2666 // Return the SpaceId for the space containing addr.  If addr is not in the
2667 // heap, last_space_id is returned.  In debug mode it expects the address to be
2668 // in the heap and asserts such.
2669 PSParallelCompact::SpaceId PSParallelCompact::space_id(HeapWord* addr) {
2670   assert(ParallelScavengeHeap::heap()->is_in_reserved(addr), "addr not in the heap");
2671 
2672   for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2673     if (_space_info[id].space()->contains(addr)) {
2674       return SpaceId(id);
2675     }
2676   }
2677 
2678   assert(false, "no space contains the addr");
2679   return last_space_id;
2680 }
2681 
2682 void PSParallelCompact::update_deferred_objects(ParCompactionManager* cm,
2683                                                 SpaceId id) {
2684   assert(id < last_space_id, "bad space id");
2685 
2686   ParallelCompactData& sd = summary_data();
2687   const SpaceInfo* const space_info = _space_info + id;
2688   ObjectStartArray* const start_array = space_info->start_array();
2689 
2690   const MutableSpace* const space = space_info->space();
2691   assert(space_info->dense_prefix() >= space->bottom(), "dense_prefix not set");
2692   HeapWord* const beg_addr = space_info->dense_prefix();
2693   HeapWord* const end_addr = sd.region_align_up(space_info->new_top());
2694 
2695   const RegionData* const beg_region = sd.addr_to_region_ptr(beg_addr);
2696   const RegionData* const end_region = sd.addr_to_region_ptr(end_addr);
2697   const RegionData* cur_region;
2698   for (cur_region = beg_region; cur_region < end_region; ++cur_region) {
2699     HeapWord* const addr = cur_region->deferred_obj_addr();
2700     if (addr != NULL) {
2701       if (start_array != NULL) {
2702         start_array->allocate_block(addr);
2703       }
2704       cm->update_contents(cast_to_oop(addr));
2705       assert(oopDesc::is_oop_or_null(cast_to_oop(addr)), "Expected an oop or NULL at " PTR_FORMAT, p2i(cast_to_oop(addr)));
2706     }
2707   }
2708 }
2709 
2710 // Skip over count live words starting from beg, and return the address of the
2711 // next live word.  Unless marked, the word corresponding to beg is assumed to
2712 // be dead.  Callers must either ensure beg does not correspond to the middle of
2713 // an object, or account for those live words in some other way.  Callers must
2714 // also ensure that there are enough live words in the range [beg, end) to skip.
2715 HeapWord*
2716 PSParallelCompact::skip_live_words(HeapWord* beg, HeapWord* end, size_t count)
2717 {
2718   assert(count > 0, "sanity");
2719 
2720   ParMarkBitMap* m = mark_bitmap();
2721   idx_t bits_to_skip = m->words_to_bits(count);
2722   idx_t cur_beg = m->addr_to_bit(beg);
2723   const idx_t search_end = m->align_range_end(m->addr_to_bit(end));
2724 
2725   do {
2726     cur_beg = m->find_obj_beg(cur_beg, search_end);
2727     idx_t cur_end = m->find_obj_end(cur_beg, search_end);
2728     const size_t obj_bits = cur_end - cur_beg + 1;
2729     if (obj_bits > bits_to_skip) {
2730       return m->bit_to_addr(cur_beg + bits_to_skip);
2731     }
2732     bits_to_skip -= obj_bits;
2733     cur_beg = cur_end + 1;
2734   } while (bits_to_skip > 0);
2735 
2736   // Skipping the desired number of words landed just past the end of an object.
2737   // Find the start of the next object.
2738   cur_beg = m->find_obj_beg(cur_beg, search_end);
2739   assert(cur_beg < m->addr_to_bit(end), "not enough live words to skip");
2740   return m->bit_to_addr(cur_beg);
2741 }
2742 
2743 HeapWord* PSParallelCompact::first_src_addr(HeapWord* const dest_addr,
2744                                             SpaceId src_space_id,
2745                                             size_t src_region_idx)
2746 {
2747   assert(summary_data().is_region_aligned(dest_addr), "not aligned");
2748 
2749   const SplitInfo& split_info = _space_info[src_space_id].split_info();
2750   if (split_info.dest_region_addr() == dest_addr) {
2751     // The partial object ending at the split point contains the first word to
2752     // be copied to dest_addr.
2753     return split_info.first_src_addr();
2754   }
2755 
2756   const ParallelCompactData& sd = summary_data();
2757   ParMarkBitMap* const bitmap = mark_bitmap();
2758   const size_t RegionSize = ParallelCompactData::RegionSize;
2759 
2760   assert(sd.is_region_aligned(dest_addr), "not aligned");
2761   const RegionData* const src_region_ptr = sd.region(src_region_idx);
2762   const size_t partial_obj_size = src_region_ptr->partial_obj_size();
2763   HeapWord* const src_region_destination = src_region_ptr->destination();
2764 
2765   assert(dest_addr >= src_region_destination, "wrong src region");
2766   assert(src_region_ptr->data_size() > 0, "src region cannot be empty");
2767 
2768   HeapWord* const src_region_beg = sd.region_to_addr(src_region_idx);
2769   HeapWord* const src_region_end = src_region_beg + RegionSize;
2770 
2771   HeapWord* addr = src_region_beg;
2772   if (dest_addr == src_region_destination) {
2773     // Return the first live word in the source region.
2774     if (partial_obj_size == 0) {
2775       addr = bitmap->find_obj_beg(addr, src_region_end);
2776       assert(addr < src_region_end, "no objects start in src region");
2777     }
2778     return addr;
2779   }
2780 
2781   // Must skip some live data.
2782   size_t words_to_skip = dest_addr - src_region_destination;
2783   assert(src_region_ptr->data_size() > words_to_skip, "wrong src region");
2784 
2785   if (partial_obj_size >= words_to_skip) {
2786     // All the live words to skip are part of the partial object.
2787     addr += words_to_skip;
2788     if (partial_obj_size == words_to_skip) {
2789       // Find the first live word past the partial object.
2790       addr = bitmap->find_obj_beg(addr, src_region_end);
2791       assert(addr < src_region_end, "wrong src region");
2792     }
2793     return addr;
2794   }
2795 
2796   // Skip over the partial object (if any).
2797   if (partial_obj_size != 0) {
2798     words_to_skip -= partial_obj_size;
2799     addr += partial_obj_size;
2800   }
2801 
2802   // Skip over live words due to objects that start in the region.
2803   addr = skip_live_words(addr, src_region_end, words_to_skip);
2804   assert(addr < src_region_end, "wrong src region");
2805   return addr;
2806 }
2807 
2808 void PSParallelCompact::decrement_destination_counts(ParCompactionManager* cm,
2809                                                      SpaceId src_space_id,
2810                                                      size_t beg_region,
2811                                                      HeapWord* end_addr)
2812 {
2813   ParallelCompactData& sd = summary_data();
2814 
2815 #ifdef ASSERT
2816   MutableSpace* const src_space = _space_info[src_space_id].space();
2817   HeapWord* const beg_addr = sd.region_to_addr(beg_region);
2818   assert(src_space->contains(beg_addr) || beg_addr == src_space->end(),
2819          "src_space_id does not match beg_addr");
2820   assert(src_space->contains(end_addr) || end_addr == src_space->end(),
2821          "src_space_id does not match end_addr");
2822 #endif // #ifdef ASSERT
2823 
2824   RegionData* const beg = sd.region(beg_region);
2825   RegionData* const end = sd.addr_to_region_ptr(sd.region_align_up(end_addr));
2826 
2827   // Regions up to new_top() are enqueued if they become available.
2828   HeapWord* const new_top = _space_info[src_space_id].new_top();
2829   RegionData* const enqueue_end =
2830     sd.addr_to_region_ptr(sd.region_align_up(new_top));
2831 
2832   for (RegionData* cur = beg; cur < end; ++cur) {
2833     assert(cur->data_size() > 0, "region must have live data");
2834     cur->decrement_destination_count();
2835     if (cur < enqueue_end && cur->available() && cur->claim()) {
2836       if (cur->mark_normal()) {
2837         cm->push_region(sd.region(cur));
2838       } else if (cur->mark_copied()) {
2839         // Try to copy the content of the shadow region back to its corresponding
2840         // heap region if the shadow region is filled. Otherwise, the GC thread
2841         // fills the shadow region will copy the data back (see
2842         // MoveAndUpdateShadowClosure::complete_region).
2843         copy_back(sd.region_to_addr(cur->shadow_region()), sd.region_to_addr(cur));
2844         ParCompactionManager::push_shadow_region_mt_safe(cur->shadow_region());
2845         cur->set_completed();
2846       }
2847     }
2848   }
2849 }
2850 
2851 size_t PSParallelCompact::next_src_region(MoveAndUpdateClosure& closure,
2852                                           SpaceId& src_space_id,
2853                                           HeapWord*& src_space_top,
2854                                           HeapWord* end_addr)
2855 {
2856   typedef ParallelCompactData::RegionData RegionData;
2857 
2858   ParallelCompactData& sd = PSParallelCompact::summary_data();
2859   const size_t region_size = ParallelCompactData::RegionSize;
2860 
2861   size_t src_region_idx = 0;
2862 
2863   // Skip empty regions (if any) up to the top of the space.
2864   HeapWord* const src_aligned_up = sd.region_align_up(end_addr);
2865   RegionData* src_region_ptr = sd.addr_to_region_ptr(src_aligned_up);
2866   HeapWord* const top_aligned_up = sd.region_align_up(src_space_top);
2867   const RegionData* const top_region_ptr =
2868     sd.addr_to_region_ptr(top_aligned_up);
2869   while (src_region_ptr < top_region_ptr && src_region_ptr->data_size() == 0) {
2870     ++src_region_ptr;
2871   }
2872 
2873   if (src_region_ptr < top_region_ptr) {
2874     // The next source region is in the current space.  Update src_region_idx
2875     // and the source address to match src_region_ptr.
2876     src_region_idx = sd.region(src_region_ptr);
2877     HeapWord* const src_region_addr = sd.region_to_addr(src_region_idx);
2878     if (src_region_addr > closure.source()) {
2879       closure.set_source(src_region_addr);
2880     }
2881     return src_region_idx;
2882   }
2883 
2884   // Switch to a new source space and find the first non-empty region.
2885   unsigned int space_id = src_space_id + 1;
2886   assert(space_id < last_space_id, "not enough spaces");
2887 
2888   HeapWord* const destination = closure.destination();
2889 
2890   do {
2891     MutableSpace* space = _space_info[space_id].space();
2892     HeapWord* const bottom = space->bottom();
2893     const RegionData* const bottom_cp = sd.addr_to_region_ptr(bottom);
2894 
2895     // Iterate over the spaces that do not compact into themselves.
2896     if (bottom_cp->destination() != bottom) {
2897       HeapWord* const top_aligned_up = sd.region_align_up(space->top());
2898       const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up);
2899 
2900       for (const RegionData* src_cp = bottom_cp; src_cp < top_cp; ++src_cp) {
2901         if (src_cp->live_obj_size() > 0) {
2902           // Found it.
2903           assert(src_cp->destination() == destination,
2904                  "first live obj in the space must match the destination");
2905           assert(src_cp->partial_obj_size() == 0,
2906                  "a space cannot begin with a partial obj");
2907 
2908           src_space_id = SpaceId(space_id);
2909           src_space_top = space->top();
2910           const size_t src_region_idx = sd.region(src_cp);
2911           closure.set_source(sd.region_to_addr(src_region_idx));
2912           return src_region_idx;
2913         } else {
2914           assert(src_cp->data_size() == 0, "sanity");
2915         }
2916       }
2917     }
2918   } while (++space_id < last_space_id);
2919 
2920   assert(false, "no source region was found");
2921   return 0;
2922 }
2923 
2924 void PSParallelCompact::fill_region(ParCompactionManager* cm, MoveAndUpdateClosure& closure, size_t region_idx)
2925 {
2926   typedef ParMarkBitMap::IterationStatus IterationStatus;
2927   ParMarkBitMap* const bitmap = mark_bitmap();
2928   ParallelCompactData& sd = summary_data();
2929   RegionData* const region_ptr = sd.region(region_idx);
2930 
2931   // Get the source region and related info.
2932   size_t src_region_idx = region_ptr->source_region();
2933   SpaceId src_space_id = space_id(sd.region_to_addr(src_region_idx));
2934   HeapWord* src_space_top = _space_info[src_space_id].space()->top();
2935   HeapWord* dest_addr = sd.region_to_addr(region_idx);
2936 
2937   closure.set_source(first_src_addr(dest_addr, src_space_id, src_region_idx));
2938 
2939   // Adjust src_region_idx to prepare for decrementing destination counts (the
2940   // destination count is not decremented when a region is copied to itself).
2941   if (src_region_idx == region_idx) {
2942     src_region_idx += 1;
2943   }
2944 
2945   if (bitmap->is_unmarked(closure.source())) {
2946     // The first source word is in the middle of an object; copy the remainder
2947     // of the object or as much as will fit.  The fact that pointer updates were
2948     // deferred will be noted when the object header is processed.
2949     HeapWord* const old_src_addr = closure.source();
2950     closure.copy_partial_obj();
2951     if (closure.is_full()) {
2952       decrement_destination_counts(cm, src_space_id, src_region_idx,
2953                                    closure.source());
2954       region_ptr->set_deferred_obj_addr(NULL);
2955       closure.complete_region(cm, dest_addr, region_ptr);
2956       return;
2957     }
2958 
2959     HeapWord* const end_addr = sd.region_align_down(closure.source());
2960     if (sd.region_align_down(old_src_addr) != end_addr) {
2961       // The partial object was copied from more than one source region.
2962       decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr);
2963 
2964       // Move to the next source region, possibly switching spaces as well.  All
2965       // args except end_addr may be modified.
2966       src_region_idx = next_src_region(closure, src_space_id, src_space_top,
2967                                        end_addr);
2968     }
2969   }
2970 
2971   do {
2972     HeapWord* const cur_addr = closure.source();
2973     HeapWord* const end_addr = MIN2(sd.region_align_up(cur_addr + 1),
2974                                     src_space_top);
2975     IterationStatus status = bitmap->iterate(&closure, cur_addr, end_addr);
2976 
2977     if (status == ParMarkBitMap::incomplete) {
2978       // The last obj that starts in the source region does not end in the
2979       // region.
2980       assert(closure.source() < end_addr, "sanity");
2981       HeapWord* const obj_beg = closure.source();
2982       HeapWord* const range_end = MIN2(obj_beg + closure.words_remaining(),
2983                                        src_space_top);
2984       HeapWord* const obj_end = bitmap->find_obj_end(obj_beg, range_end);
2985       if (obj_end < range_end) {
2986         // The end was found; the entire object will fit.
2987         status = closure.do_addr(obj_beg, bitmap->obj_size(obj_beg, obj_end));
2988         assert(status != ParMarkBitMap::would_overflow, "sanity");
2989       } else {
2990         // The end was not found; the object will not fit.
2991         assert(range_end < src_space_top, "obj cannot cross space boundary");
2992         status = ParMarkBitMap::would_overflow;
2993       }
2994     }
2995 
2996     if (status == ParMarkBitMap::would_overflow) {
2997       // The last object did not fit.  Note that interior oop updates were
2998       // deferred, then copy enough of the object to fill the region.
2999       region_ptr->set_deferred_obj_addr(closure.destination());
3000       status = closure.copy_until_full(); // copies from closure.source()
3001 
3002       decrement_destination_counts(cm, src_space_id, src_region_idx,
3003                                    closure.source());
3004       closure.complete_region(cm, dest_addr, region_ptr);
3005       return;
3006     }
3007 
3008     if (status == ParMarkBitMap::full) {
3009       decrement_destination_counts(cm, src_space_id, src_region_idx,
3010                                    closure.source());
3011       region_ptr->set_deferred_obj_addr(NULL);
3012       closure.complete_region(cm, dest_addr, region_ptr);
3013       return;
3014     }
3015 
3016     decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr);
3017 
3018     // Move to the next source region, possibly switching spaces as well.  All
3019     // args except end_addr may be modified.
3020     src_region_idx = next_src_region(closure, src_space_id, src_space_top,
3021                                      end_addr);
3022   } while (true);
3023 }
3024 
3025 void PSParallelCompact::fill_and_update_region(ParCompactionManager* cm, size_t region_idx)
3026 {
3027   MoveAndUpdateClosure cl(mark_bitmap(), cm, region_idx);
3028   fill_region(cm, cl, region_idx);
3029 }
3030 
3031 void PSParallelCompact::fill_and_update_shadow_region(ParCompactionManager* cm, size_t region_idx)
3032 {
3033   // Get a shadow region first
3034   ParallelCompactData& sd = summary_data();
3035   RegionData* const region_ptr = sd.region(region_idx);
3036   size_t shadow_region = ParCompactionManager::pop_shadow_region_mt_safe(region_ptr);
3037   // The InvalidShadow return value indicates the corresponding heap region is available,
3038   // so use MoveAndUpdateClosure to fill the normal region. Otherwise, use
3039   // MoveAndUpdateShadowClosure to fill the acquired shadow region.
3040   if (shadow_region == ParCompactionManager::InvalidShadow) {
3041     MoveAndUpdateClosure cl(mark_bitmap(), cm, region_idx);
3042     region_ptr->shadow_to_normal();
3043     return fill_region(cm, cl, region_idx);
3044   } else {
3045     MoveAndUpdateShadowClosure cl(mark_bitmap(), cm, region_idx, shadow_region);
3046     return fill_region(cm, cl, region_idx);
3047   }
3048 }
3049 
3050 void PSParallelCompact::copy_back(HeapWord *shadow_addr, HeapWord *region_addr)
3051 {
3052   Copy::aligned_conjoint_words(shadow_addr, region_addr, _summary_data.RegionSize);
3053 }
3054 
3055 bool PSParallelCompact::steal_unavailable_region(ParCompactionManager* cm, size_t &region_idx)
3056 {
3057   size_t next = cm->next_shadow_region();
3058   ParallelCompactData& sd = summary_data();
3059   size_t old_new_top = sd.addr_to_region_idx(_space_info[old_space_id].new_top());
3060   uint active_gc_threads = ParallelScavengeHeap::heap()->workers().active_workers();
3061 
3062   while (next < old_new_top) {
3063     if (sd.region(next)->mark_shadow()) {
3064       region_idx = next;
3065       return true;
3066     }
3067     next = cm->move_next_shadow_region_by(active_gc_threads);
3068   }
3069 
3070   return false;
3071 }
3072 
3073 // The shadow region is an optimization to address region dependencies in full GC. The basic
3074 // idea is making more regions available by temporally storing their live objects in empty
3075 // shadow regions to resolve dependencies between them and the destination regions. Therefore,
3076 // GC threads need not wait destination regions to be available before processing sources.
3077 //
3078 // A typical workflow would be:
3079 // After draining its own stack and failing to steal from others, a GC worker would pick an
3080 // unavailable region (destination count > 0) and get a shadow region. Then the worker fills
3081 // the shadow region by copying live objects from source regions of the unavailable one. Once
3082 // the unavailable region becomes available, the data in the shadow region will be copied back.
3083 // Shadow regions are empty regions in the to-space and regions between top and end of other spaces.
3084 void PSParallelCompact::initialize_shadow_regions(uint parallel_gc_threads)
3085 {
3086   const ParallelCompactData& sd = PSParallelCompact::summary_data();
3087 
3088   for (unsigned int id = old_space_id; id < last_space_id; ++id) {
3089     SpaceInfo* const space_info = _space_info + id;
3090     MutableSpace* const space = space_info->space();
3091 
3092     const size_t beg_region =
3093       sd.addr_to_region_idx(sd.region_align_up(MAX2(space_info->new_top(), space->top())));
3094     const size_t end_region =
3095       sd.addr_to_region_idx(sd.region_align_down(space->end()));
3096 
3097     for (size_t cur = beg_region; cur < end_region; ++cur) {
3098       ParCompactionManager::push_shadow_region(cur);
3099     }
3100   }
3101 
3102   size_t beg_region = sd.addr_to_region_idx(_space_info[old_space_id].dense_prefix());
3103   for (uint i = 0; i < parallel_gc_threads; i++) {
3104     ParCompactionManager *cm = ParCompactionManager::gc_thread_compaction_manager(i);
3105     cm->set_next_shadow_region(beg_region + i);
3106   }
3107 }
3108 
3109 void PSParallelCompact::fill_blocks(size_t region_idx)
3110 {
3111   // Fill in the block table elements for the specified region.  Each block
3112   // table element holds the number of live words in the region that are to the
3113   // left of the first object that starts in the block.  Thus only blocks in
3114   // which an object starts need to be filled.
3115   //
3116   // The algorithm scans the section of the bitmap that corresponds to the
3117   // region, keeping a running total of the live words.  When an object start is
3118   // found, if it's the first to start in the block that contains it, the
3119   // current total is written to the block table element.
3120   const size_t Log2BlockSize = ParallelCompactData::Log2BlockSize;
3121   const size_t Log2RegionSize = ParallelCompactData::Log2RegionSize;
3122   const size_t RegionSize = ParallelCompactData::RegionSize;
3123 
3124   ParallelCompactData& sd = summary_data();
3125   const size_t partial_obj_size = sd.region(region_idx)->partial_obj_size();
3126   if (partial_obj_size >= RegionSize) {
3127     return; // No objects start in this region.
3128   }
3129 
3130   // Ensure the first loop iteration decides that the block has changed.
3131   size_t cur_block = sd.block_count();
3132 
3133   const ParMarkBitMap* const bitmap = mark_bitmap();
3134 
3135   const size_t Log2BitsPerBlock = Log2BlockSize - LogMinObjAlignment;
3136   assert((size_t)1 << Log2BitsPerBlock ==
3137          bitmap->words_to_bits(ParallelCompactData::BlockSize), "sanity");
3138 
3139   size_t beg_bit = bitmap->words_to_bits(region_idx << Log2RegionSize);
3140   const size_t range_end = beg_bit + bitmap->words_to_bits(RegionSize);
3141   size_t live_bits = bitmap->words_to_bits(partial_obj_size);
3142   beg_bit = bitmap->find_obj_beg(beg_bit + live_bits, range_end);
3143   while (beg_bit < range_end) {
3144     const size_t new_block = beg_bit >> Log2BitsPerBlock;
3145     if (new_block != cur_block) {
3146       cur_block = new_block;
3147       sd.block(cur_block)->set_offset(bitmap->bits_to_words(live_bits));
3148     }
3149 
3150     const size_t end_bit = bitmap->find_obj_end(beg_bit, range_end);
3151     if (end_bit < range_end - 1) {
3152       live_bits += end_bit - beg_bit + 1;
3153       beg_bit = bitmap->find_obj_beg(end_bit + 1, range_end);
3154     } else {
3155       return;
3156     }
3157   }
3158 }
3159 
3160 ParMarkBitMap::IterationStatus MoveAndUpdateClosure::copy_until_full()
3161 {
3162   if (source() != copy_destination()) {
3163     DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3164     Copy::aligned_conjoint_words(source(), copy_destination(), words_remaining());
3165   }
3166   update_state(words_remaining());
3167   assert(is_full(), "sanity");
3168   return ParMarkBitMap::full;
3169 }
3170 
3171 void MoveAndUpdateClosure::copy_partial_obj()
3172 {
3173   size_t words = words_remaining();
3174 
3175   HeapWord* const range_end = MIN2(source() + words, bitmap()->region_end());
3176   HeapWord* const end_addr = bitmap()->find_obj_end(source(), range_end);
3177   if (end_addr < range_end) {
3178     words = bitmap()->obj_size(source(), end_addr);
3179   }
3180 
3181   // This test is necessary; if omitted, the pointer updates to a partial object
3182   // that crosses the dense prefix boundary could be overwritten.
3183   if (source() != copy_destination()) {
3184     DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3185     Copy::aligned_conjoint_words(source(), copy_destination(), words);
3186   }
3187   update_state(words);
3188 }
3189 
3190 void MoveAndUpdateClosure::complete_region(ParCompactionManager *cm, HeapWord *dest_addr,
3191                                            PSParallelCompact::RegionData *region_ptr) {
3192   assert(region_ptr->shadow_state() == ParallelCompactData::RegionData::NormalRegion, "Region should be finished");
3193   region_ptr->set_completed();
3194 }
3195 
3196 ParMarkBitMapClosure::IterationStatus
3197 MoveAndUpdateClosure::do_addr(HeapWord* addr, size_t words) {
3198   assert(destination() != NULL, "sanity");
3199   assert(bitmap()->obj_size(addr) == words, "bad size");
3200 
3201   _source = addr;
3202   assert(PSParallelCompact::summary_data().calc_new_pointer(source(), compaction_manager()) ==
3203          destination(), "wrong destination");
3204 
3205   if (words > words_remaining()) {
3206     return ParMarkBitMap::would_overflow;
3207   }
3208 
3209   // The start_array must be updated even if the object is not moving.
3210   if (_start_array != NULL) {
3211     _start_array->allocate_block(destination());
3212   }
3213 
3214   if (copy_destination() != source()) {
3215     DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3216     Copy::aligned_conjoint_words(source(), copy_destination(), words);
3217   }
3218 
3219   oop moved_oop = cast_to_oop(copy_destination());
3220   compaction_manager()->update_contents(moved_oop);
3221   assert(oopDesc::is_oop_or_null(moved_oop), "Expected an oop or NULL at " PTR_FORMAT, p2i(moved_oop));
3222 
3223   update_state(words);
3224   assert(copy_destination() == cast_from_oop<HeapWord*>(moved_oop) + moved_oop->size(), "sanity");
3225   return is_full() ? ParMarkBitMap::full : ParMarkBitMap::incomplete;
3226 }
3227 
3228 void MoveAndUpdateShadowClosure::complete_region(ParCompactionManager *cm, HeapWord *dest_addr,
3229                                                  PSParallelCompact::RegionData *region_ptr) {
3230   assert(region_ptr->shadow_state() == ParallelCompactData::RegionData::ShadowRegion, "Region should be shadow");
3231   // Record the shadow region index
3232   region_ptr->set_shadow_region(_shadow);
3233   // Mark the shadow region as filled to indicate the data is ready to be
3234   // copied back
3235   region_ptr->mark_filled();
3236   // Try to copy the content of the shadow region back to its corresponding
3237   // heap region if available; the GC thread that decreases the destination
3238   // count to zero will do the copying otherwise (see
3239   // PSParallelCompact::decrement_destination_counts).
3240   if (((region_ptr->available() && region_ptr->claim()) || region_ptr->claimed()) && region_ptr->mark_copied()) {
3241     region_ptr->set_completed();
3242     PSParallelCompact::copy_back(PSParallelCompact::summary_data().region_to_addr(_shadow), dest_addr);
3243     ParCompactionManager::push_shadow_region_mt_safe(_shadow);
3244   }
3245 }
3246 
3247 UpdateOnlyClosure::UpdateOnlyClosure(ParMarkBitMap* mbm,
3248                                      ParCompactionManager* cm,
3249                                      PSParallelCompact::SpaceId space_id) :
3250   ParMarkBitMapClosure(mbm, cm),
3251   _space_id(space_id),
3252   _start_array(PSParallelCompact::start_array(space_id))
3253 {
3254 }
3255 
3256 // Updates the references in the object to their new values.
3257 ParMarkBitMapClosure::IterationStatus
3258 UpdateOnlyClosure::do_addr(HeapWord* addr, size_t words) {
3259   do_addr(addr);
3260   return ParMarkBitMap::incomplete;
3261 }
3262 
3263 FillClosure::FillClosure(ParCompactionManager* cm, PSParallelCompact::SpaceId space_id) :
3264   ParMarkBitMapClosure(PSParallelCompact::mark_bitmap(), cm),
3265   _start_array(PSParallelCompact::start_array(space_id))
3266 {
3267   assert(space_id == PSParallelCompact::old_space_id,
3268          "cannot use FillClosure in the young gen");
3269 }
3270 
3271 ParMarkBitMapClosure::IterationStatus
3272 FillClosure::do_addr(HeapWord* addr, size_t size) {
3273   CollectedHeap::fill_with_objects(addr, size);
3274   HeapWord* const end = addr + size;
3275   do {
3276     _start_array->allocate_block(addr);
3277     addr += cast_to_oop(addr)->size();
3278   } while (addr < end);
3279   return ParMarkBitMap::incomplete;
3280 }
--- EOF ---