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/flatArrayKlass.inline.hpp"
  72 #include "oops/instanceClassLoaderKlass.inline.hpp"
  73 #include "oops/instanceKlass.inline.hpp"
  74 #include "oops/instanceMirrorKlass.inline.hpp"
  75 #include "oops/methodData.hpp"
  76 #include "oops/objArrayKlass.inline.hpp"
  77 #include "oops/oop.inline.hpp"
  78 #include "runtime/atomic.hpp"
  79 #include "runtime/handles.inline.hpp"
  80 #include "runtime/java.hpp"
  81 #include "runtime/safepoint.hpp"
  82 #include "runtime/vmThread.hpp"
  83 #include "services/memTracker.hpp"
  84 #include "services/memoryService.hpp"
  85 #include "utilities/align.hpp"
  86 #include "utilities/debug.hpp"
  87 #include "utilities/events.hpp"
  88 #include "utilities/formatBuffer.hpp"
  89 #include "utilities/macros.hpp"
  90 #include "utilities/stack.inline.hpp"
  91 #if INCLUDE_JVMCI
  92 #include "jvmci/jvmci.hpp"
  93 #endif
  94 
  95 #include <math.h>
  96 
  97 // All sizes are in HeapWords.
  98 const size_t ParallelCompactData::Log2RegionSize  = 16; // 64K words
  99 const size_t ParallelCompactData::RegionSize      = (size_t)1 << Log2RegionSize;
 100 const size_t ParallelCompactData::RegionSizeBytes =
 101   RegionSize << LogHeapWordSize;
 102 const size_t ParallelCompactData::RegionSizeOffsetMask = RegionSize - 1;
 103 const size_t ParallelCompactData::RegionAddrOffsetMask = RegionSizeBytes - 1;
 104 const size_t ParallelCompactData::RegionAddrMask       = ~RegionAddrOffsetMask;
 105 
 106 const size_t ParallelCompactData::Log2BlockSize   = 7; // 128 words
 107 const size_t ParallelCompactData::BlockSize       = (size_t)1 << Log2BlockSize;
 108 const size_t ParallelCompactData::BlockSizeBytes  =
 109   BlockSize << LogHeapWordSize;
 110 const size_t ParallelCompactData::BlockSizeOffsetMask = BlockSize - 1;
 111 const size_t ParallelCompactData::BlockAddrOffsetMask = BlockSizeBytes - 1;
 112 const size_t ParallelCompactData::BlockAddrMask       = ~BlockAddrOffsetMask;
 113 
 114 const size_t ParallelCompactData::BlocksPerRegion = RegionSize / BlockSize;
 115 const size_t ParallelCompactData::Log2BlocksPerRegion =
 116   Log2RegionSize - Log2BlockSize;
 117 
 118 const ParallelCompactData::RegionData::region_sz_t
 119 ParallelCompactData::RegionData::dc_shift = 27;
 120 
 121 const ParallelCompactData::RegionData::region_sz_t
 122 ParallelCompactData::RegionData::dc_mask = ~0U << dc_shift;
 123 
 124 const ParallelCompactData::RegionData::region_sz_t
 125 ParallelCompactData::RegionData::dc_one = 0x1U << dc_shift;
 126 
 127 const ParallelCompactData::RegionData::region_sz_t
 128 ParallelCompactData::RegionData::los_mask = ~dc_mask;
 129 
 130 const ParallelCompactData::RegionData::region_sz_t
 131 ParallelCompactData::RegionData::dc_claimed = 0x8U << dc_shift;
 132 
 133 const ParallelCompactData::RegionData::region_sz_t
 134 ParallelCompactData::RegionData::dc_completed = 0xcU << dc_shift;
 135 
 136 SpaceInfo PSParallelCompact::_space_info[PSParallelCompact::last_space_id];
 137 
 138 SpanSubjectToDiscoveryClosure PSParallelCompact::_span_based_discoverer;
 139 ReferenceProcessor* PSParallelCompact::_ref_processor = NULL;
 140 
 141 double PSParallelCompact::_dwl_mean;
 142 double PSParallelCompact::_dwl_std_dev;
 143 double PSParallelCompact::_dwl_first_term;
 144 double PSParallelCompact::_dwl_adjustment;
 145 #ifdef  ASSERT
 146 bool   PSParallelCompact::_dwl_initialized = false;
 147 #endif  // #ifdef ASSERT
 148 
 149 void SplitInfo::record(size_t src_region_idx, size_t partial_obj_size,
 150                        HeapWord* destination)
 151 {
 152   assert(src_region_idx != 0, "invalid src_region_idx");
 153   assert(partial_obj_size != 0, "invalid partial_obj_size argument");
 154   assert(destination != NULL, "invalid destination argument");
 155 
 156   _src_region_idx = src_region_idx;
 157   _partial_obj_size = partial_obj_size;
 158   _destination = destination;
 159 
 160   // These fields may not be updated below, so make sure they're clear.
 161   assert(_dest_region_addr == NULL, "should have been cleared");
 162   assert(_first_src_addr == NULL, "should have been cleared");
 163 
 164   // Determine the number of destination regions for the partial object.
 165   HeapWord* const last_word = destination + partial_obj_size - 1;
 166   const ParallelCompactData& sd = PSParallelCompact::summary_data();
 167   HeapWord* const beg_region_addr = sd.region_align_down(destination);
 168   HeapWord* const end_region_addr = sd.region_align_down(last_word);
 169 
 170   if (beg_region_addr == end_region_addr) {
 171     // One destination region.
 172     _destination_count = 1;
 173     if (end_region_addr == destination) {
 174       // The destination falls on a region boundary, thus the first word of the
 175       // partial object will be the first word copied to the destination region.
 176       _dest_region_addr = end_region_addr;
 177       _first_src_addr = sd.region_to_addr(src_region_idx);
 178     }
 179   } else {
 180     // Two destination regions.  When copied, the partial object will cross a
 181     // destination region boundary, so a word somewhere within the partial
 182     // object will be the first word copied to the second destination region.
 183     _destination_count = 2;
 184     _dest_region_addr = end_region_addr;
 185     const size_t ofs = pointer_delta(end_region_addr, destination);
 186     assert(ofs < _partial_obj_size, "sanity");
 187     _first_src_addr = sd.region_to_addr(src_region_idx) + ofs;
 188   }
 189 }
 190 
 191 void SplitInfo::clear()
 192 {
 193   _src_region_idx = 0;
 194   _partial_obj_size = 0;
 195   _destination = NULL;
 196   _destination_count = 0;
 197   _dest_region_addr = NULL;
 198   _first_src_addr = NULL;
 199   assert(!is_valid(), "sanity");
 200 }
 201 
 202 #ifdef  ASSERT
 203 void SplitInfo::verify_clear()
 204 {
 205   assert(_src_region_idx == 0, "not clear");
 206   assert(_partial_obj_size == 0, "not clear");
 207   assert(_destination == NULL, "not clear");
 208   assert(_destination_count == 0, "not clear");
 209   assert(_dest_region_addr == NULL, "not clear");
 210   assert(_first_src_addr == NULL, "not clear");
 211 }
 212 #endif  // #ifdef ASSERT
 213 
 214 
 215 void PSParallelCompact::print_on_error(outputStream* st) {
 216   _mark_bitmap.print_on_error(st);
 217 }
 218 
 219 #ifndef PRODUCT
 220 const char* PSParallelCompact::space_names[] = {
 221   "old ", "eden", "from", "to  "
 222 };
 223 
 224 void PSParallelCompact::print_region_ranges() {
 225   if (!log_develop_is_enabled(Trace, gc, compaction)) {
 226     return;
 227   }
 228   Log(gc, compaction) log;
 229   ResourceMark rm;
 230   LogStream ls(log.trace());
 231   Universe::print_on(&ls);
 232   log.trace("space  bottom     top        end        new_top");
 233   log.trace("------ ---------- ---------- ---------- ----------");
 234 
 235   for (unsigned int id = 0; id < last_space_id; ++id) {
 236     const MutableSpace* space = _space_info[id].space();
 237     log.trace("%u %s "
 238               SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " "
 239               SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " ",
 240               id, space_names[id],
 241               summary_data().addr_to_region_idx(space->bottom()),
 242               summary_data().addr_to_region_idx(space->top()),
 243               summary_data().addr_to_region_idx(space->end()),
 244               summary_data().addr_to_region_idx(_space_info[id].new_top()));
 245   }
 246 }
 247 
 248 void
 249 print_generic_summary_region(size_t i, const ParallelCompactData::RegionData* c)
 250 {
 251 #define REGION_IDX_FORMAT        SIZE_FORMAT_W(7)
 252 #define REGION_DATA_FORMAT       SIZE_FORMAT_W(5)
 253 
 254   ParallelCompactData& sd = PSParallelCompact::summary_data();
 255   size_t dci = c->destination() ? sd.addr_to_region_idx(c->destination()) : 0;
 256   log_develop_trace(gc, compaction)(
 257       REGION_IDX_FORMAT " " PTR_FORMAT " "
 258       REGION_IDX_FORMAT " " PTR_FORMAT " "
 259       REGION_DATA_FORMAT " " REGION_DATA_FORMAT " "
 260       REGION_DATA_FORMAT " " REGION_IDX_FORMAT " %d",
 261       i, p2i(c->data_location()), dci, p2i(c->destination()),
 262       c->partial_obj_size(), c->live_obj_size(),
 263       c->data_size(), c->source_region(), c->destination_count());
 264 
 265 #undef  REGION_IDX_FORMAT
 266 #undef  REGION_DATA_FORMAT
 267 }
 268 
 269 void
 270 print_generic_summary_data(ParallelCompactData& summary_data,
 271                            HeapWord* const beg_addr,
 272                            HeapWord* const end_addr)
 273 {
 274   size_t total_words = 0;
 275   size_t i = summary_data.addr_to_region_idx(beg_addr);
 276   const size_t last = summary_data.addr_to_region_idx(end_addr);
 277   HeapWord* pdest = 0;
 278 
 279   while (i < last) {
 280     ParallelCompactData::RegionData* c = summary_data.region(i);
 281     if (c->data_size() != 0 || c->destination() != pdest) {
 282       print_generic_summary_region(i, c);
 283       total_words += c->data_size();
 284       pdest = c->destination();
 285     }
 286     ++i;
 287   }
 288 
 289   log_develop_trace(gc, compaction)("summary_data_bytes=" SIZE_FORMAT, total_words * HeapWordSize);
 290 }
 291 
 292 void
 293 PSParallelCompact::print_generic_summary_data(ParallelCompactData& summary_data,
 294                                               HeapWord* const beg_addr,
 295                                               HeapWord* const end_addr) {
 296   ::print_generic_summary_data(summary_data,beg_addr, end_addr);
 297 }
 298 
 299 void
 300 print_generic_summary_data(ParallelCompactData& summary_data,
 301                            SpaceInfo* space_info)
 302 {
 303   if (!log_develop_is_enabled(Trace, gc, compaction)) {
 304     return;
 305   }
 306 
 307   for (unsigned int id = 0; id < PSParallelCompact::last_space_id; ++id) {
 308     const MutableSpace* space = space_info[id].space();
 309     print_generic_summary_data(summary_data, space->bottom(),
 310                                MAX2(space->top(), space_info[id].new_top()));
 311   }
 312 }
 313 
 314 void
 315 print_initial_summary_data(ParallelCompactData& summary_data,
 316                            const MutableSpace* space) {
 317   if (space->top() == space->bottom()) {
 318     return;
 319   }
 320 
 321   const size_t region_size = ParallelCompactData::RegionSize;
 322   typedef ParallelCompactData::RegionData RegionData;
 323   HeapWord* const top_aligned_up = summary_data.region_align_up(space->top());
 324   const size_t end_region = summary_data.addr_to_region_idx(top_aligned_up);
 325   const RegionData* c = summary_data.region(end_region - 1);
 326   HeapWord* end_addr = c->destination() + c->data_size();
 327   const size_t live_in_space = pointer_delta(end_addr, space->bottom());
 328 
 329   // Print (and count) the full regions at the beginning of the space.
 330   size_t full_region_count = 0;
 331   size_t i = summary_data.addr_to_region_idx(space->bottom());
 332   while (i < end_region && summary_data.region(i)->data_size() == region_size) {
 333     ParallelCompactData::RegionData* c = summary_data.region(i);
 334     log_develop_trace(gc, compaction)(
 335         SIZE_FORMAT_W(5) " " PTR_FORMAT " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d",
 336         i, p2i(c->destination()),
 337         c->partial_obj_size(), c->live_obj_size(),
 338         c->data_size(), c->source_region(), c->destination_count());
 339     ++full_region_count;
 340     ++i;
 341   }
 342 
 343   size_t live_to_right = live_in_space - full_region_count * region_size;
 344 
 345   double max_reclaimed_ratio = 0.0;
 346   size_t max_reclaimed_ratio_region = 0;
 347   size_t max_dead_to_right = 0;
 348   size_t max_live_to_right = 0;
 349 
 350   // Print the 'reclaimed ratio' for regions while there is something live in
 351   // the region or to the right of it.  The remaining regions are empty (and
 352   // uninteresting), and computing the ratio will result in division by 0.
 353   while (i < end_region && live_to_right > 0) {
 354     c = summary_data.region(i);
 355     HeapWord* const region_addr = summary_data.region_to_addr(i);
 356     const size_t used_to_right = pointer_delta(space->top(), region_addr);
 357     const size_t dead_to_right = used_to_right - live_to_right;
 358     const double reclaimed_ratio = double(dead_to_right) / live_to_right;
 359 
 360     if (reclaimed_ratio > max_reclaimed_ratio) {
 361             max_reclaimed_ratio = reclaimed_ratio;
 362             max_reclaimed_ratio_region = i;
 363             max_dead_to_right = dead_to_right;
 364             max_live_to_right = live_to_right;
 365     }
 366 
 367     ParallelCompactData::RegionData* c = summary_data.region(i);
 368     log_develop_trace(gc, compaction)(
 369         SIZE_FORMAT_W(5) " " PTR_FORMAT " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d"
 370         "%12.10f " SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10),
 371         i, p2i(c->destination()),
 372         c->partial_obj_size(), c->live_obj_size(),
 373         c->data_size(), c->source_region(), c->destination_count(),
 374         reclaimed_ratio, dead_to_right, live_to_right);
 375 
 376 
 377     live_to_right -= c->data_size();
 378     ++i;
 379   }
 380 
 381   // Any remaining regions are empty.  Print one more if there is one.
 382   if (i < end_region) {
 383     ParallelCompactData::RegionData* c = summary_data.region(i);
 384     log_develop_trace(gc, compaction)(
 385         SIZE_FORMAT_W(5) " " PTR_FORMAT " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d",
 386          i, p2i(c->destination()),
 387          c->partial_obj_size(), c->live_obj_size(),
 388          c->data_size(), c->source_region(), c->destination_count());
 389   }
 390 
 391   log_develop_trace(gc, compaction)("max:  " SIZE_FORMAT_W(4) " d2r=" SIZE_FORMAT_W(10) " l2r=" SIZE_FORMAT_W(10) " max_ratio=%14.12f",
 392                                     max_reclaimed_ratio_region, max_dead_to_right, max_live_to_right, max_reclaimed_ratio);
 393 }
 394 
 395 void
 396 print_initial_summary_data(ParallelCompactData& summary_data,
 397                            SpaceInfo* space_info) {
 398   if (!log_develop_is_enabled(Trace, gc, compaction)) {
 399     return;
 400   }
 401 
 402   unsigned int id = PSParallelCompact::old_space_id;
 403   const MutableSpace* space;
 404   do {
 405     space = space_info[id].space();
 406     print_initial_summary_data(summary_data, space);
 407   } while (++id < PSParallelCompact::eden_space_id);
 408 
 409   do {
 410     space = space_info[id].space();
 411     print_generic_summary_data(summary_data, space->bottom(), space->top());
 412   } while (++id < PSParallelCompact::last_space_id);
 413 }
 414 #endif  // #ifndef PRODUCT
 415 
 416 ParallelCompactData::ParallelCompactData() :
 417   _region_start(NULL),
 418   DEBUG_ONLY(_region_end(NULL) COMMA)
 419   _region_vspace(NULL),
 420   _reserved_byte_size(0),
 421   _region_data(NULL),
 422   _region_count(0),
 423   _block_vspace(NULL),
 424   _block_data(NULL),
 425   _block_count(0) {}
 426 
 427 bool ParallelCompactData::initialize(MemRegion covered_region)
 428 {
 429   _region_start = covered_region.start();
 430   const size_t region_size = covered_region.word_size();
 431   DEBUG_ONLY(_region_end = _region_start + region_size;)
 432 
 433   assert(region_align_down(_region_start) == _region_start,
 434          "region start not aligned");
 435   assert((region_size & RegionSizeOffsetMask) == 0,
 436          "region size not a multiple of RegionSize");
 437 
 438   bool result = initialize_region_data(region_size) && initialize_block_data();
 439   return result;
 440 }
 441 
 442 PSVirtualSpace*
 443 ParallelCompactData::create_vspace(size_t count, size_t element_size)
 444 {
 445   const size_t raw_bytes = count * element_size;
 446   const size_t page_sz = os::page_size_for_region_aligned(raw_bytes, 10);
 447   const size_t granularity = os::vm_allocation_granularity();
 448   _reserved_byte_size = align_up(raw_bytes, MAX2(page_sz, granularity));
 449 
 450   const size_t rs_align = page_sz == (size_t) os::vm_page_size() ? 0 :
 451     MAX2(page_sz, granularity);
 452   ReservedSpace rs(_reserved_byte_size, rs_align, page_sz);
 453   os::trace_page_sizes("Parallel Compact Data", raw_bytes, raw_bytes, page_sz, rs.base(),
 454                        rs.size());
 455 
 456   MemTracker::record_virtual_memory_type((address)rs.base(), mtGC);
 457 
 458   PSVirtualSpace* vspace = new PSVirtualSpace(rs, page_sz);
 459   if (vspace != 0) {
 460     if (vspace->expand_by(_reserved_byte_size)) {
 461       return vspace;
 462     }
 463     delete vspace;
 464     // Release memory reserved in the space.
 465     rs.release();
 466   }
 467 
 468   return 0;
 469 }
 470 
 471 bool ParallelCompactData::initialize_region_data(size_t region_size)
 472 {
 473   const size_t count = (region_size + RegionSizeOffsetMask) >> Log2RegionSize;
 474   _region_vspace = create_vspace(count, sizeof(RegionData));
 475   if (_region_vspace != 0) {
 476     _region_data = (RegionData*)_region_vspace->reserved_low_addr();
 477     _region_count = count;
 478     return true;
 479   }
 480   return false;
 481 }
 482 
 483 bool ParallelCompactData::initialize_block_data()
 484 {
 485   assert(_region_count != 0, "region data must be initialized first");
 486   const size_t count = _region_count << Log2BlocksPerRegion;
 487   _block_vspace = create_vspace(count, sizeof(BlockData));
 488   if (_block_vspace != 0) {
 489     _block_data = (BlockData*)_block_vspace->reserved_low_addr();
 490     _block_count = count;
 491     return true;
 492   }
 493   return false;
 494 }
 495 
 496 void ParallelCompactData::clear()
 497 {
 498   memset(_region_data, 0, _region_vspace->committed_size());
 499   memset(_block_data, 0, _block_vspace->committed_size());
 500 }
 501 
 502 void ParallelCompactData::clear_range(size_t beg_region, size_t end_region) {
 503   assert(beg_region <= _region_count, "beg_region out of range");
 504   assert(end_region <= _region_count, "end_region out of range");
 505   assert(RegionSize % BlockSize == 0, "RegionSize not a multiple of BlockSize");
 506 
 507   const size_t region_cnt = end_region - beg_region;
 508   memset(_region_data + beg_region, 0, region_cnt * sizeof(RegionData));
 509 
 510   const size_t beg_block = beg_region * BlocksPerRegion;
 511   const size_t block_cnt = region_cnt * BlocksPerRegion;
 512   memset(_block_data + beg_block, 0, block_cnt * sizeof(BlockData));
 513 }
 514 
 515 HeapWord* ParallelCompactData::partial_obj_end(size_t region_idx) const
 516 {
 517   const RegionData* cur_cp = region(region_idx);
 518   const RegionData* const end_cp = region(region_count() - 1);
 519 
 520   HeapWord* result = region_to_addr(region_idx);
 521   if (cur_cp < end_cp) {
 522     do {
 523       result += cur_cp->partial_obj_size();
 524     } while (cur_cp->partial_obj_size() == RegionSize && ++cur_cp < end_cp);
 525   }
 526   return result;
 527 }
 528 
 529 void ParallelCompactData::add_obj(HeapWord* addr, size_t len)
 530 {
 531   const size_t obj_ofs = pointer_delta(addr, _region_start);
 532   const size_t beg_region = obj_ofs >> Log2RegionSize;
 533   // end_region is inclusive
 534   const size_t end_region = (obj_ofs + len - 1) >> Log2RegionSize;
 535 
 536   if (beg_region == end_region) {
 537     // All in one region.
 538     _region_data[beg_region].add_live_obj(len);
 539     return;
 540   }
 541 
 542   // First region.
 543   const size_t beg_ofs = region_offset(addr);
 544   _region_data[beg_region].add_live_obj(RegionSize - beg_ofs);
 545 
 546   // Middle regions--completely spanned by this object.
 547   for (size_t region = beg_region + 1; region < end_region; ++region) {
 548     _region_data[region].set_partial_obj_size(RegionSize);
 549     _region_data[region].set_partial_obj_addr(addr);
 550   }
 551 
 552   // Last region.
 553   const size_t end_ofs = region_offset(addr + len - 1);
 554   _region_data[end_region].set_partial_obj_size(end_ofs + 1);
 555   _region_data[end_region].set_partial_obj_addr(addr);
 556 }
 557 
 558 void
 559 ParallelCompactData::summarize_dense_prefix(HeapWord* beg, HeapWord* end)
 560 {
 561   assert(is_region_aligned(beg), "not RegionSize aligned");
 562   assert(is_region_aligned(end), "not RegionSize aligned");
 563 
 564   size_t cur_region = addr_to_region_idx(beg);
 565   const size_t end_region = addr_to_region_idx(end);
 566   HeapWord* addr = beg;
 567   while (cur_region < end_region) {
 568     _region_data[cur_region].set_destination(addr);
 569     _region_data[cur_region].set_destination_count(0);
 570     _region_data[cur_region].set_source_region(cur_region);
 571     _region_data[cur_region].set_data_location(addr);
 572 
 573     // Update live_obj_size so the region appears completely full.
 574     size_t live_size = RegionSize - _region_data[cur_region].partial_obj_size();
 575     _region_data[cur_region].set_live_obj_size(live_size);
 576 
 577     ++cur_region;
 578     addr += RegionSize;
 579   }
 580 }
 581 
 582 // Find the point at which a space can be split and, if necessary, record the
 583 // split point.
 584 //
 585 // If the current src region (which overflowed the destination space) doesn't
 586 // have a partial object, the split point is at the beginning of the current src
 587 // region (an "easy" split, no extra bookkeeping required).
 588 //
 589 // If the current src region has a partial object, the split point is in the
 590 // region where that partial object starts (call it the split_region).  If
 591 // split_region has a partial object, then the split point is just after that
 592 // partial object (a "hard" split where we have to record the split data and
 593 // zero the partial_obj_size field).  With a "hard" split, we know that the
 594 // partial_obj ends within split_region because the partial object that caused
 595 // the overflow starts in split_region.  If split_region doesn't have a partial
 596 // obj, then the split is at the beginning of split_region (another "easy"
 597 // split).
 598 HeapWord*
 599 ParallelCompactData::summarize_split_space(size_t src_region,
 600                                            SplitInfo& split_info,
 601                                            HeapWord* destination,
 602                                            HeapWord* target_end,
 603                                            HeapWord** target_next)
 604 {
 605   assert(destination <= target_end, "sanity");
 606   assert(destination + _region_data[src_region].data_size() > target_end,
 607     "region should not fit into target space");
 608   assert(is_region_aligned(target_end), "sanity");
 609 
 610   size_t split_region = src_region;
 611   HeapWord* split_destination = destination;
 612   size_t partial_obj_size = _region_data[src_region].partial_obj_size();
 613 
 614   if (destination + partial_obj_size > target_end) {
 615     // The split point is just after the partial object (if any) in the
 616     // src_region that contains the start of the object that overflowed the
 617     // destination space.
 618     //
 619     // Find the start of the "overflow" object and set split_region to the
 620     // region containing it.
 621     HeapWord* const overflow_obj = _region_data[src_region].partial_obj_addr();
 622     split_region = addr_to_region_idx(overflow_obj);
 623 
 624     // Clear the source_region field of all destination regions whose first word
 625     // came from data after the split point (a non-null source_region field
 626     // implies a region must be filled).
 627     //
 628     // An alternative to the simple loop below:  clear during post_compact(),
 629     // which uses memcpy instead of individual stores, and is easy to
 630     // parallelize.  (The downside is that it clears the entire RegionData
 631     // object as opposed to just one field.)
 632     //
 633     // post_compact() would have to clear the summary data up to the highest
 634     // address that was written during the summary phase, which would be
 635     //
 636     //         max(top, max(new_top, clear_top))
 637     //
 638     // where clear_top is a new field in SpaceInfo.  Would have to set clear_top
 639     // to target_end.
 640     const RegionData* const sr = region(split_region);
 641     const size_t beg_idx =
 642       addr_to_region_idx(region_align_up(sr->destination() +
 643                                          sr->partial_obj_size()));
 644     const size_t end_idx = addr_to_region_idx(target_end);
 645 
 646     log_develop_trace(gc, compaction)("split:  clearing source_region field in [" SIZE_FORMAT ", " SIZE_FORMAT ")", beg_idx, end_idx);
 647     for (size_t idx = beg_idx; idx < end_idx; ++idx) {
 648       _region_data[idx].set_source_region(0);
 649     }
 650 
 651     // Set split_destination and partial_obj_size to reflect the split region.
 652     split_destination = sr->destination();
 653     partial_obj_size = sr->partial_obj_size();
 654   }
 655 
 656   // The split is recorded only if a partial object extends onto the region.
 657   if (partial_obj_size != 0) {
 658     _region_data[split_region].set_partial_obj_size(0);
 659     split_info.record(split_region, partial_obj_size, split_destination);
 660   }
 661 
 662   // Setup the continuation addresses.
 663   *target_next = split_destination + partial_obj_size;
 664   HeapWord* const source_next = region_to_addr(split_region) + partial_obj_size;
 665 
 666   if (log_develop_is_enabled(Trace, gc, compaction)) {
 667     const char * split_type = partial_obj_size == 0 ? "easy" : "hard";
 668     log_develop_trace(gc, compaction)("%s split:  src=" PTR_FORMAT " src_c=" SIZE_FORMAT " pos=" SIZE_FORMAT,
 669                                       split_type, p2i(source_next), split_region, partial_obj_size);
 670     log_develop_trace(gc, compaction)("%s split:  dst=" PTR_FORMAT " dst_c=" SIZE_FORMAT " tn=" PTR_FORMAT,
 671                                       split_type, p2i(split_destination),
 672                                       addr_to_region_idx(split_destination),
 673                                       p2i(*target_next));
 674 
 675     if (partial_obj_size != 0) {
 676       HeapWord* const po_beg = split_info.destination();
 677       HeapWord* const po_end = po_beg + split_info.partial_obj_size();
 678       log_develop_trace(gc, compaction)("%s split:  po_beg=" PTR_FORMAT " " SIZE_FORMAT " po_end=" PTR_FORMAT " " SIZE_FORMAT,
 679                                         split_type,
 680                                         p2i(po_beg), addr_to_region_idx(po_beg),
 681                                         p2i(po_end), addr_to_region_idx(po_end));
 682     }
 683   }
 684 
 685   return source_next;
 686 }
 687 
 688 bool ParallelCompactData::summarize(SplitInfo& split_info,
 689                                     HeapWord* source_beg, HeapWord* source_end,
 690                                     HeapWord** source_next,
 691                                     HeapWord* target_beg, HeapWord* target_end,
 692                                     HeapWord** target_next)
 693 {
 694   HeapWord* const source_next_val = source_next == NULL ? NULL : *source_next;
 695   log_develop_trace(gc, compaction)(
 696       "sb=" PTR_FORMAT " se=" PTR_FORMAT " sn=" PTR_FORMAT
 697       "tb=" PTR_FORMAT " te=" PTR_FORMAT " tn=" PTR_FORMAT,
 698       p2i(source_beg), p2i(source_end), p2i(source_next_val),
 699       p2i(target_beg), p2i(target_end), p2i(*target_next));
 700 
 701   size_t cur_region = addr_to_region_idx(source_beg);
 702   const size_t end_region = addr_to_region_idx(region_align_up(source_end));
 703 
 704   HeapWord *dest_addr = target_beg;
 705   while (cur_region < end_region) {
 706     // The destination must be set even if the region has no data.
 707     _region_data[cur_region].set_destination(dest_addr);
 708 
 709     size_t words = _region_data[cur_region].data_size();
 710     if (words > 0) {
 711       // If cur_region does not fit entirely into the target space, find a point
 712       // at which the source space can be 'split' so that part is copied to the
 713       // target space and the rest is copied elsewhere.
 714       if (dest_addr + words > target_end) {
 715         assert(source_next != NULL, "source_next is NULL when splitting");
 716         *source_next = summarize_split_space(cur_region, split_info, dest_addr,
 717                                              target_end, target_next);
 718         return false;
 719       }
 720 
 721       // Compute the destination_count for cur_region, and if necessary, update
 722       // source_region for a destination region.  The source_region field is
 723       // updated if cur_region is the first (left-most) region to be copied to a
 724       // destination region.
 725       //
 726       // The destination_count calculation is a bit subtle.  A region that has
 727       // data that compacts into itself does not count itself as a destination.
 728       // This maintains the invariant that a zero count means the region is
 729       // available and can be claimed and then filled.
 730       uint destination_count = 0;
 731       if (split_info.is_split(cur_region)) {
 732         // The current region has been split:  the partial object will be copied
 733         // to one destination space and the remaining data will be copied to
 734         // another destination space.  Adjust the initial destination_count and,
 735         // if necessary, set the source_region field if the partial object will
 736         // cross a destination region boundary.
 737         destination_count = split_info.destination_count();
 738         if (destination_count == 2) {
 739           size_t dest_idx = addr_to_region_idx(split_info.dest_region_addr());
 740           _region_data[dest_idx].set_source_region(cur_region);
 741         }
 742       }
 743 
 744       HeapWord* const last_addr = dest_addr + words - 1;
 745       const size_t dest_region_1 = addr_to_region_idx(dest_addr);
 746       const size_t dest_region_2 = addr_to_region_idx(last_addr);
 747 
 748       // Initially assume that the destination regions will be the same and
 749       // adjust the value below if necessary.  Under this assumption, if
 750       // cur_region == dest_region_2, then cur_region will be compacted
 751       // completely into itself.
 752       destination_count += cur_region == dest_region_2 ? 0 : 1;
 753       if (dest_region_1 != dest_region_2) {
 754         // Destination regions differ; adjust destination_count.
 755         destination_count += 1;
 756         // Data from cur_region will be copied to the start of dest_region_2.
 757         _region_data[dest_region_2].set_source_region(cur_region);
 758       } else if (is_region_aligned(dest_addr)) {
 759         // Data from cur_region will be copied to the start of the destination
 760         // region.
 761         _region_data[dest_region_1].set_source_region(cur_region);
 762       }
 763 
 764       _region_data[cur_region].set_destination_count(destination_count);
 765       _region_data[cur_region].set_data_location(region_to_addr(cur_region));
 766       dest_addr += words;
 767     }
 768 
 769     ++cur_region;
 770   }
 771 
 772   *target_next = dest_addr;
 773   return true;
 774 }
 775 
 776 HeapWord* ParallelCompactData::calc_new_pointer(HeapWord* addr, ParCompactionManager* cm) const {
 777   assert(addr != NULL, "Should detect NULL oop earlier");
 778   assert(ParallelScavengeHeap::heap()->is_in(addr), "not in heap");
 779   assert(PSParallelCompact::mark_bitmap()->is_marked(addr), "not marked");
 780 
 781   // Region covering the object.
 782   RegionData* const region_ptr = addr_to_region_ptr(addr);
 783   HeapWord* result = region_ptr->destination();
 784 
 785   // If the entire Region is live, the new location is region->destination + the
 786   // offset of the object within in the Region.
 787 
 788   // Run some performance tests to determine if this special case pays off.  It
 789   // is worth it for pointers into the dense prefix.  If the optimization to
 790   // avoid pointer updates in regions that only point to the dense prefix is
 791   // ever implemented, this should be revisited.
 792   if (region_ptr->data_size() == RegionSize) {
 793     result += region_offset(addr);
 794     return result;
 795   }
 796 
 797   // Otherwise, the new location is region->destination + block offset + the
 798   // number of live words in the Block that are (a) to the left of addr and (b)
 799   // due to objects that start in the Block.
 800 
 801   // Fill in the block table if necessary.  This is unsynchronized, so multiple
 802   // threads may fill the block table for a region (harmless, since it is
 803   // idempotent).
 804   if (!region_ptr->blocks_filled()) {
 805     PSParallelCompact::fill_blocks(addr_to_region_idx(addr));
 806     region_ptr->set_blocks_filled();
 807   }
 808 
 809   HeapWord* const search_start = block_align_down(addr);
 810   const size_t block_offset = addr_to_block_ptr(addr)->offset();
 811 
 812   const ParMarkBitMap* bitmap = PSParallelCompact::mark_bitmap();
 813   const size_t live = bitmap->live_words_in_range(cm, search_start, cast_to_oop(addr));
 814   result += block_offset + live;
 815   DEBUG_ONLY(PSParallelCompact::check_new_location(addr, result));
 816   return result;
 817 }
 818 
 819 #ifdef ASSERT
 820 void ParallelCompactData::verify_clear(const PSVirtualSpace* vspace)
 821 {
 822   const size_t* const beg = (const size_t*)vspace->committed_low_addr();
 823   const size_t* const end = (const size_t*)vspace->committed_high_addr();
 824   for (const size_t* p = beg; p < end; ++p) {
 825     assert(*p == 0, "not zero");
 826   }
 827 }
 828 
 829 void ParallelCompactData::verify_clear()
 830 {
 831   verify_clear(_region_vspace);
 832   verify_clear(_block_vspace);
 833 }
 834 #endif  // #ifdef ASSERT
 835 
 836 STWGCTimer          PSParallelCompact::_gc_timer;
 837 ParallelOldTracer   PSParallelCompact::_gc_tracer;
 838 elapsedTimer        PSParallelCompact::_accumulated_time;
 839 unsigned int        PSParallelCompact::_total_invocations = 0;
 840 unsigned int        PSParallelCompact::_maximum_compaction_gc_num = 0;
 841 CollectorCounters*  PSParallelCompact::_counters = NULL;
 842 ParMarkBitMap       PSParallelCompact::_mark_bitmap;
 843 ParallelCompactData PSParallelCompact::_summary_data;
 844 
 845 PSParallelCompact::IsAliveClosure PSParallelCompact::_is_alive_closure;
 846 
 847 bool PSParallelCompact::IsAliveClosure::do_object_b(oop p) { return mark_bitmap()->is_marked(p); }
 848 
 849 class PCReferenceProcessor: public ReferenceProcessor {
 850 public:
 851   PCReferenceProcessor(
 852     BoolObjectClosure* is_subject_to_discovery,
 853     BoolObjectClosure* is_alive_non_header) :
 854       ReferenceProcessor(is_subject_to_discovery,
 855       ParallelGCThreads,   // mt processing degree
 856       ParallelGCThreads,   // mt discovery degree
 857       true,                // atomic_discovery
 858       is_alive_non_header) {
 859   }
 860 
 861   template<typename T> bool discover(oop obj, ReferenceType type) {
 862     T* referent_addr = (T*) java_lang_ref_Reference::referent_addr_raw(obj);
 863     T heap_oop = RawAccess<>::oop_load(referent_addr);
 864     oop referent = CompressedOops::decode_not_null(heap_oop);
 865     return PSParallelCompact::mark_bitmap()->is_unmarked(referent)
 866         && ReferenceProcessor::discover_reference(obj, type);
 867   }
 868   virtual bool discover_reference(oop obj, ReferenceType type) {
 869     if (UseCompressedOops) {
 870       return discover<narrowOop>(obj, type);
 871     } else {
 872       return discover<oop>(obj, type);
 873     }
 874   }
 875 };
 876 
 877 void PSParallelCompact::post_initialize() {
 878   ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
 879   _span_based_discoverer.set_span(heap->reserved_region());
 880   _ref_processor =
 881     new PCReferenceProcessor(&_span_based_discoverer,
 882                              &_is_alive_closure); // non-header is alive closure
 883 
 884   _counters = new CollectorCounters("Parallel full collection pauses", 1);
 885 
 886   // Initialize static fields in ParCompactionManager.
 887   ParCompactionManager::initialize(mark_bitmap());
 888 }
 889 
 890 bool PSParallelCompact::initialize() {
 891   ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
 892   MemRegion mr = heap->reserved_region();
 893 
 894   // Was the old gen get allocated successfully?
 895   if (!heap->old_gen()->is_allocated()) {
 896     return false;
 897   }
 898 
 899   initialize_space_info();
 900   initialize_dead_wood_limiter();
 901 
 902   if (!_mark_bitmap.initialize(mr)) {
 903     vm_shutdown_during_initialization(
 904       err_msg("Unable to allocate " SIZE_FORMAT "KB bitmaps for parallel "
 905       "garbage collection for the requested " SIZE_FORMAT "KB heap.",
 906       _mark_bitmap.reserved_byte_size()/K, mr.byte_size()/K));
 907     return false;
 908   }
 909 
 910   if (!_summary_data.initialize(mr)) {
 911     vm_shutdown_during_initialization(
 912       err_msg("Unable to allocate " SIZE_FORMAT "KB card tables for parallel "
 913       "garbage collection for the requested " SIZE_FORMAT "KB heap.",
 914       _summary_data.reserved_byte_size()/K, mr.byte_size()/K));
 915     return false;
 916   }
 917 
 918   return true;
 919 }
 920 
 921 void PSParallelCompact::initialize_space_info()
 922 {
 923   memset(&_space_info, 0, sizeof(_space_info));
 924 
 925   ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
 926   PSYoungGen* young_gen = heap->young_gen();
 927 
 928   _space_info[old_space_id].set_space(heap->old_gen()->object_space());
 929   _space_info[eden_space_id].set_space(young_gen->eden_space());
 930   _space_info[from_space_id].set_space(young_gen->from_space());
 931   _space_info[to_space_id].set_space(young_gen->to_space());
 932 
 933   _space_info[old_space_id].set_start_array(heap->old_gen()->start_array());
 934 }
 935 
 936 void PSParallelCompact::initialize_dead_wood_limiter()
 937 {
 938   const size_t max = 100;
 939   _dwl_mean = double(MIN2(ParallelOldDeadWoodLimiterMean, max)) / 100.0;
 940   _dwl_std_dev = double(MIN2(ParallelOldDeadWoodLimiterStdDev, max)) / 100.0;
 941   _dwl_first_term = 1.0 / (sqrt(2.0 * M_PI) * _dwl_std_dev);
 942   DEBUG_ONLY(_dwl_initialized = true;)
 943   _dwl_adjustment = normal_distribution(1.0);
 944 }
 945 
 946 void
 947 PSParallelCompact::clear_data_covering_space(SpaceId id)
 948 {
 949   // At this point, top is the value before GC, new_top() is the value that will
 950   // be set at the end of GC.  The marking bitmap is cleared to top; nothing
 951   // should be marked above top.  The summary data is cleared to the larger of
 952   // top & new_top.
 953   MutableSpace* const space = _space_info[id].space();
 954   HeapWord* const bot = space->bottom();
 955   HeapWord* const top = space->top();
 956   HeapWord* const max_top = MAX2(top, _space_info[id].new_top());
 957 
 958   const idx_t beg_bit = _mark_bitmap.addr_to_bit(bot);
 959   const idx_t end_bit = _mark_bitmap.align_range_end(_mark_bitmap.addr_to_bit(top));
 960   _mark_bitmap.clear_range(beg_bit, end_bit);
 961 
 962   const size_t beg_region = _summary_data.addr_to_region_idx(bot);
 963   const size_t end_region =
 964     _summary_data.addr_to_region_idx(_summary_data.region_align_up(max_top));
 965   _summary_data.clear_range(beg_region, end_region);
 966 
 967   // Clear the data used to 'split' regions.
 968   SplitInfo& split_info = _space_info[id].split_info();
 969   if (split_info.is_valid()) {
 970     split_info.clear();
 971   }
 972   DEBUG_ONLY(split_info.verify_clear();)
 973 }
 974 
 975 void PSParallelCompact::pre_compact()
 976 {
 977   // Update the from & to space pointers in space_info, since they are swapped
 978   // at each young gen gc.  Do the update unconditionally (even though a
 979   // promotion failure does not swap spaces) because an unknown number of young
 980   // collections will have swapped the spaces an unknown number of times.
 981   GCTraceTime(Debug, gc, phases) tm("Pre Compact", &_gc_timer);
 982   ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
 983   _space_info[from_space_id].set_space(heap->young_gen()->from_space());
 984   _space_info[to_space_id].set_space(heap->young_gen()->to_space());
 985 
 986   // Increment the invocation count
 987   heap->increment_total_collections(true);
 988 
 989   // We need to track unique mark sweep invocations as well.
 990   _total_invocations++;
 991 
 992   heap->print_heap_before_gc();
 993   heap->trace_heap_before_gc(&_gc_tracer);
 994 
 995   // Fill in TLABs
 996   heap->ensure_parsability(true);  // retire TLABs
 997 
 998   if (VerifyBeforeGC && heap->total_collections() >= VerifyGCStartAt) {
 999     Universe::verify("Before GC");
1000   }
1001 
1002   // Verify object start arrays
1003   if (VerifyObjectStartArray &&
1004       VerifyBeforeGC) {
1005     heap->old_gen()->verify_object_start_array();
1006   }
1007 
1008   DEBUG_ONLY(mark_bitmap()->verify_clear();)
1009   DEBUG_ONLY(summary_data().verify_clear();)
1010 
1011   ParCompactionManager::reset_all_bitmap_query_caches();
1012 }
1013 
1014 void PSParallelCompact::post_compact()
1015 {
1016   GCTraceTime(Info, gc, phases) tm("Post Compact", &_gc_timer);
1017   ParCompactionManager::remove_all_shadow_regions();
1018 
1019   for (unsigned int id = old_space_id; id < last_space_id; ++id) {
1020     // Clear the marking bitmap, summary data and split info.
1021     clear_data_covering_space(SpaceId(id));
1022     // Update top().  Must be done after clearing the bitmap and summary data.
1023     _space_info[id].publish_new_top();
1024   }
1025 
1026   ParCompactionManager::flush_all_string_dedup_requests();
1027 
1028   MutableSpace* const eden_space = _space_info[eden_space_id].space();
1029   MutableSpace* const from_space = _space_info[from_space_id].space();
1030   MutableSpace* const to_space   = _space_info[to_space_id].space();
1031 
1032   ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
1033   bool eden_empty = eden_space->is_empty();
1034 
1035   // Update heap occupancy information which is used as input to the soft ref
1036   // clearing policy at the next gc.
1037   Universe::heap()->update_capacity_and_used_at_gc();
1038 
1039   bool young_gen_empty = eden_empty && from_space->is_empty() &&
1040     to_space->is_empty();
1041 
1042   PSCardTable* ct = heap->card_table();
1043   MemRegion old_mr = heap->old_gen()->reserved();
1044   if (young_gen_empty) {
1045     ct->clear(MemRegion(old_mr.start(), old_mr.end()));
1046   } else {
1047     ct->invalidate(MemRegion(old_mr.start(), old_mr.end()));
1048   }
1049 
1050   // Delete metaspaces for unloaded class loaders and clean up loader_data graph
1051   ClassLoaderDataGraph::purge(/*at_safepoint*/true);
1052   DEBUG_ONLY(MetaspaceUtils::verify();)
1053 
1054   heap->prune_scavengable_nmethods();
1055 
1056 #if COMPILER2_OR_JVMCI
1057   DerivedPointerTable::update_pointers();
1058 #endif
1059 
1060   if (ZapUnusedHeapArea) {
1061     heap->gen_mangle_unused_area();
1062   }
1063 
1064   // Signal that we have completed a visit to all live objects.
1065   Universe::heap()->record_whole_heap_examined_timestamp();
1066 }
1067 
1068 HeapWord*
1069 PSParallelCompact::compute_dense_prefix_via_density(const SpaceId id,
1070                                                     bool maximum_compaction)
1071 {
1072   const size_t region_size = ParallelCompactData::RegionSize;
1073   const ParallelCompactData& sd = summary_data();
1074 
1075   const MutableSpace* const space = _space_info[id].space();
1076   HeapWord* const top_aligned_up = sd.region_align_up(space->top());
1077   const RegionData* const beg_cp = sd.addr_to_region_ptr(space->bottom());
1078   const RegionData* const end_cp = sd.addr_to_region_ptr(top_aligned_up);
1079 
1080   // Skip full regions at the beginning of the space--they are necessarily part
1081   // of the dense prefix.
1082   size_t full_count = 0;
1083   const RegionData* cp;
1084   for (cp = beg_cp; cp < end_cp && cp->data_size() == region_size; ++cp) {
1085     ++full_count;
1086   }
1087 
1088   assert(total_invocations() >= _maximum_compaction_gc_num, "sanity");
1089   const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num;
1090   const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval;
1091   if (maximum_compaction || cp == end_cp || interval_ended) {
1092     _maximum_compaction_gc_num = total_invocations();
1093     return sd.region_to_addr(cp);
1094   }
1095 
1096   HeapWord* const new_top = _space_info[id].new_top();
1097   const size_t space_live = pointer_delta(new_top, space->bottom());
1098   const size_t space_used = space->used_in_words();
1099   const size_t space_capacity = space->capacity_in_words();
1100 
1101   const double cur_density = double(space_live) / space_capacity;
1102   const double deadwood_density =
1103     (1.0 - cur_density) * (1.0 - cur_density) * cur_density * cur_density;
1104   const size_t deadwood_goal = size_t(space_capacity * deadwood_density);
1105 
1106   log_develop_debug(gc, compaction)(
1107       "cur_dens=%5.3f dw_dens=%5.3f dw_goal=" SIZE_FORMAT,
1108       cur_density, deadwood_density, deadwood_goal);
1109   log_develop_debug(gc, compaction)(
1110       "space_live=" SIZE_FORMAT " space_used=" SIZE_FORMAT " "
1111       "space_cap=" SIZE_FORMAT,
1112       space_live, space_used,
1113       space_capacity);
1114 
1115   // XXX - Use binary search?
1116   HeapWord* dense_prefix = sd.region_to_addr(cp);
1117   const RegionData* full_cp = cp;
1118   const RegionData* const top_cp = sd.addr_to_region_ptr(space->top() - 1);
1119   while (cp < end_cp) {
1120     HeapWord* region_destination = cp->destination();
1121     const size_t cur_deadwood = pointer_delta(dense_prefix, region_destination);
1122 
1123     log_develop_trace(gc, compaction)(
1124         "c#=" SIZE_FORMAT_W(4) " dst=" PTR_FORMAT " "
1125         "dp=" PTR_FORMAT " cdw=" SIZE_FORMAT_W(8),
1126         sd.region(cp), p2i(region_destination),
1127         p2i(dense_prefix), cur_deadwood);
1128 
1129     if (cur_deadwood >= deadwood_goal) {
1130       // Found the region that has the correct amount of deadwood to the left.
1131       // This typically occurs after crossing a fairly sparse set of regions, so
1132       // iterate backwards over those sparse regions, looking for the region
1133       // that has the lowest density of live objects 'to the right.'
1134       size_t space_to_left = sd.region(cp) * region_size;
1135       size_t live_to_left = space_to_left - cur_deadwood;
1136       size_t space_to_right = space_capacity - space_to_left;
1137       size_t live_to_right = space_live - live_to_left;
1138       double density_to_right = double(live_to_right) / space_to_right;
1139       while (cp > full_cp) {
1140         --cp;
1141         const size_t prev_region_live_to_right = live_to_right -
1142           cp->data_size();
1143         const size_t prev_region_space_to_right = space_to_right + region_size;
1144         double prev_region_density_to_right =
1145           double(prev_region_live_to_right) / prev_region_space_to_right;
1146         if (density_to_right <= prev_region_density_to_right) {
1147           return dense_prefix;
1148         }
1149 
1150         log_develop_trace(gc, compaction)(
1151             "backing up from c=" SIZE_FORMAT_W(4) " d2r=%10.8f "
1152             "pc_d2r=%10.8f",
1153             sd.region(cp), density_to_right,
1154             prev_region_density_to_right);
1155 
1156         dense_prefix -= region_size;
1157         live_to_right = prev_region_live_to_right;
1158         space_to_right = prev_region_space_to_right;
1159         density_to_right = prev_region_density_to_right;
1160       }
1161       return dense_prefix;
1162     }
1163 
1164     dense_prefix += region_size;
1165     ++cp;
1166   }
1167 
1168   return dense_prefix;
1169 }
1170 
1171 #ifndef PRODUCT
1172 void PSParallelCompact::print_dense_prefix_stats(const char* const algorithm,
1173                                                  const SpaceId id,
1174                                                  const bool maximum_compaction,
1175                                                  HeapWord* const addr)
1176 {
1177   const size_t region_idx = summary_data().addr_to_region_idx(addr);
1178   RegionData* const cp = summary_data().region(region_idx);
1179   const MutableSpace* const space = _space_info[id].space();
1180   HeapWord* const new_top = _space_info[id].new_top();
1181 
1182   const size_t space_live = pointer_delta(new_top, space->bottom());
1183   const size_t dead_to_left = pointer_delta(addr, cp->destination());
1184   const size_t space_cap = space->capacity_in_words();
1185   const double dead_to_left_pct = double(dead_to_left) / space_cap;
1186   const size_t live_to_right = new_top - cp->destination();
1187   const size_t dead_to_right = space->top() - addr - live_to_right;
1188 
1189   log_develop_debug(gc, compaction)(
1190       "%s=" PTR_FORMAT " dpc=" SIZE_FORMAT_W(5) " "
1191       "spl=" SIZE_FORMAT " "
1192       "d2l=" SIZE_FORMAT " d2l%%=%6.4f "
1193       "d2r=" SIZE_FORMAT " l2r=" SIZE_FORMAT " "
1194       "ratio=%10.8f",
1195       algorithm, p2i(addr), region_idx,
1196       space_live,
1197       dead_to_left, dead_to_left_pct,
1198       dead_to_right, live_to_right,
1199       double(dead_to_right) / live_to_right);
1200 }
1201 #endif  // #ifndef PRODUCT
1202 
1203 // Return a fraction indicating how much of the generation can be treated as
1204 // "dead wood" (i.e., not reclaimed).  The function uses a normal distribution
1205 // based on the density of live objects in the generation to determine a limit,
1206 // which is then adjusted so the return value is min_percent when the density is
1207 // 1.
1208 //
1209 // The following table shows some return values for a different values of the
1210 // standard deviation (ParallelOldDeadWoodLimiterStdDev); the mean is 0.5 and
1211 // min_percent is 1.
1212 //
1213 //                          fraction allowed as dead wood
1214 //         -----------------------------------------------------------------
1215 // density std_dev=70 std_dev=75 std_dev=80 std_dev=85 std_dev=90 std_dev=95
1216 // ------- ---------- ---------- ---------- ---------- ---------- ----------
1217 // 0.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000
1218 // 0.05000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941
1219 // 0.10000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272
1220 // 0.15000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066
1221 // 0.20000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975
1222 // 0.25000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313
1223 // 0.30000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132
1224 // 0.35000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289
1225 // 0.40000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500
1226 // 0.45000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386
1227 // 0.50000 0.13832410 0.11599237 0.09847664 0.08456518 0.07338887 0.06431510
1228 // 0.55000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386
1229 // 0.60000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500
1230 // 0.65000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289
1231 // 0.70000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132
1232 // 0.75000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313
1233 // 0.80000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975
1234 // 0.85000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066
1235 // 0.90000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272
1236 // 0.95000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941
1237 // 1.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000
1238 
1239 double PSParallelCompact::dead_wood_limiter(double density, size_t min_percent)
1240 {
1241   assert(_dwl_initialized, "uninitialized");
1242 
1243   // The raw limit is the value of the normal distribution at x = density.
1244   const double raw_limit = normal_distribution(density);
1245 
1246   // Adjust the raw limit so it becomes the minimum when the density is 1.
1247   //
1248   // First subtract the adjustment value (which is simply the precomputed value
1249   // normal_distribution(1.0)); this yields a value of 0 when the density is 1.
1250   // Then add the minimum value, so the minimum is returned when the density is
1251   // 1.  Finally, prevent negative values, which occur when the mean is not 0.5.
1252   const double min = double(min_percent) / 100.0;
1253   const double limit = raw_limit - _dwl_adjustment + min;
1254   return MAX2(limit, 0.0);
1255 }
1256 
1257 ParallelCompactData::RegionData*
1258 PSParallelCompact::first_dead_space_region(const RegionData* beg,
1259                                            const RegionData* end)
1260 {
1261   const size_t region_size = ParallelCompactData::RegionSize;
1262   ParallelCompactData& sd = summary_data();
1263   size_t left = sd.region(beg);
1264   size_t right = end > beg ? sd.region(end) - 1 : left;
1265 
1266   // Binary search.
1267   while (left < right) {
1268     // Equivalent to (left + right) / 2, but does not overflow.
1269     const size_t middle = left + (right - left) / 2;
1270     RegionData* const middle_ptr = sd.region(middle);
1271     HeapWord* const dest = middle_ptr->destination();
1272     HeapWord* const addr = sd.region_to_addr(middle);
1273     assert(dest != NULL, "sanity");
1274     assert(dest <= addr, "must move left");
1275 
1276     if (middle > left && dest < addr) {
1277       right = middle - 1;
1278     } else if (middle < right && middle_ptr->data_size() == region_size) {
1279       left = middle + 1;
1280     } else {
1281       return middle_ptr;
1282     }
1283   }
1284   return sd.region(left);
1285 }
1286 
1287 ParallelCompactData::RegionData*
1288 PSParallelCompact::dead_wood_limit_region(const RegionData* beg,
1289                                           const RegionData* end,
1290                                           size_t dead_words)
1291 {
1292   ParallelCompactData& sd = summary_data();
1293   size_t left = sd.region(beg);
1294   size_t right = end > beg ? sd.region(end) - 1 : left;
1295 
1296   // Binary search.
1297   while (left < right) {
1298     // Equivalent to (left + right) / 2, but does not overflow.
1299     const size_t middle = left + (right - left) / 2;
1300     RegionData* const middle_ptr = sd.region(middle);
1301     HeapWord* const dest = middle_ptr->destination();
1302     HeapWord* const addr = sd.region_to_addr(middle);
1303     assert(dest != NULL, "sanity");
1304     assert(dest <= addr, "must move left");
1305 
1306     const size_t dead_to_left = pointer_delta(addr, dest);
1307     if (middle > left && dead_to_left > dead_words) {
1308       right = middle - 1;
1309     } else if (middle < right && dead_to_left < dead_words) {
1310       left = middle + 1;
1311     } else {
1312       return middle_ptr;
1313     }
1314   }
1315   return sd.region(left);
1316 }
1317 
1318 // The result is valid during the summary phase, after the initial summarization
1319 // of each space into itself, and before final summarization.
1320 inline double
1321 PSParallelCompact::reclaimed_ratio(const RegionData* const cp,
1322                                    HeapWord* const bottom,
1323                                    HeapWord* const top,
1324                                    HeapWord* const new_top)
1325 {
1326   ParallelCompactData& sd = summary_data();
1327 
1328   assert(cp != NULL, "sanity");
1329   assert(bottom != NULL, "sanity");
1330   assert(top != NULL, "sanity");
1331   assert(new_top != NULL, "sanity");
1332   assert(top >= new_top, "summary data problem?");
1333   assert(new_top > bottom, "space is empty; should not be here");
1334   assert(new_top >= cp->destination(), "sanity");
1335   assert(top >= sd.region_to_addr(cp), "sanity");
1336 
1337   HeapWord* const destination = cp->destination();
1338   const size_t dense_prefix_live  = pointer_delta(destination, bottom);
1339   const size_t compacted_region_live = pointer_delta(new_top, destination);
1340   const size_t compacted_region_used = pointer_delta(top,
1341                                                      sd.region_to_addr(cp));
1342   const size_t reclaimable = compacted_region_used - compacted_region_live;
1343 
1344   const double divisor = dense_prefix_live + 1.25 * compacted_region_live;
1345   return double(reclaimable) / divisor;
1346 }
1347 
1348 // Return the address of the end of the dense prefix, a.k.a. the start of the
1349 // compacted region.  The address is always on a region boundary.
1350 //
1351 // Completely full regions at the left are skipped, since no compaction can
1352 // occur in those regions.  Then the maximum amount of dead wood to allow is
1353 // computed, based on the density (amount live / capacity) of the generation;
1354 // the region with approximately that amount of dead space to the left is
1355 // identified as the limit region.  Regions between the last completely full
1356 // region and the limit region are scanned and the one that has the best
1357 // (maximum) reclaimed_ratio() is selected.
1358 HeapWord*
1359 PSParallelCompact::compute_dense_prefix(const SpaceId id,
1360                                         bool maximum_compaction)
1361 {
1362   const size_t region_size = ParallelCompactData::RegionSize;
1363   const ParallelCompactData& sd = summary_data();
1364 
1365   const MutableSpace* const space = _space_info[id].space();
1366   HeapWord* const top = space->top();
1367   HeapWord* const top_aligned_up = sd.region_align_up(top);
1368   HeapWord* const new_top = _space_info[id].new_top();
1369   HeapWord* const new_top_aligned_up = sd.region_align_up(new_top);
1370   HeapWord* const bottom = space->bottom();
1371   const RegionData* const beg_cp = sd.addr_to_region_ptr(bottom);
1372   const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up);
1373   const RegionData* const new_top_cp =
1374     sd.addr_to_region_ptr(new_top_aligned_up);
1375 
1376   // Skip full regions at the beginning of the space--they are necessarily part
1377   // of the dense prefix.
1378   const RegionData* const full_cp = first_dead_space_region(beg_cp, new_top_cp);
1379   assert(full_cp->destination() == sd.region_to_addr(full_cp) ||
1380          space->is_empty(), "no dead space allowed to the left");
1381   assert(full_cp->data_size() < region_size || full_cp == new_top_cp - 1,
1382          "region must have dead space");
1383 
1384   // The gc number is saved whenever a maximum compaction is done, and used to
1385   // determine when the maximum compaction interval has expired.  This avoids
1386   // successive max compactions for different reasons.
1387   assert(total_invocations() >= _maximum_compaction_gc_num, "sanity");
1388   const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num;
1389   const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval ||
1390     total_invocations() == HeapFirstMaximumCompactionCount;
1391   if (maximum_compaction || full_cp == top_cp || interval_ended) {
1392     _maximum_compaction_gc_num = total_invocations();
1393     return sd.region_to_addr(full_cp);
1394   }
1395 
1396   const size_t space_live = pointer_delta(new_top, bottom);
1397   const size_t space_used = space->used_in_words();
1398   const size_t space_capacity = space->capacity_in_words();
1399 
1400   const double density = double(space_live) / double(space_capacity);
1401   const size_t min_percent_free = MarkSweepDeadRatio;
1402   const double limiter = dead_wood_limiter(density, min_percent_free);
1403   const size_t dead_wood_max = space_used - space_live;
1404   const size_t dead_wood_limit = MIN2(size_t(space_capacity * limiter),
1405                                       dead_wood_max);
1406 
1407   log_develop_debug(gc, compaction)(
1408       "space_live=" SIZE_FORMAT " space_used=" SIZE_FORMAT " "
1409       "space_cap=" SIZE_FORMAT,
1410       space_live, space_used,
1411       space_capacity);
1412   log_develop_debug(gc, compaction)(
1413       "dead_wood_limiter(%6.4f, " SIZE_FORMAT ")=%6.4f "
1414       "dead_wood_max=" SIZE_FORMAT " dead_wood_limit=" SIZE_FORMAT,
1415       density, min_percent_free, limiter,
1416       dead_wood_max, dead_wood_limit);
1417 
1418   // Locate the region with the desired amount of dead space to the left.
1419   const RegionData* const limit_cp =
1420     dead_wood_limit_region(full_cp, top_cp, dead_wood_limit);
1421 
1422   // Scan from the first region with dead space to the limit region and find the
1423   // one with the best (largest) reclaimed ratio.
1424   double best_ratio = 0.0;
1425   const RegionData* best_cp = full_cp;
1426   for (const RegionData* cp = full_cp; cp < limit_cp; ++cp) {
1427     double tmp_ratio = reclaimed_ratio(cp, bottom, top, new_top);
1428     if (tmp_ratio > best_ratio) {
1429       best_cp = cp;
1430       best_ratio = tmp_ratio;
1431     }
1432   }
1433 
1434   return sd.region_to_addr(best_cp);
1435 }
1436 
1437 void PSParallelCompact::summarize_spaces_quick()
1438 {
1439   for (unsigned int i = 0; i < last_space_id; ++i) {
1440     const MutableSpace* space = _space_info[i].space();
1441     HeapWord** nta = _space_info[i].new_top_addr();
1442     bool result = _summary_data.summarize(_space_info[i].split_info(),
1443                                           space->bottom(), space->top(), NULL,
1444                                           space->bottom(), space->end(), nta);
1445     assert(result, "space must fit into itself");
1446     _space_info[i].set_dense_prefix(space->bottom());
1447   }
1448 }
1449 
1450 void PSParallelCompact::fill_dense_prefix_end(SpaceId id)
1451 {
1452   HeapWord* const dense_prefix_end = dense_prefix(id);
1453   const RegionData* region = _summary_data.addr_to_region_ptr(dense_prefix_end);
1454   const idx_t dense_prefix_bit = _mark_bitmap.addr_to_bit(dense_prefix_end);
1455   if (dead_space_crosses_boundary(region, dense_prefix_bit)) {
1456     // Only enough dead space is filled so that any remaining dead space to the
1457     // left is larger than the minimum filler object.  (The remainder is filled
1458     // during the copy/update phase.)
1459     //
1460     // The size of the dead space to the right of the boundary is not a
1461     // concern, since compaction will be able to use whatever space is
1462     // available.
1463     //
1464     // Here '||' is the boundary, 'x' represents a don't care bit and a box
1465     // surrounds the space to be filled with an object.
1466     //
1467     // In the 32-bit VM, each bit represents two 32-bit words:
1468     //                              +---+
1469     // a) beg_bits:  ...  x   x   x | 0 | ||   0   x  x  ...
1470     //    end_bits:  ...  x   x   x | 0 | ||   0   x  x  ...
1471     //                              +---+
1472     //
1473     // In the 64-bit VM, each bit represents one 64-bit word:
1474     //                              +------------+
1475     // b) beg_bits:  ...  x   x   x | 0   ||   0 | x  x  ...
1476     //    end_bits:  ...  x   x   1 | 0   ||   0 | x  x  ...
1477     //                              +------------+
1478     //                          +-------+
1479     // c) beg_bits:  ...  x   x | 0   0 | ||   0   x  x  ...
1480     //    end_bits:  ...  x   1 | 0   0 | ||   0   x  x  ...
1481     //                          +-------+
1482     //                      +-----------+
1483     // d) beg_bits:  ...  x | 0   0   0 | ||   0   x  x  ...
1484     //    end_bits:  ...  1 | 0   0   0 | ||   0   x  x  ...
1485     //                      +-----------+
1486     //                          +-------+
1487     // e) beg_bits:  ...  0   0 | 0   0 | ||   0   x  x  ...
1488     //    end_bits:  ...  0   0 | 0   0 | ||   0   x  x  ...
1489     //                          +-------+
1490 
1491     // Initially assume case a, c or e will apply.
1492     size_t obj_len = CollectedHeap::min_fill_size();
1493     HeapWord* obj_beg = dense_prefix_end - obj_len;
1494 
1495 #ifdef  _LP64
1496     if (MinObjAlignment > 1) { // object alignment > heap word size
1497       // Cases a, c or e.
1498     } else if (_mark_bitmap.is_obj_end(dense_prefix_bit - 2)) {
1499       // Case b above.
1500       obj_beg = dense_prefix_end - 1;
1501     } else if (!_mark_bitmap.is_obj_end(dense_prefix_bit - 3) &&
1502                _mark_bitmap.is_obj_end(dense_prefix_bit - 4)) {
1503       // Case d above.
1504       obj_beg = dense_prefix_end - 3;
1505       obj_len = 3;
1506     }
1507 #endif  // #ifdef _LP64
1508 
1509     CollectedHeap::fill_with_object(obj_beg, obj_len);
1510     _mark_bitmap.mark_obj(obj_beg, obj_len);
1511     _summary_data.add_obj(obj_beg, obj_len);
1512     assert(start_array(id) != NULL, "sanity");
1513     start_array(id)->allocate_block(obj_beg);
1514   }
1515 }
1516 
1517 void
1518 PSParallelCompact::summarize_space(SpaceId id, bool maximum_compaction)
1519 {
1520   assert(id < last_space_id, "id out of range");
1521   assert(_space_info[id].dense_prefix() == _space_info[id].space()->bottom(),
1522          "should have been reset in summarize_spaces_quick()");
1523 
1524   const MutableSpace* space = _space_info[id].space();
1525   if (_space_info[id].new_top() != space->bottom()) {
1526     HeapWord* dense_prefix_end = compute_dense_prefix(id, maximum_compaction);
1527     _space_info[id].set_dense_prefix(dense_prefix_end);
1528 
1529 #ifndef PRODUCT
1530     if (log_is_enabled(Debug, gc, compaction)) {
1531       print_dense_prefix_stats("ratio", id, maximum_compaction,
1532                                dense_prefix_end);
1533       HeapWord* addr = compute_dense_prefix_via_density(id, maximum_compaction);
1534       print_dense_prefix_stats("density", id, maximum_compaction, addr);
1535     }
1536 #endif  // #ifndef PRODUCT
1537 
1538     // Recompute the summary data, taking into account the dense prefix.  If
1539     // every last byte will be reclaimed, then the existing summary data which
1540     // compacts everything can be left in place.
1541     if (!maximum_compaction && dense_prefix_end != space->bottom()) {
1542       // If dead space crosses the dense prefix boundary, it is (at least
1543       // partially) filled with a dummy object, marked live and added to the
1544       // summary data.  This simplifies the copy/update phase and must be done
1545       // before the final locations of objects are determined, to prevent
1546       // leaving a fragment of dead space that is too small to fill.
1547       fill_dense_prefix_end(id);
1548 
1549       // Compute the destination of each Region, and thus each object.
1550       _summary_data.summarize_dense_prefix(space->bottom(), dense_prefix_end);
1551       _summary_data.summarize(_space_info[id].split_info(),
1552                               dense_prefix_end, space->top(), NULL,
1553                               dense_prefix_end, space->end(),
1554                               _space_info[id].new_top_addr());
1555     }
1556   }
1557 
1558   if (log_develop_is_enabled(Trace, gc, compaction)) {
1559     const size_t region_size = ParallelCompactData::RegionSize;
1560     HeapWord* const dense_prefix_end = _space_info[id].dense_prefix();
1561     const size_t dp_region = _summary_data.addr_to_region_idx(dense_prefix_end);
1562     const size_t dp_words = pointer_delta(dense_prefix_end, space->bottom());
1563     HeapWord* const new_top = _space_info[id].new_top();
1564     const HeapWord* nt_aligned_up = _summary_data.region_align_up(new_top);
1565     const size_t cr_words = pointer_delta(nt_aligned_up, dense_prefix_end);
1566     log_develop_trace(gc, compaction)(
1567         "id=%d cap=" SIZE_FORMAT " dp=" PTR_FORMAT " "
1568         "dp_region=" SIZE_FORMAT " " "dp_count=" SIZE_FORMAT " "
1569         "cr_count=" SIZE_FORMAT " " "nt=" PTR_FORMAT,
1570         id, space->capacity_in_words(), p2i(dense_prefix_end),
1571         dp_region, dp_words / region_size,
1572         cr_words / region_size, p2i(new_top));
1573   }
1574 }
1575 
1576 #ifndef PRODUCT
1577 void PSParallelCompact::summary_phase_msg(SpaceId dst_space_id,
1578                                           HeapWord* dst_beg, HeapWord* dst_end,
1579                                           SpaceId src_space_id,
1580                                           HeapWord* src_beg, HeapWord* src_end)
1581 {
1582   log_develop_trace(gc, compaction)(
1583       "Summarizing %d [%s] into %d [%s]:  "
1584       "src=" PTR_FORMAT "-" PTR_FORMAT " "
1585       SIZE_FORMAT "-" SIZE_FORMAT " "
1586       "dst=" PTR_FORMAT "-" PTR_FORMAT " "
1587       SIZE_FORMAT "-" SIZE_FORMAT,
1588       src_space_id, space_names[src_space_id],
1589       dst_space_id, space_names[dst_space_id],
1590       p2i(src_beg), p2i(src_end),
1591       _summary_data.addr_to_region_idx(src_beg),
1592       _summary_data.addr_to_region_idx(src_end),
1593       p2i(dst_beg), p2i(dst_end),
1594       _summary_data.addr_to_region_idx(dst_beg),
1595       _summary_data.addr_to_region_idx(dst_end));
1596 }
1597 #endif  // #ifndef PRODUCT
1598 
1599 void PSParallelCompact::summary_phase(ParCompactionManager* cm,
1600                                       bool maximum_compaction)
1601 {
1602   GCTraceTime(Info, gc, phases) tm("Summary Phase", &_gc_timer);
1603 
1604   // Quick summarization of each space into itself, to see how much is live.
1605   summarize_spaces_quick();
1606 
1607   log_develop_trace(gc, compaction)("summary phase:  after summarizing each space to self");
1608   NOT_PRODUCT(print_region_ranges());
1609   NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info));
1610 
1611   // The amount of live data that will end up in old space (assuming it fits).
1612   size_t old_space_total_live = 0;
1613   for (unsigned int id = old_space_id; id < last_space_id; ++id) {
1614     old_space_total_live += pointer_delta(_space_info[id].new_top(),
1615                                           _space_info[id].space()->bottom());
1616   }
1617 
1618   MutableSpace* const old_space = _space_info[old_space_id].space();
1619   const size_t old_capacity = old_space->capacity_in_words();
1620   if (old_space_total_live > old_capacity) {
1621     // XXX - should also try to expand
1622     maximum_compaction = true;
1623   }
1624 
1625   // Old generations.
1626   summarize_space(old_space_id, maximum_compaction);
1627 
1628   // Summarize the remaining spaces in the young gen.  The initial target space
1629   // is the old gen.  If a space does not fit entirely into the target, then the
1630   // remainder is compacted into the space itself and that space becomes the new
1631   // target.
1632   SpaceId dst_space_id = old_space_id;
1633   HeapWord* dst_space_end = old_space->end();
1634   HeapWord** new_top_addr = _space_info[dst_space_id].new_top_addr();
1635   for (unsigned int id = eden_space_id; id < last_space_id; ++id) {
1636     const MutableSpace* space = _space_info[id].space();
1637     const size_t live = pointer_delta(_space_info[id].new_top(),
1638                                       space->bottom());
1639     const size_t available = pointer_delta(dst_space_end, *new_top_addr);
1640 
1641     NOT_PRODUCT(summary_phase_msg(dst_space_id, *new_top_addr, dst_space_end,
1642                                   SpaceId(id), space->bottom(), space->top());)
1643     if (live > 0 && live <= available) {
1644       // All the live data will fit.
1645       bool done = _summary_data.summarize(_space_info[id].split_info(),
1646                                           space->bottom(), space->top(),
1647                                           NULL,
1648                                           *new_top_addr, dst_space_end,
1649                                           new_top_addr);
1650       assert(done, "space must fit into old gen");
1651 
1652       // Reset the new_top value for the space.
1653       _space_info[id].set_new_top(space->bottom());
1654     } else if (live > 0) {
1655       // Attempt to fit part of the source space into the target space.
1656       HeapWord* next_src_addr = NULL;
1657       bool done = _summary_data.summarize(_space_info[id].split_info(),
1658                                           space->bottom(), space->top(),
1659                                           &next_src_addr,
1660                                           *new_top_addr, dst_space_end,
1661                                           new_top_addr);
1662       assert(!done, "space should not fit into old gen");
1663       assert(next_src_addr != NULL, "sanity");
1664 
1665       // The source space becomes the new target, so the remainder is compacted
1666       // within the space itself.
1667       dst_space_id = SpaceId(id);
1668       dst_space_end = space->end();
1669       new_top_addr = _space_info[id].new_top_addr();
1670       NOT_PRODUCT(summary_phase_msg(dst_space_id,
1671                                     space->bottom(), dst_space_end,
1672                                     SpaceId(id), next_src_addr, space->top());)
1673       done = _summary_data.summarize(_space_info[id].split_info(),
1674                                      next_src_addr, space->top(),
1675                                      NULL,
1676                                      space->bottom(), dst_space_end,
1677                                      new_top_addr);
1678       assert(done, "space must fit when compacted into itself");
1679       assert(*new_top_addr <= space->top(), "usage should not grow");
1680     }
1681   }
1682 
1683   log_develop_trace(gc, compaction)("Summary_phase:  after final summarization");
1684   NOT_PRODUCT(print_region_ranges());
1685   NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info));
1686 }
1687 
1688 // This method should contain all heap-specific policy for invoking a full
1689 // collection.  invoke_no_policy() will only attempt to compact the heap; it
1690 // will do nothing further.  If we need to bail out for policy reasons, scavenge
1691 // before full gc, or any other specialized behavior, it needs to be added here.
1692 //
1693 // Note that this method should only be called from the vm_thread while at a
1694 // safepoint.
1695 //
1696 // Note that the all_soft_refs_clear flag in the soft ref policy
1697 // may be true because this method can be called without intervening
1698 // activity.  For example when the heap space is tight and full measure
1699 // are being taken to free space.
1700 void PSParallelCompact::invoke(bool maximum_heap_compaction) {
1701   assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
1702   assert(Thread::current() == (Thread*)VMThread::vm_thread(),
1703          "should be in vm thread");
1704 
1705   ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
1706   GCCause::Cause gc_cause = heap->gc_cause();
1707   assert(!heap->is_gc_active(), "not reentrant");
1708 
1709   PSAdaptiveSizePolicy* policy = heap->size_policy();
1710   IsGCActiveMark mark;
1711 
1712   if (ScavengeBeforeFullGC) {
1713     PSScavenge::invoke_no_policy();
1714   }
1715 
1716   const bool clear_all_soft_refs =
1717     heap->soft_ref_policy()->should_clear_all_soft_refs();
1718 
1719   PSParallelCompact::invoke_no_policy(clear_all_soft_refs ||
1720                                       maximum_heap_compaction);
1721 }
1722 
1723 // This method contains no policy. You should probably
1724 // be calling invoke() instead.
1725 bool PSParallelCompact::invoke_no_policy(bool maximum_heap_compaction) {
1726   assert(SafepointSynchronize::is_at_safepoint(), "must be at a safepoint");
1727   assert(ref_processor() != NULL, "Sanity");
1728 
1729   if (GCLocker::check_active_before_gc()) {
1730     return false;
1731   }
1732 
1733   ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
1734 
1735   GCIdMark gc_id_mark;
1736   _gc_timer.register_gc_start();
1737   _gc_tracer.report_gc_start(heap->gc_cause(), _gc_timer.gc_start());
1738 
1739   TimeStamp marking_start;
1740   TimeStamp compaction_start;
1741   TimeStamp collection_exit;
1742 
1743   GCCause::Cause gc_cause = heap->gc_cause();
1744   PSYoungGen* young_gen = heap->young_gen();
1745   PSOldGen* old_gen = heap->old_gen();
1746   PSAdaptiveSizePolicy* size_policy = heap->size_policy();
1747 
1748   // The scope of casr should end after code that can change
1749   // SoftRefPolicy::_should_clear_all_soft_refs.
1750   ClearedAllSoftRefs casr(maximum_heap_compaction,
1751                           heap->soft_ref_policy());
1752 
1753   if (ZapUnusedHeapArea) {
1754     // Save information needed to minimize mangling
1755     heap->record_gen_tops_before_GC();
1756   }
1757 
1758   // Make sure data structures are sane, make the heap parsable, and do other
1759   // miscellaneous bookkeeping.
1760   pre_compact();
1761 
1762   const PreGenGCValues pre_gc_values = heap->get_pre_gc_values();
1763 
1764   // Get the compaction manager reserved for the VM thread.
1765   ParCompactionManager* const vmthread_cm = ParCompactionManager::get_vmthread_cm();
1766 
1767   {
1768     const uint active_workers =
1769       WorkerPolicy::calc_active_workers(ParallelScavengeHeap::heap()->workers().max_workers(),
1770                                         ParallelScavengeHeap::heap()->workers().active_workers(),
1771                                         Threads::number_of_non_daemon_threads());
1772     ParallelScavengeHeap::heap()->workers().set_active_workers(active_workers);
1773 
1774     GCTraceCPUTime tcpu;
1775     GCTraceTime(Info, gc) tm("Pause Full", NULL, gc_cause, true);
1776 
1777     heap->pre_full_gc_dump(&_gc_timer);
1778 
1779     TraceCollectorStats tcs(counters());
1780     TraceMemoryManagerStats tms(heap->old_gc_manager(), gc_cause);
1781 
1782     if (log_is_enabled(Debug, gc, heap, exit)) {
1783       accumulated_time()->start();
1784     }
1785 
1786     // Let the size policy know we're starting
1787     size_policy->major_collection_begin();
1788 
1789 #if COMPILER2_OR_JVMCI
1790     DerivedPointerTable::clear();
1791 #endif
1792 
1793     ref_processor()->start_discovery(maximum_heap_compaction);
1794 
1795     marking_start.update();
1796     marking_phase(vmthread_cm, &_gc_tracer);
1797 
1798     bool max_on_system_gc = UseMaximumCompactionOnSystemGC
1799       && GCCause::is_user_requested_gc(gc_cause);
1800     summary_phase(vmthread_cm, maximum_heap_compaction || max_on_system_gc);
1801 
1802 #if COMPILER2_OR_JVMCI
1803     assert(DerivedPointerTable::is_active(), "Sanity");
1804     DerivedPointerTable::set_active(false);
1805 #endif
1806 
1807     // adjust_roots() updates Universe::_intArrayKlassObj which is
1808     // needed by the compaction for filling holes in the dense prefix.
1809     adjust_roots();
1810 
1811     compaction_start.update();
1812     compact();
1813 
1814     ParCompactionManager::verify_all_region_stack_empty();
1815 
1816     // Reset the mark bitmap, summary data, and do other bookkeeping.  Must be
1817     // done before resizing.
1818     post_compact();
1819 
1820     // Let the size policy know we're done
1821     size_policy->major_collection_end(old_gen->used_in_bytes(), gc_cause);
1822 
1823     if (UseAdaptiveSizePolicy) {
1824       log_debug(gc, ergo)("AdaptiveSizeStart: collection: %d ", heap->total_collections());
1825       log_trace(gc, ergo)("old_gen_capacity: " SIZE_FORMAT " young_gen_capacity: " SIZE_FORMAT,
1826                           old_gen->capacity_in_bytes(), young_gen->capacity_in_bytes());
1827 
1828       // Don't check if the size_policy is ready here.  Let
1829       // the size_policy check that internally.
1830       if (UseAdaptiveGenerationSizePolicyAtMajorCollection &&
1831           AdaptiveSizePolicy::should_update_promo_stats(gc_cause)) {
1832         // Swap the survivor spaces if from_space is empty. The
1833         // resize_young_gen() called below is normally used after
1834         // a successful young GC and swapping of survivor spaces;
1835         // otherwise, it will fail to resize the young gen with
1836         // the current implementation.
1837         if (young_gen->from_space()->is_empty()) {
1838           young_gen->from_space()->clear(SpaceDecorator::Mangle);
1839           young_gen->swap_spaces();
1840         }
1841 
1842         // Calculate optimal free space amounts
1843         assert(young_gen->max_gen_size() >
1844           young_gen->from_space()->capacity_in_bytes() +
1845           young_gen->to_space()->capacity_in_bytes(),
1846           "Sizes of space in young gen are out-of-bounds");
1847 
1848         size_t young_live = young_gen->used_in_bytes();
1849         size_t eden_live = young_gen->eden_space()->used_in_bytes();
1850         size_t old_live = old_gen->used_in_bytes();
1851         size_t cur_eden = young_gen->eden_space()->capacity_in_bytes();
1852         size_t max_old_gen_size = old_gen->max_gen_size();
1853         size_t max_eden_size = young_gen->max_gen_size() -
1854           young_gen->from_space()->capacity_in_bytes() -
1855           young_gen->to_space()->capacity_in_bytes();
1856 
1857         // Used for diagnostics
1858         size_policy->clear_generation_free_space_flags();
1859 
1860         size_policy->compute_generations_free_space(young_live,
1861                                                     eden_live,
1862                                                     old_live,
1863                                                     cur_eden,
1864                                                     max_old_gen_size,
1865                                                     max_eden_size,
1866                                                     true /* full gc*/);
1867 
1868         size_policy->check_gc_overhead_limit(eden_live,
1869                                              max_old_gen_size,
1870                                              max_eden_size,
1871                                              true /* full gc*/,
1872                                              gc_cause,
1873                                              heap->soft_ref_policy());
1874 
1875         size_policy->decay_supplemental_growth(true /* full gc*/);
1876 
1877         heap->resize_old_gen(
1878           size_policy->calculated_old_free_size_in_bytes());
1879 
1880         heap->resize_young_gen(size_policy->calculated_eden_size_in_bytes(),
1881                                size_policy->calculated_survivor_size_in_bytes());
1882       }
1883 
1884       log_debug(gc, ergo)("AdaptiveSizeStop: collection: %d ", heap->total_collections());
1885     }
1886 
1887     if (UsePerfData) {
1888       PSGCAdaptivePolicyCounters* const counters = heap->gc_policy_counters();
1889       counters->update_counters();
1890       counters->update_old_capacity(old_gen->capacity_in_bytes());
1891       counters->update_young_capacity(young_gen->capacity_in_bytes());
1892     }
1893 
1894     heap->resize_all_tlabs();
1895 
1896     // Resize the metaspace capacity after a collection
1897     MetaspaceGC::compute_new_size();
1898 
1899     if (log_is_enabled(Debug, gc, heap, exit)) {
1900       accumulated_time()->stop();
1901     }
1902 
1903     heap->print_heap_change(pre_gc_values);
1904 
1905     // Track memory usage and detect low memory
1906     MemoryService::track_memory_usage();
1907     heap->update_counters();
1908 
1909     heap->post_full_gc_dump(&_gc_timer);
1910   }
1911 
1912   if (VerifyAfterGC && heap->total_collections() >= VerifyGCStartAt) {
1913     Universe::verify("After GC");
1914   }
1915 
1916   // Re-verify object start arrays
1917   if (VerifyObjectStartArray &&
1918       VerifyAfterGC) {
1919     old_gen->verify_object_start_array();
1920   }
1921 
1922   if (ZapUnusedHeapArea) {
1923     old_gen->object_space()->check_mangled_unused_area_complete();
1924   }
1925 
1926   collection_exit.update();
1927 
1928   heap->print_heap_after_gc();
1929   heap->trace_heap_after_gc(&_gc_tracer);
1930 
1931   log_debug(gc, task, time)("VM-Thread " JLONG_FORMAT " " JLONG_FORMAT " " JLONG_FORMAT,
1932                          marking_start.ticks(), compaction_start.ticks(),
1933                          collection_exit.ticks());
1934 
1935   AdaptiveSizePolicyOutput::print(size_policy, heap->total_collections());
1936 
1937   _gc_timer.register_gc_end();
1938 
1939   _gc_tracer.report_dense_prefix(dense_prefix(old_space_id));
1940   _gc_tracer.report_gc_end(_gc_timer.gc_end(), _gc_timer.time_partitions());
1941 
1942   return true;
1943 }
1944 
1945 class PCAddThreadRootsMarkingTaskClosure : public ThreadClosure {
1946 private:
1947   uint _worker_id;
1948 
1949 public:
1950   PCAddThreadRootsMarkingTaskClosure(uint worker_id) : _worker_id(worker_id) { }
1951   void do_thread(Thread* thread) {
1952     assert(ParallelScavengeHeap::heap()->is_gc_active(), "called outside gc");
1953 
1954     ResourceMark rm;
1955 
1956     ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(_worker_id);
1957 
1958     PCMarkAndPushClosure mark_and_push_closure(cm);
1959     MarkingCodeBlobClosure mark_and_push_in_blobs(&mark_and_push_closure, !CodeBlobToOopClosure::FixRelocations);
1960 
1961     thread->oops_do(&mark_and_push_closure, &mark_and_push_in_blobs);
1962 
1963     // Do the real work
1964     cm->follow_marking_stacks();
1965   }
1966 };
1967 
1968 static void mark_from_roots_work(ParallelRootType::Value root_type, uint worker_id) {
1969   assert(ParallelScavengeHeap::heap()->is_gc_active(), "called outside gc");
1970 
1971   ParCompactionManager* cm =
1972     ParCompactionManager::gc_thread_compaction_manager(worker_id);
1973   PCMarkAndPushClosure mark_and_push_closure(cm);
1974 
1975   switch (root_type) {
1976     case ParallelRootType::class_loader_data:
1977       {
1978         CLDToOopClosure cld_closure(&mark_and_push_closure, ClassLoaderData::_claim_strong);
1979         ClassLoaderDataGraph::always_strong_cld_do(&cld_closure);
1980       }
1981       break;
1982 
1983     case ParallelRootType::code_cache:
1984       // Do not treat nmethods as strong roots for mark/sweep, since we can unload them.
1985       //ScavengableNMethods::scavengable_nmethods_do(CodeBlobToOopClosure(&mark_and_push_closure));
1986       break;
1987 
1988     case ParallelRootType::sentinel:
1989     DEBUG_ONLY(default:) // DEBUG_ONLY hack will create compile error on release builds (-Wswitch) and runtime check on debug builds
1990       fatal("Bad enumeration value: %u", root_type);
1991       break;
1992   }
1993 
1994   // Do the real work
1995   cm->follow_marking_stacks();
1996 }
1997 
1998 void steal_marking_work(TaskTerminator& terminator, uint worker_id) {
1999   assert(ParallelScavengeHeap::heap()->is_gc_active(), "called outside gc");
2000 
2001   ParCompactionManager* cm =
2002     ParCompactionManager::gc_thread_compaction_manager(worker_id);
2003 
2004   oop obj = NULL;
2005   ObjArrayTask task;
2006   do {
2007     while (ParCompactionManager::steal_objarray(worker_id,  task)) {
2008       cm->follow_array((objArrayOop)task.obj(), task.index());
2009       cm->follow_marking_stacks();
2010     }
2011     while (ParCompactionManager::steal(worker_id, obj)) {
2012       cm->follow_contents(obj);
2013       cm->follow_marking_stacks();
2014     }
2015   } while (!terminator.offer_termination());
2016 }
2017 
2018 class MarkFromRootsTask : public WorkerTask {
2019   StrongRootsScope _strong_roots_scope; // needed for Threads::possibly_parallel_threads_do
2020   OopStorageSetStrongParState<false /* concurrent */, false /* is_const */> _oop_storage_set_par_state;
2021   SequentialSubTasksDone _subtasks;
2022   TaskTerminator _terminator;
2023   uint _active_workers;
2024 
2025 public:
2026   MarkFromRootsTask(uint active_workers) :
2027       WorkerTask("MarkFromRootsTask"),
2028       _strong_roots_scope(active_workers),
2029       _subtasks(ParallelRootType::sentinel),
2030       _terminator(active_workers, ParCompactionManager::oop_task_queues()),
2031       _active_workers(active_workers) {
2032   }
2033 
2034   virtual void work(uint worker_id) {
2035     for (uint task = 0; _subtasks.try_claim_task(task); /*empty*/ ) {
2036       mark_from_roots_work(static_cast<ParallelRootType::Value>(task), worker_id);
2037     }
2038 
2039     PCAddThreadRootsMarkingTaskClosure closure(worker_id);
2040     Threads::possibly_parallel_threads_do(true /*parallel */, &closure);
2041 
2042     // Mark from OopStorages
2043     {
2044       ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(worker_id);
2045       PCMarkAndPushClosure closure(cm);
2046       _oop_storage_set_par_state.oops_do(&closure);
2047       // Do the real work
2048       cm->follow_marking_stacks();
2049     }
2050 
2051     if (_active_workers > 1) {
2052       steal_marking_work(_terminator, worker_id);
2053     }
2054   }
2055 };
2056 
2057 class ParallelCompactRefProcProxyTask : public RefProcProxyTask {
2058   TaskTerminator _terminator;
2059 
2060 public:
2061   ParallelCompactRefProcProxyTask(uint max_workers)
2062     : RefProcProxyTask("ParallelCompactRefProcProxyTask", max_workers),
2063       _terminator(_max_workers, ParCompactionManager::oop_task_queues()) {}
2064 
2065   void work(uint worker_id) override {
2066     assert(worker_id < _max_workers, "sanity");
2067     ParCompactionManager* cm = (_tm == RefProcThreadModel::Single) ? ParCompactionManager::get_vmthread_cm() : ParCompactionManager::gc_thread_compaction_manager(worker_id);
2068     PCMarkAndPushClosure keep_alive(cm);
2069     BarrierEnqueueDiscoveredFieldClosure enqueue;
2070     ParCompactionManager::FollowStackClosure complete_gc(cm, (_tm == RefProcThreadModel::Single) ? nullptr : &_terminator, worker_id);
2071     _rp_task->rp_work(worker_id, PSParallelCompact::is_alive_closure(), &keep_alive, &enqueue, &complete_gc);
2072   }
2073 
2074   void prepare_run_task_hook() override {
2075     _terminator.reset_for_reuse(_queue_count);
2076   }
2077 };
2078 
2079 void PSParallelCompact::marking_phase(ParCompactionManager* cm,
2080                                       ParallelOldTracer *gc_tracer) {
2081   // Recursively traverse all live objects and mark them
2082   GCTraceTime(Info, gc, phases) tm("Marking Phase", &_gc_timer);
2083 
2084   uint active_gc_threads = ParallelScavengeHeap::heap()->workers().active_workers();
2085 
2086   // Need new claim bits before marking starts.
2087   ClassLoaderDataGraph::clear_claimed_marks();
2088 
2089   {
2090     GCTraceTime(Debug, gc, phases) tm("Par Mark", &_gc_timer);
2091 
2092     MarkFromRootsTask task(active_gc_threads);
2093     ParallelScavengeHeap::heap()->workers().run_task(&task);
2094   }
2095 
2096   // Process reference objects found during marking
2097   {
2098     GCTraceTime(Debug, gc, phases) tm("Reference Processing", &_gc_timer);
2099 
2100     ReferenceProcessorStats stats;
2101     ReferenceProcessorPhaseTimes pt(&_gc_timer, ref_processor()->max_num_queues());
2102 
2103     ref_processor()->set_active_mt_degree(active_gc_threads);
2104     ParallelCompactRefProcProxyTask task(ref_processor()->max_num_queues());
2105     stats = ref_processor()->process_discovered_references(task, pt);
2106 
2107     gc_tracer->report_gc_reference_stats(stats);
2108     pt.print_all_references();
2109   }
2110 
2111   // This is the point where the entire marking should have completed.
2112   ParCompactionManager::verify_all_marking_stack_empty();
2113 
2114   {
2115     GCTraceTime(Debug, gc, phases) tm("Weak Processing", &_gc_timer);
2116     WeakProcessor::weak_oops_do(&ParallelScavengeHeap::heap()->workers(),
2117                                 is_alive_closure(),
2118                                 &do_nothing_cl,
2119                                 1);
2120   }
2121 
2122   {
2123     GCTraceTime(Debug, gc, phases) tm_m("Class Unloading", &_gc_timer);
2124 
2125     // Follow system dictionary roots and unload classes.
2126     bool purged_class = SystemDictionary::do_unloading(&_gc_timer);
2127 
2128     // Unload nmethods.
2129     CodeCache::do_unloading(is_alive_closure(), purged_class);
2130 
2131     // Prune dead klasses from subklass/sibling/implementor lists.
2132     Klass::clean_weak_klass_links(purged_class);
2133 
2134     // Clean JVMCI metadata handles.
2135     JVMCI_ONLY(JVMCI::do_unloading(purged_class));
2136   }
2137 
2138   _gc_tracer.report_object_count_after_gc(is_alive_closure());
2139 }
2140 
2141 #ifdef ASSERT
2142 void PCAdjustPointerClosure::verify_cm(ParCompactionManager* cm) {
2143   assert(cm != NULL, "associate ParCompactionManage should not be NULL");
2144   auto vmthread_cm = ParCompactionManager::get_vmthread_cm();
2145   if (Thread::current()->is_VM_thread()) {
2146     assert(cm == vmthread_cm, "VM threads should use ParCompactionManager from get_vmthread_cm()");
2147   } else {
2148     assert(Thread::current()->is_Worker_thread(), "Must be a GC thread");
2149     assert(cm != vmthread_cm, "GC threads should use ParCompactionManager from gc_thread_compaction_manager()");
2150   }
2151 }
2152 #endif
2153 
2154 class PSAdjustTask final : public WorkerTask {
2155   SubTasksDone                               _sub_tasks;
2156   WeakProcessor::Task                        _weak_proc_task;
2157   OopStorageSetStrongParState<false, false>  _oop_storage_iter;
2158   uint                                       _nworkers;
2159 
2160   enum PSAdjustSubTask {
2161     PSAdjustSubTask_code_cache,
2162 
2163     PSAdjustSubTask_num_elements
2164   };
2165 
2166 public:
2167   PSAdjustTask(uint nworkers) :
2168     WorkerTask("PSAdjust task"),
2169     _sub_tasks(PSAdjustSubTask_num_elements),
2170     _weak_proc_task(nworkers),
2171     _nworkers(nworkers) {
2172     // Need new claim bits when tracing through and adjusting pointers.
2173     ClassLoaderDataGraph::clear_claimed_marks();
2174     if (nworkers > 1) {
2175       Threads::change_thread_claim_token();
2176     }
2177   }
2178 
2179   ~PSAdjustTask() {
2180     Threads::assert_all_threads_claimed();
2181   }
2182 
2183   void work(uint worker_id) {
2184     ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(worker_id);
2185     PCAdjustPointerClosure adjust(cm);
2186     {
2187       ResourceMark rm;
2188       Threads::possibly_parallel_oops_do(_nworkers > 1, &adjust, nullptr);
2189     }
2190     _oop_storage_iter.oops_do(&adjust);
2191     {
2192       CLDToOopClosure cld_closure(&adjust, ClassLoaderData::_claim_strong);
2193       ClassLoaderDataGraph::cld_do(&cld_closure);
2194     }
2195     {
2196       AlwaysTrueClosure always_alive;
2197       _weak_proc_task.work(worker_id, &always_alive, &adjust);
2198     }
2199     if (_sub_tasks.try_claim_task(PSAdjustSubTask_code_cache)) {
2200       CodeBlobToOopClosure adjust_code(&adjust, CodeBlobToOopClosure::FixRelocations);
2201       CodeCache::blobs_do(&adjust_code);
2202     }
2203     _sub_tasks.all_tasks_claimed();
2204   }
2205 };
2206 
2207 void PSParallelCompact::adjust_roots() {
2208   // Adjust the pointers to reflect the new locations
2209   GCTraceTime(Info, gc, phases) tm("Adjust Roots", &_gc_timer);
2210   uint nworkers = ParallelScavengeHeap::heap()->workers().active_workers();
2211   PSAdjustTask task(nworkers);
2212   ParallelScavengeHeap::heap()->workers().run_task(&task);
2213 }
2214 
2215 // Helper class to print 8 region numbers per line and then print the total at the end.
2216 class FillableRegionLogger : public StackObj {
2217 private:
2218   Log(gc, compaction) log;
2219   static const int LineLength = 8;
2220   size_t _regions[LineLength];
2221   int _next_index;
2222   bool _enabled;
2223   size_t _total_regions;
2224 public:
2225   FillableRegionLogger() : _next_index(0), _enabled(log_develop_is_enabled(Trace, gc, compaction)), _total_regions(0) { }
2226   ~FillableRegionLogger() {
2227     log.trace(SIZE_FORMAT " initially fillable regions", _total_regions);
2228   }
2229 
2230   void print_line() {
2231     if (!_enabled || _next_index == 0) {
2232       return;
2233     }
2234     FormatBuffer<> line("Fillable: ");
2235     for (int i = 0; i < _next_index; i++) {
2236       line.append(" " SIZE_FORMAT_W(7), _regions[i]);
2237     }
2238     log.trace("%s", line.buffer());
2239     _next_index = 0;
2240   }
2241 
2242   void handle(size_t region) {
2243     if (!_enabled) {
2244       return;
2245     }
2246     _regions[_next_index++] = region;
2247     if (_next_index == LineLength) {
2248       print_line();
2249     }
2250     _total_regions++;
2251   }
2252 };
2253 
2254 void PSParallelCompact::prepare_region_draining_tasks(uint parallel_gc_threads)
2255 {
2256   GCTraceTime(Trace, gc, phases) tm("Drain Task Setup", &_gc_timer);
2257 
2258   // Find the threads that are active
2259   uint worker_id = 0;
2260 
2261   // Find all regions that are available (can be filled immediately) and
2262   // distribute them to the thread stacks.  The iteration is done in reverse
2263   // order (high to low) so the regions will be removed in ascending order.
2264 
2265   const ParallelCompactData& sd = PSParallelCompact::summary_data();
2266 
2267   // id + 1 is used to test termination so unsigned  can
2268   // be used with an old_space_id == 0.
2269   FillableRegionLogger region_logger;
2270   for (unsigned int id = to_space_id; id + 1 > old_space_id; --id) {
2271     SpaceInfo* const space_info = _space_info + id;
2272     HeapWord* const new_top = space_info->new_top();
2273 
2274     const size_t beg_region = sd.addr_to_region_idx(space_info->dense_prefix());
2275     const size_t end_region =
2276       sd.addr_to_region_idx(sd.region_align_up(new_top));
2277 
2278     for (size_t cur = end_region - 1; cur + 1 > beg_region; --cur) {
2279       if (sd.region(cur)->claim_unsafe()) {
2280         ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(worker_id);
2281         bool result = sd.region(cur)->mark_normal();
2282         assert(result, "Must succeed at this point.");
2283         cm->region_stack()->push(cur);
2284         region_logger.handle(cur);
2285         // Assign regions to tasks in round-robin fashion.
2286         if (++worker_id == parallel_gc_threads) {
2287           worker_id = 0;
2288         }
2289       }
2290     }
2291     region_logger.print_line();
2292   }
2293 }
2294 
2295 class TaskQueue : StackObj {
2296   volatile uint _counter;
2297   uint _size;
2298   uint _insert_index;
2299   PSParallelCompact::UpdateDensePrefixTask* _backing_array;
2300 public:
2301   explicit TaskQueue(uint size) : _counter(0), _size(size), _insert_index(0), _backing_array(NULL) {
2302     _backing_array = NEW_C_HEAP_ARRAY(PSParallelCompact::UpdateDensePrefixTask, _size, mtGC);
2303   }
2304   ~TaskQueue() {
2305     assert(_counter >= _insert_index, "not all queue elements were claimed");
2306     FREE_C_HEAP_ARRAY(T, _backing_array);
2307   }
2308 
2309   void push(const PSParallelCompact::UpdateDensePrefixTask& value) {
2310     assert(_insert_index < _size, "too small backing array");
2311     _backing_array[_insert_index++] = value;
2312   }
2313 
2314   bool try_claim(PSParallelCompact::UpdateDensePrefixTask& reference) {
2315     uint claimed = Atomic::fetch_and_add(&_counter, 1u);
2316     if (claimed < _insert_index) {
2317       reference = _backing_array[claimed];
2318       return true;
2319     } else {
2320       return false;
2321     }
2322   }
2323 };
2324 
2325 #define PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING 4
2326 
2327 void PSParallelCompact::enqueue_dense_prefix_tasks(TaskQueue& task_queue,
2328                                                    uint parallel_gc_threads) {
2329   GCTraceTime(Trace, gc, phases) tm("Dense Prefix Task Setup", &_gc_timer);
2330 
2331   ParallelCompactData& sd = PSParallelCompact::summary_data();
2332 
2333   // Iterate over all the spaces adding tasks for updating
2334   // regions in the dense prefix.  Assume that 1 gc thread
2335   // will work on opening the gaps and the remaining gc threads
2336   // will work on the dense prefix.
2337   unsigned int space_id;
2338   for (space_id = old_space_id; space_id < last_space_id; ++ space_id) {
2339     HeapWord* const dense_prefix_end = _space_info[space_id].dense_prefix();
2340     const MutableSpace* const space = _space_info[space_id].space();
2341 
2342     if (dense_prefix_end == space->bottom()) {
2343       // There is no dense prefix for this space.
2344       continue;
2345     }
2346 
2347     // The dense prefix is before this region.
2348     size_t region_index_end_dense_prefix =
2349         sd.addr_to_region_idx(dense_prefix_end);
2350     RegionData* const dense_prefix_cp =
2351       sd.region(region_index_end_dense_prefix);
2352     assert(dense_prefix_end == space->end() ||
2353            dense_prefix_cp->available() ||
2354            dense_prefix_cp->claimed(),
2355            "The region after the dense prefix should always be ready to fill");
2356 
2357     size_t region_index_start = sd.addr_to_region_idx(space->bottom());
2358 
2359     // Is there dense prefix work?
2360     size_t total_dense_prefix_regions =
2361       region_index_end_dense_prefix - region_index_start;
2362     // How many regions of the dense prefix should be given to
2363     // each thread?
2364     if (total_dense_prefix_regions > 0) {
2365       uint tasks_for_dense_prefix = 1;
2366       if (total_dense_prefix_regions <=
2367           (parallel_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING)) {
2368         // Don't over partition.  This assumes that
2369         // PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING is a small integer value
2370         // so there are not many regions to process.
2371         tasks_for_dense_prefix = parallel_gc_threads;
2372       } else {
2373         // Over partition
2374         tasks_for_dense_prefix = parallel_gc_threads *
2375           PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING;
2376       }
2377       size_t regions_per_thread = total_dense_prefix_regions /
2378         tasks_for_dense_prefix;
2379       // Give each thread at least 1 region.
2380       if (regions_per_thread == 0) {
2381         regions_per_thread = 1;
2382       }
2383 
2384       for (uint k = 0; k < tasks_for_dense_prefix; k++) {
2385         if (region_index_start >= region_index_end_dense_prefix) {
2386           break;
2387         }
2388         // region_index_end is not processed
2389         size_t region_index_end = MIN2(region_index_start + regions_per_thread,
2390                                        region_index_end_dense_prefix);
2391         task_queue.push(UpdateDensePrefixTask(SpaceId(space_id),
2392                                               region_index_start,
2393                                               region_index_end));
2394         region_index_start = region_index_end;
2395       }
2396     }
2397     // This gets any part of the dense prefix that did not
2398     // fit evenly.
2399     if (region_index_start < region_index_end_dense_prefix) {
2400       task_queue.push(UpdateDensePrefixTask(SpaceId(space_id),
2401                                             region_index_start,
2402                                             region_index_end_dense_prefix));
2403     }
2404   }
2405 }
2406 
2407 #ifdef ASSERT
2408 // Write a histogram of the number of times the block table was filled for a
2409 // region.
2410 void PSParallelCompact::write_block_fill_histogram()
2411 {
2412   if (!log_develop_is_enabled(Trace, gc, compaction)) {
2413     return;
2414   }
2415 
2416   Log(gc, compaction) log;
2417   ResourceMark rm;
2418   LogStream ls(log.trace());
2419   outputStream* out = &ls;
2420 
2421   typedef ParallelCompactData::RegionData rd_t;
2422   ParallelCompactData& sd = summary_data();
2423 
2424   for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2425     MutableSpace* const spc = _space_info[id].space();
2426     if (spc->bottom() != spc->top()) {
2427       const rd_t* const beg = sd.addr_to_region_ptr(spc->bottom());
2428       HeapWord* const top_aligned_up = sd.region_align_up(spc->top());
2429       const rd_t* const end = sd.addr_to_region_ptr(top_aligned_up);
2430 
2431       size_t histo[5] = { 0, 0, 0, 0, 0 };
2432       const size_t histo_len = sizeof(histo) / sizeof(size_t);
2433       const size_t region_cnt = pointer_delta(end, beg, sizeof(rd_t));
2434 
2435       for (const rd_t* cur = beg; cur < end; ++cur) {
2436         ++histo[MIN2(cur->blocks_filled_count(), histo_len - 1)];
2437       }
2438       out->print("Block fill histogram: %u %-4s" SIZE_FORMAT_W(5), id, space_names[id], region_cnt);
2439       for (size_t i = 0; i < histo_len; ++i) {
2440         out->print(" " SIZE_FORMAT_W(5) " %5.1f%%",
2441                    histo[i], 100.0 * histo[i] / region_cnt);
2442       }
2443       out->cr();
2444     }
2445   }
2446 }
2447 #endif // #ifdef ASSERT
2448 
2449 static void compaction_with_stealing_work(TaskTerminator* terminator, uint worker_id) {
2450   assert(ParallelScavengeHeap::heap()->is_gc_active(), "called outside gc");
2451 
2452   ParCompactionManager* cm =
2453     ParCompactionManager::gc_thread_compaction_manager(worker_id);
2454 
2455   // Drain the stacks that have been preloaded with regions
2456   // that are ready to fill.
2457 
2458   cm->drain_region_stacks();
2459 
2460   guarantee(cm->region_stack()->is_empty(), "Not empty");
2461 
2462   size_t region_index = 0;
2463 
2464   while (true) {
2465     if (ParCompactionManager::steal(worker_id, region_index)) {
2466       PSParallelCompact::fill_and_update_region(cm, region_index);
2467       cm->drain_region_stacks();
2468     } else if (PSParallelCompact::steal_unavailable_region(cm, region_index)) {
2469       // Fill and update an unavailable region with the help of a shadow region
2470       PSParallelCompact::fill_and_update_shadow_region(cm, region_index);
2471       cm->drain_region_stacks();
2472     } else {
2473       if (terminator->offer_termination()) {
2474         break;
2475       }
2476       // Go around again.
2477     }
2478   }
2479 }
2480 
2481 class UpdateDensePrefixAndCompactionTask: public WorkerTask {
2482   TaskQueue& _tq;
2483   TaskTerminator _terminator;
2484   uint _active_workers;
2485 
2486 public:
2487   UpdateDensePrefixAndCompactionTask(TaskQueue& tq, uint active_workers) :
2488       WorkerTask("UpdateDensePrefixAndCompactionTask"),
2489       _tq(tq),
2490       _terminator(active_workers, ParCompactionManager::region_task_queues()),
2491       _active_workers(active_workers) {
2492   }
2493   virtual void work(uint worker_id) {
2494     ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(worker_id);
2495 
2496     for (PSParallelCompact::UpdateDensePrefixTask task; _tq.try_claim(task); /* empty */) {
2497       PSParallelCompact::update_and_deadwood_in_dense_prefix(cm,
2498                                                              task._space_id,
2499                                                              task._region_index_start,
2500                                                              task._region_index_end);
2501     }
2502 
2503     // Once a thread has drained it's stack, it should try to steal regions from
2504     // other threads.
2505     compaction_with_stealing_work(&_terminator, worker_id);
2506   }
2507 };
2508 
2509 void PSParallelCompact::compact() {
2510   GCTraceTime(Info, gc, phases) tm("Compaction Phase", &_gc_timer);
2511 
2512   ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
2513   PSOldGen* old_gen = heap->old_gen();
2514   old_gen->start_array()->reset();
2515   uint active_gc_threads = ParallelScavengeHeap::heap()->workers().active_workers();
2516 
2517   // for [0..last_space_id)
2518   //     for [0..active_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING)
2519   //         push
2520   //     push
2521   //
2522   // max push count is thus: last_space_id * (active_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING + 1)
2523   TaskQueue task_queue(last_space_id * (active_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING + 1));
2524   initialize_shadow_regions(active_gc_threads);
2525   prepare_region_draining_tasks(active_gc_threads);
2526   enqueue_dense_prefix_tasks(task_queue, active_gc_threads);
2527 
2528   {
2529     GCTraceTime(Trace, gc, phases) tm("Par Compact", &_gc_timer);
2530 
2531     UpdateDensePrefixAndCompactionTask task(task_queue, active_gc_threads);
2532     ParallelScavengeHeap::heap()->workers().run_task(&task);
2533 
2534 #ifdef  ASSERT
2535     // Verify that all regions have been processed before the deferred updates.
2536     for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2537       verify_complete(SpaceId(id));
2538     }
2539 #endif
2540   }
2541 
2542   {
2543     GCTraceTime(Trace, gc, phases) tm("Deferred Updates", &_gc_timer);
2544     // Update the deferred objects, if any. In principle, any compaction
2545     // manager can be used. However, since the current thread is VM thread, we
2546     // use the rightful one to keep the verification logic happy.
2547     ParCompactionManager* cm = ParCompactionManager::get_vmthread_cm();
2548     for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2549       update_deferred_objects(cm, SpaceId(id));
2550     }
2551   }
2552 
2553   DEBUG_ONLY(write_block_fill_histogram());
2554 }
2555 
2556 #ifdef  ASSERT
2557 void PSParallelCompact::verify_complete(SpaceId space_id) {
2558   // All Regions between space bottom() to new_top() should be marked as filled
2559   // and all Regions between new_top() and top() should be available (i.e.,
2560   // should have been emptied).
2561   ParallelCompactData& sd = summary_data();
2562   SpaceInfo si = _space_info[space_id];
2563   HeapWord* new_top_addr = sd.region_align_up(si.new_top());
2564   HeapWord* old_top_addr = sd.region_align_up(si.space()->top());
2565   const size_t beg_region = sd.addr_to_region_idx(si.space()->bottom());
2566   const size_t new_top_region = sd.addr_to_region_idx(new_top_addr);
2567   const size_t old_top_region = sd.addr_to_region_idx(old_top_addr);
2568 
2569   bool issued_a_warning = false;
2570 
2571   size_t cur_region;
2572   for (cur_region = beg_region; cur_region < new_top_region; ++cur_region) {
2573     const RegionData* const c = sd.region(cur_region);
2574     if (!c->completed()) {
2575       log_warning(gc)("region " SIZE_FORMAT " not filled: destination_count=%u",
2576                       cur_region, c->destination_count());
2577       issued_a_warning = true;
2578     }
2579   }
2580 
2581   for (cur_region = new_top_region; cur_region < old_top_region; ++cur_region) {
2582     const RegionData* const c = sd.region(cur_region);
2583     if (!c->available()) {
2584       log_warning(gc)("region " SIZE_FORMAT " not empty: destination_count=%u",
2585                       cur_region, c->destination_count());
2586       issued_a_warning = true;
2587     }
2588   }
2589 
2590   if (issued_a_warning) {
2591     print_region_ranges();
2592   }
2593 }
2594 #endif  // #ifdef ASSERT
2595 
2596 inline void UpdateOnlyClosure::do_addr(HeapWord* addr) {
2597   _start_array->allocate_block(addr);
2598   compaction_manager()->update_contents(cast_to_oop(addr));
2599 }
2600 
2601 // Update interior oops in the ranges of regions [beg_region, end_region).
2602 void
2603 PSParallelCompact::update_and_deadwood_in_dense_prefix(ParCompactionManager* cm,
2604                                                        SpaceId space_id,
2605                                                        size_t beg_region,
2606                                                        size_t end_region) {
2607   ParallelCompactData& sd = summary_data();
2608   ParMarkBitMap* const mbm = mark_bitmap();
2609 
2610   HeapWord* beg_addr = sd.region_to_addr(beg_region);
2611   HeapWord* const end_addr = sd.region_to_addr(end_region);
2612   assert(beg_region <= end_region, "bad region range");
2613   assert(end_addr <= dense_prefix(space_id), "not in the dense prefix");
2614 
2615 #ifdef  ASSERT
2616   // Claim the regions to avoid triggering an assert when they are marked as
2617   // filled.
2618   for (size_t claim_region = beg_region; claim_region < end_region; ++claim_region) {
2619     assert(sd.region(claim_region)->claim_unsafe(), "claim() failed");
2620   }
2621 #endif  // #ifdef ASSERT
2622 
2623   if (beg_addr != space(space_id)->bottom()) {
2624     // Find the first live object or block of dead space that *starts* in this
2625     // range of regions.  If a partial object crosses onto the region, skip it;
2626     // it will be marked for 'deferred update' when the object head is
2627     // processed.  If dead space crosses onto the region, it is also skipped; it
2628     // will be filled when the prior region is processed.  If neither of those
2629     // apply, the first word in the region is the start of a live object or dead
2630     // space.
2631     assert(beg_addr > space(space_id)->bottom(), "sanity");
2632     const RegionData* const cp = sd.region(beg_region);
2633     if (cp->partial_obj_size() != 0) {
2634       beg_addr = sd.partial_obj_end(beg_region);
2635     } else if (dead_space_crosses_boundary(cp, mbm->addr_to_bit(beg_addr))) {
2636       beg_addr = mbm->find_obj_beg(beg_addr, end_addr);
2637     }
2638   }
2639 
2640   if (beg_addr < end_addr) {
2641     // A live object or block of dead space starts in this range of Regions.
2642      HeapWord* const dense_prefix_end = dense_prefix(space_id);
2643 
2644     // Create closures and iterate.
2645     UpdateOnlyClosure update_closure(mbm, cm, space_id);
2646     FillClosure fill_closure(cm, space_id);
2647     ParMarkBitMap::IterationStatus status;
2648     status = mbm->iterate(&update_closure, &fill_closure, beg_addr, end_addr,
2649                           dense_prefix_end);
2650     if (status == ParMarkBitMap::incomplete) {
2651       update_closure.do_addr(update_closure.source());
2652     }
2653   }
2654 
2655   // Mark the regions as filled.
2656   RegionData* const beg_cp = sd.region(beg_region);
2657   RegionData* const end_cp = sd.region(end_region);
2658   for (RegionData* cp = beg_cp; cp < end_cp; ++cp) {
2659     cp->set_completed();
2660   }
2661 }
2662 
2663 // Return the SpaceId for the space containing addr.  If addr is not in the
2664 // heap, last_space_id is returned.  In debug mode it expects the address to be
2665 // in the heap and asserts such.
2666 PSParallelCompact::SpaceId PSParallelCompact::space_id(HeapWord* addr) {
2667   assert(ParallelScavengeHeap::heap()->is_in_reserved(addr), "addr not in the heap");
2668 
2669   for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2670     if (_space_info[id].space()->contains(addr)) {
2671       return SpaceId(id);
2672     }
2673   }
2674 
2675   assert(false, "no space contains the addr");
2676   return last_space_id;
2677 }
2678 
2679 void PSParallelCompact::update_deferred_objects(ParCompactionManager* cm,
2680                                                 SpaceId id) {
2681   assert(id < last_space_id, "bad space id");
2682 
2683   ParallelCompactData& sd = summary_data();
2684   const SpaceInfo* const space_info = _space_info + id;
2685   ObjectStartArray* const start_array = space_info->start_array();
2686 
2687   const MutableSpace* const space = space_info->space();
2688   assert(space_info->dense_prefix() >= space->bottom(), "dense_prefix not set");
2689   HeapWord* const beg_addr = space_info->dense_prefix();
2690   HeapWord* const end_addr = sd.region_align_up(space_info->new_top());
2691 
2692   const RegionData* const beg_region = sd.addr_to_region_ptr(beg_addr);
2693   const RegionData* const end_region = sd.addr_to_region_ptr(end_addr);
2694   const RegionData* cur_region;
2695   for (cur_region = beg_region; cur_region < end_region; ++cur_region) {
2696     HeapWord* const addr = cur_region->deferred_obj_addr();
2697     if (addr != NULL) {
2698       if (start_array != NULL) {
2699         start_array->allocate_block(addr);
2700       }
2701       cm->update_contents(cast_to_oop(addr));
2702       assert(oopDesc::is_oop_or_null(cast_to_oop(addr)), "Expected an oop or NULL at " PTR_FORMAT, p2i(cast_to_oop(addr)));
2703     }
2704   }
2705 }
2706 
2707 // Skip over count live words starting from beg, and return the address of the
2708 // next live word.  Unless marked, the word corresponding to beg is assumed to
2709 // be dead.  Callers must either ensure beg does not correspond to the middle of
2710 // an object, or account for those live words in some other way.  Callers must
2711 // also ensure that there are enough live words in the range [beg, end) to skip.
2712 HeapWord*
2713 PSParallelCompact::skip_live_words(HeapWord* beg, HeapWord* end, size_t count)
2714 {
2715   assert(count > 0, "sanity");
2716 
2717   ParMarkBitMap* m = mark_bitmap();
2718   idx_t bits_to_skip = m->words_to_bits(count);
2719   idx_t cur_beg = m->addr_to_bit(beg);
2720   const idx_t search_end = m->align_range_end(m->addr_to_bit(end));
2721 
2722   do {
2723     cur_beg = m->find_obj_beg(cur_beg, search_end);
2724     idx_t cur_end = m->find_obj_end(cur_beg, search_end);
2725     const size_t obj_bits = cur_end - cur_beg + 1;
2726     if (obj_bits > bits_to_skip) {
2727       return m->bit_to_addr(cur_beg + bits_to_skip);
2728     }
2729     bits_to_skip -= obj_bits;
2730     cur_beg = cur_end + 1;
2731   } while (bits_to_skip > 0);
2732 
2733   // Skipping the desired number of words landed just past the end of an object.
2734   // Find the start of the next object.
2735   cur_beg = m->find_obj_beg(cur_beg, search_end);
2736   assert(cur_beg < m->addr_to_bit(end), "not enough live words to skip");
2737   return m->bit_to_addr(cur_beg);
2738 }
2739 
2740 HeapWord* PSParallelCompact::first_src_addr(HeapWord* const dest_addr,
2741                                             SpaceId src_space_id,
2742                                             size_t src_region_idx)
2743 {
2744   assert(summary_data().is_region_aligned(dest_addr), "not aligned");
2745 
2746   const SplitInfo& split_info = _space_info[src_space_id].split_info();
2747   if (split_info.dest_region_addr() == dest_addr) {
2748     // The partial object ending at the split point contains the first word to
2749     // be copied to dest_addr.
2750     return split_info.first_src_addr();
2751   }
2752 
2753   const ParallelCompactData& sd = summary_data();
2754   ParMarkBitMap* const bitmap = mark_bitmap();
2755   const size_t RegionSize = ParallelCompactData::RegionSize;
2756 
2757   assert(sd.is_region_aligned(dest_addr), "not aligned");
2758   const RegionData* const src_region_ptr = sd.region(src_region_idx);
2759   const size_t partial_obj_size = src_region_ptr->partial_obj_size();
2760   HeapWord* const src_region_destination = src_region_ptr->destination();
2761 
2762   assert(dest_addr >= src_region_destination, "wrong src region");
2763   assert(src_region_ptr->data_size() > 0, "src region cannot be empty");
2764 
2765   HeapWord* const src_region_beg = sd.region_to_addr(src_region_idx);
2766   HeapWord* const src_region_end = src_region_beg + RegionSize;
2767 
2768   HeapWord* addr = src_region_beg;
2769   if (dest_addr == src_region_destination) {
2770     // Return the first live word in the source region.
2771     if (partial_obj_size == 0) {
2772       addr = bitmap->find_obj_beg(addr, src_region_end);
2773       assert(addr < src_region_end, "no objects start in src region");
2774     }
2775     return addr;
2776   }
2777 
2778   // Must skip some live data.
2779   size_t words_to_skip = dest_addr - src_region_destination;
2780   assert(src_region_ptr->data_size() > words_to_skip, "wrong src region");
2781 
2782   if (partial_obj_size >= words_to_skip) {
2783     // All the live words to skip are part of the partial object.
2784     addr += words_to_skip;
2785     if (partial_obj_size == words_to_skip) {
2786       // Find the first live word past the partial object.
2787       addr = bitmap->find_obj_beg(addr, src_region_end);
2788       assert(addr < src_region_end, "wrong src region");
2789     }
2790     return addr;
2791   }
2792 
2793   // Skip over the partial object (if any).
2794   if (partial_obj_size != 0) {
2795     words_to_skip -= partial_obj_size;
2796     addr += partial_obj_size;
2797   }
2798 
2799   // Skip over live words due to objects that start in the region.
2800   addr = skip_live_words(addr, src_region_end, words_to_skip);
2801   assert(addr < src_region_end, "wrong src region");
2802   return addr;
2803 }
2804 
2805 void PSParallelCompact::decrement_destination_counts(ParCompactionManager* cm,
2806                                                      SpaceId src_space_id,
2807                                                      size_t beg_region,
2808                                                      HeapWord* end_addr)
2809 {
2810   ParallelCompactData& sd = summary_data();
2811 
2812 #ifdef ASSERT
2813   MutableSpace* const src_space = _space_info[src_space_id].space();
2814   HeapWord* const beg_addr = sd.region_to_addr(beg_region);
2815   assert(src_space->contains(beg_addr) || beg_addr == src_space->end(),
2816          "src_space_id does not match beg_addr");
2817   assert(src_space->contains(end_addr) || end_addr == src_space->end(),
2818          "src_space_id does not match end_addr");
2819 #endif // #ifdef ASSERT
2820 
2821   RegionData* const beg = sd.region(beg_region);
2822   RegionData* const end = sd.addr_to_region_ptr(sd.region_align_up(end_addr));
2823 
2824   // Regions up to new_top() are enqueued if they become available.
2825   HeapWord* const new_top = _space_info[src_space_id].new_top();
2826   RegionData* const enqueue_end =
2827     sd.addr_to_region_ptr(sd.region_align_up(new_top));
2828 
2829   for (RegionData* cur = beg; cur < end; ++cur) {
2830     assert(cur->data_size() > 0, "region must have live data");
2831     cur->decrement_destination_count();
2832     if (cur < enqueue_end && cur->available() && cur->claim()) {
2833       if (cur->mark_normal()) {
2834         cm->push_region(sd.region(cur));
2835       } else if (cur->mark_copied()) {
2836         // Try to copy the content of the shadow region back to its corresponding
2837         // heap region if the shadow region is filled. Otherwise, the GC thread
2838         // fills the shadow region will copy the data back (see
2839         // MoveAndUpdateShadowClosure::complete_region).
2840         copy_back(sd.region_to_addr(cur->shadow_region()), sd.region_to_addr(cur));
2841         ParCompactionManager::push_shadow_region_mt_safe(cur->shadow_region());
2842         cur->set_completed();
2843       }
2844     }
2845   }
2846 }
2847 
2848 size_t PSParallelCompact::next_src_region(MoveAndUpdateClosure& closure,
2849                                           SpaceId& src_space_id,
2850                                           HeapWord*& src_space_top,
2851                                           HeapWord* end_addr)
2852 {
2853   typedef ParallelCompactData::RegionData RegionData;
2854 
2855   ParallelCompactData& sd = PSParallelCompact::summary_data();
2856   const size_t region_size = ParallelCompactData::RegionSize;
2857 
2858   size_t src_region_idx = 0;
2859 
2860   // Skip empty regions (if any) up to the top of the space.
2861   HeapWord* const src_aligned_up = sd.region_align_up(end_addr);
2862   RegionData* src_region_ptr = sd.addr_to_region_ptr(src_aligned_up);
2863   HeapWord* const top_aligned_up = sd.region_align_up(src_space_top);
2864   const RegionData* const top_region_ptr =
2865     sd.addr_to_region_ptr(top_aligned_up);
2866   while (src_region_ptr < top_region_ptr && src_region_ptr->data_size() == 0) {
2867     ++src_region_ptr;
2868   }
2869 
2870   if (src_region_ptr < top_region_ptr) {
2871     // The next source region is in the current space.  Update src_region_idx
2872     // and the source address to match src_region_ptr.
2873     src_region_idx = sd.region(src_region_ptr);
2874     HeapWord* const src_region_addr = sd.region_to_addr(src_region_idx);
2875     if (src_region_addr > closure.source()) {
2876       closure.set_source(src_region_addr);
2877     }
2878     return src_region_idx;
2879   }
2880 
2881   // Switch to a new source space and find the first non-empty region.
2882   unsigned int space_id = src_space_id + 1;
2883   assert(space_id < last_space_id, "not enough spaces");
2884 
2885   HeapWord* const destination = closure.destination();
2886 
2887   do {
2888     MutableSpace* space = _space_info[space_id].space();
2889     HeapWord* const bottom = space->bottom();
2890     const RegionData* const bottom_cp = sd.addr_to_region_ptr(bottom);
2891 
2892     // Iterate over the spaces that do not compact into themselves.
2893     if (bottom_cp->destination() != bottom) {
2894       HeapWord* const top_aligned_up = sd.region_align_up(space->top());
2895       const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up);
2896 
2897       for (const RegionData* src_cp = bottom_cp; src_cp < top_cp; ++src_cp) {
2898         if (src_cp->live_obj_size() > 0) {
2899           // Found it.
2900           assert(src_cp->destination() == destination,
2901                  "first live obj in the space must match the destination");
2902           assert(src_cp->partial_obj_size() == 0,
2903                  "a space cannot begin with a partial obj");
2904 
2905           src_space_id = SpaceId(space_id);
2906           src_space_top = space->top();
2907           const size_t src_region_idx = sd.region(src_cp);
2908           closure.set_source(sd.region_to_addr(src_region_idx));
2909           return src_region_idx;
2910         } else {
2911           assert(src_cp->data_size() == 0, "sanity");
2912         }
2913       }
2914     }
2915   } while (++space_id < last_space_id);
2916 
2917   assert(false, "no source region was found");
2918   return 0;
2919 }
2920 
2921 void PSParallelCompact::fill_region(ParCompactionManager* cm, MoveAndUpdateClosure& closure, size_t region_idx)
2922 {
2923   typedef ParMarkBitMap::IterationStatus IterationStatus;
2924   ParMarkBitMap* const bitmap = mark_bitmap();
2925   ParallelCompactData& sd = summary_data();
2926   RegionData* const region_ptr = sd.region(region_idx);
2927 
2928   // Get the source region and related info.
2929   size_t src_region_idx = region_ptr->source_region();
2930   SpaceId src_space_id = space_id(sd.region_to_addr(src_region_idx));
2931   HeapWord* src_space_top = _space_info[src_space_id].space()->top();
2932   HeapWord* dest_addr = sd.region_to_addr(region_idx);
2933 
2934   closure.set_source(first_src_addr(dest_addr, src_space_id, src_region_idx));
2935 
2936   // Adjust src_region_idx to prepare for decrementing destination counts (the
2937   // destination count is not decremented when a region is copied to itself).
2938   if (src_region_idx == region_idx) {
2939     src_region_idx += 1;
2940   }
2941 
2942   if (bitmap->is_unmarked(closure.source())) {
2943     // The first source word is in the middle of an object; copy the remainder
2944     // of the object or as much as will fit.  The fact that pointer updates were
2945     // deferred will be noted when the object header is processed.
2946     HeapWord* const old_src_addr = closure.source();
2947     closure.copy_partial_obj();
2948     if (closure.is_full()) {
2949       decrement_destination_counts(cm, src_space_id, src_region_idx,
2950                                    closure.source());
2951       region_ptr->set_deferred_obj_addr(NULL);
2952       closure.complete_region(cm, dest_addr, region_ptr);
2953       return;
2954     }
2955 
2956     HeapWord* const end_addr = sd.region_align_down(closure.source());
2957     if (sd.region_align_down(old_src_addr) != end_addr) {
2958       // The partial object was copied from more than one source region.
2959       decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr);
2960 
2961       // Move to the next source region, possibly switching spaces as well.  All
2962       // args except end_addr may be modified.
2963       src_region_idx = next_src_region(closure, src_space_id, src_space_top,
2964                                        end_addr);
2965     }
2966   }
2967 
2968   do {
2969     HeapWord* const cur_addr = closure.source();
2970     HeapWord* const end_addr = MIN2(sd.region_align_up(cur_addr + 1),
2971                                     src_space_top);
2972     IterationStatus status = bitmap->iterate(&closure, cur_addr, end_addr);
2973 
2974     if (status == ParMarkBitMap::incomplete) {
2975       // The last obj that starts in the source region does not end in the
2976       // region.
2977       assert(closure.source() < end_addr, "sanity");
2978       HeapWord* const obj_beg = closure.source();
2979       HeapWord* const range_end = MIN2(obj_beg + closure.words_remaining(),
2980                                        src_space_top);
2981       HeapWord* const obj_end = bitmap->find_obj_end(obj_beg, range_end);
2982       if (obj_end < range_end) {
2983         // The end was found; the entire object will fit.
2984         status = closure.do_addr(obj_beg, bitmap->obj_size(obj_beg, obj_end));
2985         assert(status != ParMarkBitMap::would_overflow, "sanity");
2986       } else {
2987         // The end was not found; the object will not fit.
2988         assert(range_end < src_space_top, "obj cannot cross space boundary");
2989         status = ParMarkBitMap::would_overflow;
2990       }
2991     }
2992 
2993     if (status == ParMarkBitMap::would_overflow) {
2994       // The last object did not fit.  Note that interior oop updates were
2995       // deferred, then copy enough of the object to fill the region.
2996       region_ptr->set_deferred_obj_addr(closure.destination());
2997       status = closure.copy_until_full(); // copies from closure.source()
2998 
2999       decrement_destination_counts(cm, src_space_id, src_region_idx,
3000                                    closure.source());
3001       closure.complete_region(cm, dest_addr, region_ptr);
3002       return;
3003     }
3004 
3005     if (status == ParMarkBitMap::full) {
3006       decrement_destination_counts(cm, src_space_id, src_region_idx,
3007                                    closure.source());
3008       region_ptr->set_deferred_obj_addr(NULL);
3009       closure.complete_region(cm, dest_addr, region_ptr);
3010       return;
3011     }
3012 
3013     decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr);
3014 
3015     // Move to the next source region, possibly switching spaces as well.  All
3016     // args except end_addr may be modified.
3017     src_region_idx = next_src_region(closure, src_space_id, src_space_top,
3018                                      end_addr);
3019   } while (true);
3020 }
3021 
3022 void PSParallelCompact::fill_and_update_region(ParCompactionManager* cm, size_t region_idx)
3023 {
3024   MoveAndUpdateClosure cl(mark_bitmap(), cm, region_idx);
3025   fill_region(cm, cl, region_idx);
3026 }
3027 
3028 void PSParallelCompact::fill_and_update_shadow_region(ParCompactionManager* cm, size_t region_idx)
3029 {
3030   // Get a shadow region first
3031   ParallelCompactData& sd = summary_data();
3032   RegionData* const region_ptr = sd.region(region_idx);
3033   size_t shadow_region = ParCompactionManager::pop_shadow_region_mt_safe(region_ptr);
3034   // The InvalidShadow return value indicates the corresponding heap region is available,
3035   // so use MoveAndUpdateClosure to fill the normal region. Otherwise, use
3036   // MoveAndUpdateShadowClosure to fill the acquired shadow region.
3037   if (shadow_region == ParCompactionManager::InvalidShadow) {
3038     MoveAndUpdateClosure cl(mark_bitmap(), cm, region_idx);
3039     region_ptr->shadow_to_normal();
3040     return fill_region(cm, cl, region_idx);
3041   } else {
3042     MoveAndUpdateShadowClosure cl(mark_bitmap(), cm, region_idx, shadow_region);
3043     return fill_region(cm, cl, region_idx);
3044   }
3045 }
3046 
3047 void PSParallelCompact::copy_back(HeapWord *shadow_addr, HeapWord *region_addr)
3048 {
3049   Copy::aligned_conjoint_words(shadow_addr, region_addr, _summary_data.RegionSize);
3050 }
3051 
3052 bool PSParallelCompact::steal_unavailable_region(ParCompactionManager* cm, size_t &region_idx)
3053 {
3054   size_t next = cm->next_shadow_region();
3055   ParallelCompactData& sd = summary_data();
3056   size_t old_new_top = sd.addr_to_region_idx(_space_info[old_space_id].new_top());
3057   uint active_gc_threads = ParallelScavengeHeap::heap()->workers().active_workers();
3058 
3059   while (next < old_new_top) {
3060     if (sd.region(next)->mark_shadow()) {
3061       region_idx = next;
3062       return true;
3063     }
3064     next = cm->move_next_shadow_region_by(active_gc_threads);
3065   }
3066 
3067   return false;
3068 }
3069 
3070 // The shadow region is an optimization to address region dependencies in full GC. The basic
3071 // idea is making more regions available by temporally storing their live objects in empty
3072 // shadow regions to resolve dependencies between them and the destination regions. Therefore,
3073 // GC threads need not wait destination regions to be available before processing sources.
3074 //
3075 // A typical workflow would be:
3076 // After draining its own stack and failing to steal from others, a GC worker would pick an
3077 // unavailable region (destination count > 0) and get a shadow region. Then the worker fills
3078 // the shadow region by copying live objects from source regions of the unavailable one. Once
3079 // the unavailable region becomes available, the data in the shadow region will be copied back.
3080 // Shadow regions are empty regions in the to-space and regions between top and end of other spaces.
3081 void PSParallelCompact::initialize_shadow_regions(uint parallel_gc_threads)
3082 {
3083   const ParallelCompactData& sd = PSParallelCompact::summary_data();
3084 
3085   for (unsigned int id = old_space_id; id < last_space_id; ++id) {
3086     SpaceInfo* const space_info = _space_info + id;
3087     MutableSpace* const space = space_info->space();
3088 
3089     const size_t beg_region =
3090       sd.addr_to_region_idx(sd.region_align_up(MAX2(space_info->new_top(), space->top())));
3091     const size_t end_region =
3092       sd.addr_to_region_idx(sd.region_align_down(space->end()));
3093 
3094     for (size_t cur = beg_region; cur < end_region; ++cur) {
3095       ParCompactionManager::push_shadow_region(cur);
3096     }
3097   }
3098 
3099   size_t beg_region = sd.addr_to_region_idx(_space_info[old_space_id].dense_prefix());
3100   for (uint i = 0; i < parallel_gc_threads; i++) {
3101     ParCompactionManager *cm = ParCompactionManager::gc_thread_compaction_manager(i);
3102     cm->set_next_shadow_region(beg_region + i);
3103   }
3104 }
3105 
3106 void PSParallelCompact::fill_blocks(size_t region_idx)
3107 {
3108   // Fill in the block table elements for the specified region.  Each block
3109   // table element holds the number of live words in the region that are to the
3110   // left of the first object that starts in the block.  Thus only blocks in
3111   // which an object starts need to be filled.
3112   //
3113   // The algorithm scans the section of the bitmap that corresponds to the
3114   // region, keeping a running total of the live words.  When an object start is
3115   // found, if it's the first to start in the block that contains it, the
3116   // current total is written to the block table element.
3117   const size_t Log2BlockSize = ParallelCompactData::Log2BlockSize;
3118   const size_t Log2RegionSize = ParallelCompactData::Log2RegionSize;
3119   const size_t RegionSize = ParallelCompactData::RegionSize;
3120 
3121   ParallelCompactData& sd = summary_data();
3122   const size_t partial_obj_size = sd.region(region_idx)->partial_obj_size();
3123   if (partial_obj_size >= RegionSize) {
3124     return; // No objects start in this region.
3125   }
3126 
3127   // Ensure the first loop iteration decides that the block has changed.
3128   size_t cur_block = sd.block_count();
3129 
3130   const ParMarkBitMap* const bitmap = mark_bitmap();
3131 
3132   const size_t Log2BitsPerBlock = Log2BlockSize - LogMinObjAlignment;
3133   assert((size_t)1 << Log2BitsPerBlock ==
3134          bitmap->words_to_bits(ParallelCompactData::BlockSize), "sanity");
3135 
3136   size_t beg_bit = bitmap->words_to_bits(region_idx << Log2RegionSize);
3137   const size_t range_end = beg_bit + bitmap->words_to_bits(RegionSize);
3138   size_t live_bits = bitmap->words_to_bits(partial_obj_size);
3139   beg_bit = bitmap->find_obj_beg(beg_bit + live_bits, range_end);
3140   while (beg_bit < range_end) {
3141     const size_t new_block = beg_bit >> Log2BitsPerBlock;
3142     if (new_block != cur_block) {
3143       cur_block = new_block;
3144       sd.block(cur_block)->set_offset(bitmap->bits_to_words(live_bits));
3145     }
3146 
3147     const size_t end_bit = bitmap->find_obj_end(beg_bit, range_end);
3148     if (end_bit < range_end - 1) {
3149       live_bits += end_bit - beg_bit + 1;
3150       beg_bit = bitmap->find_obj_beg(end_bit + 1, range_end);
3151     } else {
3152       return;
3153     }
3154   }
3155 }
3156 
3157 ParMarkBitMap::IterationStatus MoveAndUpdateClosure::copy_until_full()
3158 {
3159   if (source() != copy_destination()) {
3160     DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3161     Copy::aligned_conjoint_words(source(), copy_destination(), words_remaining());
3162   }
3163   update_state(words_remaining());
3164   assert(is_full(), "sanity");
3165   return ParMarkBitMap::full;
3166 }
3167 
3168 void MoveAndUpdateClosure::copy_partial_obj()
3169 {
3170   size_t words = words_remaining();
3171 
3172   HeapWord* const range_end = MIN2(source() + words, bitmap()->region_end());
3173   HeapWord* const end_addr = bitmap()->find_obj_end(source(), range_end);
3174   if (end_addr < range_end) {
3175     words = bitmap()->obj_size(source(), end_addr);
3176   }
3177 
3178   // This test is necessary; if omitted, the pointer updates to a partial object
3179   // that crosses the dense prefix boundary could be overwritten.
3180   if (source() != copy_destination()) {
3181     DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3182     Copy::aligned_conjoint_words(source(), copy_destination(), words);
3183   }
3184   update_state(words);
3185 }
3186 
3187 void MoveAndUpdateClosure::complete_region(ParCompactionManager *cm, HeapWord *dest_addr,
3188                                            PSParallelCompact::RegionData *region_ptr) {
3189   assert(region_ptr->shadow_state() == ParallelCompactData::RegionData::NormalRegion, "Region should be finished");
3190   region_ptr->set_completed();
3191 }
3192 
3193 ParMarkBitMapClosure::IterationStatus
3194 MoveAndUpdateClosure::do_addr(HeapWord* addr, size_t words) {
3195   assert(destination() != NULL, "sanity");
3196   assert(bitmap()->obj_size(addr) == words, "bad size");
3197 
3198   _source = addr;
3199   assert(PSParallelCompact::summary_data().calc_new_pointer(source(), compaction_manager()) ==
3200          destination(), "wrong destination");
3201 
3202   if (words > words_remaining()) {
3203     return ParMarkBitMap::would_overflow;
3204   }
3205 
3206   // The start_array must be updated even if the object is not moving.
3207   if (_start_array != NULL) {
3208     _start_array->allocate_block(destination());
3209   }
3210 
3211   if (copy_destination() != source()) {
3212     DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3213     Copy::aligned_conjoint_words(source(), copy_destination(), words);
3214   }
3215 
3216   oop moved_oop = cast_to_oop(copy_destination());
3217   compaction_manager()->update_contents(moved_oop);
3218   assert(oopDesc::is_oop_or_null(moved_oop), "Expected an oop or NULL at " PTR_FORMAT, p2i(moved_oop));
3219 
3220   update_state(words);
3221   assert(copy_destination() == cast_from_oop<HeapWord*>(moved_oop) + moved_oop->size(), "sanity");
3222   return is_full() ? ParMarkBitMap::full : ParMarkBitMap::incomplete;
3223 }
3224 
3225 void MoveAndUpdateShadowClosure::complete_region(ParCompactionManager *cm, HeapWord *dest_addr,
3226                                                  PSParallelCompact::RegionData *region_ptr) {
3227   assert(region_ptr->shadow_state() == ParallelCompactData::RegionData::ShadowRegion, "Region should be shadow");
3228   // Record the shadow region index
3229   region_ptr->set_shadow_region(_shadow);
3230   // Mark the shadow region as filled to indicate the data is ready to be
3231   // copied back
3232   region_ptr->mark_filled();
3233   // Try to copy the content of the shadow region back to its corresponding
3234   // heap region if available; the GC thread that decreases the destination
3235   // count to zero will do the copying otherwise (see
3236   // PSParallelCompact::decrement_destination_counts).
3237   if (((region_ptr->available() && region_ptr->claim()) || region_ptr->claimed()) && region_ptr->mark_copied()) {
3238     region_ptr->set_completed();
3239     PSParallelCompact::copy_back(PSParallelCompact::summary_data().region_to_addr(_shadow), dest_addr);
3240     ParCompactionManager::push_shadow_region_mt_safe(_shadow);
3241   }
3242 }
3243 
3244 UpdateOnlyClosure::UpdateOnlyClosure(ParMarkBitMap* mbm,
3245                                      ParCompactionManager* cm,
3246                                      PSParallelCompact::SpaceId space_id) :
3247   ParMarkBitMapClosure(mbm, cm),
3248   _space_id(space_id),
3249   _start_array(PSParallelCompact::start_array(space_id))
3250 {
3251 }
3252 
3253 // Updates the references in the object to their new values.
3254 ParMarkBitMapClosure::IterationStatus
3255 UpdateOnlyClosure::do_addr(HeapWord* addr, size_t words) {
3256   do_addr(addr);
3257   return ParMarkBitMap::incomplete;
3258 }
3259 
3260 FillClosure::FillClosure(ParCompactionManager* cm, PSParallelCompact::SpaceId space_id) :
3261   ParMarkBitMapClosure(PSParallelCompact::mark_bitmap(), cm),
3262   _start_array(PSParallelCompact::start_array(space_id))
3263 {
3264   assert(space_id == PSParallelCompact::old_space_id,
3265          "cannot use FillClosure in the young gen");
3266 }
3267 
3268 ParMarkBitMapClosure::IterationStatus
3269 FillClosure::do_addr(HeapWord* addr, size_t size) {
3270   CollectedHeap::fill_with_objects(addr, size);
3271   HeapWord* const end = addr + size;
3272   do {
3273     _start_array->allocate_block(addr);
3274     addr += cast_to_oop(addr)->size();
3275   } while (addr < end);
3276   return ParMarkBitMap::incomplete;
3277 }