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/workgroup.hpp"
  64 #include "logging/log.hpp"
  65 #include "memory/iterator.inline.hpp"
  66 #include "memory/metaspaceUtils.hpp"
  67 #include "memory/resourceArea.hpp"
  68 #include "memory/universe.hpp"
  69 #include "oops/access.inline.hpp"
  70 #include "oops/instanceClassLoaderKlass.inline.hpp"
  71 #include "oops/instanceKlass.inline.hpp"
  72 #include "oops/instanceMirrorKlass.inline.hpp"
  73 #include "oops/methodData.hpp"
  74 #include "oops/objArrayKlass.inline.hpp"
  75 #include "oops/oop.inline.hpp"
  76 #include "runtime/atomic.hpp"
  77 #include "runtime/handles.inline.hpp"
  78 #include "runtime/java.hpp"
  79 #include "runtime/safepoint.hpp"
  80 #include "runtime/vmThread.hpp"
  81 #include "services/memTracker.hpp"
  82 #include "services/memoryService.hpp"
  83 #include "utilities/align.hpp"
  84 #include "utilities/debug.hpp"
  85 #include "utilities/events.hpp"
  86 #include "utilities/formatBuffer.hpp"
  87 #include "utilities/macros.hpp"
  88 #include "utilities/stack.inline.hpp"
  89 #if INCLUDE_JVMCI
  90 #include "jvmci/jvmci.hpp"
  91 #endif
  92 
  93 #include <math.h>
  94 
  95 // All sizes are in HeapWords.
  96 const size_t ParallelCompactData::Log2RegionSize  = 16; // 64K words
  97 const size_t ParallelCompactData::RegionSize      = (size_t)1 << Log2RegionSize;
  98 const size_t ParallelCompactData::RegionSizeBytes =
  99   RegionSize << LogHeapWordSize;
 100 const size_t ParallelCompactData::RegionSizeOffsetMask = RegionSize - 1;
 101 const size_t ParallelCompactData::RegionAddrOffsetMask = RegionSizeBytes - 1;
 102 const size_t ParallelCompactData::RegionAddrMask       = ~RegionAddrOffsetMask;
 103 
 104 const size_t ParallelCompactData::Log2BlockSize   = 7; // 128 words
 105 const size_t ParallelCompactData::BlockSize       = (size_t)1 << Log2BlockSize;
 106 const size_t ParallelCompactData::BlockSizeBytes  =
 107   BlockSize << LogHeapWordSize;
 108 const size_t ParallelCompactData::BlockSizeOffsetMask = BlockSize - 1;
 109 const size_t ParallelCompactData::BlockAddrOffsetMask = BlockSizeBytes - 1;
 110 const size_t ParallelCompactData::BlockAddrMask       = ~BlockAddrOffsetMask;
 111 
 112 const size_t ParallelCompactData::BlocksPerRegion = RegionSize / BlockSize;
 113 const size_t ParallelCompactData::Log2BlocksPerRegion =
 114   Log2RegionSize - Log2BlockSize;
 115 
 116 const ParallelCompactData::RegionData::region_sz_t
 117 ParallelCompactData::RegionData::dc_shift = 27;
 118 
 119 const ParallelCompactData::RegionData::region_sz_t
 120 ParallelCompactData::RegionData::dc_mask = ~0U << dc_shift;
 121 
 122 const ParallelCompactData::RegionData::region_sz_t
 123 ParallelCompactData::RegionData::dc_one = 0x1U << dc_shift;
 124 
 125 const ParallelCompactData::RegionData::region_sz_t
 126 ParallelCompactData::RegionData::los_mask = ~dc_mask;
 127 
 128 const ParallelCompactData::RegionData::region_sz_t
 129 ParallelCompactData::RegionData::dc_claimed = 0x8U << dc_shift;
 130 
 131 const ParallelCompactData::RegionData::region_sz_t
 132 ParallelCompactData::RegionData::dc_completed = 0xcU << dc_shift;
 133 
 134 SpaceInfo PSParallelCompact::_space_info[PSParallelCompact::last_space_id];
 135 
 136 SpanSubjectToDiscoveryClosure PSParallelCompact::_span_based_discoverer;
 137 ReferenceProcessor* PSParallelCompact::_ref_processor = NULL;
 138 
 139 double PSParallelCompact::_dwl_mean;
 140 double PSParallelCompact::_dwl_std_dev;
 141 double PSParallelCompact::_dwl_first_term;
 142 double PSParallelCompact::_dwl_adjustment;
 143 #ifdef  ASSERT
 144 bool   PSParallelCompact::_dwl_initialized = false;
 145 #endif  // #ifdef ASSERT
 146 
 147 void SplitInfo::record(size_t src_region_idx, size_t partial_obj_size,
 148                        HeapWord* destination)
 149 {
 150   assert(src_region_idx != 0, "invalid src_region_idx");
 151   assert(partial_obj_size != 0, "invalid partial_obj_size argument");
 152   assert(destination != NULL, "invalid destination argument");
 153 
 154   _src_region_idx = src_region_idx;
 155   _partial_obj_size = partial_obj_size;
 156   _destination = destination;
 157 
 158   // These fields may not be updated below, so make sure they're clear.
 159   assert(_dest_region_addr == NULL, "should have been cleared");
 160   assert(_first_src_addr == NULL, "should have been cleared");
 161 
 162   // Determine the number of destination regions for the partial object.
 163   HeapWord* const last_word = destination + partial_obj_size - 1;
 164   const ParallelCompactData& sd = PSParallelCompact::summary_data();
 165   HeapWord* const beg_region_addr = sd.region_align_down(destination);
 166   HeapWord* const end_region_addr = sd.region_align_down(last_word);
 167 
 168   if (beg_region_addr == end_region_addr) {
 169     // One destination region.
 170     _destination_count = 1;
 171     if (end_region_addr == destination) {
 172       // The destination falls on a region boundary, thus the first word of the
 173       // partial object will be the first word copied to the destination region.
 174       _dest_region_addr = end_region_addr;
 175       _first_src_addr = sd.region_to_addr(src_region_idx);
 176     }
 177   } else {
 178     // Two destination regions.  When copied, the partial object will cross a
 179     // destination region boundary, so a word somewhere within the partial
 180     // object will be the first word copied to the second destination region.
 181     _destination_count = 2;
 182     _dest_region_addr = end_region_addr;
 183     const size_t ofs = pointer_delta(end_region_addr, destination);
 184     assert(ofs < _partial_obj_size, "sanity");
 185     _first_src_addr = sd.region_to_addr(src_region_idx) + ofs;
 186   }
 187 }
 188 
 189 void SplitInfo::clear()
 190 {
 191   _src_region_idx = 0;
 192   _partial_obj_size = 0;
 193   _destination = NULL;
 194   _destination_count = 0;
 195   _dest_region_addr = NULL;
 196   _first_src_addr = NULL;
 197   assert(!is_valid(), "sanity");
 198 }
 199 
 200 #ifdef  ASSERT
 201 void SplitInfo::verify_clear()
 202 {
 203   assert(_src_region_idx == 0, "not clear");
 204   assert(_partial_obj_size == 0, "not clear");
 205   assert(_destination == NULL, "not clear");
 206   assert(_destination_count == 0, "not clear");
 207   assert(_dest_region_addr == NULL, "not clear");
 208   assert(_first_src_addr == NULL, "not clear");
 209 }
 210 #endif  // #ifdef ASSERT
 211 
 212 
 213 void PSParallelCompact::print_on_error(outputStream* st) {
 214   _mark_bitmap.print_on_error(st);
 215 }
 216 
 217 #ifndef PRODUCT
 218 const char* PSParallelCompact::space_names[] = {
 219   "old ", "eden", "from", "to  "
 220 };
 221 
 222 void PSParallelCompact::print_region_ranges() {
 223   if (!log_develop_is_enabled(Trace, gc, compaction)) {
 224     return;
 225   }
 226   Log(gc, compaction) log;
 227   ResourceMark rm;
 228   LogStream ls(log.trace());
 229   Universe::print_on(&ls);
 230   log.trace("space  bottom     top        end        new_top");
 231   log.trace("------ ---------- ---------- ---------- ----------");
 232 
 233   for (unsigned int id = 0; id < last_space_id; ++id) {
 234     const MutableSpace* space = _space_info[id].space();
 235     log.trace("%u %s "
 236               SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " "
 237               SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " ",
 238               id, space_names[id],
 239               summary_data().addr_to_region_idx(space->bottom()),
 240               summary_data().addr_to_region_idx(space->top()),
 241               summary_data().addr_to_region_idx(space->end()),
 242               summary_data().addr_to_region_idx(_space_info[id].new_top()));
 243   }
 244 }
 245 
 246 void
 247 print_generic_summary_region(size_t i, const ParallelCompactData::RegionData* c)
 248 {
 249 #define REGION_IDX_FORMAT        SIZE_FORMAT_W(7)
 250 #define REGION_DATA_FORMAT       SIZE_FORMAT_W(5)
 251 
 252   ParallelCompactData& sd = PSParallelCompact::summary_data();
 253   size_t dci = c->destination() ? sd.addr_to_region_idx(c->destination()) : 0;
 254   log_develop_trace(gc, compaction)(
 255       REGION_IDX_FORMAT " " PTR_FORMAT " "
 256       REGION_IDX_FORMAT " " PTR_FORMAT " "
 257       REGION_DATA_FORMAT " " REGION_DATA_FORMAT " "
 258       REGION_DATA_FORMAT " " REGION_IDX_FORMAT " %d",
 259       i, p2i(c->data_location()), dci, p2i(c->destination()),
 260       c->partial_obj_size(), c->live_obj_size(),
 261       c->data_size(), c->source_region(), c->destination_count());
 262 
 263 #undef  REGION_IDX_FORMAT
 264 #undef  REGION_DATA_FORMAT
 265 }
 266 
 267 void
 268 print_generic_summary_data(ParallelCompactData& summary_data,
 269                            HeapWord* const beg_addr,
 270                            HeapWord* const end_addr)
 271 {
 272   size_t total_words = 0;
 273   size_t i = summary_data.addr_to_region_idx(beg_addr);
 274   const size_t last = summary_data.addr_to_region_idx(end_addr);
 275   HeapWord* pdest = 0;
 276 
 277   while (i < last) {
 278     ParallelCompactData::RegionData* c = summary_data.region(i);
 279     if (c->data_size() != 0 || c->destination() != pdest) {
 280       print_generic_summary_region(i, c);
 281       total_words += c->data_size();
 282       pdest = c->destination();
 283     }
 284     ++i;
 285   }
 286 
 287   log_develop_trace(gc, compaction)("summary_data_bytes=" SIZE_FORMAT, total_words * HeapWordSize);
 288 }
 289 
 290 void
 291 PSParallelCompact::print_generic_summary_data(ParallelCompactData& summary_data,
 292                                               HeapWord* const beg_addr,
 293                                               HeapWord* const end_addr) {
 294   ::print_generic_summary_data(summary_data,beg_addr, end_addr);
 295 }
 296 
 297 void
 298 print_generic_summary_data(ParallelCompactData& summary_data,
 299                            SpaceInfo* space_info)
 300 {
 301   if (!log_develop_is_enabled(Trace, gc, compaction)) {
 302     return;
 303   }
 304 
 305   for (unsigned int id = 0; id < PSParallelCompact::last_space_id; ++id) {
 306     const MutableSpace* space = space_info[id].space();
 307     print_generic_summary_data(summary_data, space->bottom(),
 308                                MAX2(space->top(), space_info[id].new_top()));
 309   }
 310 }
 311 
 312 void
 313 print_initial_summary_data(ParallelCompactData& summary_data,
 314                            const MutableSpace* space) {
 315   if (space->top() == space->bottom()) {
 316     return;
 317   }
 318 
 319   const size_t region_size = ParallelCompactData::RegionSize;
 320   typedef ParallelCompactData::RegionData RegionData;
 321   HeapWord* const top_aligned_up = summary_data.region_align_up(space->top());
 322   const size_t end_region = summary_data.addr_to_region_idx(top_aligned_up);
 323   const RegionData* c = summary_data.region(end_region - 1);
 324   HeapWord* end_addr = c->destination() + c->data_size();
 325   const size_t live_in_space = pointer_delta(end_addr, space->bottom());
 326 
 327   // Print (and count) the full regions at the beginning of the space.
 328   size_t full_region_count = 0;
 329   size_t i = summary_data.addr_to_region_idx(space->bottom());
 330   while (i < end_region && summary_data.region(i)->data_size() == region_size) {
 331     ParallelCompactData::RegionData* c = summary_data.region(i);
 332     log_develop_trace(gc, compaction)(
 333         SIZE_FORMAT_W(5) " " PTR_FORMAT " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d",
 334         i, p2i(c->destination()),
 335         c->partial_obj_size(), c->live_obj_size(),
 336         c->data_size(), c->source_region(), c->destination_count());
 337     ++full_region_count;
 338     ++i;
 339   }
 340 
 341   size_t live_to_right = live_in_space - full_region_count * region_size;
 342 
 343   double max_reclaimed_ratio = 0.0;
 344   size_t max_reclaimed_ratio_region = 0;
 345   size_t max_dead_to_right = 0;
 346   size_t max_live_to_right = 0;
 347 
 348   // Print the 'reclaimed ratio' for regions while there is something live in
 349   // the region or to the right of it.  The remaining regions are empty (and
 350   // uninteresting), and computing the ratio will result in division by 0.
 351   while (i < end_region && live_to_right > 0) {
 352     c = summary_data.region(i);
 353     HeapWord* const region_addr = summary_data.region_to_addr(i);
 354     const size_t used_to_right = pointer_delta(space->top(), region_addr);
 355     const size_t dead_to_right = used_to_right - live_to_right;
 356     const double reclaimed_ratio = double(dead_to_right) / live_to_right;
 357 
 358     if (reclaimed_ratio > max_reclaimed_ratio) {
 359             max_reclaimed_ratio = reclaimed_ratio;
 360             max_reclaimed_ratio_region = i;
 361             max_dead_to_right = dead_to_right;
 362             max_live_to_right = live_to_right;
 363     }
 364 
 365     ParallelCompactData::RegionData* c = summary_data.region(i);
 366     log_develop_trace(gc, compaction)(
 367         SIZE_FORMAT_W(5) " " PTR_FORMAT " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d"
 368         "%12.10f " SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10),
 369         i, p2i(c->destination()),
 370         c->partial_obj_size(), c->live_obj_size(),
 371         c->data_size(), c->source_region(), c->destination_count(),
 372         reclaimed_ratio, dead_to_right, live_to_right);
 373 
 374 
 375     live_to_right -= c->data_size();
 376     ++i;
 377   }
 378 
 379   // Any remaining regions are empty.  Print one more if there is one.
 380   if (i < end_region) {
 381     ParallelCompactData::RegionData* c = summary_data.region(i);
 382     log_develop_trace(gc, compaction)(
 383         SIZE_FORMAT_W(5) " " PTR_FORMAT " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d",
 384          i, p2i(c->destination()),
 385          c->partial_obj_size(), c->live_obj_size(),
 386          c->data_size(), c->source_region(), c->destination_count());
 387   }
 388 
 389   log_develop_trace(gc, compaction)("max:  " SIZE_FORMAT_W(4) " d2r=" SIZE_FORMAT_W(10) " l2r=" SIZE_FORMAT_W(10) " max_ratio=%14.12f",
 390                                     max_reclaimed_ratio_region, max_dead_to_right, max_live_to_right, max_reclaimed_ratio);
 391 }
 392 
 393 void
 394 print_initial_summary_data(ParallelCompactData& summary_data,
 395                            SpaceInfo* space_info) {
 396   if (!log_develop_is_enabled(Trace, gc, compaction)) {
 397     return;
 398   }
 399 
 400   unsigned int id = PSParallelCompact::old_space_id;
 401   const MutableSpace* space;
 402   do {
 403     space = space_info[id].space();
 404     print_initial_summary_data(summary_data, space);
 405   } while (++id < PSParallelCompact::eden_space_id);
 406 
 407   do {
 408     space = space_info[id].space();
 409     print_generic_summary_data(summary_data, space->bottom(), space->top());
 410   } while (++id < PSParallelCompact::last_space_id);
 411 }
 412 #endif  // #ifndef PRODUCT
 413 
 414 ParallelCompactData::ParallelCompactData() :
 415   _region_start(NULL),
 416   DEBUG_ONLY(_region_end(NULL) COMMA)
 417   _region_vspace(NULL),
 418   _reserved_byte_size(0),
 419   _region_data(NULL),
 420   _region_count(0),
 421   _block_vspace(NULL),
 422   _block_data(NULL),
 423   _block_count(0) {}
 424 
 425 bool ParallelCompactData::initialize(MemRegion covered_region)
 426 {
 427   _region_start = covered_region.start();
 428   const size_t region_size = covered_region.word_size();
 429   DEBUG_ONLY(_region_end = _region_start + region_size;)
 430 
 431   assert(region_align_down(_region_start) == _region_start,
 432          "region start not aligned");
 433   assert((region_size & RegionSizeOffsetMask) == 0,
 434          "region size not a multiple of RegionSize");
 435 
 436   bool result = initialize_region_data(region_size) && initialize_block_data();
 437   return result;
 438 }
 439 
 440 PSVirtualSpace*
 441 ParallelCompactData::create_vspace(size_t count, size_t element_size)
 442 {
 443   const size_t raw_bytes = count * element_size;
 444   const size_t page_sz = os::page_size_for_region_aligned(raw_bytes, 10);
 445   const size_t granularity = os::vm_allocation_granularity();
 446   _reserved_byte_size = align_up(raw_bytes, MAX2(page_sz, granularity));
 447 
 448   const size_t rs_align = page_sz == (size_t) os::vm_page_size() ? 0 :
 449     MAX2(page_sz, granularity);
 450   ReservedSpace rs(_reserved_byte_size, rs_align, page_sz);
 451   os::trace_page_sizes("Parallel Compact Data", raw_bytes, raw_bytes, page_sz, rs.base(),
 452                        rs.size());
 453 
 454   MemTracker::record_virtual_memory_type((address)rs.base(), mtGC);
 455 
 456   PSVirtualSpace* vspace = new PSVirtualSpace(rs, page_sz);
 457   if (vspace != 0) {
 458     if (vspace->expand_by(_reserved_byte_size)) {
 459       return vspace;
 460     }
 461     delete vspace;
 462     // Release memory reserved in the space.
 463     rs.release();
 464   }
 465 
 466   return 0;
 467 }
 468 
 469 bool ParallelCompactData::initialize_region_data(size_t region_size)
 470 {
 471   const size_t count = (region_size + RegionSizeOffsetMask) >> Log2RegionSize;
 472   _region_vspace = create_vspace(count, sizeof(RegionData));
 473   if (_region_vspace != 0) {
 474     _region_data = (RegionData*)_region_vspace->reserved_low_addr();
 475     _region_count = count;
 476     return true;
 477   }
 478   return false;
 479 }
 480 
 481 bool ParallelCompactData::initialize_block_data()
 482 {
 483   assert(_region_count != 0, "region data must be initialized first");
 484   const size_t count = _region_count << Log2BlocksPerRegion;
 485   _block_vspace = create_vspace(count, sizeof(BlockData));
 486   if (_block_vspace != 0) {
 487     _block_data = (BlockData*)_block_vspace->reserved_low_addr();
 488     _block_count = count;
 489     return true;
 490   }
 491   return false;
 492 }
 493 
 494 void ParallelCompactData::clear()
 495 {
 496   memset(_region_data, 0, _region_vspace->committed_size());
 497   memset(_block_data, 0, _block_vspace->committed_size());
 498 }
 499 
 500 void ParallelCompactData::clear_range(size_t beg_region, size_t end_region) {
 501   assert(beg_region <= _region_count, "beg_region out of range");
 502   assert(end_region <= _region_count, "end_region out of range");
 503   assert(RegionSize % BlockSize == 0, "RegionSize not a multiple of BlockSize");
 504 
 505   const size_t region_cnt = end_region - beg_region;
 506   memset(_region_data + beg_region, 0, region_cnt * sizeof(RegionData));
 507 
 508   const size_t beg_block = beg_region * BlocksPerRegion;
 509   const size_t block_cnt = region_cnt * BlocksPerRegion;
 510   memset(_block_data + beg_block, 0, block_cnt * sizeof(BlockData));
 511 }
 512 
 513 HeapWord* ParallelCompactData::partial_obj_end(size_t region_idx) const
 514 {
 515   const RegionData* cur_cp = region(region_idx);
 516   const RegionData* const end_cp = region(region_count() - 1);
 517 
 518   HeapWord* result = region_to_addr(region_idx);
 519   if (cur_cp < end_cp) {
 520     do {
 521       result += cur_cp->partial_obj_size();
 522     } while (cur_cp->partial_obj_size() == RegionSize && ++cur_cp < end_cp);
 523   }
 524   return result;
 525 }
 526 
 527 void ParallelCompactData::add_obj(HeapWord* addr, size_t len)
 528 {
 529   const size_t obj_ofs = pointer_delta(addr, _region_start);
 530   const size_t beg_region = obj_ofs >> Log2RegionSize;
 531   // end_region is inclusive
 532   const size_t end_region = (obj_ofs + len - 1) >> Log2RegionSize;
 533 
 534   if (beg_region == end_region) {
 535     // All in one region.
 536     _region_data[beg_region].add_live_obj(len);
 537     return;
 538   }
 539 
 540   // First region.
 541   const size_t beg_ofs = region_offset(addr);
 542   _region_data[beg_region].add_live_obj(RegionSize - beg_ofs);
 543 
 544   // Middle regions--completely spanned by this object.
 545   for (size_t region = beg_region + 1; region < end_region; ++region) {
 546     _region_data[region].set_partial_obj_size(RegionSize);
 547     _region_data[region].set_partial_obj_addr(addr);
 548   }
 549 
 550   // Last region.
 551   const size_t end_ofs = region_offset(addr + len - 1);
 552   _region_data[end_region].set_partial_obj_size(end_ofs + 1);
 553   _region_data[end_region].set_partial_obj_addr(addr);
 554 }
 555 
 556 void
 557 ParallelCompactData::summarize_dense_prefix(HeapWord* beg, HeapWord* end)
 558 {
 559   assert(is_region_aligned(beg), "not RegionSize aligned");
 560   assert(is_region_aligned(end), "not RegionSize aligned");
 561 
 562   size_t cur_region = addr_to_region_idx(beg);
 563   const size_t end_region = addr_to_region_idx(end);
 564   HeapWord* addr = beg;
 565   while (cur_region < end_region) {
 566     _region_data[cur_region].set_destination(addr);
 567     _region_data[cur_region].set_destination_count(0);
 568     _region_data[cur_region].set_source_region(cur_region);
 569     _region_data[cur_region].set_data_location(addr);
 570 
 571     // Update live_obj_size so the region appears completely full.
 572     size_t live_size = RegionSize - _region_data[cur_region].partial_obj_size();
 573     _region_data[cur_region].set_live_obj_size(live_size);
 574 
 575     ++cur_region;
 576     addr += RegionSize;
 577   }
 578 }
 579 
 580 // Find the point at which a space can be split and, if necessary, record the
 581 // split point.
 582 //
 583 // If the current src region (which overflowed the destination space) doesn't
 584 // have a partial object, the split point is at the beginning of the current src
 585 // region (an "easy" split, no extra bookkeeping required).
 586 //
 587 // If the current src region has a partial object, the split point is in the
 588 // region where that partial object starts (call it the split_region).  If
 589 // split_region has a partial object, then the split point is just after that
 590 // partial object (a "hard" split where we have to record the split data and
 591 // zero the partial_obj_size field).  With a "hard" split, we know that the
 592 // partial_obj ends within split_region because the partial object that caused
 593 // the overflow starts in split_region.  If split_region doesn't have a partial
 594 // obj, then the split is at the beginning of split_region (another "easy"
 595 // split).
 596 HeapWord*
 597 ParallelCompactData::summarize_split_space(size_t src_region,
 598                                            SplitInfo& split_info,
 599                                            HeapWord* destination,
 600                                            HeapWord* target_end,
 601                                            HeapWord** target_next)
 602 {
 603   assert(destination <= target_end, "sanity");
 604   assert(destination + _region_data[src_region].data_size() > target_end,
 605     "region should not fit into target space");
 606   assert(is_region_aligned(target_end), "sanity");
 607 
 608   size_t split_region = src_region;
 609   HeapWord* split_destination = destination;
 610   size_t partial_obj_size = _region_data[src_region].partial_obj_size();
 611 
 612   if (destination + partial_obj_size > target_end) {
 613     // The split point is just after the partial object (if any) in the
 614     // src_region that contains the start of the object that overflowed the
 615     // destination space.
 616     //
 617     // Find the start of the "overflow" object and set split_region to the
 618     // region containing it.
 619     HeapWord* const overflow_obj = _region_data[src_region].partial_obj_addr();
 620     split_region = addr_to_region_idx(overflow_obj);
 621 
 622     // Clear the source_region field of all destination regions whose first word
 623     // came from data after the split point (a non-null source_region field
 624     // implies a region must be filled).
 625     //
 626     // An alternative to the simple loop below:  clear during post_compact(),
 627     // which uses memcpy instead of individual stores, and is easy to
 628     // parallelize.  (The downside is that it clears the entire RegionData
 629     // object as opposed to just one field.)
 630     //
 631     // post_compact() would have to clear the summary data up to the highest
 632     // address that was written during the summary phase, which would be
 633     //
 634     //         max(top, max(new_top, clear_top))
 635     //
 636     // where clear_top is a new field in SpaceInfo.  Would have to set clear_top
 637     // to target_end.
 638     const RegionData* const sr = region(split_region);
 639     const size_t beg_idx =
 640       addr_to_region_idx(region_align_up(sr->destination() +
 641                                          sr->partial_obj_size()));
 642     const size_t end_idx = addr_to_region_idx(target_end);
 643 
 644     log_develop_trace(gc, compaction)("split:  clearing source_region field in [" SIZE_FORMAT ", " SIZE_FORMAT ")", beg_idx, end_idx);
 645     for (size_t idx = beg_idx; idx < end_idx; ++idx) {
 646       _region_data[idx].set_source_region(0);
 647     }
 648 
 649     // Set split_destination and partial_obj_size to reflect the split region.
 650     split_destination = sr->destination();
 651     partial_obj_size = sr->partial_obj_size();
 652   }
 653 
 654   // The split is recorded only if a partial object extends onto the region.
 655   if (partial_obj_size != 0) {
 656     _region_data[split_region].set_partial_obj_size(0);
 657     split_info.record(split_region, partial_obj_size, split_destination);
 658   }
 659 
 660   // Setup the continuation addresses.
 661   *target_next = split_destination + partial_obj_size;
 662   HeapWord* const source_next = region_to_addr(split_region) + partial_obj_size;
 663 
 664   if (log_develop_is_enabled(Trace, gc, compaction)) {
 665     const char * split_type = partial_obj_size == 0 ? "easy" : "hard";
 666     log_develop_trace(gc, compaction)("%s split:  src=" PTR_FORMAT " src_c=" SIZE_FORMAT " pos=" SIZE_FORMAT,
 667                                       split_type, p2i(source_next), split_region, partial_obj_size);
 668     log_develop_trace(gc, compaction)("%s split:  dst=" PTR_FORMAT " dst_c=" SIZE_FORMAT " tn=" PTR_FORMAT,
 669                                       split_type, p2i(split_destination),
 670                                       addr_to_region_idx(split_destination),
 671                                       p2i(*target_next));
 672 
 673     if (partial_obj_size != 0) {
 674       HeapWord* const po_beg = split_info.destination();
 675       HeapWord* const po_end = po_beg + split_info.partial_obj_size();
 676       log_develop_trace(gc, compaction)("%s split:  po_beg=" PTR_FORMAT " " SIZE_FORMAT " po_end=" PTR_FORMAT " " SIZE_FORMAT,
 677                                         split_type,
 678                                         p2i(po_beg), addr_to_region_idx(po_beg),
 679                                         p2i(po_end), addr_to_region_idx(po_end));
 680     }
 681   }
 682 
 683   return source_next;
 684 }
 685 
 686 bool ParallelCompactData::summarize(SplitInfo& split_info,
 687                                     HeapWord* source_beg, HeapWord* source_end,
 688                                     HeapWord** source_next,
 689                                     HeapWord* target_beg, HeapWord* target_end,
 690                                     HeapWord** target_next)
 691 {
 692   HeapWord* const source_next_val = source_next == NULL ? NULL : *source_next;
 693   log_develop_trace(gc, compaction)(
 694       "sb=" PTR_FORMAT " se=" PTR_FORMAT " sn=" PTR_FORMAT
 695       "tb=" PTR_FORMAT " te=" PTR_FORMAT " tn=" PTR_FORMAT,
 696       p2i(source_beg), p2i(source_end), p2i(source_next_val),
 697       p2i(target_beg), p2i(target_end), p2i(*target_next));
 698 
 699   size_t cur_region = addr_to_region_idx(source_beg);
 700   const size_t end_region = addr_to_region_idx(region_align_up(source_end));
 701 
 702   HeapWord *dest_addr = target_beg;
 703   while (cur_region < end_region) {
 704     // The destination must be set even if the region has no data.
 705     _region_data[cur_region].set_destination(dest_addr);
 706 
 707     size_t words = _region_data[cur_region].data_size();
 708     if (words > 0) {
 709       // If cur_region does not fit entirely into the target space, find a point
 710       // at which the source space can be 'split' so that part is copied to the
 711       // target space and the rest is copied elsewhere.
 712       if (dest_addr + words > target_end) {
 713         assert(source_next != NULL, "source_next is NULL when splitting");
 714         *source_next = summarize_split_space(cur_region, split_info, dest_addr,
 715                                              target_end, target_next);
 716         return false;
 717       }
 718 
 719       // Compute the destination_count for cur_region, and if necessary, update
 720       // source_region for a destination region.  The source_region field is
 721       // updated if cur_region is the first (left-most) region to be copied to a
 722       // destination region.
 723       //
 724       // The destination_count calculation is a bit subtle.  A region that has
 725       // data that compacts into itself does not count itself as a destination.
 726       // This maintains the invariant that a zero count means the region is
 727       // available and can be claimed and then filled.
 728       uint destination_count = 0;
 729       if (split_info.is_split(cur_region)) {
 730         // The current region has been split:  the partial object will be copied
 731         // to one destination space and the remaining data will be copied to
 732         // another destination space.  Adjust the initial destination_count and,
 733         // if necessary, set the source_region field if the partial object will
 734         // cross a destination region boundary.
 735         destination_count = split_info.destination_count();
 736         if (destination_count == 2) {
 737           size_t dest_idx = addr_to_region_idx(split_info.dest_region_addr());
 738           _region_data[dest_idx].set_source_region(cur_region);
 739         }
 740       }
 741 
 742       HeapWord* const last_addr = dest_addr + words - 1;
 743       const size_t dest_region_1 = addr_to_region_idx(dest_addr);
 744       const size_t dest_region_2 = addr_to_region_idx(last_addr);
 745 
 746       // Initially assume that the destination regions will be the same and
 747       // adjust the value below if necessary.  Under this assumption, if
 748       // cur_region == dest_region_2, then cur_region will be compacted
 749       // completely into itself.
 750       destination_count += cur_region == dest_region_2 ? 0 : 1;
 751       if (dest_region_1 != dest_region_2) {
 752         // Destination regions differ; adjust destination_count.
 753         destination_count += 1;
 754         // Data from cur_region will be copied to the start of dest_region_2.
 755         _region_data[dest_region_2].set_source_region(cur_region);
 756       } else if (is_region_aligned(dest_addr)) {
 757         // Data from cur_region will be copied to the start of the destination
 758         // region.
 759         _region_data[dest_region_1].set_source_region(cur_region);
 760       }
 761 
 762       _region_data[cur_region].set_destination_count(destination_count);
 763       _region_data[cur_region].set_data_location(region_to_addr(cur_region));
 764       dest_addr += words;
 765     }
 766 
 767     ++cur_region;
 768   }
 769 
 770   *target_next = dest_addr;
 771   return true;
 772 }
 773 
 774 HeapWord* ParallelCompactData::calc_new_pointer(HeapWord* addr, ParCompactionManager* cm) const {
 775   assert(addr != NULL, "Should detect NULL oop earlier");
 776   assert(ParallelScavengeHeap::heap()->is_in(addr), "not in heap");
 777   assert(PSParallelCompact::mark_bitmap()->is_marked(addr), "not marked");
 778 
 779   // Region covering the object.
 780   RegionData* const region_ptr = addr_to_region_ptr(addr);
 781   HeapWord* result = region_ptr->destination();
 782 
 783   // If the entire Region is live, the new location is region->destination + the
 784   // offset of the object within in the Region.
 785 
 786   // Run some performance tests to determine if this special case pays off.  It
 787   // is worth it for pointers into the dense prefix.  If the optimization to
 788   // avoid pointer updates in regions that only point to the dense prefix is
 789   // ever implemented, this should be revisited.
 790   if (region_ptr->data_size() == RegionSize) {
 791     result += region_offset(addr);
 792     return result;
 793   }
 794 
 795   // Otherwise, the new location is region->destination + block offset + the
 796   // number of live words in the Block that are (a) to the left of addr and (b)
 797   // due to objects that start in the Block.
 798 
 799   // Fill in the block table if necessary.  This is unsynchronized, so multiple
 800   // threads may fill the block table for a region (harmless, since it is
 801   // idempotent).
 802   if (!region_ptr->blocks_filled()) {
 803     PSParallelCompact::fill_blocks(addr_to_region_idx(addr));
 804     region_ptr->set_blocks_filled();
 805   }
 806 
 807   HeapWord* const search_start = block_align_down(addr);
 808   const size_t block_offset = addr_to_block_ptr(addr)->offset();
 809 
 810   const ParMarkBitMap* bitmap = PSParallelCompact::mark_bitmap();
 811   const size_t live = bitmap->live_words_in_range(cm, search_start, cast_to_oop(addr));
 812   result += block_offset + live;
 813   DEBUG_ONLY(PSParallelCompact::check_new_location(addr, result));
 814   return result;
 815 }
 816 
 817 #ifdef ASSERT
 818 void ParallelCompactData::verify_clear(const PSVirtualSpace* vspace)
 819 {
 820   const size_t* const beg = (const size_t*)vspace->committed_low_addr();
 821   const size_t* const end = (const size_t*)vspace->committed_high_addr();
 822   for (const size_t* p = beg; p < end; ++p) {
 823     assert(*p == 0, "not zero");
 824   }
 825 }
 826 
 827 void ParallelCompactData::verify_clear()
 828 {
 829   verify_clear(_region_vspace);
 830   verify_clear(_block_vspace);
 831 }
 832 #endif  // #ifdef ASSERT
 833 
 834 STWGCTimer          PSParallelCompact::_gc_timer;
 835 ParallelOldTracer   PSParallelCompact::_gc_tracer;
 836 elapsedTimer        PSParallelCompact::_accumulated_time;
 837 unsigned int        PSParallelCompact::_total_invocations = 0;
 838 unsigned int        PSParallelCompact::_maximum_compaction_gc_num = 0;
 839 CollectorCounters*  PSParallelCompact::_counters = NULL;
 840 ParMarkBitMap       PSParallelCompact::_mark_bitmap;
 841 ParallelCompactData PSParallelCompact::_summary_data;
 842 
 843 PSParallelCompact::IsAliveClosure PSParallelCompact::_is_alive_closure;
 844 
 845 bool PSParallelCompact::IsAliveClosure::do_object_b(oop p) { return mark_bitmap()->is_marked(p); }
 846 
 847 class PCReferenceProcessor: public ReferenceProcessor {
 848 public:
 849   PCReferenceProcessor(
 850     BoolObjectClosure* is_subject_to_discovery,
 851     BoolObjectClosure* is_alive_non_header) :
 852       ReferenceProcessor(is_subject_to_discovery,
 853       ParallelGCThreads,   // mt processing degree
 854       true,                // mt discovery
 855       ParallelGCThreads,   // mt discovery degree
 856       true,                // atomic_discovery
 857       is_alive_non_header) {
 858   }
 859 
 860   template<typename T> bool discover(oop obj, ReferenceType type) {
 861     T* referent_addr = (T*) java_lang_ref_Reference::referent_addr_raw(obj);
 862     T heap_oop = RawAccess<>::oop_load(referent_addr);
 863     oop referent = CompressedOops::decode_not_null(heap_oop);
 864     return PSParallelCompact::mark_bitmap()->is_unmarked(referent)
 865         && ReferenceProcessor::discover_reference(obj, type);
 866   }
 867   virtual bool discover_reference(oop obj, ReferenceType type) {
 868     if (UseCompressedOops) {
 869       return discover<narrowOop>(obj, type);
 870     } else {
 871       return discover<oop>(obj, type);
 872     }
 873   }
 874 };
 875 
 876 void PSParallelCompact::post_initialize() {
 877   ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
 878   _span_based_discoverer.set_span(heap->reserved_region());
 879   _ref_processor =
 880     new PCReferenceProcessor(&_span_based_discoverer,
 881                              &_is_alive_closure); // non-header is alive closure
 882 
 883   _counters = new CollectorCounters("Parallel full collection pauses", 1);
 884 
 885   // Initialize static fields in ParCompactionManager.
 886   ParCompactionManager::initialize(mark_bitmap());
 887 }
 888 
 889 bool PSParallelCompact::initialize() {
 890   ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
 891   MemRegion mr = heap->reserved_region();
 892 
 893   // Was the old gen get allocated successfully?
 894   if (!heap->old_gen()->is_allocated()) {
 895     return false;
 896   }
 897 
 898   initialize_space_info();
 899   initialize_dead_wood_limiter();
 900 
 901   if (!_mark_bitmap.initialize(mr)) {
 902     vm_shutdown_during_initialization(
 903       err_msg("Unable to allocate " SIZE_FORMAT "KB bitmaps for parallel "
 904       "garbage collection for the requested " SIZE_FORMAT "KB heap.",
 905       _mark_bitmap.reserved_byte_size()/K, mr.byte_size()/K));
 906     return false;
 907   }
 908 
 909   if (!_summary_data.initialize(mr)) {
 910     vm_shutdown_during_initialization(
 911       err_msg("Unable to allocate " SIZE_FORMAT "KB card tables for parallel "
 912       "garbage collection for the requested " SIZE_FORMAT "KB heap.",
 913       _summary_data.reserved_byte_size()/K, mr.byte_size()/K));
 914     return false;
 915   }
 916 
 917   return true;
 918 }
 919 
 920 void PSParallelCompact::initialize_space_info()
 921 {
 922   memset(&_space_info, 0, sizeof(_space_info));
 923 
 924   ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
 925   PSYoungGen* young_gen = heap->young_gen();
 926 
 927   _space_info[old_space_id].set_space(heap->old_gen()->object_space());
 928   _space_info[eden_space_id].set_space(young_gen->eden_space());
 929   _space_info[from_space_id].set_space(young_gen->from_space());
 930   _space_info[to_space_id].set_space(young_gen->to_space());
 931 
 932   _space_info[old_space_id].set_start_array(heap->old_gen()->start_array());
 933 }
 934 
 935 void PSParallelCompact::initialize_dead_wood_limiter()
 936 {
 937   const size_t max = 100;
 938   _dwl_mean = double(MIN2(ParallelOldDeadWoodLimiterMean, max)) / 100.0;
 939   _dwl_std_dev = double(MIN2(ParallelOldDeadWoodLimiterStdDev, max)) / 100.0;
 940   _dwl_first_term = 1.0 / (sqrt(2.0 * M_PI) * _dwl_std_dev);
 941   DEBUG_ONLY(_dwl_initialized = true;)
 942   _dwl_adjustment = normal_distribution(1.0);
 943 }
 944 
 945 void
 946 PSParallelCompact::clear_data_covering_space(SpaceId id)
 947 {
 948   // At this point, top is the value before GC, new_top() is the value that will
 949   // be set at the end of GC.  The marking bitmap is cleared to top; nothing
 950   // should be marked above top.  The summary data is cleared to the larger of
 951   // top & new_top.
 952   MutableSpace* const space = _space_info[id].space();
 953   HeapWord* const bot = space->bottom();
 954   HeapWord* const top = space->top();
 955   HeapWord* const max_top = MAX2(top, _space_info[id].new_top());
 956 
 957   const idx_t beg_bit = _mark_bitmap.addr_to_bit(bot);
 958   const idx_t end_bit = _mark_bitmap.align_range_end(_mark_bitmap.addr_to_bit(top));
 959   _mark_bitmap.clear_range(beg_bit, end_bit);
 960 
 961   const size_t beg_region = _summary_data.addr_to_region_idx(bot);
 962   const size_t end_region =
 963     _summary_data.addr_to_region_idx(_summary_data.region_align_up(max_top));
 964   _summary_data.clear_range(beg_region, end_region);
 965 
 966   // Clear the data used to 'split' regions.
 967   SplitInfo& split_info = _space_info[id].split_info();
 968   if (split_info.is_valid()) {
 969     split_info.clear();
 970   }
 971   DEBUG_ONLY(split_info.verify_clear();)
 972 }
 973 
 974 void PSParallelCompact::pre_compact()
 975 {
 976   // Update the from & to space pointers in space_info, since they are swapped
 977   // at each young gen gc.  Do the update unconditionally (even though a
 978   // promotion failure does not swap spaces) because an unknown number of young
 979   // collections will have swapped the spaces an unknown number of times.
 980   GCTraceTime(Debug, gc, phases) tm("Pre Compact", &_gc_timer);
 981   ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
 982   _space_info[from_space_id].set_space(heap->young_gen()->from_space());
 983   _space_info[to_space_id].set_space(heap->young_gen()->to_space());
 984 
 985   // Increment the invocation count
 986   heap->increment_total_collections(true);
 987 
 988   // We need to track unique mark sweep invocations as well.
 989   _total_invocations++;
 990 
 991   heap->print_heap_before_gc();
 992   heap->trace_heap_before_gc(&_gc_tracer);
 993 
 994   // Fill in TLABs
 995   heap->ensure_parsability(true);  // retire TLABs
 996 
 997   if (VerifyBeforeGC && heap->total_collections() >= VerifyGCStartAt) {
 998     Universe::verify("Before GC");
 999   }
1000 
1001   // Verify object start arrays
1002   if (VerifyObjectStartArray &&
1003       VerifyBeforeGC) {
1004     heap->old_gen()->verify_object_start_array();
1005   }
1006 
1007   DEBUG_ONLY(mark_bitmap()->verify_clear();)
1008   DEBUG_ONLY(summary_data().verify_clear();)
1009 
1010   ParCompactionManager::reset_all_bitmap_query_caches();
1011 }
1012 
1013 void PSParallelCompact::post_compact()
1014 {
1015   GCTraceTime(Info, gc, phases) tm("Post Compact", &_gc_timer);
1016   ParCompactionManager::remove_all_shadow_regions();
1017 
1018   for (unsigned int id = old_space_id; id < last_space_id; ++id) {
1019     // Clear the marking bitmap, summary data and split info.
1020     clear_data_covering_space(SpaceId(id));
1021     // Update top().  Must be done after clearing the bitmap and summary data.
1022     _space_info[id].publish_new_top();
1023   }
1024 
1025   ParCompactionManager::flush_all_string_dedup_requests();
1026 
1027   MutableSpace* const eden_space = _space_info[eden_space_id].space();
1028   MutableSpace* const from_space = _space_info[from_space_id].space();
1029   MutableSpace* const to_space   = _space_info[to_space_id].space();
1030 
1031   ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
1032   bool eden_empty = eden_space->is_empty();
1033 
1034   // Update heap occupancy information which is used as input to the soft ref
1035   // clearing policy at the next gc.
1036   Universe::heap()->update_capacity_and_used_at_gc();
1037 
1038   bool young_gen_empty = eden_empty && from_space->is_empty() &&
1039     to_space->is_empty();
1040 
1041   PSCardTable* ct = heap->card_table();
1042   MemRegion old_mr = heap->old_gen()->reserved();
1043   if (young_gen_empty) {
1044     ct->clear(MemRegion(old_mr.start(), old_mr.end()));
1045   } else {
1046     ct->invalidate(MemRegion(old_mr.start(), old_mr.end()));
1047   }
1048 
1049   // Delete metaspaces for unloaded class loaders and clean up loader_data graph
1050   ClassLoaderDataGraph::purge(/*at_safepoint*/true);
1051   DEBUG_ONLY(MetaspaceUtils::verify();)
1052 
1053   heap->prune_scavengable_nmethods();
1054 
1055 #if COMPILER2_OR_JVMCI
1056   DerivedPointerTable::update_pointers();
1057 #endif
1058 
1059   if (ZapUnusedHeapArea) {
1060     heap->gen_mangle_unused_area();
1061   }
1062 
1063   // Signal that we have completed a visit to all live objects.
1064   Universe::heap()->record_whole_heap_examined_timestamp();
1065 }
1066 
1067 HeapWord*
1068 PSParallelCompact::compute_dense_prefix_via_density(const SpaceId id,
1069                                                     bool maximum_compaction)
1070 {
1071   const size_t region_size = ParallelCompactData::RegionSize;
1072   const ParallelCompactData& sd = summary_data();
1073 
1074   const MutableSpace* const space = _space_info[id].space();
1075   HeapWord* const top_aligned_up = sd.region_align_up(space->top());
1076   const RegionData* const beg_cp = sd.addr_to_region_ptr(space->bottom());
1077   const RegionData* const end_cp = sd.addr_to_region_ptr(top_aligned_up);
1078 
1079   // Skip full regions at the beginning of the space--they are necessarily part
1080   // of the dense prefix.
1081   size_t full_count = 0;
1082   const RegionData* cp;
1083   for (cp = beg_cp; cp < end_cp && cp->data_size() == region_size; ++cp) {
1084     ++full_count;
1085   }
1086 
1087   assert(total_invocations() >= _maximum_compaction_gc_num, "sanity");
1088   const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num;
1089   const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval;
1090   if (maximum_compaction || cp == end_cp || interval_ended) {
1091     _maximum_compaction_gc_num = total_invocations();
1092     return sd.region_to_addr(cp);
1093   }
1094 
1095   HeapWord* const new_top = _space_info[id].new_top();
1096   const size_t space_live = pointer_delta(new_top, space->bottom());
1097   const size_t space_used = space->used_in_words();
1098   const size_t space_capacity = space->capacity_in_words();
1099 
1100   const double cur_density = double(space_live) / space_capacity;
1101   const double deadwood_density =
1102     (1.0 - cur_density) * (1.0 - cur_density) * cur_density * cur_density;
1103   const size_t deadwood_goal = size_t(space_capacity * deadwood_density);
1104 
1105   log_develop_debug(gc, compaction)(
1106       "cur_dens=%5.3f dw_dens=%5.3f dw_goal=" SIZE_FORMAT,
1107       cur_density, deadwood_density, deadwood_goal);
1108   log_develop_debug(gc, compaction)(
1109       "space_live=" SIZE_FORMAT " space_used=" SIZE_FORMAT " "
1110       "space_cap=" SIZE_FORMAT,
1111       space_live, space_used,
1112       space_capacity);
1113 
1114   // XXX - Use binary search?
1115   HeapWord* dense_prefix = sd.region_to_addr(cp);
1116   const RegionData* full_cp = cp;
1117   const RegionData* const top_cp = sd.addr_to_region_ptr(space->top() - 1);
1118   while (cp < end_cp) {
1119     HeapWord* region_destination = cp->destination();
1120     const size_t cur_deadwood = pointer_delta(dense_prefix, region_destination);
1121 
1122     log_develop_trace(gc, compaction)(
1123         "c#=" SIZE_FORMAT_W(4) " dst=" PTR_FORMAT " "
1124         "dp=" PTR_FORMAT " cdw=" SIZE_FORMAT_W(8),
1125         sd.region(cp), p2i(region_destination),
1126         p2i(dense_prefix), cur_deadwood);
1127 
1128     if (cur_deadwood >= deadwood_goal) {
1129       // Found the region that has the correct amount of deadwood to the left.
1130       // This typically occurs after crossing a fairly sparse set of regions, so
1131       // iterate backwards over those sparse regions, looking for the region
1132       // that has the lowest density of live objects 'to the right.'
1133       size_t space_to_left = sd.region(cp) * region_size;
1134       size_t live_to_left = space_to_left - cur_deadwood;
1135       size_t space_to_right = space_capacity - space_to_left;
1136       size_t live_to_right = space_live - live_to_left;
1137       double density_to_right = double(live_to_right) / space_to_right;
1138       while (cp > full_cp) {
1139         --cp;
1140         const size_t prev_region_live_to_right = live_to_right -
1141           cp->data_size();
1142         const size_t prev_region_space_to_right = space_to_right + region_size;
1143         double prev_region_density_to_right =
1144           double(prev_region_live_to_right) / prev_region_space_to_right;
1145         if (density_to_right <= prev_region_density_to_right) {
1146           return dense_prefix;
1147         }
1148 
1149         log_develop_trace(gc, compaction)(
1150             "backing up from c=" SIZE_FORMAT_W(4) " d2r=%10.8f "
1151             "pc_d2r=%10.8f",
1152             sd.region(cp), density_to_right,
1153             prev_region_density_to_right);
1154 
1155         dense_prefix -= region_size;
1156         live_to_right = prev_region_live_to_right;
1157         space_to_right = prev_region_space_to_right;
1158         density_to_right = prev_region_density_to_right;
1159       }
1160       return dense_prefix;
1161     }
1162 
1163     dense_prefix += region_size;
1164     ++cp;
1165   }
1166 
1167   return dense_prefix;
1168 }
1169 
1170 #ifndef PRODUCT
1171 void PSParallelCompact::print_dense_prefix_stats(const char* const algorithm,
1172                                                  const SpaceId id,
1173                                                  const bool maximum_compaction,
1174                                                  HeapWord* const addr)
1175 {
1176   const size_t region_idx = summary_data().addr_to_region_idx(addr);
1177   RegionData* const cp = summary_data().region(region_idx);
1178   const MutableSpace* const space = _space_info[id].space();
1179   HeapWord* const new_top = _space_info[id].new_top();
1180 
1181   const size_t space_live = pointer_delta(new_top, space->bottom());
1182   const size_t dead_to_left = pointer_delta(addr, cp->destination());
1183   const size_t space_cap = space->capacity_in_words();
1184   const double dead_to_left_pct = double(dead_to_left) / space_cap;
1185   const size_t live_to_right = new_top - cp->destination();
1186   const size_t dead_to_right = space->top() - addr - live_to_right;
1187 
1188   log_develop_debug(gc, compaction)(
1189       "%s=" PTR_FORMAT " dpc=" SIZE_FORMAT_W(5) " "
1190       "spl=" SIZE_FORMAT " "
1191       "d2l=" SIZE_FORMAT " d2l%%=%6.4f "
1192       "d2r=" SIZE_FORMAT " l2r=" SIZE_FORMAT " "
1193       "ratio=%10.8f",
1194       algorithm, p2i(addr), region_idx,
1195       space_live,
1196       dead_to_left, dead_to_left_pct,
1197       dead_to_right, live_to_right,
1198       double(dead_to_right) / live_to_right);
1199 }
1200 #endif  // #ifndef PRODUCT
1201 
1202 // Return a fraction indicating how much of the generation can be treated as
1203 // "dead wood" (i.e., not reclaimed).  The function uses a normal distribution
1204 // based on the density of live objects in the generation to determine a limit,
1205 // which is then adjusted so the return value is min_percent when the density is
1206 // 1.
1207 //
1208 // The following table shows some return values for a different values of the
1209 // standard deviation (ParallelOldDeadWoodLimiterStdDev); the mean is 0.5 and
1210 // min_percent is 1.
1211 //
1212 //                          fraction allowed as dead wood
1213 //         -----------------------------------------------------------------
1214 // density std_dev=70 std_dev=75 std_dev=80 std_dev=85 std_dev=90 std_dev=95
1215 // ------- ---------- ---------- ---------- ---------- ---------- ----------
1216 // 0.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000
1217 // 0.05000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941
1218 // 0.10000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272
1219 // 0.15000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066
1220 // 0.20000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975
1221 // 0.25000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313
1222 // 0.30000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132
1223 // 0.35000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289
1224 // 0.40000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500
1225 // 0.45000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386
1226 // 0.50000 0.13832410 0.11599237 0.09847664 0.08456518 0.07338887 0.06431510
1227 // 0.55000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386
1228 // 0.60000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500
1229 // 0.65000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289
1230 // 0.70000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132
1231 // 0.75000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313
1232 // 0.80000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975
1233 // 0.85000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066
1234 // 0.90000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272
1235 // 0.95000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941
1236 // 1.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000
1237 
1238 double PSParallelCompact::dead_wood_limiter(double density, size_t min_percent)
1239 {
1240   assert(_dwl_initialized, "uninitialized");
1241 
1242   // The raw limit is the value of the normal distribution at x = density.
1243   const double raw_limit = normal_distribution(density);
1244 
1245   // Adjust the raw limit so it becomes the minimum when the density is 1.
1246   //
1247   // First subtract the adjustment value (which is simply the precomputed value
1248   // normal_distribution(1.0)); this yields a value of 0 when the density is 1.
1249   // Then add the minimum value, so the minimum is returned when the density is
1250   // 1.  Finally, prevent negative values, which occur when the mean is not 0.5.
1251   const double min = double(min_percent) / 100.0;
1252   const double limit = raw_limit - _dwl_adjustment + min;
1253   return MAX2(limit, 0.0);
1254 }
1255 
1256 ParallelCompactData::RegionData*
1257 PSParallelCompact::first_dead_space_region(const RegionData* beg,
1258                                            const RegionData* end)
1259 {
1260   const size_t region_size = ParallelCompactData::RegionSize;
1261   ParallelCompactData& sd = summary_data();
1262   size_t left = sd.region(beg);
1263   size_t right = end > beg ? sd.region(end) - 1 : left;
1264 
1265   // Binary search.
1266   while (left < right) {
1267     // Equivalent to (left + right) / 2, but does not overflow.
1268     const size_t middle = left + (right - left) / 2;
1269     RegionData* const middle_ptr = sd.region(middle);
1270     HeapWord* const dest = middle_ptr->destination();
1271     HeapWord* const addr = sd.region_to_addr(middle);
1272     assert(dest != NULL, "sanity");
1273     assert(dest <= addr, "must move left");
1274 
1275     if (middle > left && dest < addr) {
1276       right = middle - 1;
1277     } else if (middle < right && middle_ptr->data_size() == region_size) {
1278       left = middle + 1;
1279     } else {
1280       return middle_ptr;
1281     }
1282   }
1283   return sd.region(left);
1284 }
1285 
1286 ParallelCompactData::RegionData*
1287 PSParallelCompact::dead_wood_limit_region(const RegionData* beg,
1288                                           const RegionData* end,
1289                                           size_t dead_words)
1290 {
1291   ParallelCompactData& sd = summary_data();
1292   size_t left = sd.region(beg);
1293   size_t right = end > beg ? sd.region(end) - 1 : left;
1294 
1295   // Binary search.
1296   while (left < right) {
1297     // Equivalent to (left + right) / 2, but does not overflow.
1298     const size_t middle = left + (right - left) / 2;
1299     RegionData* const middle_ptr = sd.region(middle);
1300     HeapWord* const dest = middle_ptr->destination();
1301     HeapWord* const addr = sd.region_to_addr(middle);
1302     assert(dest != NULL, "sanity");
1303     assert(dest <= addr, "must move left");
1304 
1305     const size_t dead_to_left = pointer_delta(addr, dest);
1306     if (middle > left && dead_to_left > dead_words) {
1307       right = middle - 1;
1308     } else if (middle < right && dead_to_left < dead_words) {
1309       left = middle + 1;
1310     } else {
1311       return middle_ptr;
1312     }
1313   }
1314   return sd.region(left);
1315 }
1316 
1317 // The result is valid during the summary phase, after the initial summarization
1318 // of each space into itself, and before final summarization.
1319 inline double
1320 PSParallelCompact::reclaimed_ratio(const RegionData* const cp,
1321                                    HeapWord* const bottom,
1322                                    HeapWord* const top,
1323                                    HeapWord* const new_top)
1324 {
1325   ParallelCompactData& sd = summary_data();
1326 
1327   assert(cp != NULL, "sanity");
1328   assert(bottom != NULL, "sanity");
1329   assert(top != NULL, "sanity");
1330   assert(new_top != NULL, "sanity");
1331   assert(top >= new_top, "summary data problem?");
1332   assert(new_top > bottom, "space is empty; should not be here");
1333   assert(new_top >= cp->destination(), "sanity");
1334   assert(top >= sd.region_to_addr(cp), "sanity");
1335 
1336   HeapWord* const destination = cp->destination();
1337   const size_t dense_prefix_live  = pointer_delta(destination, bottom);
1338   const size_t compacted_region_live = pointer_delta(new_top, destination);
1339   const size_t compacted_region_used = pointer_delta(top,
1340                                                      sd.region_to_addr(cp));
1341   const size_t reclaimable = compacted_region_used - compacted_region_live;
1342 
1343   const double divisor = dense_prefix_live + 1.25 * compacted_region_live;
1344   return double(reclaimable) / divisor;
1345 }
1346 
1347 // Return the address of the end of the dense prefix, a.k.a. the start of the
1348 // compacted region.  The address is always on a region boundary.
1349 //
1350 // Completely full regions at the left are skipped, since no compaction can
1351 // occur in those regions.  Then the maximum amount of dead wood to allow is
1352 // computed, based on the density (amount live / capacity) of the generation;
1353 // the region with approximately that amount of dead space to the left is
1354 // identified as the limit region.  Regions between the last completely full
1355 // region and the limit region are scanned and the one that has the best
1356 // (maximum) reclaimed_ratio() is selected.
1357 HeapWord*
1358 PSParallelCompact::compute_dense_prefix(const SpaceId id,
1359                                         bool maximum_compaction)
1360 {
1361   const size_t region_size = ParallelCompactData::RegionSize;
1362   const ParallelCompactData& sd = summary_data();
1363 
1364   const MutableSpace* const space = _space_info[id].space();
1365   HeapWord* const top = space->top();
1366   HeapWord* const top_aligned_up = sd.region_align_up(top);
1367   HeapWord* const new_top = _space_info[id].new_top();
1368   HeapWord* const new_top_aligned_up = sd.region_align_up(new_top);
1369   HeapWord* const bottom = space->bottom();
1370   const RegionData* const beg_cp = sd.addr_to_region_ptr(bottom);
1371   const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up);
1372   const RegionData* const new_top_cp =
1373     sd.addr_to_region_ptr(new_top_aligned_up);
1374 
1375   // Skip full regions at the beginning of the space--they are necessarily part
1376   // of the dense prefix.
1377   const RegionData* const full_cp = first_dead_space_region(beg_cp, new_top_cp);
1378   assert(full_cp->destination() == sd.region_to_addr(full_cp) ||
1379          space->is_empty(), "no dead space allowed to the left");
1380   assert(full_cp->data_size() < region_size || full_cp == new_top_cp - 1,
1381          "region must have dead space");
1382 
1383   // The gc number is saved whenever a maximum compaction is done, and used to
1384   // determine when the maximum compaction interval has expired.  This avoids
1385   // successive max compactions for different reasons.
1386   assert(total_invocations() >= _maximum_compaction_gc_num, "sanity");
1387   const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num;
1388   const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval ||
1389     total_invocations() == HeapFirstMaximumCompactionCount;
1390   if (maximum_compaction || full_cp == top_cp || interval_ended) {
1391     _maximum_compaction_gc_num = total_invocations();
1392     return sd.region_to_addr(full_cp);
1393   }
1394 
1395   const size_t space_live = pointer_delta(new_top, bottom);
1396   const size_t space_used = space->used_in_words();
1397   const size_t space_capacity = space->capacity_in_words();
1398 
1399   const double density = double(space_live) / double(space_capacity);
1400   const size_t min_percent_free = MarkSweepDeadRatio;
1401   const double limiter = dead_wood_limiter(density, min_percent_free);
1402   const size_t dead_wood_max = space_used - space_live;
1403   const size_t dead_wood_limit = MIN2(size_t(space_capacity * limiter),
1404                                       dead_wood_max);
1405 
1406   log_develop_debug(gc, compaction)(
1407       "space_live=" SIZE_FORMAT " space_used=" SIZE_FORMAT " "
1408       "space_cap=" SIZE_FORMAT,
1409       space_live, space_used,
1410       space_capacity);
1411   log_develop_debug(gc, compaction)(
1412       "dead_wood_limiter(%6.4f, " SIZE_FORMAT ")=%6.4f "
1413       "dead_wood_max=" SIZE_FORMAT " dead_wood_limit=" SIZE_FORMAT,
1414       density, min_percent_free, limiter,
1415       dead_wood_max, dead_wood_limit);
1416 
1417   // Locate the region with the desired amount of dead space to the left.
1418   const RegionData* const limit_cp =
1419     dead_wood_limit_region(full_cp, top_cp, dead_wood_limit);
1420 
1421   // Scan from the first region with dead space to the limit region and find the
1422   // one with the best (largest) reclaimed ratio.
1423   double best_ratio = 0.0;
1424   const RegionData* best_cp = full_cp;
1425   for (const RegionData* cp = full_cp; cp < limit_cp; ++cp) {
1426     double tmp_ratio = reclaimed_ratio(cp, bottom, top, new_top);
1427     if (tmp_ratio > best_ratio) {
1428       best_cp = cp;
1429       best_ratio = tmp_ratio;
1430     }
1431   }
1432 
1433   return sd.region_to_addr(best_cp);
1434 }
1435 
1436 void PSParallelCompact::summarize_spaces_quick()
1437 {
1438   for (unsigned int i = 0; i < last_space_id; ++i) {
1439     const MutableSpace* space = _space_info[i].space();
1440     HeapWord** nta = _space_info[i].new_top_addr();
1441     bool result = _summary_data.summarize(_space_info[i].split_info(),
1442                                           space->bottom(), space->top(), NULL,
1443                                           space->bottom(), space->end(), nta);
1444     assert(result, "space must fit into itself");
1445     _space_info[i].set_dense_prefix(space->bottom());
1446   }
1447 }
1448 
1449 void PSParallelCompact::fill_dense_prefix_end(SpaceId id)
1450 {
1451   HeapWord* const dense_prefix_end = dense_prefix(id);
1452   const RegionData* region = _summary_data.addr_to_region_ptr(dense_prefix_end);
1453   const idx_t dense_prefix_bit = _mark_bitmap.addr_to_bit(dense_prefix_end);
1454   if (dead_space_crosses_boundary(region, dense_prefix_bit)) {
1455     // Only enough dead space is filled so that any remaining dead space to the
1456     // left is larger than the minimum filler object.  (The remainder is filled
1457     // during the copy/update phase.)
1458     //
1459     // The size of the dead space to the right of the boundary is not a
1460     // concern, since compaction will be able to use whatever space is
1461     // available.
1462     //
1463     // Here '||' is the boundary, 'x' represents a don't care bit and a box
1464     // surrounds the space to be filled with an object.
1465     //
1466     // In the 32-bit VM, each bit represents two 32-bit words:
1467     //                              +---+
1468     // a) beg_bits:  ...  x   x   x | 0 | ||   0   x  x  ...
1469     //    end_bits:  ...  x   x   x | 0 | ||   0   x  x  ...
1470     //                              +---+
1471     //
1472     // In the 64-bit VM, each bit represents one 64-bit word:
1473     //                              +------------+
1474     // b) beg_bits:  ...  x   x   x | 0   ||   0 | x  x  ...
1475     //    end_bits:  ...  x   x   1 | 0   ||   0 | x  x  ...
1476     //                              +------------+
1477     //                          +-------+
1478     // c) beg_bits:  ...  x   x | 0   0 | ||   0   x  x  ...
1479     //    end_bits:  ...  x   1 | 0   0 | ||   0   x  x  ...
1480     //                          +-------+
1481     //                      +-----------+
1482     // d) beg_bits:  ...  x | 0   0   0 | ||   0   x  x  ...
1483     //    end_bits:  ...  1 | 0   0   0 | ||   0   x  x  ...
1484     //                      +-----------+
1485     //                          +-------+
1486     // e) beg_bits:  ...  0   0 | 0   0 | ||   0   x  x  ...
1487     //    end_bits:  ...  0   0 | 0   0 | ||   0   x  x  ...
1488     //                          +-------+
1489 
1490     // Initially assume case a, c or e will apply.
1491     size_t obj_len = CollectedHeap::min_fill_size();
1492     HeapWord* obj_beg = dense_prefix_end - obj_len;
1493 
1494 #ifdef  _LP64
1495     if (MinObjAlignment > 1) { // object alignment > heap word size
1496       // Cases a, c or e.
1497     } else if (_mark_bitmap.is_obj_end(dense_prefix_bit - 2)) {
1498       // Case b above.
1499       obj_beg = dense_prefix_end - 1;
1500     } else if (!_mark_bitmap.is_obj_end(dense_prefix_bit - 3) &&
1501                _mark_bitmap.is_obj_end(dense_prefix_bit - 4)) {
1502       // Case d above.
1503       obj_beg = dense_prefix_end - 3;
1504       obj_len = 3;
1505     }
1506 #endif  // #ifdef _LP64
1507 
1508     CollectedHeap::fill_with_object(obj_beg, obj_len);
1509     _mark_bitmap.mark_obj(obj_beg, obj_len);
1510     _summary_data.add_obj(obj_beg, obj_len);
1511     assert(start_array(id) != NULL, "sanity");
1512     start_array(id)->allocate_block(obj_beg);
1513   }
1514 }
1515 
1516 void
1517 PSParallelCompact::summarize_space(SpaceId id, bool maximum_compaction)
1518 {
1519   assert(id < last_space_id, "id out of range");
1520   assert(_space_info[id].dense_prefix() == _space_info[id].space()->bottom(),
1521          "should have been reset in summarize_spaces_quick()");
1522 
1523   const MutableSpace* space = _space_info[id].space();
1524   if (_space_info[id].new_top() != space->bottom()) {
1525     HeapWord* dense_prefix_end = compute_dense_prefix(id, maximum_compaction);
1526     _space_info[id].set_dense_prefix(dense_prefix_end);
1527 
1528 #ifndef PRODUCT
1529     if (log_is_enabled(Debug, gc, compaction)) {
1530       print_dense_prefix_stats("ratio", id, maximum_compaction,
1531                                dense_prefix_end);
1532       HeapWord* addr = compute_dense_prefix_via_density(id, maximum_compaction);
1533       print_dense_prefix_stats("density", id, maximum_compaction, addr);
1534     }
1535 #endif  // #ifndef PRODUCT
1536 
1537     // Recompute the summary data, taking into account the dense prefix.  If
1538     // every last byte will be reclaimed, then the existing summary data which
1539     // compacts everything can be left in place.
1540     if (!maximum_compaction && dense_prefix_end != space->bottom()) {
1541       // If dead space crosses the dense prefix boundary, it is (at least
1542       // partially) filled with a dummy object, marked live and added to the
1543       // summary data.  This simplifies the copy/update phase and must be done
1544       // before the final locations of objects are determined, to prevent
1545       // leaving a fragment of dead space that is too small to fill.
1546       fill_dense_prefix_end(id);
1547 
1548       // Compute the destination of each Region, and thus each object.
1549       _summary_data.summarize_dense_prefix(space->bottom(), dense_prefix_end);
1550       _summary_data.summarize(_space_info[id].split_info(),
1551                               dense_prefix_end, space->top(), NULL,
1552                               dense_prefix_end, space->end(),
1553                               _space_info[id].new_top_addr());
1554     }
1555   }
1556 
1557   if (log_develop_is_enabled(Trace, gc, compaction)) {
1558     const size_t region_size = ParallelCompactData::RegionSize;
1559     HeapWord* const dense_prefix_end = _space_info[id].dense_prefix();
1560     const size_t dp_region = _summary_data.addr_to_region_idx(dense_prefix_end);
1561     const size_t dp_words = pointer_delta(dense_prefix_end, space->bottom());
1562     HeapWord* const new_top = _space_info[id].new_top();
1563     const HeapWord* nt_aligned_up = _summary_data.region_align_up(new_top);
1564     const size_t cr_words = pointer_delta(nt_aligned_up, dense_prefix_end);
1565     log_develop_trace(gc, compaction)(
1566         "id=%d cap=" SIZE_FORMAT " dp=" PTR_FORMAT " "
1567         "dp_region=" SIZE_FORMAT " " "dp_count=" SIZE_FORMAT " "
1568         "cr_count=" SIZE_FORMAT " " "nt=" PTR_FORMAT,
1569         id, space->capacity_in_words(), p2i(dense_prefix_end),
1570         dp_region, dp_words / region_size,
1571         cr_words / region_size, p2i(new_top));
1572   }
1573 }
1574 
1575 #ifndef PRODUCT
1576 void PSParallelCompact::summary_phase_msg(SpaceId dst_space_id,
1577                                           HeapWord* dst_beg, HeapWord* dst_end,
1578                                           SpaceId src_space_id,
1579                                           HeapWord* src_beg, HeapWord* src_end)
1580 {
1581   log_develop_trace(gc, compaction)(
1582       "Summarizing %d [%s] into %d [%s]:  "
1583       "src=" PTR_FORMAT "-" PTR_FORMAT " "
1584       SIZE_FORMAT "-" SIZE_FORMAT " "
1585       "dst=" PTR_FORMAT "-" PTR_FORMAT " "
1586       SIZE_FORMAT "-" SIZE_FORMAT,
1587       src_space_id, space_names[src_space_id],
1588       dst_space_id, space_names[dst_space_id],
1589       p2i(src_beg), p2i(src_end),
1590       _summary_data.addr_to_region_idx(src_beg),
1591       _summary_data.addr_to_region_idx(src_end),
1592       p2i(dst_beg), p2i(dst_end),
1593       _summary_data.addr_to_region_idx(dst_beg),
1594       _summary_data.addr_to_region_idx(dst_end));
1595 }
1596 #endif  // #ifndef PRODUCT
1597 
1598 void PSParallelCompact::summary_phase(ParCompactionManager* cm,
1599                                       bool maximum_compaction)
1600 {
1601   GCTraceTime(Info, gc, phases) tm("Summary Phase", &_gc_timer);
1602 
1603   // Quick summarization of each space into itself, to see how much is live.
1604   summarize_spaces_quick();
1605 
1606   log_develop_trace(gc, compaction)("summary phase:  after summarizing each space to self");
1607   NOT_PRODUCT(print_region_ranges());
1608   NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info));
1609 
1610   // The amount of live data that will end up in old space (assuming it fits).
1611   size_t old_space_total_live = 0;
1612   for (unsigned int id = old_space_id; id < last_space_id; ++id) {
1613     old_space_total_live += pointer_delta(_space_info[id].new_top(),
1614                                           _space_info[id].space()->bottom());
1615   }
1616 
1617   MutableSpace* const old_space = _space_info[old_space_id].space();
1618   const size_t old_capacity = old_space->capacity_in_words();
1619   if (old_space_total_live > old_capacity) {
1620     // XXX - should also try to expand
1621     maximum_compaction = true;
1622   }
1623 
1624   // Old generations.
1625   summarize_space(old_space_id, maximum_compaction);
1626 
1627   // Summarize the remaining spaces in the young gen.  The initial target space
1628   // is the old gen.  If a space does not fit entirely into the target, then the
1629   // remainder is compacted into the space itself and that space becomes the new
1630   // target.
1631   SpaceId dst_space_id = old_space_id;
1632   HeapWord* dst_space_end = old_space->end();
1633   HeapWord** new_top_addr = _space_info[dst_space_id].new_top_addr();
1634   for (unsigned int id = eden_space_id; id < last_space_id; ++id) {
1635     const MutableSpace* space = _space_info[id].space();
1636     const size_t live = pointer_delta(_space_info[id].new_top(),
1637                                       space->bottom());
1638     const size_t available = pointer_delta(dst_space_end, *new_top_addr);
1639 
1640     NOT_PRODUCT(summary_phase_msg(dst_space_id, *new_top_addr, dst_space_end,
1641                                   SpaceId(id), space->bottom(), space->top());)
1642     if (live > 0 && live <= available) {
1643       // All the live data will fit.
1644       bool done = _summary_data.summarize(_space_info[id].split_info(),
1645                                           space->bottom(), space->top(),
1646                                           NULL,
1647                                           *new_top_addr, dst_space_end,
1648                                           new_top_addr);
1649       assert(done, "space must fit into old gen");
1650 
1651       // Reset the new_top value for the space.
1652       _space_info[id].set_new_top(space->bottom());
1653     } else if (live > 0) {
1654       // Attempt to fit part of the source space into the target space.
1655       HeapWord* next_src_addr = NULL;
1656       bool done = _summary_data.summarize(_space_info[id].split_info(),
1657                                           space->bottom(), space->top(),
1658                                           &next_src_addr,
1659                                           *new_top_addr, dst_space_end,
1660                                           new_top_addr);
1661       assert(!done, "space should not fit into old gen");
1662       assert(next_src_addr != NULL, "sanity");
1663 
1664       // The source space becomes the new target, so the remainder is compacted
1665       // within the space itself.
1666       dst_space_id = SpaceId(id);
1667       dst_space_end = space->end();
1668       new_top_addr = _space_info[id].new_top_addr();
1669       NOT_PRODUCT(summary_phase_msg(dst_space_id,
1670                                     space->bottom(), dst_space_end,
1671                                     SpaceId(id), next_src_addr, space->top());)
1672       done = _summary_data.summarize(_space_info[id].split_info(),
1673                                      next_src_addr, space->top(),
1674                                      NULL,
1675                                      space->bottom(), dst_space_end,
1676                                      new_top_addr);
1677       assert(done, "space must fit when compacted into itself");
1678       assert(*new_top_addr <= space->top(), "usage should not grow");
1679     }
1680   }
1681 
1682   log_develop_trace(gc, compaction)("Summary_phase:  after final summarization");
1683   NOT_PRODUCT(print_region_ranges());
1684   NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info));
1685 }
1686 
1687 // This method should contain all heap-specific policy for invoking a full
1688 // collection.  invoke_no_policy() will only attempt to compact the heap; it
1689 // will do nothing further.  If we need to bail out for policy reasons, scavenge
1690 // before full gc, or any other specialized behavior, it needs to be added here.
1691 //
1692 // Note that this method should only be called from the vm_thread while at a
1693 // safepoint.
1694 //
1695 // Note that the all_soft_refs_clear flag in the soft ref policy
1696 // may be true because this method can be called without intervening
1697 // activity.  For example when the heap space is tight and full measure
1698 // are being taken to free space.
1699 void PSParallelCompact::invoke(bool maximum_heap_compaction) {
1700   assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
1701   assert(Thread::current() == (Thread*)VMThread::vm_thread(),
1702          "should be in vm thread");
1703 
1704   ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
1705   GCCause::Cause gc_cause = heap->gc_cause();
1706   assert(!heap->is_gc_active(), "not reentrant");
1707 
1708   PSAdaptiveSizePolicy* policy = heap->size_policy();
1709   IsGCActiveMark mark;
1710 
1711   if (ScavengeBeforeFullGC) {
1712     PSScavenge::invoke_no_policy();
1713   }
1714 
1715   const bool clear_all_soft_refs =
1716     heap->soft_ref_policy()->should_clear_all_soft_refs();
1717 
1718   PSParallelCompact::invoke_no_policy(clear_all_soft_refs ||
1719                                       maximum_heap_compaction);
1720 }
1721 
1722 // This method contains no policy. You should probably
1723 // be calling invoke() instead.
1724 bool PSParallelCompact::invoke_no_policy(bool maximum_heap_compaction) {
1725   assert(SafepointSynchronize::is_at_safepoint(), "must be at a safepoint");
1726   assert(ref_processor() != NULL, "Sanity");
1727 
1728   if (GCLocker::check_active_before_gc()) {
1729     return false;
1730   }
1731 
1732   ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
1733 
1734   GCIdMark gc_id_mark;
1735   _gc_timer.register_gc_start();
1736   _gc_tracer.report_gc_start(heap->gc_cause(), _gc_timer.gc_start());
1737 
1738   TimeStamp marking_start;
1739   TimeStamp compaction_start;
1740   TimeStamp collection_exit;
1741 
1742   GCCause::Cause gc_cause = heap->gc_cause();
1743   PSYoungGen* young_gen = heap->young_gen();
1744   PSOldGen* old_gen = heap->old_gen();
1745   PSAdaptiveSizePolicy* size_policy = heap->size_policy();
1746 
1747   // The scope of casr should end after code that can change
1748   // SoftRefPolicy::_should_clear_all_soft_refs.
1749   ClearedAllSoftRefs casr(maximum_heap_compaction,
1750                           heap->soft_ref_policy());
1751 
1752   if (ZapUnusedHeapArea) {
1753     // Save information needed to minimize mangling
1754     heap->record_gen_tops_before_GC();
1755   }
1756 
1757   // Make sure data structures are sane, make the heap parsable, and do other
1758   // miscellaneous bookkeeping.
1759   pre_compact();
1760 
1761   const PreGenGCValues pre_gc_values = heap->get_pre_gc_values();
1762 
1763   // Get the compaction manager reserved for the VM thread.
1764   ParCompactionManager* const vmthread_cm = ParCompactionManager::get_vmthread_cm();
1765 
1766   {
1767     const uint active_workers =
1768       WorkerPolicy::calc_active_workers(ParallelScavengeHeap::heap()->workers().total_workers(),
1769                                         ParallelScavengeHeap::heap()->workers().active_workers(),
1770                                         Threads::number_of_non_daemon_threads());
1771     ParallelScavengeHeap::heap()->workers().update_active_workers(active_workers);
1772 
1773     GCTraceCPUTime tcpu;
1774     GCTraceTime(Info, gc) tm("Pause Full", NULL, gc_cause, true);
1775 
1776     heap->pre_full_gc_dump(&_gc_timer);
1777 
1778     TraceCollectorStats tcs(counters());
1779     TraceMemoryManagerStats tms(heap->old_gc_manager(), gc_cause);
1780 
1781     if (log_is_enabled(Debug, gc, heap, exit)) {
1782       accumulated_time()->start();
1783     }
1784 
1785     // Let the size policy know we're starting
1786     size_policy->major_collection_begin();
1787 
1788 #if COMPILER2_OR_JVMCI
1789     DerivedPointerTable::clear();
1790 #endif
1791 
1792     ref_processor()->start_discovery(maximum_heap_compaction);
1793 
1794     marking_start.update();
1795     marking_phase(vmthread_cm, &_gc_tracer);
1796 
1797     bool max_on_system_gc = UseMaximumCompactionOnSystemGC
1798       && GCCause::is_user_requested_gc(gc_cause);
1799     summary_phase(vmthread_cm, maximum_heap_compaction || max_on_system_gc);
1800 
1801 #if COMPILER2_OR_JVMCI
1802     assert(DerivedPointerTable::is_active(), "Sanity");
1803     DerivedPointerTable::set_active(false);
1804 #endif
1805 
1806     // adjust_roots() updates Universe::_intArrayKlassObj which is
1807     // needed by the compaction for filling holes in the dense prefix.
1808     adjust_roots();
1809 
1810     compaction_start.update();
1811     compact();
1812 
1813     ParCompactionManager::verify_all_region_stack_empty();
1814 
1815     // Reset the mark bitmap, summary data, and do other bookkeeping.  Must be
1816     // done before resizing.
1817     post_compact();
1818 
1819     // Let the size policy know we're done
1820     size_policy->major_collection_end(old_gen->used_in_bytes(), gc_cause);
1821 
1822     if (UseAdaptiveSizePolicy) {
1823       log_debug(gc, ergo)("AdaptiveSizeStart: collection: %d ", heap->total_collections());
1824       log_trace(gc, ergo)("old_gen_capacity: " SIZE_FORMAT " young_gen_capacity: " SIZE_FORMAT,
1825                           old_gen->capacity_in_bytes(), young_gen->capacity_in_bytes());
1826 
1827       // Don't check if the size_policy is ready here.  Let
1828       // the size_policy check that internally.
1829       if (UseAdaptiveGenerationSizePolicyAtMajorCollection &&
1830           AdaptiveSizePolicy::should_update_promo_stats(gc_cause)) {
1831         // Swap the survivor spaces if from_space is empty. The
1832         // resize_young_gen() called below is normally used after
1833         // a successful young GC and swapping of survivor spaces;
1834         // otherwise, it will fail to resize the young gen with
1835         // the current implementation.
1836         if (young_gen->from_space()->is_empty()) {
1837           young_gen->from_space()->clear(SpaceDecorator::Mangle);
1838           young_gen->swap_spaces();
1839         }
1840 
1841         // Calculate optimal free space amounts
1842         assert(young_gen->max_gen_size() >
1843           young_gen->from_space()->capacity_in_bytes() +
1844           young_gen->to_space()->capacity_in_bytes(),
1845           "Sizes of space in young gen are out-of-bounds");
1846 
1847         size_t young_live = young_gen->used_in_bytes();
1848         size_t eden_live = young_gen->eden_space()->used_in_bytes();
1849         size_t old_live = old_gen->used_in_bytes();
1850         size_t cur_eden = young_gen->eden_space()->capacity_in_bytes();
1851         size_t max_old_gen_size = old_gen->max_gen_size();
1852         size_t max_eden_size = young_gen->max_gen_size() -
1853           young_gen->from_space()->capacity_in_bytes() -
1854           young_gen->to_space()->capacity_in_bytes();
1855 
1856         // Used for diagnostics
1857         size_policy->clear_generation_free_space_flags();
1858 
1859         size_policy->compute_generations_free_space(young_live,
1860                                                     eden_live,
1861                                                     old_live,
1862                                                     cur_eden,
1863                                                     max_old_gen_size,
1864                                                     max_eden_size,
1865                                                     true /* full gc*/);
1866 
1867         size_policy->check_gc_overhead_limit(eden_live,
1868                                              max_old_gen_size,
1869                                              max_eden_size,
1870                                              true /* full gc*/,
1871                                              gc_cause,
1872                                              heap->soft_ref_policy());
1873 
1874         size_policy->decay_supplemental_growth(true /* full gc*/);
1875 
1876         heap->resize_old_gen(
1877           size_policy->calculated_old_free_size_in_bytes());
1878 
1879         heap->resize_young_gen(size_policy->calculated_eden_size_in_bytes(),
1880                                size_policy->calculated_survivor_size_in_bytes());
1881       }
1882 
1883       log_debug(gc, ergo)("AdaptiveSizeStop: collection: %d ", heap->total_collections());
1884     }
1885 
1886     if (UsePerfData) {
1887       PSGCAdaptivePolicyCounters* const counters = heap->gc_policy_counters();
1888       counters->update_counters();
1889       counters->update_old_capacity(old_gen->capacity_in_bytes());
1890       counters->update_young_capacity(young_gen->capacity_in_bytes());
1891     }
1892 
1893     heap->resize_all_tlabs();
1894 
1895     // Resize the metaspace capacity after a collection
1896     MetaspaceGC::compute_new_size();
1897 
1898     if (log_is_enabled(Debug, gc, heap, exit)) {
1899       accumulated_time()->stop();
1900     }
1901 
1902     heap->print_heap_change(pre_gc_values);
1903 
1904     // Track memory usage and detect low memory
1905     MemoryService::track_memory_usage();
1906     heap->update_counters();
1907 
1908     heap->post_full_gc_dump(&_gc_timer);
1909   }
1910 
1911   if (VerifyAfterGC && heap->total_collections() >= VerifyGCStartAt) {
1912     Universe::verify("After GC");
1913   }
1914 
1915   // Re-verify object start arrays
1916   if (VerifyObjectStartArray &&
1917       VerifyAfterGC) {
1918     old_gen->verify_object_start_array();
1919   }
1920 
1921   if (ZapUnusedHeapArea) {
1922     old_gen->object_space()->check_mangled_unused_area_complete();
1923   }
1924 
1925   NOT_PRODUCT(ref_processor()->verify_no_references_recorded());
1926 
1927   collection_exit.update();
1928 
1929   heap->print_heap_after_gc();
1930   heap->trace_heap_after_gc(&_gc_tracer);
1931 
1932   log_debug(gc, task, time)("VM-Thread " JLONG_FORMAT " " JLONG_FORMAT " " JLONG_FORMAT,
1933                          marking_start.ticks(), compaction_start.ticks(),
1934                          collection_exit.ticks());
1935 
1936   AdaptiveSizePolicyOutput::print(size_policy, heap->total_collections());
1937 
1938   _gc_timer.register_gc_end();
1939 
1940   _gc_tracer.report_dense_prefix(dense_prefix(old_space_id));
1941   _gc_tracer.report_gc_end(_gc_timer.gc_end(), _gc_timer.time_partitions());
1942 
1943   return true;
1944 }
1945 
1946 class PCAddThreadRootsMarkingTaskClosure : public ThreadClosure {
1947 private:
1948   uint _worker_id;
1949 
1950 public:
1951   PCAddThreadRootsMarkingTaskClosure(uint worker_id) : _worker_id(worker_id) { }
1952   void do_thread(Thread* thread) {
1953     assert(ParallelScavengeHeap::heap()->is_gc_active(), "called outside gc");
1954 
1955     ResourceMark rm;
1956 
1957     ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(_worker_id);
1958 
1959     PCMarkAndPushClosure mark_and_push_closure(cm);
1960     MarkingCodeBlobClosure mark_and_push_in_blobs(&mark_and_push_closure, !CodeBlobToOopClosure::FixRelocations);
1961 
1962     thread->oops_do(&mark_and_push_closure, &mark_and_push_in_blobs);
1963 
1964     // Do the real work
1965     cm->follow_marking_stacks();
1966   }
1967 };
1968 
1969 static void mark_from_roots_work(ParallelRootType::Value root_type, uint worker_id) {
1970   assert(ParallelScavengeHeap::heap()->is_gc_active(), "called outside gc");
1971 
1972   ParCompactionManager* cm =
1973     ParCompactionManager::gc_thread_compaction_manager(worker_id);
1974   PCMarkAndPushClosure mark_and_push_closure(cm);
1975 
1976   switch (root_type) {
1977     case ParallelRootType::class_loader_data:
1978       {
1979         CLDToOopClosure cld_closure(&mark_and_push_closure, ClassLoaderData::_claim_strong);
1980         ClassLoaderDataGraph::always_strong_cld_do(&cld_closure);
1981       }
1982       break;
1983 
1984     case ParallelRootType::code_cache:
1985       // Do not treat nmethods as strong roots for mark/sweep, since we can unload them.
1986       //ScavengableNMethods::scavengable_nmethods_do(CodeBlobToOopClosure(&mark_and_push_closure));
1987       break;
1988 
1989     case ParallelRootType::sentinel:
1990     DEBUG_ONLY(default:) // DEBUG_ONLY hack will create compile error on release builds (-Wswitch) and runtime check on debug builds
1991       fatal("Bad enumeration value: %u", root_type);
1992       break;
1993   }
1994 
1995   // Do the real work
1996   cm->follow_marking_stacks();
1997 }
1998 
1999 void steal_marking_work(TaskTerminator& terminator, uint worker_id) {
2000   assert(ParallelScavengeHeap::heap()->is_gc_active(), "called outside gc");
2001 
2002   ParCompactionManager* cm =
2003     ParCompactionManager::gc_thread_compaction_manager(worker_id);
2004 
2005   oop obj = NULL;
2006   ObjArrayTask task;
2007   do {
2008     while (ParCompactionManager::steal_objarray(worker_id,  task)) {
2009       cm->follow_array((objArrayOop)task.obj(), task.index());
2010       cm->follow_marking_stacks();
2011     }
2012     while (ParCompactionManager::steal(worker_id, obj)) {
2013       cm->follow_contents(obj);
2014       cm->follow_marking_stacks();
2015     }
2016   } while (!terminator.offer_termination());
2017 }
2018 
2019 class MarkFromRootsTask : public AbstractGangTask {
2020   StrongRootsScope _strong_roots_scope; // needed for Threads::possibly_parallel_threads_do
2021   OopStorageSetStrongParState<false /* concurrent */, false /* is_const */> _oop_storage_set_par_state;
2022   SequentialSubTasksDone _subtasks;
2023   TaskTerminator _terminator;
2024   uint _active_workers;
2025 
2026 public:
2027   MarkFromRootsTask(uint active_workers) :
2028       AbstractGangTask("MarkFromRootsTask"),
2029       _strong_roots_scope(active_workers),
2030       _subtasks(ParallelRootType::sentinel),
2031       _terminator(active_workers, ParCompactionManager::oop_task_queues()),
2032       _active_workers(active_workers) {
2033   }
2034 
2035   virtual void work(uint worker_id) {
2036     for (uint task = 0; _subtasks.try_claim_task(task); /*empty*/ ) {
2037       mark_from_roots_work(static_cast<ParallelRootType::Value>(task), worker_id);
2038     }
2039 
2040     PCAddThreadRootsMarkingTaskClosure closure(worker_id);
2041     Threads::possibly_parallel_threads_do(true /*parallel */, &closure);
2042 
2043     // Mark from OopStorages
2044     {
2045       ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(worker_id);
2046       PCMarkAndPushClosure closure(cm);
2047       _oop_storage_set_par_state.oops_do(&closure);
2048       // Do the real work
2049       cm->follow_marking_stacks();
2050     }
2051 
2052     if (_active_workers > 1) {
2053       steal_marking_work(_terminator, worker_id);
2054     }
2055   }
2056 };
2057 
2058 class ParallelCompactRefProcProxyTask : public RefProcProxyTask {
2059   TaskTerminator _terminator;
2060 
2061 public:
2062   ParallelCompactRefProcProxyTask(uint max_workers)
2063     : RefProcProxyTask("ParallelCompactRefProcProxyTask", max_workers),
2064       _terminator(_max_workers, ParCompactionManager::oop_task_queues()) {}
2065 
2066   void work(uint worker_id) override {
2067     assert(worker_id < _max_workers, "sanity");
2068     ParCompactionManager* cm = (_tm == RefProcThreadModel::Single) ? ParCompactionManager::get_vmthread_cm() : ParCompactionManager::gc_thread_compaction_manager(worker_id);
2069     PCMarkAndPushClosure keep_alive(cm);
2070     BarrierEnqueueDiscoveredFieldClosure enqueue;
2071     ParCompactionManager::FollowStackClosure complete_gc(cm, (_tm == RefProcThreadModel::Single) ? nullptr : &_terminator, worker_id);
2072     _rp_task->rp_work(worker_id, PSParallelCompact::is_alive_closure(), &keep_alive, &enqueue, &complete_gc);
2073   }
2074 
2075   void prepare_run_task_hook() override {
2076     _terminator.reset_for_reuse(_queue_count);
2077   }
2078 };
2079 
2080 void PSParallelCompact::marking_phase(ParCompactionManager* cm,
2081                                       ParallelOldTracer *gc_tracer) {
2082   // Recursively traverse all live objects and mark them
2083   GCTraceTime(Info, gc, phases) tm("Marking Phase", &_gc_timer);
2084 
2085   uint active_gc_threads = ParallelScavengeHeap::heap()->workers().active_workers();
2086 
2087   // Need new claim bits before marking starts.
2088   ClassLoaderDataGraph::clear_claimed_marks();
2089 
2090   {
2091     GCTraceTime(Debug, gc, phases) tm("Par Mark", &_gc_timer);
2092 
2093     MarkFromRootsTask task(active_gc_threads);
2094     ParallelScavengeHeap::heap()->workers().run_task(&task);
2095   }
2096 
2097   // Process reference objects found during marking
2098   {
2099     GCTraceTime(Debug, gc, phases) tm("Reference Processing", &_gc_timer);
2100 
2101     ReferenceProcessorStats stats;
2102     ReferenceProcessorPhaseTimes pt(&_gc_timer, ref_processor()->max_num_queues());
2103 
2104     ref_processor()->set_active_mt_degree(active_gc_threads);
2105     ParallelCompactRefProcProxyTask task(ref_processor()->max_num_queues());
2106     stats = ref_processor()->process_discovered_references(task, pt);
2107 
2108     gc_tracer->report_gc_reference_stats(stats);
2109     pt.print_all_references();
2110   }
2111 
2112   // This is the point where the entire marking should have completed.
2113   ParCompactionManager::verify_all_marking_stack_empty();
2114 
2115   {
2116     GCTraceTime(Debug, gc, phases) tm("Weak Processing", &_gc_timer);
2117     WeakProcessor::weak_oops_do(&ParallelScavengeHeap::heap()->workers(),
2118                                 is_alive_closure(),
2119                                 &do_nothing_cl,
2120                                 1);
2121   }
2122 
2123   {
2124     GCTraceTime(Debug, gc, phases) tm_m("Class Unloading", &_gc_timer);
2125 
2126     // Follow system dictionary roots and unload classes.
2127     bool purged_class = SystemDictionary::do_unloading(&_gc_timer);
2128 
2129     // Unload nmethods.
2130     CodeCache::do_unloading(is_alive_closure(), purged_class);
2131 
2132     // Prune dead klasses from subklass/sibling/implementor lists.
2133     Klass::clean_weak_klass_links(purged_class);
2134 
2135     // Clean JVMCI metadata handles.
2136     JVMCI_ONLY(JVMCI::do_unloading(purged_class));
2137   }
2138 
2139   _gc_tracer.report_object_count_after_gc(is_alive_closure());
2140 }
2141 
2142 #ifdef ASSERT
2143 void PCAdjustPointerClosure::verify_cm(ParCompactionManager* cm) {
2144   assert(cm != NULL, "associate ParCompactionManage should not be NULL");
2145   auto vmthread_cm = ParCompactionManager::get_vmthread_cm();
2146   if (Thread::current()->is_VM_thread()) {
2147     assert(cm == vmthread_cm, "VM threads should use ParCompactionManager from get_vmthread_cm()");
2148   } else {
2149     assert(Thread::current()->is_Worker_thread(), "Must be a GC thread");
2150     assert(cm != vmthread_cm, "GC threads should use ParCompactionManager from gc_thread_compaction_manager()");
2151   }
2152 }
2153 #endif
2154 
2155 class PSAdjustTask final : public AbstractGangTask {
2156   SubTasksDone                               _sub_tasks;
2157   WeakProcessor::Task                        _weak_proc_task;
2158   OopStorageSetStrongParState<false, false>  _oop_storage_iter;
2159   uint                                       _nworkers;
2160 
2161   enum PSAdjustSubTask {
2162     PSAdjustSubTask_code_cache,
2163     PSAdjustSubTask_old_ref_process,
2164     PSAdjustSubTask_young_ref_process,
2165 
2166     PSAdjustSubTask_num_elements
2167   };
2168 
2169 public:
2170   PSAdjustTask(uint nworkers) :
2171     AbstractGangTask("PSAdjust task"),
2172     _sub_tasks(PSAdjustSubTask_num_elements),
2173     _weak_proc_task(nworkers),
2174     _nworkers(nworkers) {
2175     // Need new claim bits when tracing through and adjusting pointers.
2176     ClassLoaderDataGraph::clear_claimed_marks();
2177     if (nworkers > 1) {
2178       Threads::change_thread_claim_token();
2179     }
2180   }
2181 
2182   ~PSAdjustTask() {
2183     Threads::assert_all_threads_claimed();
2184   }
2185 
2186   void work(uint worker_id) {
2187     ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(worker_id);
2188     PCAdjustPointerClosure adjust(cm);
2189     {
2190       ResourceMark rm;
2191       Threads::possibly_parallel_oops_do(_nworkers > 1, &adjust, nullptr);
2192     }
2193     _oop_storage_iter.oops_do(&adjust);
2194     {
2195       CLDToOopClosure cld_closure(&adjust, ClassLoaderData::_claim_strong);
2196       ClassLoaderDataGraph::cld_do(&cld_closure);
2197     }
2198     {
2199       AlwaysTrueClosure always_alive;
2200       _weak_proc_task.work(worker_id, &always_alive, &adjust);
2201     }
2202     if (_sub_tasks.try_claim_task(PSAdjustSubTask_code_cache)) {
2203       CodeBlobToOopClosure adjust_code(&adjust, CodeBlobToOopClosure::FixRelocations);
2204       CodeCache::blobs_do(&adjust_code);
2205     }
2206     if (_sub_tasks.try_claim_task(PSAdjustSubTask_old_ref_process)) {
2207       PSParallelCompact::ref_processor()->weak_oops_do(&adjust);
2208     }
2209     if (_sub_tasks.try_claim_task(PSAdjustSubTask_young_ref_process)) {
2210       // Roots were visited so references into the young gen in roots
2211       // may have been scanned.  Process them also.
2212       // Should the reference processor have a span that excludes
2213       // young gen objects?
2214       PSScavenge::reference_processor()->weak_oops_do(&adjust);
2215     }
2216     _sub_tasks.all_tasks_claimed();
2217   }
2218 };
2219 
2220 void PSParallelCompact::adjust_roots() {
2221   // Adjust the pointers to reflect the new locations
2222   GCTraceTime(Info, gc, phases) tm("Adjust Roots", &_gc_timer);
2223   uint nworkers = ParallelScavengeHeap::heap()->workers().active_workers();
2224   PSAdjustTask task(nworkers);
2225   ParallelScavengeHeap::heap()->workers().run_task(&task);
2226 }
2227 
2228 // Helper class to print 8 region numbers per line and then print the total at the end.
2229 class FillableRegionLogger : public StackObj {
2230 private:
2231   Log(gc, compaction) log;
2232   static const int LineLength = 8;
2233   size_t _regions[LineLength];
2234   int _next_index;
2235   bool _enabled;
2236   size_t _total_regions;
2237 public:
2238   FillableRegionLogger() : _next_index(0), _enabled(log_develop_is_enabled(Trace, gc, compaction)), _total_regions(0) { }
2239   ~FillableRegionLogger() {
2240     log.trace(SIZE_FORMAT " initially fillable regions", _total_regions);
2241   }
2242 
2243   void print_line() {
2244     if (!_enabled || _next_index == 0) {
2245       return;
2246     }
2247     FormatBuffer<> line("Fillable: ");
2248     for (int i = 0; i < _next_index; i++) {
2249       line.append(" " SIZE_FORMAT_W(7), _regions[i]);
2250     }
2251     log.trace("%s", line.buffer());
2252     _next_index = 0;
2253   }
2254 
2255   void handle(size_t region) {
2256     if (!_enabled) {
2257       return;
2258     }
2259     _regions[_next_index++] = region;
2260     if (_next_index == LineLength) {
2261       print_line();
2262     }
2263     _total_regions++;
2264   }
2265 };
2266 
2267 void PSParallelCompact::prepare_region_draining_tasks(uint parallel_gc_threads)
2268 {
2269   GCTraceTime(Trace, gc, phases) tm("Drain Task Setup", &_gc_timer);
2270 
2271   // Find the threads that are active
2272   uint worker_id = 0;
2273 
2274   // Find all regions that are available (can be filled immediately) and
2275   // distribute them to the thread stacks.  The iteration is done in reverse
2276   // order (high to low) so the regions will be removed in ascending order.
2277 
2278   const ParallelCompactData& sd = PSParallelCompact::summary_data();
2279 
2280   // id + 1 is used to test termination so unsigned  can
2281   // be used with an old_space_id == 0.
2282   FillableRegionLogger region_logger;
2283   for (unsigned int id = to_space_id; id + 1 > old_space_id; --id) {
2284     SpaceInfo* const space_info = _space_info + id;
2285     HeapWord* const new_top = space_info->new_top();
2286 
2287     const size_t beg_region = sd.addr_to_region_idx(space_info->dense_prefix());
2288     const size_t end_region =
2289       sd.addr_to_region_idx(sd.region_align_up(new_top));
2290 
2291     for (size_t cur = end_region - 1; cur + 1 > beg_region; --cur) {
2292       if (sd.region(cur)->claim_unsafe()) {
2293         ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(worker_id);
2294         bool result = sd.region(cur)->mark_normal();
2295         assert(result, "Must succeed at this point.");
2296         cm->region_stack()->push(cur);
2297         region_logger.handle(cur);
2298         // Assign regions to tasks in round-robin fashion.
2299         if (++worker_id == parallel_gc_threads) {
2300           worker_id = 0;
2301         }
2302       }
2303     }
2304     region_logger.print_line();
2305   }
2306 }
2307 
2308 class TaskQueue : StackObj {
2309   volatile uint _counter;
2310   uint _size;
2311   uint _insert_index;
2312   PSParallelCompact::UpdateDensePrefixTask* _backing_array;
2313 public:
2314   explicit TaskQueue(uint size) : _counter(0), _size(size), _insert_index(0), _backing_array(NULL) {
2315     _backing_array = NEW_C_HEAP_ARRAY(PSParallelCompact::UpdateDensePrefixTask, _size, mtGC);
2316   }
2317   ~TaskQueue() {
2318     assert(_counter >= _insert_index, "not all queue elements were claimed");
2319     FREE_C_HEAP_ARRAY(T, _backing_array);
2320   }
2321 
2322   void push(const PSParallelCompact::UpdateDensePrefixTask& value) {
2323     assert(_insert_index < _size, "too small backing array");
2324     _backing_array[_insert_index++] = value;
2325   }
2326 
2327   bool try_claim(PSParallelCompact::UpdateDensePrefixTask& reference) {
2328     uint claimed = Atomic::fetch_and_add(&_counter, 1u);
2329     if (claimed < _insert_index) {
2330       reference = _backing_array[claimed];
2331       return true;
2332     } else {
2333       return false;
2334     }
2335   }
2336 };
2337 
2338 #define PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING 4
2339 
2340 void PSParallelCompact::enqueue_dense_prefix_tasks(TaskQueue& task_queue,
2341                                                    uint parallel_gc_threads) {
2342   GCTraceTime(Trace, gc, phases) tm("Dense Prefix Task Setup", &_gc_timer);
2343 
2344   ParallelCompactData& sd = PSParallelCompact::summary_data();
2345 
2346   // Iterate over all the spaces adding tasks for updating
2347   // regions in the dense prefix.  Assume that 1 gc thread
2348   // will work on opening the gaps and the remaining gc threads
2349   // will work on the dense prefix.
2350   unsigned int space_id;
2351   for (space_id = old_space_id; space_id < last_space_id; ++ space_id) {
2352     HeapWord* const dense_prefix_end = _space_info[space_id].dense_prefix();
2353     const MutableSpace* const space = _space_info[space_id].space();
2354 
2355     if (dense_prefix_end == space->bottom()) {
2356       // There is no dense prefix for this space.
2357       continue;
2358     }
2359 
2360     // The dense prefix is before this region.
2361     size_t region_index_end_dense_prefix =
2362         sd.addr_to_region_idx(dense_prefix_end);
2363     RegionData* const dense_prefix_cp =
2364       sd.region(region_index_end_dense_prefix);
2365     assert(dense_prefix_end == space->end() ||
2366            dense_prefix_cp->available() ||
2367            dense_prefix_cp->claimed(),
2368            "The region after the dense prefix should always be ready to fill");
2369 
2370     size_t region_index_start = sd.addr_to_region_idx(space->bottom());
2371 
2372     // Is there dense prefix work?
2373     size_t total_dense_prefix_regions =
2374       region_index_end_dense_prefix - region_index_start;
2375     // How many regions of the dense prefix should be given to
2376     // each thread?
2377     if (total_dense_prefix_regions > 0) {
2378       uint tasks_for_dense_prefix = 1;
2379       if (total_dense_prefix_regions <=
2380           (parallel_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING)) {
2381         // Don't over partition.  This assumes that
2382         // PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING is a small integer value
2383         // so there are not many regions to process.
2384         tasks_for_dense_prefix = parallel_gc_threads;
2385       } else {
2386         // Over partition
2387         tasks_for_dense_prefix = parallel_gc_threads *
2388           PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING;
2389       }
2390       size_t regions_per_thread = total_dense_prefix_regions /
2391         tasks_for_dense_prefix;
2392       // Give each thread at least 1 region.
2393       if (regions_per_thread == 0) {
2394         regions_per_thread = 1;
2395       }
2396 
2397       for (uint k = 0; k < tasks_for_dense_prefix; k++) {
2398         if (region_index_start >= region_index_end_dense_prefix) {
2399           break;
2400         }
2401         // region_index_end is not processed
2402         size_t region_index_end = MIN2(region_index_start + regions_per_thread,
2403                                        region_index_end_dense_prefix);
2404         task_queue.push(UpdateDensePrefixTask(SpaceId(space_id),
2405                                               region_index_start,
2406                                               region_index_end));
2407         region_index_start = region_index_end;
2408       }
2409     }
2410     // This gets any part of the dense prefix that did not
2411     // fit evenly.
2412     if (region_index_start < region_index_end_dense_prefix) {
2413       task_queue.push(UpdateDensePrefixTask(SpaceId(space_id),
2414                                             region_index_start,
2415                                             region_index_end_dense_prefix));
2416     }
2417   }
2418 }
2419 
2420 #ifdef ASSERT
2421 // Write a histogram of the number of times the block table was filled for a
2422 // region.
2423 void PSParallelCompact::write_block_fill_histogram()
2424 {
2425   if (!log_develop_is_enabled(Trace, gc, compaction)) {
2426     return;
2427   }
2428 
2429   Log(gc, compaction) log;
2430   ResourceMark rm;
2431   LogStream ls(log.trace());
2432   outputStream* out = &ls;
2433 
2434   typedef ParallelCompactData::RegionData rd_t;
2435   ParallelCompactData& sd = summary_data();
2436 
2437   for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2438     MutableSpace* const spc = _space_info[id].space();
2439     if (spc->bottom() != spc->top()) {
2440       const rd_t* const beg = sd.addr_to_region_ptr(spc->bottom());
2441       HeapWord* const top_aligned_up = sd.region_align_up(spc->top());
2442       const rd_t* const end = sd.addr_to_region_ptr(top_aligned_up);
2443 
2444       size_t histo[5] = { 0, 0, 0, 0, 0 };
2445       const size_t histo_len = sizeof(histo) / sizeof(size_t);
2446       const size_t region_cnt = pointer_delta(end, beg, sizeof(rd_t));
2447 
2448       for (const rd_t* cur = beg; cur < end; ++cur) {
2449         ++histo[MIN2(cur->blocks_filled_count(), histo_len - 1)];
2450       }
2451       out->print("Block fill histogram: %u %-4s" SIZE_FORMAT_W(5), id, space_names[id], region_cnt);
2452       for (size_t i = 0; i < histo_len; ++i) {
2453         out->print(" " SIZE_FORMAT_W(5) " %5.1f%%",
2454                    histo[i], 100.0 * histo[i] / region_cnt);
2455       }
2456       out->cr();
2457     }
2458   }
2459 }
2460 #endif // #ifdef ASSERT
2461 
2462 static void compaction_with_stealing_work(TaskTerminator* terminator, uint worker_id) {
2463   assert(ParallelScavengeHeap::heap()->is_gc_active(), "called outside gc");
2464 
2465   ParCompactionManager* cm =
2466     ParCompactionManager::gc_thread_compaction_manager(worker_id);
2467 
2468   // Drain the stacks that have been preloaded with regions
2469   // that are ready to fill.
2470 
2471   cm->drain_region_stacks();
2472 
2473   guarantee(cm->region_stack()->is_empty(), "Not empty");
2474 
2475   size_t region_index = 0;
2476 
2477   while (true) {
2478     if (ParCompactionManager::steal(worker_id, region_index)) {
2479       PSParallelCompact::fill_and_update_region(cm, region_index);
2480       cm->drain_region_stacks();
2481     } else if (PSParallelCompact::steal_unavailable_region(cm, region_index)) {
2482       // Fill and update an unavailable region with the help of a shadow region
2483       PSParallelCompact::fill_and_update_shadow_region(cm, region_index);
2484       cm->drain_region_stacks();
2485     } else {
2486       if (terminator->offer_termination()) {
2487         break;
2488       }
2489       // Go around again.
2490     }
2491   }
2492 }
2493 
2494 class UpdateDensePrefixAndCompactionTask: public AbstractGangTask {
2495   TaskQueue& _tq;
2496   TaskTerminator _terminator;
2497   uint _active_workers;
2498 
2499 public:
2500   UpdateDensePrefixAndCompactionTask(TaskQueue& tq, uint active_workers) :
2501       AbstractGangTask("UpdateDensePrefixAndCompactionTask"),
2502       _tq(tq),
2503       _terminator(active_workers, ParCompactionManager::region_task_queues()),
2504       _active_workers(active_workers) {
2505   }
2506   virtual void work(uint worker_id) {
2507     ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(worker_id);
2508 
2509     for (PSParallelCompact::UpdateDensePrefixTask task; _tq.try_claim(task); /* empty */) {
2510       PSParallelCompact::update_and_deadwood_in_dense_prefix(cm,
2511                                                              task._space_id,
2512                                                              task._region_index_start,
2513                                                              task._region_index_end);
2514     }
2515 
2516     // Once a thread has drained it's stack, it should try to steal regions from
2517     // other threads.
2518     compaction_with_stealing_work(&_terminator, worker_id);
2519   }
2520 };
2521 
2522 void PSParallelCompact::compact() {
2523   GCTraceTime(Info, gc, phases) tm("Compaction Phase", &_gc_timer);
2524 
2525   ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
2526   PSOldGen* old_gen = heap->old_gen();
2527   old_gen->start_array()->reset();
2528   uint active_gc_threads = ParallelScavengeHeap::heap()->workers().active_workers();
2529 
2530   // for [0..last_space_id)
2531   //     for [0..active_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING)
2532   //         push
2533   //     push
2534   //
2535   // max push count is thus: last_space_id * (active_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING + 1)
2536   TaskQueue task_queue(last_space_id * (active_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING + 1));
2537   initialize_shadow_regions(active_gc_threads);
2538   prepare_region_draining_tasks(active_gc_threads);
2539   enqueue_dense_prefix_tasks(task_queue, active_gc_threads);
2540 
2541   {
2542     GCTraceTime(Trace, gc, phases) tm("Par Compact", &_gc_timer);
2543 
2544     UpdateDensePrefixAndCompactionTask task(task_queue, active_gc_threads);
2545     ParallelScavengeHeap::heap()->workers().run_task(&task);
2546 
2547 #ifdef  ASSERT
2548     // Verify that all regions have been processed before the deferred updates.
2549     for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2550       verify_complete(SpaceId(id));
2551     }
2552 #endif
2553   }
2554 
2555   {
2556     GCTraceTime(Trace, gc, phases) tm("Deferred Updates", &_gc_timer);
2557     // Update the deferred objects, if any. In principle, any compaction
2558     // manager can be used. However, since the current thread is VM thread, we
2559     // use the rightful one to keep the verification logic happy.
2560     ParCompactionManager* cm = ParCompactionManager::get_vmthread_cm();
2561     for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2562       update_deferred_objects(cm, SpaceId(id));
2563     }
2564   }
2565 
2566   DEBUG_ONLY(write_block_fill_histogram());
2567 }
2568 
2569 #ifdef  ASSERT
2570 void PSParallelCompact::verify_complete(SpaceId space_id) {
2571   // All Regions between space bottom() to new_top() should be marked as filled
2572   // and all Regions between new_top() and top() should be available (i.e.,
2573   // should have been emptied).
2574   ParallelCompactData& sd = summary_data();
2575   SpaceInfo si = _space_info[space_id];
2576   HeapWord* new_top_addr = sd.region_align_up(si.new_top());
2577   HeapWord* old_top_addr = sd.region_align_up(si.space()->top());
2578   const size_t beg_region = sd.addr_to_region_idx(si.space()->bottom());
2579   const size_t new_top_region = sd.addr_to_region_idx(new_top_addr);
2580   const size_t old_top_region = sd.addr_to_region_idx(old_top_addr);
2581 
2582   bool issued_a_warning = false;
2583 
2584   size_t cur_region;
2585   for (cur_region = beg_region; cur_region < new_top_region; ++cur_region) {
2586     const RegionData* const c = sd.region(cur_region);
2587     if (!c->completed()) {
2588       log_warning(gc)("region " SIZE_FORMAT " not filled: destination_count=%u",
2589                       cur_region, c->destination_count());
2590       issued_a_warning = true;
2591     }
2592   }
2593 
2594   for (cur_region = new_top_region; cur_region < old_top_region; ++cur_region) {
2595     const RegionData* const c = sd.region(cur_region);
2596     if (!c->available()) {
2597       log_warning(gc)("region " SIZE_FORMAT " not empty: destination_count=%u",
2598                       cur_region, c->destination_count());
2599       issued_a_warning = true;
2600     }
2601   }
2602 
2603   if (issued_a_warning) {
2604     print_region_ranges();
2605   }
2606 }
2607 #endif  // #ifdef ASSERT
2608 
2609 inline void UpdateOnlyClosure::do_addr(HeapWord* addr) {
2610   _start_array->allocate_block(addr);
2611   compaction_manager()->update_contents(cast_to_oop(addr));
2612 }
2613 
2614 // Update interior oops in the ranges of regions [beg_region, end_region).
2615 void
2616 PSParallelCompact::update_and_deadwood_in_dense_prefix(ParCompactionManager* cm,
2617                                                        SpaceId space_id,
2618                                                        size_t beg_region,
2619                                                        size_t end_region) {
2620   ParallelCompactData& sd = summary_data();
2621   ParMarkBitMap* const mbm = mark_bitmap();
2622 
2623   HeapWord* beg_addr = sd.region_to_addr(beg_region);
2624   HeapWord* const end_addr = sd.region_to_addr(end_region);
2625   assert(beg_region <= end_region, "bad region range");
2626   assert(end_addr <= dense_prefix(space_id), "not in the dense prefix");
2627 
2628 #ifdef  ASSERT
2629   // Claim the regions to avoid triggering an assert when they are marked as
2630   // filled.
2631   for (size_t claim_region = beg_region; claim_region < end_region; ++claim_region) {
2632     assert(sd.region(claim_region)->claim_unsafe(), "claim() failed");
2633   }
2634 #endif  // #ifdef ASSERT
2635 
2636   if (beg_addr != space(space_id)->bottom()) {
2637     // Find the first live object or block of dead space that *starts* in this
2638     // range of regions.  If a partial object crosses onto the region, skip it;
2639     // it will be marked for 'deferred update' when the object head is
2640     // processed.  If dead space crosses onto the region, it is also skipped; it
2641     // will be filled when the prior region is processed.  If neither of those
2642     // apply, the first word in the region is the start of a live object or dead
2643     // space.
2644     assert(beg_addr > space(space_id)->bottom(), "sanity");
2645     const RegionData* const cp = sd.region(beg_region);
2646     if (cp->partial_obj_size() != 0) {
2647       beg_addr = sd.partial_obj_end(beg_region);
2648     } else if (dead_space_crosses_boundary(cp, mbm->addr_to_bit(beg_addr))) {
2649       beg_addr = mbm->find_obj_beg(beg_addr, end_addr);
2650     }
2651   }
2652 
2653   if (beg_addr < end_addr) {
2654     // A live object or block of dead space starts in this range of Regions.
2655      HeapWord* const dense_prefix_end = dense_prefix(space_id);
2656 
2657     // Create closures and iterate.
2658     UpdateOnlyClosure update_closure(mbm, cm, space_id);
2659     FillClosure fill_closure(cm, space_id);
2660     ParMarkBitMap::IterationStatus status;
2661     status = mbm->iterate(&update_closure, &fill_closure, beg_addr, end_addr,
2662                           dense_prefix_end);
2663     if (status == ParMarkBitMap::incomplete) {
2664       update_closure.do_addr(update_closure.source());
2665     }
2666   }
2667 
2668   // Mark the regions as filled.
2669   RegionData* const beg_cp = sd.region(beg_region);
2670   RegionData* const end_cp = sd.region(end_region);
2671   for (RegionData* cp = beg_cp; cp < end_cp; ++cp) {
2672     cp->set_completed();
2673   }
2674 }
2675 
2676 // Return the SpaceId for the space containing addr.  If addr is not in the
2677 // heap, last_space_id is returned.  In debug mode it expects the address to be
2678 // in the heap and asserts such.
2679 PSParallelCompact::SpaceId PSParallelCompact::space_id(HeapWord* addr) {
2680   assert(ParallelScavengeHeap::heap()->is_in_reserved(addr), "addr not in the heap");
2681 
2682   for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2683     if (_space_info[id].space()->contains(addr)) {
2684       return SpaceId(id);
2685     }
2686   }
2687 
2688   assert(false, "no space contains the addr");
2689   return last_space_id;
2690 }
2691 
2692 void PSParallelCompact::update_deferred_objects(ParCompactionManager* cm,
2693                                                 SpaceId id) {
2694   assert(id < last_space_id, "bad space id");
2695 
2696   ParallelCompactData& sd = summary_data();
2697   const SpaceInfo* const space_info = _space_info + id;
2698   ObjectStartArray* const start_array = space_info->start_array();
2699 
2700   const MutableSpace* const space = space_info->space();
2701   assert(space_info->dense_prefix() >= space->bottom(), "dense_prefix not set");
2702   HeapWord* const beg_addr = space_info->dense_prefix();
2703   HeapWord* const end_addr = sd.region_align_up(space_info->new_top());
2704 
2705   const RegionData* const beg_region = sd.addr_to_region_ptr(beg_addr);
2706   const RegionData* const end_region = sd.addr_to_region_ptr(end_addr);
2707   const RegionData* cur_region;
2708   for (cur_region = beg_region; cur_region < end_region; ++cur_region) {
2709     HeapWord* const addr = cur_region->deferred_obj_addr();
2710     if (addr != NULL) {
2711       if (start_array != NULL) {
2712         start_array->allocate_block(addr);
2713       }
2714       cm->update_contents(cast_to_oop(addr));
2715       assert(oopDesc::is_oop_or_null(cast_to_oop(addr)), "Expected an oop or NULL at " PTR_FORMAT, p2i(cast_to_oop(addr)));
2716     }
2717   }
2718 }
2719 
2720 // Skip over count live words starting from beg, and return the address of the
2721 // next live word.  Unless marked, the word corresponding to beg is assumed to
2722 // be dead.  Callers must either ensure beg does not correspond to the middle of
2723 // an object, or account for those live words in some other way.  Callers must
2724 // also ensure that there are enough live words in the range [beg, end) to skip.
2725 HeapWord*
2726 PSParallelCompact::skip_live_words(HeapWord* beg, HeapWord* end, size_t count)
2727 {
2728   assert(count > 0, "sanity");
2729 
2730   ParMarkBitMap* m = mark_bitmap();
2731   idx_t bits_to_skip = m->words_to_bits(count);
2732   idx_t cur_beg = m->addr_to_bit(beg);
2733   const idx_t search_end = m->align_range_end(m->addr_to_bit(end));
2734 
2735   do {
2736     cur_beg = m->find_obj_beg(cur_beg, search_end);
2737     idx_t cur_end = m->find_obj_end(cur_beg, search_end);
2738     const size_t obj_bits = cur_end - cur_beg + 1;
2739     if (obj_bits > bits_to_skip) {
2740       return m->bit_to_addr(cur_beg + bits_to_skip);
2741     }
2742     bits_to_skip -= obj_bits;
2743     cur_beg = cur_end + 1;
2744   } while (bits_to_skip > 0);
2745 
2746   // Skipping the desired number of words landed just past the end of an object.
2747   // Find the start of the next object.
2748   cur_beg = m->find_obj_beg(cur_beg, search_end);
2749   assert(cur_beg < m->addr_to_bit(end), "not enough live words to skip");
2750   return m->bit_to_addr(cur_beg);
2751 }
2752 
2753 HeapWord* PSParallelCompact::first_src_addr(HeapWord* const dest_addr,
2754                                             SpaceId src_space_id,
2755                                             size_t src_region_idx)
2756 {
2757   assert(summary_data().is_region_aligned(dest_addr), "not aligned");
2758 
2759   const SplitInfo& split_info = _space_info[src_space_id].split_info();
2760   if (split_info.dest_region_addr() == dest_addr) {
2761     // The partial object ending at the split point contains the first word to
2762     // be copied to dest_addr.
2763     return split_info.first_src_addr();
2764   }
2765 
2766   const ParallelCompactData& sd = summary_data();
2767   ParMarkBitMap* const bitmap = mark_bitmap();
2768   const size_t RegionSize = ParallelCompactData::RegionSize;
2769 
2770   assert(sd.is_region_aligned(dest_addr), "not aligned");
2771   const RegionData* const src_region_ptr = sd.region(src_region_idx);
2772   const size_t partial_obj_size = src_region_ptr->partial_obj_size();
2773   HeapWord* const src_region_destination = src_region_ptr->destination();
2774 
2775   assert(dest_addr >= src_region_destination, "wrong src region");
2776   assert(src_region_ptr->data_size() > 0, "src region cannot be empty");
2777 
2778   HeapWord* const src_region_beg = sd.region_to_addr(src_region_idx);
2779   HeapWord* const src_region_end = src_region_beg + RegionSize;
2780 
2781   HeapWord* addr = src_region_beg;
2782   if (dest_addr == src_region_destination) {
2783     // Return the first live word in the source region.
2784     if (partial_obj_size == 0) {
2785       addr = bitmap->find_obj_beg(addr, src_region_end);
2786       assert(addr < src_region_end, "no objects start in src region");
2787     }
2788     return addr;
2789   }
2790 
2791   // Must skip some live data.
2792   size_t words_to_skip = dest_addr - src_region_destination;
2793   assert(src_region_ptr->data_size() > words_to_skip, "wrong src region");
2794 
2795   if (partial_obj_size >= words_to_skip) {
2796     // All the live words to skip are part of the partial object.
2797     addr += words_to_skip;
2798     if (partial_obj_size == words_to_skip) {
2799       // Find the first live word past the partial object.
2800       addr = bitmap->find_obj_beg(addr, src_region_end);
2801       assert(addr < src_region_end, "wrong src region");
2802     }
2803     return addr;
2804   }
2805 
2806   // Skip over the partial object (if any).
2807   if (partial_obj_size != 0) {
2808     words_to_skip -= partial_obj_size;
2809     addr += partial_obj_size;
2810   }
2811 
2812   // Skip over live words due to objects that start in the region.
2813   addr = skip_live_words(addr, src_region_end, words_to_skip);
2814   assert(addr < src_region_end, "wrong src region");
2815   return addr;
2816 }
2817 
2818 void PSParallelCompact::decrement_destination_counts(ParCompactionManager* cm,
2819                                                      SpaceId src_space_id,
2820                                                      size_t beg_region,
2821                                                      HeapWord* end_addr)
2822 {
2823   ParallelCompactData& sd = summary_data();
2824 
2825 #ifdef ASSERT
2826   MutableSpace* const src_space = _space_info[src_space_id].space();
2827   HeapWord* const beg_addr = sd.region_to_addr(beg_region);
2828   assert(src_space->contains(beg_addr) || beg_addr == src_space->end(),
2829          "src_space_id does not match beg_addr");
2830   assert(src_space->contains(end_addr) || end_addr == src_space->end(),
2831          "src_space_id does not match end_addr");
2832 #endif // #ifdef ASSERT
2833 
2834   RegionData* const beg = sd.region(beg_region);
2835   RegionData* const end = sd.addr_to_region_ptr(sd.region_align_up(end_addr));
2836 
2837   // Regions up to new_top() are enqueued if they become available.
2838   HeapWord* const new_top = _space_info[src_space_id].new_top();
2839   RegionData* const enqueue_end =
2840     sd.addr_to_region_ptr(sd.region_align_up(new_top));
2841 
2842   for (RegionData* cur = beg; cur < end; ++cur) {
2843     assert(cur->data_size() > 0, "region must have live data");
2844     cur->decrement_destination_count();
2845     if (cur < enqueue_end && cur->available() && cur->claim()) {
2846       if (cur->mark_normal()) {
2847         cm->push_region(sd.region(cur));
2848       } else if (cur->mark_copied()) {
2849         // Try to copy the content of the shadow region back to its corresponding
2850         // heap region if the shadow region is filled. Otherwise, the GC thread
2851         // fills the shadow region will copy the data back (see
2852         // MoveAndUpdateShadowClosure::complete_region).
2853         copy_back(sd.region_to_addr(cur->shadow_region()), sd.region_to_addr(cur));
2854         ParCompactionManager::push_shadow_region_mt_safe(cur->shadow_region());
2855         cur->set_completed();
2856       }
2857     }
2858   }
2859 }
2860 
2861 size_t PSParallelCompact::next_src_region(MoveAndUpdateClosure& closure,
2862                                           SpaceId& src_space_id,
2863                                           HeapWord*& src_space_top,
2864                                           HeapWord* end_addr)
2865 {
2866   typedef ParallelCompactData::RegionData RegionData;
2867 
2868   ParallelCompactData& sd = PSParallelCompact::summary_data();
2869   const size_t region_size = ParallelCompactData::RegionSize;
2870 
2871   size_t src_region_idx = 0;
2872 
2873   // Skip empty regions (if any) up to the top of the space.
2874   HeapWord* const src_aligned_up = sd.region_align_up(end_addr);
2875   RegionData* src_region_ptr = sd.addr_to_region_ptr(src_aligned_up);
2876   HeapWord* const top_aligned_up = sd.region_align_up(src_space_top);
2877   const RegionData* const top_region_ptr =
2878     sd.addr_to_region_ptr(top_aligned_up);
2879   while (src_region_ptr < top_region_ptr && src_region_ptr->data_size() == 0) {
2880     ++src_region_ptr;
2881   }
2882 
2883   if (src_region_ptr < top_region_ptr) {
2884     // The next source region is in the current space.  Update src_region_idx
2885     // and the source address to match src_region_ptr.
2886     src_region_idx = sd.region(src_region_ptr);
2887     HeapWord* const src_region_addr = sd.region_to_addr(src_region_idx);
2888     if (src_region_addr > closure.source()) {
2889       closure.set_source(src_region_addr);
2890     }
2891     return src_region_idx;
2892   }
2893 
2894   // Switch to a new source space and find the first non-empty region.
2895   unsigned int space_id = src_space_id + 1;
2896   assert(space_id < last_space_id, "not enough spaces");
2897 
2898   HeapWord* const destination = closure.destination();
2899 
2900   do {
2901     MutableSpace* space = _space_info[space_id].space();
2902     HeapWord* const bottom = space->bottom();
2903     const RegionData* const bottom_cp = sd.addr_to_region_ptr(bottom);
2904 
2905     // Iterate over the spaces that do not compact into themselves.
2906     if (bottom_cp->destination() != bottom) {
2907       HeapWord* const top_aligned_up = sd.region_align_up(space->top());
2908       const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up);
2909 
2910       for (const RegionData* src_cp = bottom_cp; src_cp < top_cp; ++src_cp) {
2911         if (src_cp->live_obj_size() > 0) {
2912           // Found it.
2913           assert(src_cp->destination() == destination,
2914                  "first live obj in the space must match the destination");
2915           assert(src_cp->partial_obj_size() == 0,
2916                  "a space cannot begin with a partial obj");
2917 
2918           src_space_id = SpaceId(space_id);
2919           src_space_top = space->top();
2920           const size_t src_region_idx = sd.region(src_cp);
2921           closure.set_source(sd.region_to_addr(src_region_idx));
2922           return src_region_idx;
2923         } else {
2924           assert(src_cp->data_size() == 0, "sanity");
2925         }
2926       }
2927     }
2928   } while (++space_id < last_space_id);
2929 
2930   assert(false, "no source region was found");
2931   return 0;
2932 }
2933 
2934 void PSParallelCompact::fill_region(ParCompactionManager* cm, MoveAndUpdateClosure& closure, size_t region_idx)
2935 {
2936   typedef ParMarkBitMap::IterationStatus IterationStatus;
2937   ParMarkBitMap* const bitmap = mark_bitmap();
2938   ParallelCompactData& sd = summary_data();
2939   RegionData* const region_ptr = sd.region(region_idx);
2940 
2941   // Get the source region and related info.
2942   size_t src_region_idx = region_ptr->source_region();
2943   SpaceId src_space_id = space_id(sd.region_to_addr(src_region_idx));
2944   HeapWord* src_space_top = _space_info[src_space_id].space()->top();
2945   HeapWord* dest_addr = sd.region_to_addr(region_idx);
2946 
2947   closure.set_source(first_src_addr(dest_addr, src_space_id, src_region_idx));
2948 
2949   // Adjust src_region_idx to prepare for decrementing destination counts (the
2950   // destination count is not decremented when a region is copied to itself).
2951   if (src_region_idx == region_idx) {
2952     src_region_idx += 1;
2953   }
2954 
2955   if (bitmap->is_unmarked(closure.source())) {
2956     // The first source word is in the middle of an object; copy the remainder
2957     // of the object or as much as will fit.  The fact that pointer updates were
2958     // deferred will be noted when the object header is processed.
2959     HeapWord* const old_src_addr = closure.source();
2960     closure.copy_partial_obj();
2961     if (closure.is_full()) {
2962       decrement_destination_counts(cm, src_space_id, src_region_idx,
2963                                    closure.source());
2964       region_ptr->set_deferred_obj_addr(NULL);
2965       closure.complete_region(cm, dest_addr, region_ptr);
2966       return;
2967     }
2968 
2969     HeapWord* const end_addr = sd.region_align_down(closure.source());
2970     if (sd.region_align_down(old_src_addr) != end_addr) {
2971       // The partial object was copied from more than one source region.
2972       decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr);
2973 
2974       // Move to the next source region, possibly switching spaces as well.  All
2975       // args except end_addr may be modified.
2976       src_region_idx = next_src_region(closure, src_space_id, src_space_top,
2977                                        end_addr);
2978     }
2979   }
2980 
2981   do {
2982     HeapWord* const cur_addr = closure.source();
2983     HeapWord* const end_addr = MIN2(sd.region_align_up(cur_addr + 1),
2984                                     src_space_top);
2985     IterationStatus status = bitmap->iterate(&closure, cur_addr, end_addr);
2986 
2987     if (status == ParMarkBitMap::incomplete) {
2988       // The last obj that starts in the source region does not end in the
2989       // region.
2990       assert(closure.source() < end_addr, "sanity");
2991       HeapWord* const obj_beg = closure.source();
2992       HeapWord* const range_end = MIN2(obj_beg + closure.words_remaining(),
2993                                        src_space_top);
2994       HeapWord* const obj_end = bitmap->find_obj_end(obj_beg, range_end);
2995       if (obj_end < range_end) {
2996         // The end was found; the entire object will fit.
2997         status = closure.do_addr(obj_beg, bitmap->obj_size(obj_beg, obj_end));
2998         assert(status != ParMarkBitMap::would_overflow, "sanity");
2999       } else {
3000         // The end was not found; the object will not fit.
3001         assert(range_end < src_space_top, "obj cannot cross space boundary");
3002         status = ParMarkBitMap::would_overflow;
3003       }
3004     }
3005 
3006     if (status == ParMarkBitMap::would_overflow) {
3007       // The last object did not fit.  Note that interior oop updates were
3008       // deferred, then copy enough of the object to fill the region.
3009       region_ptr->set_deferred_obj_addr(closure.destination());
3010       status = closure.copy_until_full(); // copies from closure.source()
3011 
3012       decrement_destination_counts(cm, src_space_id, src_region_idx,
3013                                    closure.source());
3014       closure.complete_region(cm, dest_addr, region_ptr);
3015       return;
3016     }
3017 
3018     if (status == ParMarkBitMap::full) {
3019       decrement_destination_counts(cm, src_space_id, src_region_idx,
3020                                    closure.source());
3021       region_ptr->set_deferred_obj_addr(NULL);
3022       closure.complete_region(cm, dest_addr, region_ptr);
3023       return;
3024     }
3025 
3026     decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr);
3027 
3028     // Move to the next source region, possibly switching spaces as well.  All
3029     // args except end_addr may be modified.
3030     src_region_idx = next_src_region(closure, src_space_id, src_space_top,
3031                                      end_addr);
3032   } while (true);
3033 }
3034 
3035 void PSParallelCompact::fill_and_update_region(ParCompactionManager* cm, size_t region_idx)
3036 {
3037   MoveAndUpdateClosure cl(mark_bitmap(), cm, region_idx);
3038   fill_region(cm, cl, region_idx);
3039 }
3040 
3041 void PSParallelCompact::fill_and_update_shadow_region(ParCompactionManager* cm, size_t region_idx)
3042 {
3043   // Get a shadow region first
3044   ParallelCompactData& sd = summary_data();
3045   RegionData* const region_ptr = sd.region(region_idx);
3046   size_t shadow_region = ParCompactionManager::pop_shadow_region_mt_safe(region_ptr);
3047   // The InvalidShadow return value indicates the corresponding heap region is available,
3048   // so use MoveAndUpdateClosure to fill the normal region. Otherwise, use
3049   // MoveAndUpdateShadowClosure to fill the acquired shadow region.
3050   if (shadow_region == ParCompactionManager::InvalidShadow) {
3051     MoveAndUpdateClosure cl(mark_bitmap(), cm, region_idx);
3052     region_ptr->shadow_to_normal();
3053     return fill_region(cm, cl, region_idx);
3054   } else {
3055     MoveAndUpdateShadowClosure cl(mark_bitmap(), cm, region_idx, shadow_region);
3056     return fill_region(cm, cl, region_idx);
3057   }
3058 }
3059 
3060 void PSParallelCompact::copy_back(HeapWord *shadow_addr, HeapWord *region_addr)
3061 {
3062   Copy::aligned_conjoint_words(shadow_addr, region_addr, _summary_data.RegionSize);
3063 }
3064 
3065 bool PSParallelCompact::steal_unavailable_region(ParCompactionManager* cm, size_t &region_idx)
3066 {
3067   size_t next = cm->next_shadow_region();
3068   ParallelCompactData& sd = summary_data();
3069   size_t old_new_top = sd.addr_to_region_idx(_space_info[old_space_id].new_top());
3070   uint active_gc_threads = ParallelScavengeHeap::heap()->workers().active_workers();
3071 
3072   while (next < old_new_top) {
3073     if (sd.region(next)->mark_shadow()) {
3074       region_idx = next;
3075       return true;
3076     }
3077     next = cm->move_next_shadow_region_by(active_gc_threads);
3078   }
3079 
3080   return false;
3081 }
3082 
3083 // The shadow region is an optimization to address region dependencies in full GC. The basic
3084 // idea is making more regions available by temporally storing their live objects in empty
3085 // shadow regions to resolve dependencies between them and the destination regions. Therefore,
3086 // GC threads need not wait destination regions to be available before processing sources.
3087 //
3088 // A typical workflow would be:
3089 // After draining its own stack and failing to steal from others, a GC worker would pick an
3090 // unavailable region (destination count > 0) and get a shadow region. Then the worker fills
3091 // the shadow region by copying live objects from source regions of the unavailable one. Once
3092 // the unavailable region becomes available, the data in the shadow region will be copied back.
3093 // Shadow regions are empty regions in the to-space and regions between top and end of other spaces.
3094 void PSParallelCompact::initialize_shadow_regions(uint parallel_gc_threads)
3095 {
3096   const ParallelCompactData& sd = PSParallelCompact::summary_data();
3097 
3098   for (unsigned int id = old_space_id; id < last_space_id; ++id) {
3099     SpaceInfo* const space_info = _space_info + id;
3100     MutableSpace* const space = space_info->space();
3101 
3102     const size_t beg_region =
3103       sd.addr_to_region_idx(sd.region_align_up(MAX2(space_info->new_top(), space->top())));
3104     const size_t end_region =
3105       sd.addr_to_region_idx(sd.region_align_down(space->end()));
3106 
3107     for (size_t cur = beg_region; cur < end_region; ++cur) {
3108       ParCompactionManager::push_shadow_region(cur);
3109     }
3110   }
3111 
3112   size_t beg_region = sd.addr_to_region_idx(_space_info[old_space_id].dense_prefix());
3113   for (uint i = 0; i < parallel_gc_threads; i++) {
3114     ParCompactionManager *cm = ParCompactionManager::gc_thread_compaction_manager(i);
3115     cm->set_next_shadow_region(beg_region + i);
3116   }
3117 }
3118 
3119 void PSParallelCompact::fill_blocks(size_t region_idx)
3120 {
3121   // Fill in the block table elements for the specified region.  Each block
3122   // table element holds the number of live words in the region that are to the
3123   // left of the first object that starts in the block.  Thus only blocks in
3124   // which an object starts need to be filled.
3125   //
3126   // The algorithm scans the section of the bitmap that corresponds to the
3127   // region, keeping a running total of the live words.  When an object start is
3128   // found, if it's the first to start in the block that contains it, the
3129   // current total is written to the block table element.
3130   const size_t Log2BlockSize = ParallelCompactData::Log2BlockSize;
3131   const size_t Log2RegionSize = ParallelCompactData::Log2RegionSize;
3132   const size_t RegionSize = ParallelCompactData::RegionSize;
3133 
3134   ParallelCompactData& sd = summary_data();
3135   const size_t partial_obj_size = sd.region(region_idx)->partial_obj_size();
3136   if (partial_obj_size >= RegionSize) {
3137     return; // No objects start in this region.
3138   }
3139 
3140   // Ensure the first loop iteration decides that the block has changed.
3141   size_t cur_block = sd.block_count();
3142 
3143   const ParMarkBitMap* const bitmap = mark_bitmap();
3144 
3145   const size_t Log2BitsPerBlock = Log2BlockSize - LogMinObjAlignment;
3146   assert((size_t)1 << Log2BitsPerBlock ==
3147          bitmap->words_to_bits(ParallelCompactData::BlockSize), "sanity");
3148 
3149   size_t beg_bit = bitmap->words_to_bits(region_idx << Log2RegionSize);
3150   const size_t range_end = beg_bit + bitmap->words_to_bits(RegionSize);
3151   size_t live_bits = bitmap->words_to_bits(partial_obj_size);
3152   beg_bit = bitmap->find_obj_beg(beg_bit + live_bits, range_end);
3153   while (beg_bit < range_end) {
3154     const size_t new_block = beg_bit >> Log2BitsPerBlock;
3155     if (new_block != cur_block) {
3156       cur_block = new_block;
3157       sd.block(cur_block)->set_offset(bitmap->bits_to_words(live_bits));
3158     }
3159 
3160     const size_t end_bit = bitmap->find_obj_end(beg_bit, range_end);
3161     if (end_bit < range_end - 1) {
3162       live_bits += end_bit - beg_bit + 1;
3163       beg_bit = bitmap->find_obj_beg(end_bit + 1, range_end);
3164     } else {
3165       return;
3166     }
3167   }
3168 }
3169 
3170 ParMarkBitMap::IterationStatus MoveAndUpdateClosure::copy_until_full()
3171 {
3172   if (source() != copy_destination()) {
3173     DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3174     Copy::aligned_conjoint_words(source(), copy_destination(), words_remaining());
3175   }
3176   update_state(words_remaining());
3177   assert(is_full(), "sanity");
3178   return ParMarkBitMap::full;
3179 }
3180 
3181 void MoveAndUpdateClosure::copy_partial_obj()
3182 {
3183   size_t words = words_remaining();
3184 
3185   HeapWord* const range_end = MIN2(source() + words, bitmap()->region_end());
3186   HeapWord* const end_addr = bitmap()->find_obj_end(source(), range_end);
3187   if (end_addr < range_end) {
3188     words = bitmap()->obj_size(source(), end_addr);
3189   }
3190 
3191   // This test is necessary; if omitted, the pointer updates to a partial object
3192   // that crosses the dense prefix boundary could be overwritten.
3193   if (source() != copy_destination()) {
3194     DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3195     Copy::aligned_conjoint_words(source(), copy_destination(), words);
3196   }
3197   update_state(words);
3198 }
3199 
3200 void MoveAndUpdateClosure::complete_region(ParCompactionManager *cm, HeapWord *dest_addr,
3201                                            PSParallelCompact::RegionData *region_ptr) {
3202   assert(region_ptr->shadow_state() == ParallelCompactData::RegionData::NormalRegion, "Region should be finished");
3203   region_ptr->set_completed();
3204 }
3205 
3206 ParMarkBitMapClosure::IterationStatus
3207 MoveAndUpdateClosure::do_addr(HeapWord* addr, size_t words) {
3208   assert(destination() != NULL, "sanity");
3209   assert(bitmap()->obj_size(addr) == words, "bad size");
3210 
3211   _source = addr;
3212   assert(PSParallelCompact::summary_data().calc_new_pointer(source(), compaction_manager()) ==
3213          destination(), "wrong destination");
3214 
3215   if (words > words_remaining()) {
3216     return ParMarkBitMap::would_overflow;
3217   }
3218 
3219   // The start_array must be updated even if the object is not moving.
3220   if (_start_array != NULL) {
3221     _start_array->allocate_block(destination());
3222   }
3223 
3224   if (copy_destination() != source()) {
3225     DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3226     Copy::aligned_conjoint_words(source(), copy_destination(), words);
3227   }
3228 
3229   oop moved_oop = cast_to_oop(copy_destination());
3230   compaction_manager()->update_contents(moved_oop);
3231   assert(oopDesc::is_oop_or_null(moved_oop), "Expected an oop or NULL at " PTR_FORMAT, p2i(moved_oop));
3232 
3233   update_state(words);
3234   assert(copy_destination() == cast_from_oop<HeapWord*>(moved_oop) + moved_oop->size(), "sanity");
3235   return is_full() ? ParMarkBitMap::full : ParMarkBitMap::incomplete;
3236 }
3237 
3238 void MoveAndUpdateShadowClosure::complete_region(ParCompactionManager *cm, HeapWord *dest_addr,
3239                                                  PSParallelCompact::RegionData *region_ptr) {
3240   assert(region_ptr->shadow_state() == ParallelCompactData::RegionData::ShadowRegion, "Region should be shadow");
3241   // Record the shadow region index
3242   region_ptr->set_shadow_region(_shadow);
3243   // Mark the shadow region as filled to indicate the data is ready to be
3244   // copied back
3245   region_ptr->mark_filled();
3246   // Try to copy the content of the shadow region back to its corresponding
3247   // heap region if available; the GC thread that decreases the destination
3248   // count to zero will do the copying otherwise (see
3249   // PSParallelCompact::decrement_destination_counts).
3250   if (((region_ptr->available() && region_ptr->claim()) || region_ptr->claimed()) && region_ptr->mark_copied()) {
3251     region_ptr->set_completed();
3252     PSParallelCompact::copy_back(PSParallelCompact::summary_data().region_to_addr(_shadow), dest_addr);
3253     ParCompactionManager::push_shadow_region_mt_safe(_shadow);
3254   }
3255 }
3256 
3257 UpdateOnlyClosure::UpdateOnlyClosure(ParMarkBitMap* mbm,
3258                                      ParCompactionManager* cm,
3259                                      PSParallelCompact::SpaceId space_id) :
3260   ParMarkBitMapClosure(mbm, cm),
3261   _space_id(space_id),
3262   _start_array(PSParallelCompact::start_array(space_id))
3263 {
3264 }
3265 
3266 // Updates the references in the object to their new values.
3267 ParMarkBitMapClosure::IterationStatus
3268 UpdateOnlyClosure::do_addr(HeapWord* addr, size_t words) {
3269   do_addr(addr);
3270   return ParMarkBitMap::incomplete;
3271 }
3272 
3273 FillClosure::FillClosure(ParCompactionManager* cm, PSParallelCompact::SpaceId space_id) :
3274   ParMarkBitMapClosure(PSParallelCompact::mark_bitmap(), cm),
3275   _start_array(PSParallelCompact::start_array(space_id))
3276 {
3277   assert(space_id == PSParallelCompact::old_space_id,
3278          "cannot use FillClosure in the young gen");
3279 }
3280 
3281 ParMarkBitMapClosure::IterationStatus
3282 FillClosure::do_addr(HeapWord* addr, size_t size) {
3283   CollectedHeap::fill_with_objects(addr, size);
3284   HeapWord* const end = addr + size;
3285   do {
3286     _start_array->allocate_block(addr);
3287     addr += cast_to_oop(addr)->size();
3288   } while (addr < end);
3289   return ParMarkBitMap::incomplete;
3290 }