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   CodeCache::increment_marking_cycle();
 989 
 990   // We need to track unique mark sweep invocations as well.
 991   _total_invocations++;
 992 
 993   heap->print_heap_before_gc();
 994   heap->trace_heap_before_gc(&_gc_tracer);
 995 
 996   // Fill in TLABs
 997   heap->ensure_parsability(true);  // retire TLABs
 998 
 999   if (VerifyBeforeGC && heap->total_collections() >= VerifyGCStartAt) {
1000     Universe::verify("Before GC");
1001   }
1002 
1003   // Verify object start arrays
1004   if (VerifyObjectStartArray &&
1005       VerifyBeforeGC) {
1006     heap->old_gen()->verify_object_start_array();
1007   }
1008 
1009   DEBUG_ONLY(mark_bitmap()->verify_clear();)
1010   DEBUG_ONLY(summary_data().verify_clear();)
1011 
1012   ParCompactionManager::reset_all_bitmap_query_caches();
1013 }
1014 
1015 void PSParallelCompact::post_compact()
1016 {
1017   GCTraceTime(Info, gc, phases) tm("Post Compact", &_gc_timer);
1018   ParCompactionManager::remove_all_shadow_regions();
1019 
1020   CodeCache::increment_marking_cycle();
1021 
1022   for (unsigned int id = old_space_id; id < last_space_id; ++id) {
1023     // Clear the marking bitmap, summary data and split info.
1024     clear_data_covering_space(SpaceId(id));
1025     // Update top().  Must be done after clearing the bitmap and summary data.
1026     _space_info[id].publish_new_top();
1027   }
1028 
1029   ParCompactionManager::flush_all_string_dedup_requests();
1030 
1031   MutableSpace* const eden_space = _space_info[eden_space_id].space();
1032   MutableSpace* const from_space = _space_info[from_space_id].space();
1033   MutableSpace* const to_space   = _space_info[to_space_id].space();
1034 
1035   ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
1036   bool eden_empty = eden_space->is_empty();
1037 
1038   // Update heap occupancy information which is used as input to the soft ref
1039   // clearing policy at the next gc.
1040   Universe::heap()->update_capacity_and_used_at_gc();
1041 
1042   bool young_gen_empty = eden_empty && from_space->is_empty() &&
1043     to_space->is_empty();
1044 
1045   PSCardTable* ct = heap->card_table();
1046   MemRegion old_mr = heap->old_gen()->reserved();
1047   if (young_gen_empty) {
1048     ct->clear(MemRegion(old_mr.start(), old_mr.end()));
1049   } else {
1050     ct->invalidate(MemRegion(old_mr.start(), old_mr.end()));
1051   }
1052 
1053   // Delete metaspaces for unloaded class loaders and clean up loader_data graph
1054   ClassLoaderDataGraph::purge(/*at_safepoint*/true);
1055   DEBUG_ONLY(MetaspaceUtils::verify();)
1056 
1057   heap->prune_scavengable_nmethods();
1058 
1059 #if COMPILER2_OR_JVMCI
1060   DerivedPointerTable::update_pointers();
1061 #endif
1062 
1063   if (ZapUnusedHeapArea) {
1064     heap->gen_mangle_unused_area();
1065   }
1066 
1067   // Signal that we have completed a visit to all live objects.
1068   Universe::heap()->record_whole_heap_examined_timestamp();
1069 }
1070 
1071 HeapWord*
1072 PSParallelCompact::compute_dense_prefix_via_density(const SpaceId id,
1073                                                     bool maximum_compaction)
1074 {
1075   const size_t region_size = ParallelCompactData::RegionSize;
1076   const ParallelCompactData& sd = summary_data();
1077 
1078   const MutableSpace* const space = _space_info[id].space();
1079   HeapWord* const top_aligned_up = sd.region_align_up(space->top());
1080   const RegionData* const beg_cp = sd.addr_to_region_ptr(space->bottom());
1081   const RegionData* const end_cp = sd.addr_to_region_ptr(top_aligned_up);
1082 
1083   // Skip full regions at the beginning of the space--they are necessarily part
1084   // of the dense prefix.
1085   size_t full_count = 0;
1086   const RegionData* cp;
1087   for (cp = beg_cp; cp < end_cp && cp->data_size() == region_size; ++cp) {
1088     ++full_count;
1089   }
1090 
1091   assert(total_invocations() >= _maximum_compaction_gc_num, "sanity");
1092   const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num;
1093   const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval;
1094   if (maximum_compaction || cp == end_cp || interval_ended) {
1095     _maximum_compaction_gc_num = total_invocations();
1096     return sd.region_to_addr(cp);
1097   }
1098 
1099   HeapWord* const new_top = _space_info[id].new_top();
1100   const size_t space_live = pointer_delta(new_top, space->bottom());
1101   const size_t space_used = space->used_in_words();
1102   const size_t space_capacity = space->capacity_in_words();
1103 
1104   const double cur_density = double(space_live) / space_capacity;
1105   const double deadwood_density =
1106     (1.0 - cur_density) * (1.0 - cur_density) * cur_density * cur_density;
1107   const size_t deadwood_goal = size_t(space_capacity * deadwood_density);
1108 
1109   log_develop_debug(gc, compaction)(
1110       "cur_dens=%5.3f dw_dens=%5.3f dw_goal=" SIZE_FORMAT,
1111       cur_density, deadwood_density, deadwood_goal);
1112   log_develop_debug(gc, compaction)(
1113       "space_live=" SIZE_FORMAT " space_used=" SIZE_FORMAT " "
1114       "space_cap=" SIZE_FORMAT,
1115       space_live, space_used,
1116       space_capacity);
1117 
1118   // XXX - Use binary search?
1119   HeapWord* dense_prefix = sd.region_to_addr(cp);
1120   const RegionData* full_cp = cp;
1121   const RegionData* const top_cp = sd.addr_to_region_ptr(space->top() - 1);
1122   while (cp < end_cp) {
1123     HeapWord* region_destination = cp->destination();
1124     const size_t cur_deadwood = pointer_delta(dense_prefix, region_destination);
1125 
1126     log_develop_trace(gc, compaction)(
1127         "c#=" SIZE_FORMAT_W(4) " dst=" PTR_FORMAT " "
1128         "dp=" PTR_FORMAT " cdw=" SIZE_FORMAT_W(8),
1129         sd.region(cp), p2i(region_destination),
1130         p2i(dense_prefix), cur_deadwood);
1131 
1132     if (cur_deadwood >= deadwood_goal) {
1133       // Found the region that has the correct amount of deadwood to the left.
1134       // This typically occurs after crossing a fairly sparse set of regions, so
1135       // iterate backwards over those sparse regions, looking for the region
1136       // that has the lowest density of live objects 'to the right.'
1137       size_t space_to_left = sd.region(cp) * region_size;
1138       size_t live_to_left = space_to_left - cur_deadwood;
1139       size_t space_to_right = space_capacity - space_to_left;
1140       size_t live_to_right = space_live - live_to_left;
1141       double density_to_right = double(live_to_right) / space_to_right;
1142       while (cp > full_cp) {
1143         --cp;
1144         const size_t prev_region_live_to_right = live_to_right -
1145           cp->data_size();
1146         const size_t prev_region_space_to_right = space_to_right + region_size;
1147         double prev_region_density_to_right =
1148           double(prev_region_live_to_right) / prev_region_space_to_right;
1149         if (density_to_right <= prev_region_density_to_right) {
1150           return dense_prefix;
1151         }
1152 
1153         log_develop_trace(gc, compaction)(
1154             "backing up from c=" SIZE_FORMAT_W(4) " d2r=%10.8f "
1155             "pc_d2r=%10.8f",
1156             sd.region(cp), density_to_right,
1157             prev_region_density_to_right);
1158 
1159         dense_prefix -= region_size;
1160         live_to_right = prev_region_live_to_right;
1161         space_to_right = prev_region_space_to_right;
1162         density_to_right = prev_region_density_to_right;
1163       }
1164       return dense_prefix;
1165     }
1166 
1167     dense_prefix += region_size;
1168     ++cp;
1169   }
1170 
1171   return dense_prefix;
1172 }
1173 
1174 #ifndef PRODUCT
1175 void PSParallelCompact::print_dense_prefix_stats(const char* const algorithm,
1176                                                  const SpaceId id,
1177                                                  const bool maximum_compaction,
1178                                                  HeapWord* const addr)
1179 {
1180   const size_t region_idx = summary_data().addr_to_region_idx(addr);
1181   RegionData* const cp = summary_data().region(region_idx);
1182   const MutableSpace* const space = _space_info[id].space();
1183   HeapWord* const new_top = _space_info[id].new_top();
1184 
1185   const size_t space_live = pointer_delta(new_top, space->bottom());
1186   const size_t dead_to_left = pointer_delta(addr, cp->destination());
1187   const size_t space_cap = space->capacity_in_words();
1188   const double dead_to_left_pct = double(dead_to_left) / space_cap;
1189   const size_t live_to_right = new_top - cp->destination();
1190   const size_t dead_to_right = space->top() - addr - live_to_right;
1191 
1192   log_develop_debug(gc, compaction)(
1193       "%s=" PTR_FORMAT " dpc=" SIZE_FORMAT_W(5) " "
1194       "spl=" SIZE_FORMAT " "
1195       "d2l=" SIZE_FORMAT " d2l%%=%6.4f "
1196       "d2r=" SIZE_FORMAT " l2r=" SIZE_FORMAT " "
1197       "ratio=%10.8f",
1198       algorithm, p2i(addr), region_idx,
1199       space_live,
1200       dead_to_left, dead_to_left_pct,
1201       dead_to_right, live_to_right,
1202       double(dead_to_right) / live_to_right);
1203 }
1204 #endif  // #ifndef PRODUCT
1205 
1206 // Return a fraction indicating how much of the generation can be treated as
1207 // "dead wood" (i.e., not reclaimed).  The function uses a normal distribution
1208 // based on the density of live objects in the generation to determine a limit,
1209 // which is then adjusted so the return value is min_percent when the density is
1210 // 1.
1211 //
1212 // The following table shows some return values for a different values of the
1213 // standard deviation (ParallelOldDeadWoodLimiterStdDev); the mean is 0.5 and
1214 // min_percent is 1.
1215 //
1216 //                          fraction allowed as dead wood
1217 //         -----------------------------------------------------------------
1218 // density std_dev=70 std_dev=75 std_dev=80 std_dev=85 std_dev=90 std_dev=95
1219 // ------- ---------- ---------- ---------- ---------- ---------- ----------
1220 // 0.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000
1221 // 0.05000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941
1222 // 0.10000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272
1223 // 0.15000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066
1224 // 0.20000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975
1225 // 0.25000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313
1226 // 0.30000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132
1227 // 0.35000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289
1228 // 0.40000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500
1229 // 0.45000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386
1230 // 0.50000 0.13832410 0.11599237 0.09847664 0.08456518 0.07338887 0.06431510
1231 // 0.55000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386
1232 // 0.60000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500
1233 // 0.65000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289
1234 // 0.70000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132
1235 // 0.75000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313
1236 // 0.80000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975
1237 // 0.85000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066
1238 // 0.90000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272
1239 // 0.95000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941
1240 // 1.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000
1241 
1242 double PSParallelCompact::dead_wood_limiter(double density, size_t min_percent)
1243 {
1244   assert(_dwl_initialized, "uninitialized");
1245 
1246   // The raw limit is the value of the normal distribution at x = density.
1247   const double raw_limit = normal_distribution(density);
1248 
1249   // Adjust the raw limit so it becomes the minimum when the density is 1.
1250   //
1251   // First subtract the adjustment value (which is simply the precomputed value
1252   // normal_distribution(1.0)); this yields a value of 0 when the density is 1.
1253   // Then add the minimum value, so the minimum is returned when the density is
1254   // 1.  Finally, prevent negative values, which occur when the mean is not 0.5.
1255   const double min = double(min_percent) / 100.0;
1256   const double limit = raw_limit - _dwl_adjustment + min;
1257   return MAX2(limit, 0.0);
1258 }
1259 
1260 ParallelCompactData::RegionData*
1261 PSParallelCompact::first_dead_space_region(const RegionData* beg,
1262                                            const RegionData* end)
1263 {
1264   const size_t region_size = ParallelCompactData::RegionSize;
1265   ParallelCompactData& sd = summary_data();
1266   size_t left = sd.region(beg);
1267   size_t right = end > beg ? sd.region(end) - 1 : left;
1268 
1269   // Binary search.
1270   while (left < right) {
1271     // Equivalent to (left + right) / 2, but does not overflow.
1272     const size_t middle = left + (right - left) / 2;
1273     RegionData* const middle_ptr = sd.region(middle);
1274     HeapWord* const dest = middle_ptr->destination();
1275     HeapWord* const addr = sd.region_to_addr(middle);
1276     assert(dest != NULL, "sanity");
1277     assert(dest <= addr, "must move left");
1278 
1279     if (middle > left && dest < addr) {
1280       right = middle - 1;
1281     } else if (middle < right && middle_ptr->data_size() == region_size) {
1282       left = middle + 1;
1283     } else {
1284       return middle_ptr;
1285     }
1286   }
1287   return sd.region(left);
1288 }
1289 
1290 ParallelCompactData::RegionData*
1291 PSParallelCompact::dead_wood_limit_region(const RegionData* beg,
1292                                           const RegionData* end,
1293                                           size_t dead_words)
1294 {
1295   ParallelCompactData& sd = summary_data();
1296   size_t left = sd.region(beg);
1297   size_t right = end > beg ? sd.region(end) - 1 : left;
1298 
1299   // Binary search.
1300   while (left < right) {
1301     // Equivalent to (left + right) / 2, but does not overflow.
1302     const size_t middle = left + (right - left) / 2;
1303     RegionData* const middle_ptr = sd.region(middle);
1304     HeapWord* const dest = middle_ptr->destination();
1305     HeapWord* const addr = sd.region_to_addr(middle);
1306     assert(dest != NULL, "sanity");
1307     assert(dest <= addr, "must move left");
1308 
1309     const size_t dead_to_left = pointer_delta(addr, dest);
1310     if (middle > left && dead_to_left > dead_words) {
1311       right = middle - 1;
1312     } else if (middle < right && dead_to_left < dead_words) {
1313       left = middle + 1;
1314     } else {
1315       return middle_ptr;
1316     }
1317   }
1318   return sd.region(left);
1319 }
1320 
1321 // The result is valid during the summary phase, after the initial summarization
1322 // of each space into itself, and before final summarization.
1323 inline double
1324 PSParallelCompact::reclaimed_ratio(const RegionData* const cp,
1325                                    HeapWord* const bottom,
1326                                    HeapWord* const top,
1327                                    HeapWord* const new_top)
1328 {
1329   ParallelCompactData& sd = summary_data();
1330 
1331   assert(cp != NULL, "sanity");
1332   assert(bottom != NULL, "sanity");
1333   assert(top != NULL, "sanity");
1334   assert(new_top != NULL, "sanity");
1335   assert(top >= new_top, "summary data problem?");
1336   assert(new_top > bottom, "space is empty; should not be here");
1337   assert(new_top >= cp->destination(), "sanity");
1338   assert(top >= sd.region_to_addr(cp), "sanity");
1339 
1340   HeapWord* const destination = cp->destination();
1341   const size_t dense_prefix_live  = pointer_delta(destination, bottom);
1342   const size_t compacted_region_live = pointer_delta(new_top, destination);
1343   const size_t compacted_region_used = pointer_delta(top,
1344                                                      sd.region_to_addr(cp));
1345   const size_t reclaimable = compacted_region_used - compacted_region_live;
1346 
1347   const double divisor = dense_prefix_live + 1.25 * compacted_region_live;
1348   return double(reclaimable) / divisor;
1349 }
1350 
1351 // Return the address of the end of the dense prefix, a.k.a. the start of the
1352 // compacted region.  The address is always on a region boundary.
1353 //
1354 // Completely full regions at the left are skipped, since no compaction can
1355 // occur in those regions.  Then the maximum amount of dead wood to allow is
1356 // computed, based on the density (amount live / capacity) of the generation;
1357 // the region with approximately that amount of dead space to the left is
1358 // identified as the limit region.  Regions between the last completely full
1359 // region and the limit region are scanned and the one that has the best
1360 // (maximum) reclaimed_ratio() is selected.
1361 HeapWord*
1362 PSParallelCompact::compute_dense_prefix(const SpaceId id,
1363                                         bool maximum_compaction)
1364 {
1365   const size_t region_size = ParallelCompactData::RegionSize;
1366   const ParallelCompactData& sd = summary_data();
1367 
1368   const MutableSpace* const space = _space_info[id].space();
1369   HeapWord* const top = space->top();
1370   HeapWord* const top_aligned_up = sd.region_align_up(top);
1371   HeapWord* const new_top = _space_info[id].new_top();
1372   HeapWord* const new_top_aligned_up = sd.region_align_up(new_top);
1373   HeapWord* const bottom = space->bottom();
1374   const RegionData* const beg_cp = sd.addr_to_region_ptr(bottom);
1375   const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up);
1376   const RegionData* const new_top_cp =
1377     sd.addr_to_region_ptr(new_top_aligned_up);
1378 
1379   // Skip full regions at the beginning of the space--they are necessarily part
1380   // of the dense prefix.
1381   const RegionData* const full_cp = first_dead_space_region(beg_cp, new_top_cp);
1382   assert(full_cp->destination() == sd.region_to_addr(full_cp) ||
1383          space->is_empty(), "no dead space allowed to the left");
1384   assert(full_cp->data_size() < region_size || full_cp == new_top_cp - 1,
1385          "region must have dead space");
1386 
1387   // The gc number is saved whenever a maximum compaction is done, and used to
1388   // determine when the maximum compaction interval has expired.  This avoids
1389   // successive max compactions for different reasons.
1390   assert(total_invocations() >= _maximum_compaction_gc_num, "sanity");
1391   const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num;
1392   const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval ||
1393     total_invocations() == HeapFirstMaximumCompactionCount;
1394   if (maximum_compaction || full_cp == top_cp || interval_ended) {
1395     _maximum_compaction_gc_num = total_invocations();
1396     return sd.region_to_addr(full_cp);
1397   }
1398 
1399   const size_t space_live = pointer_delta(new_top, bottom);
1400   const size_t space_used = space->used_in_words();
1401   const size_t space_capacity = space->capacity_in_words();
1402 
1403   const double density = double(space_live) / double(space_capacity);
1404   const size_t min_percent_free = MarkSweepDeadRatio;
1405   const double limiter = dead_wood_limiter(density, min_percent_free);
1406   const size_t dead_wood_max = space_used - space_live;
1407   const size_t dead_wood_limit = MIN2(size_t(space_capacity * limiter),
1408                                       dead_wood_max);
1409 
1410   log_develop_debug(gc, compaction)(
1411       "space_live=" SIZE_FORMAT " space_used=" SIZE_FORMAT " "
1412       "space_cap=" SIZE_FORMAT,
1413       space_live, space_used,
1414       space_capacity);
1415   log_develop_debug(gc, compaction)(
1416       "dead_wood_limiter(%6.4f, " SIZE_FORMAT ")=%6.4f "
1417       "dead_wood_max=" SIZE_FORMAT " dead_wood_limit=" SIZE_FORMAT,
1418       density, min_percent_free, limiter,
1419       dead_wood_max, dead_wood_limit);
1420 
1421   // Locate the region with the desired amount of dead space to the left.
1422   const RegionData* const limit_cp =
1423     dead_wood_limit_region(full_cp, top_cp, dead_wood_limit);
1424 
1425   // Scan from the first region with dead space to the limit region and find the
1426   // one with the best (largest) reclaimed ratio.
1427   double best_ratio = 0.0;
1428   const RegionData* best_cp = full_cp;
1429   for (const RegionData* cp = full_cp; cp < limit_cp; ++cp) {
1430     double tmp_ratio = reclaimed_ratio(cp, bottom, top, new_top);
1431     if (tmp_ratio > best_ratio) {
1432       best_cp = cp;
1433       best_ratio = tmp_ratio;
1434     }
1435   }
1436 
1437   return sd.region_to_addr(best_cp);
1438 }
1439 
1440 void PSParallelCompact::summarize_spaces_quick()
1441 {
1442   for (unsigned int i = 0; i < last_space_id; ++i) {
1443     const MutableSpace* space = _space_info[i].space();
1444     HeapWord** nta = _space_info[i].new_top_addr();
1445     bool result = _summary_data.summarize(_space_info[i].split_info(),
1446                                           space->bottom(), space->top(), NULL,
1447                                           space->bottom(), space->end(), nta);
1448     assert(result, "space must fit into itself");
1449     _space_info[i].set_dense_prefix(space->bottom());
1450   }
1451 }
1452 
1453 void PSParallelCompact::fill_dense_prefix_end(SpaceId id)
1454 {
1455   HeapWord* const dense_prefix_end = dense_prefix(id);
1456   const RegionData* region = _summary_data.addr_to_region_ptr(dense_prefix_end);
1457   const idx_t dense_prefix_bit = _mark_bitmap.addr_to_bit(dense_prefix_end);
1458   if (dead_space_crosses_boundary(region, dense_prefix_bit)) {
1459     // Only enough dead space is filled so that any remaining dead space to the
1460     // left is larger than the minimum filler object.  (The remainder is filled
1461     // during the copy/update phase.)
1462     //
1463     // The size of the dead space to the right of the boundary is not a
1464     // concern, since compaction will be able to use whatever space is
1465     // available.
1466     //
1467     // Here '||' is the boundary, 'x' represents a don't care bit and a box
1468     // surrounds the space to be filled with an object.
1469     //
1470     // In the 32-bit VM, each bit represents two 32-bit words:
1471     //                              +---+
1472     // a) beg_bits:  ...  x   x   x | 0 | ||   0   x  x  ...
1473     //    end_bits:  ...  x   x   x | 0 | ||   0   x  x  ...
1474     //                              +---+
1475     //
1476     // In the 64-bit VM, each bit represents one 64-bit word:
1477     //                              +------------+
1478     // b) beg_bits:  ...  x   x   x | 0   ||   0 | x  x  ...
1479     //    end_bits:  ...  x   x   1 | 0   ||   0 | x  x  ...
1480     //                              +------------+
1481     //                          +-------+
1482     // c) beg_bits:  ...  x   x | 0   0 | ||   0   x  x  ...
1483     //    end_bits:  ...  x   1 | 0   0 | ||   0   x  x  ...
1484     //                          +-------+
1485     //                      +-----------+
1486     // d) beg_bits:  ...  x | 0   0   0 | ||   0   x  x  ...
1487     //    end_bits:  ...  1 | 0   0   0 | ||   0   x  x  ...
1488     //                      +-----------+
1489     //                          +-------+
1490     // e) beg_bits:  ...  0   0 | 0   0 | ||   0   x  x  ...
1491     //    end_bits:  ...  0   0 | 0   0 | ||   0   x  x  ...
1492     //                          +-------+
1493 
1494     // Initially assume case a, c or e will apply.
1495     size_t obj_len = CollectedHeap::min_fill_size();
1496     HeapWord* obj_beg = dense_prefix_end - obj_len;
1497 
1498 #ifdef  _LP64
1499     if (MinObjAlignment > 1) { // object alignment > heap word size
1500       // Cases a, c or e.
1501     } else if (_mark_bitmap.is_obj_end(dense_prefix_bit - 2)) {
1502       // Case b above.
1503       obj_beg = dense_prefix_end - 1;
1504     } else if (!_mark_bitmap.is_obj_end(dense_prefix_bit - 3) &&
1505                _mark_bitmap.is_obj_end(dense_prefix_bit - 4)) {
1506       // Case d above.
1507       obj_beg = dense_prefix_end - 3;
1508       obj_len = 3;
1509     }
1510 #endif  // #ifdef _LP64
1511 
1512     CollectedHeap::fill_with_object(obj_beg, obj_len);
1513     _mark_bitmap.mark_obj(obj_beg, obj_len);
1514     _summary_data.add_obj(obj_beg, obj_len);
1515     assert(start_array(id) != NULL, "sanity");
1516     start_array(id)->allocate_block(obj_beg);
1517   }
1518 }
1519 
1520 void
1521 PSParallelCompact::summarize_space(SpaceId id, bool maximum_compaction)
1522 {
1523   assert(id < last_space_id, "id out of range");
1524   assert(_space_info[id].dense_prefix() == _space_info[id].space()->bottom(),
1525          "should have been reset in summarize_spaces_quick()");
1526 
1527   const MutableSpace* space = _space_info[id].space();
1528   if (_space_info[id].new_top() != space->bottom()) {
1529     HeapWord* dense_prefix_end = compute_dense_prefix(id, maximum_compaction);
1530     _space_info[id].set_dense_prefix(dense_prefix_end);
1531 
1532 #ifndef PRODUCT
1533     if (log_is_enabled(Debug, gc, compaction)) {
1534       print_dense_prefix_stats("ratio", id, maximum_compaction,
1535                                dense_prefix_end);
1536       HeapWord* addr = compute_dense_prefix_via_density(id, maximum_compaction);
1537       print_dense_prefix_stats("density", id, maximum_compaction, addr);
1538     }
1539 #endif  // #ifndef PRODUCT
1540 
1541     // Recompute the summary data, taking into account the dense prefix.  If
1542     // every last byte will be reclaimed, then the existing summary data which
1543     // compacts everything can be left in place.
1544     if (!maximum_compaction && dense_prefix_end != space->bottom()) {
1545       // If dead space crosses the dense prefix boundary, it is (at least
1546       // partially) filled with a dummy object, marked live and added to the
1547       // summary data.  This simplifies the copy/update phase and must be done
1548       // before the final locations of objects are determined, to prevent
1549       // leaving a fragment of dead space that is too small to fill.
1550       fill_dense_prefix_end(id);
1551 
1552       // Compute the destination of each Region, and thus each object.
1553       _summary_data.summarize_dense_prefix(space->bottom(), dense_prefix_end);
1554       _summary_data.summarize(_space_info[id].split_info(),
1555                               dense_prefix_end, space->top(), NULL,
1556                               dense_prefix_end, space->end(),
1557                               _space_info[id].new_top_addr());
1558     }
1559   }
1560 
1561   if (log_develop_is_enabled(Trace, gc, compaction)) {
1562     const size_t region_size = ParallelCompactData::RegionSize;
1563     HeapWord* const dense_prefix_end = _space_info[id].dense_prefix();
1564     const size_t dp_region = _summary_data.addr_to_region_idx(dense_prefix_end);
1565     const size_t dp_words = pointer_delta(dense_prefix_end, space->bottom());
1566     HeapWord* const new_top = _space_info[id].new_top();
1567     const HeapWord* nt_aligned_up = _summary_data.region_align_up(new_top);
1568     const size_t cr_words = pointer_delta(nt_aligned_up, dense_prefix_end);
1569     log_develop_trace(gc, compaction)(
1570         "id=%d cap=" SIZE_FORMAT " dp=" PTR_FORMAT " "
1571         "dp_region=" SIZE_FORMAT " " "dp_count=" SIZE_FORMAT " "
1572         "cr_count=" SIZE_FORMAT " " "nt=" PTR_FORMAT,
1573         id, space->capacity_in_words(), p2i(dense_prefix_end),
1574         dp_region, dp_words / region_size,
1575         cr_words / region_size, p2i(new_top));
1576   }
1577 }
1578 
1579 #ifndef PRODUCT
1580 void PSParallelCompact::summary_phase_msg(SpaceId dst_space_id,
1581                                           HeapWord* dst_beg, HeapWord* dst_end,
1582                                           SpaceId src_space_id,
1583                                           HeapWord* src_beg, HeapWord* src_end)
1584 {
1585   log_develop_trace(gc, compaction)(
1586       "Summarizing %d [%s] into %d [%s]:  "
1587       "src=" PTR_FORMAT "-" PTR_FORMAT " "
1588       SIZE_FORMAT "-" SIZE_FORMAT " "
1589       "dst=" PTR_FORMAT "-" PTR_FORMAT " "
1590       SIZE_FORMAT "-" SIZE_FORMAT,
1591       src_space_id, space_names[src_space_id],
1592       dst_space_id, space_names[dst_space_id],
1593       p2i(src_beg), p2i(src_end),
1594       _summary_data.addr_to_region_idx(src_beg),
1595       _summary_data.addr_to_region_idx(src_end),
1596       p2i(dst_beg), p2i(dst_end),
1597       _summary_data.addr_to_region_idx(dst_beg),
1598       _summary_data.addr_to_region_idx(dst_end));
1599 }
1600 #endif  // #ifndef PRODUCT
1601 
1602 void PSParallelCompact::summary_phase(ParCompactionManager* cm,
1603                                       bool maximum_compaction)
1604 {
1605   GCTraceTime(Info, gc, phases) tm("Summary Phase", &_gc_timer);
1606 
1607   // Quick summarization of each space into itself, to see how much is live.
1608   summarize_spaces_quick();
1609 
1610   log_develop_trace(gc, compaction)("summary phase:  after summarizing each space to self");
1611   NOT_PRODUCT(print_region_ranges());
1612   NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info));
1613 
1614   // The amount of live data that will end up in old space (assuming it fits).
1615   size_t old_space_total_live = 0;
1616   for (unsigned int id = old_space_id; id < last_space_id; ++id) {
1617     old_space_total_live += pointer_delta(_space_info[id].new_top(),
1618                                           _space_info[id].space()->bottom());
1619   }
1620 
1621   MutableSpace* const old_space = _space_info[old_space_id].space();
1622   const size_t old_capacity = old_space->capacity_in_words();
1623   if (old_space_total_live > old_capacity) {
1624     // XXX - should also try to expand
1625     maximum_compaction = true;
1626   }
1627 
1628   // Old generations.
1629   summarize_space(old_space_id, maximum_compaction);
1630 
1631   // Summarize the remaining spaces in the young gen.  The initial target space
1632   // is the old gen.  If a space does not fit entirely into the target, then the
1633   // remainder is compacted into the space itself and that space becomes the new
1634   // target.
1635   SpaceId dst_space_id = old_space_id;
1636   HeapWord* dst_space_end = old_space->end();
1637   HeapWord** new_top_addr = _space_info[dst_space_id].new_top_addr();
1638   for (unsigned int id = eden_space_id; id < last_space_id; ++id) {
1639     const MutableSpace* space = _space_info[id].space();
1640     const size_t live = pointer_delta(_space_info[id].new_top(),
1641                                       space->bottom());
1642     const size_t available = pointer_delta(dst_space_end, *new_top_addr);
1643 
1644     NOT_PRODUCT(summary_phase_msg(dst_space_id, *new_top_addr, dst_space_end,
1645                                   SpaceId(id), space->bottom(), space->top());)
1646     if (live > 0 && live <= available) {
1647       // All the live data will fit.
1648       bool done = _summary_data.summarize(_space_info[id].split_info(),
1649                                           space->bottom(), space->top(),
1650                                           NULL,
1651                                           *new_top_addr, dst_space_end,
1652                                           new_top_addr);
1653       assert(done, "space must fit into old gen");
1654 
1655       // Reset the new_top value for the space.
1656       _space_info[id].set_new_top(space->bottom());
1657     } else if (live > 0) {
1658       // Attempt to fit part of the source space into the target space.
1659       HeapWord* next_src_addr = NULL;
1660       bool done = _summary_data.summarize(_space_info[id].split_info(),
1661                                           space->bottom(), space->top(),
1662                                           &next_src_addr,
1663                                           *new_top_addr, dst_space_end,
1664                                           new_top_addr);
1665       assert(!done, "space should not fit into old gen");
1666       assert(next_src_addr != NULL, "sanity");
1667 
1668       // The source space becomes the new target, so the remainder is compacted
1669       // within the space itself.
1670       dst_space_id = SpaceId(id);
1671       dst_space_end = space->end();
1672       new_top_addr = _space_info[id].new_top_addr();
1673       NOT_PRODUCT(summary_phase_msg(dst_space_id,
1674                                     space->bottom(), dst_space_end,
1675                                     SpaceId(id), next_src_addr, space->top());)
1676       done = _summary_data.summarize(_space_info[id].split_info(),
1677                                      next_src_addr, space->top(),
1678                                      NULL,
1679                                      space->bottom(), dst_space_end,
1680                                      new_top_addr);
1681       assert(done, "space must fit when compacted into itself");
1682       assert(*new_top_addr <= space->top(), "usage should not grow");
1683     }
1684   }
1685 
1686   log_develop_trace(gc, compaction)("Summary_phase:  after final summarization");
1687   NOT_PRODUCT(print_region_ranges());
1688   NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info));
1689 }
1690 
1691 // This method should contain all heap-specific policy for invoking a full
1692 // collection.  invoke_no_policy() will only attempt to compact the heap; it
1693 // will do nothing further.  If we need to bail out for policy reasons, scavenge
1694 // before full gc, or any other specialized behavior, it needs to be added here.
1695 //
1696 // Note that this method should only be called from the vm_thread while at a
1697 // safepoint.
1698 //
1699 // Note that the all_soft_refs_clear flag in the soft ref policy
1700 // may be true because this method can be called without intervening
1701 // activity.  For example when the heap space is tight and full measure
1702 // are being taken to free space.
1703 void PSParallelCompact::invoke(bool maximum_heap_compaction) {
1704   assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
1705   assert(Thread::current() == (Thread*)VMThread::vm_thread(),
1706          "should be in vm thread");
1707 
1708   ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
1709   GCCause::Cause gc_cause = heap->gc_cause();
1710   assert(!heap->is_gc_active(), "not reentrant");
1711 
1712   PSAdaptiveSizePolicy* policy = heap->size_policy();
1713   IsGCActiveMark mark;
1714 
1715   if (ScavengeBeforeFullGC) {
1716     PSScavenge::invoke_no_policy();
1717   }
1718 
1719   const bool clear_all_soft_refs =
1720     heap->soft_ref_policy()->should_clear_all_soft_refs();
1721 
1722   PSParallelCompact::invoke_no_policy(clear_all_soft_refs ||
1723                                       maximum_heap_compaction);
1724 }
1725 
1726 // This method contains no policy. You should probably
1727 // be calling invoke() instead.
1728 bool PSParallelCompact::invoke_no_policy(bool maximum_heap_compaction) {
1729   assert(SafepointSynchronize::is_at_safepoint(), "must be at a safepoint");
1730   assert(ref_processor() != NULL, "Sanity");
1731 
1732   if (GCLocker::check_active_before_gc()) {
1733     return false;
1734   }
1735 
1736   ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
1737 
1738   GCIdMark gc_id_mark;
1739   _gc_timer.register_gc_start();
1740   _gc_tracer.report_gc_start(heap->gc_cause(), _gc_timer.gc_start());
1741 
1742   TimeStamp marking_start;
1743   TimeStamp compaction_start;
1744   TimeStamp collection_exit;
1745 
1746   GCCause::Cause gc_cause = heap->gc_cause();
1747   PSYoungGen* young_gen = heap->young_gen();
1748   PSOldGen* old_gen = heap->old_gen();
1749   PSAdaptiveSizePolicy* size_policy = heap->size_policy();
1750 
1751   // The scope of casr should end after code that can change
1752   // SoftRefPolicy::_should_clear_all_soft_refs.
1753   ClearedAllSoftRefs casr(maximum_heap_compaction,
1754                           heap->soft_ref_policy());
1755 
1756   if (ZapUnusedHeapArea) {
1757     // Save information needed to minimize mangling
1758     heap->record_gen_tops_before_GC();
1759   }
1760 
1761   // Make sure data structures are sane, make the heap parsable, and do other
1762   // miscellaneous bookkeeping.
1763   pre_compact();
1764 
1765   const PreGenGCValues pre_gc_values = heap->get_pre_gc_values();
1766 
1767   // Get the compaction manager reserved for the VM thread.
1768   ParCompactionManager* const vmthread_cm = ParCompactionManager::get_vmthread_cm();
1769 
1770   {
1771     const uint active_workers =
1772       WorkerPolicy::calc_active_workers(ParallelScavengeHeap::heap()->workers().total_workers(),
1773                                         ParallelScavengeHeap::heap()->workers().active_workers(),
1774                                         Threads::number_of_non_daemon_threads());
1775     ParallelScavengeHeap::heap()->workers().update_active_workers(active_workers);
1776 
1777     GCTraceCPUTime tcpu;
1778     GCTraceTime(Info, gc) tm("Pause Full", NULL, gc_cause, true);
1779 
1780     heap->pre_full_gc_dump(&_gc_timer);
1781 
1782     TraceCollectorStats tcs(counters());
1783     TraceMemoryManagerStats tms(heap->old_gc_manager(), gc_cause);
1784 
1785     if (log_is_enabled(Debug, gc, heap, exit)) {
1786       accumulated_time()->start();
1787     }
1788 
1789     // Let the size policy know we're starting
1790     size_policy->major_collection_begin();
1791 
1792 #if COMPILER2_OR_JVMCI
1793     DerivedPointerTable::clear();
1794 #endif
1795 
1796     ref_processor()->start_discovery(maximum_heap_compaction);
1797 
1798     marking_start.update();
1799     marking_phase(vmthread_cm, &_gc_tracer);
1800 
1801     bool max_on_system_gc = UseMaximumCompactionOnSystemGC
1802       && GCCause::is_user_requested_gc(gc_cause);
1803     summary_phase(vmthread_cm, maximum_heap_compaction || max_on_system_gc);
1804 
1805 #if COMPILER2_OR_JVMCI
1806     assert(DerivedPointerTable::is_active(), "Sanity");
1807     DerivedPointerTable::set_active(false);
1808 #endif
1809 
1810     // adjust_roots() updates Universe::_intArrayKlassObj which is
1811     // needed by the compaction for filling holes in the dense prefix.
1812     adjust_roots();
1813 
1814     compaction_start.update();
1815     compact();
1816 
1817     ParCompactionManager::verify_all_region_stack_empty();
1818 
1819     // Reset the mark bitmap, summary data, and do other bookkeeping.  Must be
1820     // done before resizing.
1821     post_compact();
1822 
1823     // Let the size policy know we're done
1824     size_policy->major_collection_end(old_gen->used_in_bytes(), gc_cause);
1825 
1826     if (UseAdaptiveSizePolicy) {
1827       log_debug(gc, ergo)("AdaptiveSizeStart: collection: %d ", heap->total_collections());
1828       log_trace(gc, ergo)("old_gen_capacity: " SIZE_FORMAT " young_gen_capacity: " SIZE_FORMAT,
1829                           old_gen->capacity_in_bytes(), young_gen->capacity_in_bytes());
1830 
1831       // Don't check if the size_policy is ready here.  Let
1832       // the size_policy check that internally.
1833       if (UseAdaptiveGenerationSizePolicyAtMajorCollection &&
1834           AdaptiveSizePolicy::should_update_promo_stats(gc_cause)) {
1835         // Swap the survivor spaces if from_space is empty. The
1836         // resize_young_gen() called below is normally used after
1837         // a successful young GC and swapping of survivor spaces;
1838         // otherwise, it will fail to resize the young gen with
1839         // the current implementation.
1840         if (young_gen->from_space()->is_empty()) {
1841           young_gen->from_space()->clear(SpaceDecorator::Mangle);
1842           young_gen->swap_spaces();
1843         }
1844 
1845         // Calculate optimal free space amounts
1846         assert(young_gen->max_gen_size() >
1847           young_gen->from_space()->capacity_in_bytes() +
1848           young_gen->to_space()->capacity_in_bytes(),
1849           "Sizes of space in young gen are out-of-bounds");
1850 
1851         size_t young_live = young_gen->used_in_bytes();
1852         size_t eden_live = young_gen->eden_space()->used_in_bytes();
1853         size_t old_live = old_gen->used_in_bytes();
1854         size_t cur_eden = young_gen->eden_space()->capacity_in_bytes();
1855         size_t max_old_gen_size = old_gen->max_gen_size();
1856         size_t max_eden_size = young_gen->max_gen_size() -
1857           young_gen->from_space()->capacity_in_bytes() -
1858           young_gen->to_space()->capacity_in_bytes();
1859 
1860         // Used for diagnostics
1861         size_policy->clear_generation_free_space_flags();
1862 
1863         size_policy->compute_generations_free_space(young_live,
1864                                                     eden_live,
1865                                                     old_live,
1866                                                     cur_eden,
1867                                                     max_old_gen_size,
1868                                                     max_eden_size,
1869                                                     true /* full gc*/);
1870 
1871         size_policy->check_gc_overhead_limit(eden_live,
1872                                              max_old_gen_size,
1873                                              max_eden_size,
1874                                              true /* full gc*/,
1875                                              gc_cause,
1876                                              heap->soft_ref_policy());
1877 
1878         size_policy->decay_supplemental_growth(true /* full gc*/);
1879 
1880         heap->resize_old_gen(
1881           size_policy->calculated_old_free_size_in_bytes());
1882 
1883         heap->resize_young_gen(size_policy->calculated_eden_size_in_bytes(),
1884                                size_policy->calculated_survivor_size_in_bytes());
1885       }
1886 
1887       log_debug(gc, ergo)("AdaptiveSizeStop: collection: %d ", heap->total_collections());
1888     }
1889 
1890     if (UsePerfData) {
1891       PSGCAdaptivePolicyCounters* const counters = heap->gc_policy_counters();
1892       counters->update_counters();
1893       counters->update_old_capacity(old_gen->capacity_in_bytes());
1894       counters->update_young_capacity(young_gen->capacity_in_bytes());
1895     }
1896 
1897     heap->resize_all_tlabs();
1898 
1899     // Resize the metaspace capacity after a collection
1900     MetaspaceGC::compute_new_size();
1901 
1902     if (log_is_enabled(Debug, gc, heap, exit)) {
1903       accumulated_time()->stop();
1904     }
1905 
1906     heap->print_heap_change(pre_gc_values);
1907 
1908     // Track memory usage and detect low memory
1909     MemoryService::track_memory_usage();
1910     heap->update_counters();
1911 
1912     heap->post_full_gc_dump(&_gc_timer);
1913   }
1914 
1915   if (VerifyAfterGC && heap->total_collections() >= VerifyGCStartAt) {
1916     Universe::verify("After GC");
1917   }
1918 
1919   // Re-verify object start arrays
1920   if (VerifyObjectStartArray &&
1921       VerifyAfterGC) {
1922     old_gen->verify_object_start_array();
1923   }
1924 
1925   if (ZapUnusedHeapArea) {
1926     old_gen->object_space()->check_mangled_unused_area_complete();
1927   }
1928 
1929   NOT_PRODUCT(ref_processor()->verify_no_references_recorded());
1930 
1931   collection_exit.update();
1932 
1933   heap->print_heap_after_gc();
1934   heap->trace_heap_after_gc(&_gc_tracer);
1935 
1936   log_debug(gc, task, time)("VM-Thread " JLONG_FORMAT " " JLONG_FORMAT " " JLONG_FORMAT,
1937                          marking_start.ticks(), compaction_start.ticks(),
1938                          collection_exit.ticks());
1939 
1940   AdaptiveSizePolicyOutput::print(size_policy, heap->total_collections());
1941 
1942   _gc_timer.register_gc_end();
1943 
1944   _gc_tracer.report_dense_prefix(dense_prefix(old_space_id));
1945   _gc_tracer.report_gc_end(_gc_timer.gc_end(), _gc_timer.time_partitions());
1946 
1947   return true;
1948 }
1949 
1950 class PCAddThreadRootsMarkingTaskClosure : public ThreadClosure {
1951 private:
1952   uint _worker_id;
1953 
1954 public:
1955   PCAddThreadRootsMarkingTaskClosure(uint worker_id) : _worker_id(worker_id) { }
1956   void do_thread(Thread* thread) {
1957     assert(ParallelScavengeHeap::heap()->is_gc_active(), "called outside gc");
1958 
1959     ResourceMark rm;
1960 
1961     ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(_worker_id);
1962 
1963     PCMarkAndPushClosure mark_and_push_closure(cm);
1964     MarkingCodeBlobClosure mark_and_push_in_blobs(&mark_and_push_closure, !CodeBlobToOopClosure::FixRelocations, true /* keepalive nmethods */);
1965 
1966     thread->oops_do(&mark_and_push_closure, &mark_and_push_in_blobs);
1967 
1968     // Do the real work
1969     cm->follow_marking_stacks();
1970   }
1971 };
1972 
1973 static void mark_from_roots_work(ParallelRootType::Value root_type, uint worker_id) {
1974   assert(ParallelScavengeHeap::heap()->is_gc_active(), "called outside gc");
1975 
1976   ParCompactionManager* cm =
1977     ParCompactionManager::gc_thread_compaction_manager(worker_id);
1978   PCMarkAndPushClosure mark_and_push_closure(cm);
1979 
1980   switch (root_type) {
1981     case ParallelRootType::class_loader_data:
1982       {
1983         CLDToOopClosure cld_closure(&mark_and_push_closure, ClassLoaderData::_claim_strong);
1984         ClassLoaderDataGraph::always_strong_cld_do(&cld_closure);
1985       }
1986       break;
1987 
1988     case ParallelRootType::code_cache:
1989       // Do not treat nmethods as strong roots for mark/sweep, since we can unload them.
1990       //ScavengableNMethods::scavengable_nmethods_do(CodeBlobToOopClosure(&mark_and_push_closure));
1991       break;
1992 
1993     case ParallelRootType::sentinel:
1994     DEBUG_ONLY(default:) // DEBUG_ONLY hack will create compile error on release builds (-Wswitch) and runtime check on debug builds
1995       fatal("Bad enumeration value: %u", root_type);
1996       break;
1997   }
1998 
1999   // Do the real work
2000   cm->follow_marking_stacks();
2001 }
2002 
2003 void steal_marking_work(TaskTerminator& terminator, uint worker_id) {
2004   assert(ParallelScavengeHeap::heap()->is_gc_active(), "called outside gc");
2005 
2006   ParCompactionManager* cm =
2007     ParCompactionManager::gc_thread_compaction_manager(worker_id);
2008 
2009   oop obj = NULL;
2010   ObjArrayTask task;
2011   do {
2012     while (ParCompactionManager::steal_objarray(worker_id,  task)) {
2013       cm->follow_array((objArrayOop)task.obj(), task.index());
2014       cm->follow_marking_stacks();
2015     }
2016     while (ParCompactionManager::steal(worker_id, obj)) {
2017       cm->follow_contents(obj);
2018       cm->follow_marking_stacks();
2019     }
2020   } while (!terminator.offer_termination());
2021 }
2022 
2023 class MarkFromRootsTask : public AbstractGangTask {
2024   StrongRootsScope _strong_roots_scope; // needed for Threads::possibly_parallel_threads_do
2025   OopStorageSetStrongParState<false /* concurrent */, false /* is_const */> _oop_storage_set_par_state;
2026   SequentialSubTasksDone _subtasks;
2027   TaskTerminator _terminator;
2028   uint _active_workers;
2029 
2030 public:
2031   MarkFromRootsTask(uint active_workers) :
2032       AbstractGangTask("MarkFromRootsTask"),
2033       _strong_roots_scope(active_workers),
2034       _subtasks(ParallelRootType::sentinel),
2035       _terminator(active_workers, ParCompactionManager::oop_task_queues()),
2036       _active_workers(active_workers) {
2037   }
2038 
2039   virtual void work(uint worker_id) {
2040     for (uint task = 0; _subtasks.try_claim_task(task); /*empty*/ ) {
2041       mark_from_roots_work(static_cast<ParallelRootType::Value>(task), worker_id);
2042     }
2043 
2044     PCAddThreadRootsMarkingTaskClosure closure(worker_id);
2045     Threads::possibly_parallel_threads_do(true /*parallel */, &closure);
2046 
2047     // Mark from OopStorages
2048     {
2049       ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(worker_id);
2050       PCMarkAndPushClosure closure(cm);
2051       _oop_storage_set_par_state.oops_do(&closure);
2052       // Do the real work
2053       cm->follow_marking_stacks();
2054     }
2055 
2056     if (_active_workers > 1) {
2057       steal_marking_work(_terminator, worker_id);
2058     }
2059   }
2060 };
2061 
2062 class ParallelCompactRefProcProxyTask : public RefProcProxyTask {
2063   TaskTerminator _terminator;
2064 
2065 public:
2066   ParallelCompactRefProcProxyTask(uint max_workers)
2067     : RefProcProxyTask("ParallelCompactRefProcProxyTask", max_workers),
2068       _terminator(_max_workers, ParCompactionManager::oop_task_queues()) {}
2069 
2070   void work(uint worker_id) override {
2071     assert(worker_id < _max_workers, "sanity");
2072     ParCompactionManager* cm = (_tm == RefProcThreadModel::Single) ? ParCompactionManager::get_vmthread_cm() : ParCompactionManager::gc_thread_compaction_manager(worker_id);
2073     PCMarkAndPushClosure keep_alive(cm);
2074     BarrierEnqueueDiscoveredFieldClosure enqueue;
2075     ParCompactionManager::FollowStackClosure complete_gc(cm, (_tm == RefProcThreadModel::Single) ? nullptr : &_terminator, worker_id);
2076     _rp_task->rp_work(worker_id, PSParallelCompact::is_alive_closure(), &keep_alive, &enqueue, &complete_gc);
2077   }
2078 
2079   void prepare_run_task_hook() override {
2080     _terminator.reset_for_reuse(_queue_count);
2081   }
2082 };
2083 
2084 void PSParallelCompact::marking_phase(ParCompactionManager* cm,
2085                                       ParallelOldTracer *gc_tracer) {
2086   // Recursively traverse all live objects and mark them
2087   GCTraceTime(Info, gc, phases) tm("Marking Phase", &_gc_timer);
2088 
2089   uint active_gc_threads = ParallelScavengeHeap::heap()->workers().active_workers();
2090 
2091   // Need new claim bits before marking starts.
2092   ClassLoaderDataGraph::clear_claimed_marks();
2093 
2094   {
2095     GCTraceTime(Debug, gc, phases) tm("Par Mark", &_gc_timer);
2096 
2097     MarkFromRootsTask task(active_gc_threads);
2098     ParallelScavengeHeap::heap()->workers().run_task(&task);
2099   }
2100 
2101   // Process reference objects found during marking
2102   {
2103     GCTraceTime(Debug, gc, phases) tm("Reference Processing", &_gc_timer);
2104 
2105     ReferenceProcessorStats stats;
2106     ReferenceProcessorPhaseTimes pt(&_gc_timer, ref_processor()->max_num_queues());
2107 
2108     ref_processor()->set_active_mt_degree(active_gc_threads);
2109     ParallelCompactRefProcProxyTask task(ref_processor()->max_num_queues());
2110     stats = ref_processor()->process_discovered_references(task, pt);
2111 
2112     gc_tracer->report_gc_reference_stats(stats);
2113     pt.print_all_references();
2114   }
2115 
2116   // This is the point where the entire marking should have completed.
2117   ParCompactionManager::verify_all_marking_stack_empty();
2118 
2119   {
2120     GCTraceTime(Debug, gc, phases) tm("Weak Processing", &_gc_timer);
2121     WeakProcessor::weak_oops_do(&ParallelScavengeHeap::heap()->workers(),
2122                                 is_alive_closure(),
2123                                 &do_nothing_cl,
2124                                 1);
2125   }
2126 
2127   {
2128     GCTraceTime(Debug, gc, phases) tm_m("Class Unloading", &_gc_timer);
2129 
2130     // Follow system dictionary roots and unload classes.
2131     bool purged_class = SystemDictionary::do_unloading(&_gc_timer);
2132 
2133     // Unload nmethods.
2134     CodeCache::do_unloading(is_alive_closure(), purged_class);
2135 
2136     // Prune dead klasses from subklass/sibling/implementor lists.
2137     Klass::clean_weak_klass_links(purged_class);
2138 
2139     // Clean JVMCI metadata handles.
2140     JVMCI_ONLY(JVMCI::do_unloading(purged_class));
2141   }
2142 
2143   _gc_tracer.report_object_count_after_gc(is_alive_closure());
2144 }
2145 
2146 #ifdef ASSERT
2147 void PCAdjustPointerClosure::verify_cm(ParCompactionManager* cm) {
2148   assert(cm != NULL, "associate ParCompactionManage should not be NULL");
2149   auto vmthread_cm = ParCompactionManager::get_vmthread_cm();
2150   if (Thread::current()->is_VM_thread()) {
2151     assert(cm == vmthread_cm, "VM threads should use ParCompactionManager from get_vmthread_cm()");
2152   } else {
2153     assert(Thread::current()->is_Worker_thread(), "Must be a GC thread");
2154     assert(cm != vmthread_cm, "GC threads should use ParCompactionManager from gc_thread_compaction_manager()");
2155   }
2156 }
2157 #endif
2158 
2159 class PSAdjustTask final : public AbstractGangTask {
2160   SubTasksDone                               _sub_tasks;
2161   WeakProcessor::Task                        _weak_proc_task;
2162   OopStorageSetStrongParState<false, false>  _oop_storage_iter;
2163   uint                                       _nworkers;
2164 
2165   enum PSAdjustSubTask {
2166     PSAdjustSubTask_code_cache,
2167     PSAdjustSubTask_old_ref_process,
2168     PSAdjustSubTask_young_ref_process,
2169 
2170     PSAdjustSubTask_num_elements
2171   };
2172 
2173 public:
2174   PSAdjustTask(uint nworkers) :
2175     AbstractGangTask("PSAdjust task"),
2176     _sub_tasks(PSAdjustSubTask_num_elements),
2177     _weak_proc_task(nworkers),
2178     _nworkers(nworkers) {
2179     // Need new claim bits when tracing through and adjusting pointers.
2180     ClassLoaderDataGraph::clear_claimed_marks();
2181     if (nworkers > 1) {
2182       Threads::change_thread_claim_token();
2183     }
2184   }
2185 
2186   ~PSAdjustTask() {
2187     Threads::assert_all_threads_claimed();
2188   }
2189 
2190   void work(uint worker_id) {
2191     ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(worker_id);
2192     PCAdjustPointerClosure adjust(cm);
2193     {
2194       ResourceMark rm;
2195       Threads::possibly_parallel_oops_do(_nworkers > 1, &adjust, nullptr);
2196     }
2197     _oop_storage_iter.oops_do(&adjust);
2198     {
2199       CLDToOopClosure cld_closure(&adjust, ClassLoaderData::_claim_strong);
2200       ClassLoaderDataGraph::cld_do(&cld_closure);
2201     }
2202     {
2203       AlwaysTrueClosure always_alive;
2204       _weak_proc_task.work(worker_id, &always_alive, &adjust);
2205     }
2206     if (_sub_tasks.try_claim_task(PSAdjustSubTask_code_cache)) {
2207       CodeBlobToOopClosure adjust_code(&adjust, CodeBlobToOopClosure::FixRelocations);
2208       CodeCache::blobs_do(&adjust_code);
2209     }
2210     if (_sub_tasks.try_claim_task(PSAdjustSubTask_old_ref_process)) {
2211       PSParallelCompact::ref_processor()->weak_oops_do(&adjust);
2212     }
2213     if (_sub_tasks.try_claim_task(PSAdjustSubTask_young_ref_process)) {
2214       // Roots were visited so references into the young gen in roots
2215       // may have been scanned.  Process them also.
2216       // Should the reference processor have a span that excludes
2217       // young gen objects?
2218       PSScavenge::reference_processor()->weak_oops_do(&adjust);
2219     }
2220     _sub_tasks.all_tasks_claimed();
2221   }
2222 };
2223 
2224 void PSParallelCompact::adjust_roots() {
2225   // Adjust the pointers to reflect the new locations
2226   GCTraceTime(Info, gc, phases) tm("Adjust Roots", &_gc_timer);
2227   uint nworkers = ParallelScavengeHeap::heap()->workers().active_workers();
2228   PSAdjustTask task(nworkers);
2229   ParallelScavengeHeap::heap()->workers().run_task(&task);
2230 }
2231 
2232 // Helper class to print 8 region numbers per line and then print the total at the end.
2233 class FillableRegionLogger : public StackObj {
2234 private:
2235   Log(gc, compaction) log;
2236   static const int LineLength = 8;
2237   size_t _regions[LineLength];
2238   int _next_index;
2239   bool _enabled;
2240   size_t _total_regions;
2241 public:
2242   FillableRegionLogger() : _next_index(0), _enabled(log_develop_is_enabled(Trace, gc, compaction)), _total_regions(0) { }
2243   ~FillableRegionLogger() {
2244     log.trace(SIZE_FORMAT " initially fillable regions", _total_regions);
2245   }
2246 
2247   void print_line() {
2248     if (!_enabled || _next_index == 0) {
2249       return;
2250     }
2251     FormatBuffer<> line("Fillable: ");
2252     for (int i = 0; i < _next_index; i++) {
2253       line.append(" " SIZE_FORMAT_W(7), _regions[i]);
2254     }
2255     log.trace("%s", line.buffer());
2256     _next_index = 0;
2257   }
2258 
2259   void handle(size_t region) {
2260     if (!_enabled) {
2261       return;
2262     }
2263     _regions[_next_index++] = region;
2264     if (_next_index == LineLength) {
2265       print_line();
2266     }
2267     _total_regions++;
2268   }
2269 };
2270 
2271 void PSParallelCompact::prepare_region_draining_tasks(uint parallel_gc_threads)
2272 {
2273   GCTraceTime(Trace, gc, phases) tm("Drain Task Setup", &_gc_timer);
2274 
2275   // Find the threads that are active
2276   uint worker_id = 0;
2277 
2278   // Find all regions that are available (can be filled immediately) and
2279   // distribute them to the thread stacks.  The iteration is done in reverse
2280   // order (high to low) so the regions will be removed in ascending order.
2281 
2282   const ParallelCompactData& sd = PSParallelCompact::summary_data();
2283 
2284   // id + 1 is used to test termination so unsigned  can
2285   // be used with an old_space_id == 0.
2286   FillableRegionLogger region_logger;
2287   for (unsigned int id = to_space_id; id + 1 > old_space_id; --id) {
2288     SpaceInfo* const space_info = _space_info + id;
2289     HeapWord* const new_top = space_info->new_top();
2290 
2291     const size_t beg_region = sd.addr_to_region_idx(space_info->dense_prefix());
2292     const size_t end_region =
2293       sd.addr_to_region_idx(sd.region_align_up(new_top));
2294 
2295     for (size_t cur = end_region - 1; cur + 1 > beg_region; --cur) {
2296       if (sd.region(cur)->claim_unsafe()) {
2297         ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(worker_id);
2298         bool result = sd.region(cur)->mark_normal();
2299         assert(result, "Must succeed at this point.");
2300         cm->region_stack()->push(cur);
2301         region_logger.handle(cur);
2302         // Assign regions to tasks in round-robin fashion.
2303         if (++worker_id == parallel_gc_threads) {
2304           worker_id = 0;
2305         }
2306       }
2307     }
2308     region_logger.print_line();
2309   }
2310 }
2311 
2312 class TaskQueue : StackObj {
2313   volatile uint _counter;
2314   uint _size;
2315   uint _insert_index;
2316   PSParallelCompact::UpdateDensePrefixTask* _backing_array;
2317 public:
2318   explicit TaskQueue(uint size) : _counter(0), _size(size), _insert_index(0), _backing_array(NULL) {
2319     _backing_array = NEW_C_HEAP_ARRAY(PSParallelCompact::UpdateDensePrefixTask, _size, mtGC);
2320   }
2321   ~TaskQueue() {
2322     assert(_counter >= _insert_index, "not all queue elements were claimed");
2323     FREE_C_HEAP_ARRAY(T, _backing_array);
2324   }
2325 
2326   void push(const PSParallelCompact::UpdateDensePrefixTask& value) {
2327     assert(_insert_index < _size, "too small backing array");
2328     _backing_array[_insert_index++] = value;
2329   }
2330 
2331   bool try_claim(PSParallelCompact::UpdateDensePrefixTask& reference) {
2332     uint claimed = Atomic::fetch_and_add(&_counter, 1u);
2333     if (claimed < _insert_index) {
2334       reference = _backing_array[claimed];
2335       return true;
2336     } else {
2337       return false;
2338     }
2339   }
2340 };
2341 
2342 #define PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING 4
2343 
2344 void PSParallelCompact::enqueue_dense_prefix_tasks(TaskQueue& task_queue,
2345                                                    uint parallel_gc_threads) {
2346   GCTraceTime(Trace, gc, phases) tm("Dense Prefix Task Setup", &_gc_timer);
2347 
2348   ParallelCompactData& sd = PSParallelCompact::summary_data();
2349 
2350   // Iterate over all the spaces adding tasks for updating
2351   // regions in the dense prefix.  Assume that 1 gc thread
2352   // will work on opening the gaps and the remaining gc threads
2353   // will work on the dense prefix.
2354   unsigned int space_id;
2355   for (space_id = old_space_id; space_id < last_space_id; ++ space_id) {
2356     HeapWord* const dense_prefix_end = _space_info[space_id].dense_prefix();
2357     const MutableSpace* const space = _space_info[space_id].space();
2358 
2359     if (dense_prefix_end == space->bottom()) {
2360       // There is no dense prefix for this space.
2361       continue;
2362     }
2363 
2364     // The dense prefix is before this region.
2365     size_t region_index_end_dense_prefix =
2366         sd.addr_to_region_idx(dense_prefix_end);
2367     RegionData* const dense_prefix_cp =
2368       sd.region(region_index_end_dense_prefix);
2369     assert(dense_prefix_end == space->end() ||
2370            dense_prefix_cp->available() ||
2371            dense_prefix_cp->claimed(),
2372            "The region after the dense prefix should always be ready to fill");
2373 
2374     size_t region_index_start = sd.addr_to_region_idx(space->bottom());
2375 
2376     // Is there dense prefix work?
2377     size_t total_dense_prefix_regions =
2378       region_index_end_dense_prefix - region_index_start;
2379     // How many regions of the dense prefix should be given to
2380     // each thread?
2381     if (total_dense_prefix_regions > 0) {
2382       uint tasks_for_dense_prefix = 1;
2383       if (total_dense_prefix_regions <=
2384           (parallel_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING)) {
2385         // Don't over partition.  This assumes that
2386         // PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING is a small integer value
2387         // so there are not many regions to process.
2388         tasks_for_dense_prefix = parallel_gc_threads;
2389       } else {
2390         // Over partition
2391         tasks_for_dense_prefix = parallel_gc_threads *
2392           PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING;
2393       }
2394       size_t regions_per_thread = total_dense_prefix_regions /
2395         tasks_for_dense_prefix;
2396       // Give each thread at least 1 region.
2397       if (regions_per_thread == 0) {
2398         regions_per_thread = 1;
2399       }
2400 
2401       for (uint k = 0; k < tasks_for_dense_prefix; k++) {
2402         if (region_index_start >= region_index_end_dense_prefix) {
2403           break;
2404         }
2405         // region_index_end is not processed
2406         size_t region_index_end = MIN2(region_index_start + regions_per_thread,
2407                                        region_index_end_dense_prefix);
2408         task_queue.push(UpdateDensePrefixTask(SpaceId(space_id),
2409                                               region_index_start,
2410                                               region_index_end));
2411         region_index_start = region_index_end;
2412       }
2413     }
2414     // This gets any part of the dense prefix that did not
2415     // fit evenly.
2416     if (region_index_start < region_index_end_dense_prefix) {
2417       task_queue.push(UpdateDensePrefixTask(SpaceId(space_id),
2418                                             region_index_start,
2419                                             region_index_end_dense_prefix));
2420     }
2421   }
2422 }
2423 
2424 #ifdef ASSERT
2425 // Write a histogram of the number of times the block table was filled for a
2426 // region.
2427 void PSParallelCompact::write_block_fill_histogram()
2428 {
2429   if (!log_develop_is_enabled(Trace, gc, compaction)) {
2430     return;
2431   }
2432 
2433   Log(gc, compaction) log;
2434   ResourceMark rm;
2435   LogStream ls(log.trace());
2436   outputStream* out = &ls;
2437 
2438   typedef ParallelCompactData::RegionData rd_t;
2439   ParallelCompactData& sd = summary_data();
2440 
2441   for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2442     MutableSpace* const spc = _space_info[id].space();
2443     if (spc->bottom() != spc->top()) {
2444       const rd_t* const beg = sd.addr_to_region_ptr(spc->bottom());
2445       HeapWord* const top_aligned_up = sd.region_align_up(spc->top());
2446       const rd_t* const end = sd.addr_to_region_ptr(top_aligned_up);
2447 
2448       size_t histo[5] = { 0, 0, 0, 0, 0 };
2449       const size_t histo_len = sizeof(histo) / sizeof(size_t);
2450       const size_t region_cnt = pointer_delta(end, beg, sizeof(rd_t));
2451 
2452       for (const rd_t* cur = beg; cur < end; ++cur) {
2453         ++histo[MIN2(cur->blocks_filled_count(), histo_len - 1)];
2454       }
2455       out->print("Block fill histogram: %u %-4s" SIZE_FORMAT_W(5), id, space_names[id], region_cnt);
2456       for (size_t i = 0; i < histo_len; ++i) {
2457         out->print(" " SIZE_FORMAT_W(5) " %5.1f%%",
2458                    histo[i], 100.0 * histo[i] / region_cnt);
2459       }
2460       out->cr();
2461     }
2462   }
2463 }
2464 #endif // #ifdef ASSERT
2465 
2466 static void compaction_with_stealing_work(TaskTerminator* terminator, uint worker_id) {
2467   assert(ParallelScavengeHeap::heap()->is_gc_active(), "called outside gc");
2468 
2469   ParCompactionManager* cm =
2470     ParCompactionManager::gc_thread_compaction_manager(worker_id);
2471 
2472   // Drain the stacks that have been preloaded with regions
2473   // that are ready to fill.
2474 
2475   cm->drain_region_stacks();
2476 
2477   guarantee(cm->region_stack()->is_empty(), "Not empty");
2478 
2479   size_t region_index = 0;
2480 
2481   while (true) {
2482     if (ParCompactionManager::steal(worker_id, region_index)) {
2483       PSParallelCompact::fill_and_update_region(cm, region_index);
2484       cm->drain_region_stacks();
2485     } else if (PSParallelCompact::steal_unavailable_region(cm, region_index)) {
2486       // Fill and update an unavailable region with the help of a shadow region
2487       PSParallelCompact::fill_and_update_shadow_region(cm, region_index);
2488       cm->drain_region_stacks();
2489     } else {
2490       if (terminator->offer_termination()) {
2491         break;
2492       }
2493       // Go around again.
2494     }
2495   }
2496 }
2497 
2498 class UpdateDensePrefixAndCompactionTask: public AbstractGangTask {
2499   TaskQueue& _tq;
2500   TaskTerminator _terminator;
2501   uint _active_workers;
2502 
2503 public:
2504   UpdateDensePrefixAndCompactionTask(TaskQueue& tq, uint active_workers) :
2505       AbstractGangTask("UpdateDensePrefixAndCompactionTask"),
2506       _tq(tq),
2507       _terminator(active_workers, ParCompactionManager::region_task_queues()),
2508       _active_workers(active_workers) {
2509   }
2510   virtual void work(uint worker_id) {
2511     ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(worker_id);
2512 
2513     for (PSParallelCompact::UpdateDensePrefixTask task; _tq.try_claim(task); /* empty */) {
2514       PSParallelCompact::update_and_deadwood_in_dense_prefix(cm,
2515                                                              task._space_id,
2516                                                              task._region_index_start,
2517                                                              task._region_index_end);
2518     }
2519 
2520     // Once a thread has drained it's stack, it should try to steal regions from
2521     // other threads.
2522     compaction_with_stealing_work(&_terminator, worker_id);
2523   }
2524 };
2525 
2526 void PSParallelCompact::compact() {
2527   GCTraceTime(Info, gc, phases) tm("Compaction Phase", &_gc_timer);
2528 
2529   ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
2530   PSOldGen* old_gen = heap->old_gen();
2531   old_gen->start_array()->reset();
2532   uint active_gc_threads = ParallelScavengeHeap::heap()->workers().active_workers();
2533 
2534   // for [0..last_space_id)
2535   //     for [0..active_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING)
2536   //         push
2537   //     push
2538   //
2539   // max push count is thus: last_space_id * (active_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING + 1)
2540   TaskQueue task_queue(last_space_id * (active_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING + 1));
2541   initialize_shadow_regions(active_gc_threads);
2542   prepare_region_draining_tasks(active_gc_threads);
2543   enqueue_dense_prefix_tasks(task_queue, active_gc_threads);
2544 
2545   {
2546     GCTraceTime(Trace, gc, phases) tm("Par Compact", &_gc_timer);
2547 
2548     UpdateDensePrefixAndCompactionTask task(task_queue, active_gc_threads);
2549     ParallelScavengeHeap::heap()->workers().run_task(&task);
2550 
2551 #ifdef  ASSERT
2552     // Verify that all regions have been processed before the deferred updates.
2553     for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2554       verify_complete(SpaceId(id));
2555     }
2556 #endif
2557   }
2558 
2559   {
2560     GCTraceTime(Trace, gc, phases) tm("Deferred Updates", &_gc_timer);
2561     // Update the deferred objects, if any. In principle, any compaction
2562     // manager can be used. However, since the current thread is VM thread, we
2563     // use the rightful one to keep the verification logic happy.
2564     ParCompactionManager* cm = ParCompactionManager::get_vmthread_cm();
2565     for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2566       update_deferred_objects(cm, SpaceId(id));
2567     }
2568   }
2569 
2570   DEBUG_ONLY(write_block_fill_histogram());
2571 }
2572 
2573 #ifdef  ASSERT
2574 void PSParallelCompact::verify_complete(SpaceId space_id) {
2575   // All Regions between space bottom() to new_top() should be marked as filled
2576   // and all Regions between new_top() and top() should be available (i.e.,
2577   // should have been emptied).
2578   ParallelCompactData& sd = summary_data();
2579   SpaceInfo si = _space_info[space_id];
2580   HeapWord* new_top_addr = sd.region_align_up(si.new_top());
2581   HeapWord* old_top_addr = sd.region_align_up(si.space()->top());
2582   const size_t beg_region = sd.addr_to_region_idx(si.space()->bottom());
2583   const size_t new_top_region = sd.addr_to_region_idx(new_top_addr);
2584   const size_t old_top_region = sd.addr_to_region_idx(old_top_addr);
2585 
2586   bool issued_a_warning = false;
2587 
2588   size_t cur_region;
2589   for (cur_region = beg_region; cur_region < new_top_region; ++cur_region) {
2590     const RegionData* const c = sd.region(cur_region);
2591     if (!c->completed()) {
2592       log_warning(gc)("region " SIZE_FORMAT " not filled: destination_count=%u",
2593                       cur_region, c->destination_count());
2594       issued_a_warning = true;
2595     }
2596   }
2597 
2598   for (cur_region = new_top_region; cur_region < old_top_region; ++cur_region) {
2599     const RegionData* const c = sd.region(cur_region);
2600     if (!c->available()) {
2601       log_warning(gc)("region " SIZE_FORMAT " not empty: destination_count=%u",
2602                       cur_region, c->destination_count());
2603       issued_a_warning = true;
2604     }
2605   }
2606 
2607   if (issued_a_warning) {
2608     print_region_ranges();
2609   }
2610 }
2611 #endif  // #ifdef ASSERT
2612 
2613 inline void UpdateOnlyClosure::do_addr(HeapWord* addr) {
2614   _start_array->allocate_block(addr);
2615   compaction_manager()->update_contents(cast_to_oop(addr));
2616 }
2617 
2618 // Update interior oops in the ranges of regions [beg_region, end_region).
2619 void
2620 PSParallelCompact::update_and_deadwood_in_dense_prefix(ParCompactionManager* cm,
2621                                                        SpaceId space_id,
2622                                                        size_t beg_region,
2623                                                        size_t end_region) {
2624   ParallelCompactData& sd = summary_data();
2625   ParMarkBitMap* const mbm = mark_bitmap();
2626 
2627   HeapWord* beg_addr = sd.region_to_addr(beg_region);
2628   HeapWord* const end_addr = sd.region_to_addr(end_region);
2629   assert(beg_region <= end_region, "bad region range");
2630   assert(end_addr <= dense_prefix(space_id), "not in the dense prefix");
2631 
2632 #ifdef  ASSERT
2633   // Claim the regions to avoid triggering an assert when they are marked as
2634   // filled.
2635   for (size_t claim_region = beg_region; claim_region < end_region; ++claim_region) {
2636     assert(sd.region(claim_region)->claim_unsafe(), "claim() failed");
2637   }
2638 #endif  // #ifdef ASSERT
2639 
2640   if (beg_addr != space(space_id)->bottom()) {
2641     // Find the first live object or block of dead space that *starts* in this
2642     // range of regions.  If a partial object crosses onto the region, skip it;
2643     // it will be marked for 'deferred update' when the object head is
2644     // processed.  If dead space crosses onto the region, it is also skipped; it
2645     // will be filled when the prior region is processed.  If neither of those
2646     // apply, the first word in the region is the start of a live object or dead
2647     // space.
2648     assert(beg_addr > space(space_id)->bottom(), "sanity");
2649     const RegionData* const cp = sd.region(beg_region);
2650     if (cp->partial_obj_size() != 0) {
2651       beg_addr = sd.partial_obj_end(beg_region);
2652     } else if (dead_space_crosses_boundary(cp, mbm->addr_to_bit(beg_addr))) {
2653       beg_addr = mbm->find_obj_beg(beg_addr, end_addr);
2654     }
2655   }
2656 
2657   if (beg_addr < end_addr) {
2658     // A live object or block of dead space starts in this range of Regions.
2659      HeapWord* const dense_prefix_end = dense_prefix(space_id);
2660 
2661     // Create closures and iterate.
2662     UpdateOnlyClosure update_closure(mbm, cm, space_id);
2663     FillClosure fill_closure(cm, space_id);
2664     ParMarkBitMap::IterationStatus status;
2665     status = mbm->iterate(&update_closure, &fill_closure, beg_addr, end_addr,
2666                           dense_prefix_end);
2667     if (status == ParMarkBitMap::incomplete) {
2668       update_closure.do_addr(update_closure.source());
2669     }
2670   }
2671 
2672   // Mark the regions as filled.
2673   RegionData* const beg_cp = sd.region(beg_region);
2674   RegionData* const end_cp = sd.region(end_region);
2675   for (RegionData* cp = beg_cp; cp < end_cp; ++cp) {
2676     cp->set_completed();
2677   }
2678 }
2679 
2680 // Return the SpaceId for the space containing addr.  If addr is not in the
2681 // heap, last_space_id is returned.  In debug mode it expects the address to be
2682 // in the heap and asserts such.
2683 PSParallelCompact::SpaceId PSParallelCompact::space_id(HeapWord* addr) {
2684   assert(ParallelScavengeHeap::heap()->is_in_reserved(addr), "addr not in the heap");
2685 
2686   for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2687     if (_space_info[id].space()->contains(addr)) {
2688       return SpaceId(id);
2689     }
2690   }
2691 
2692   assert(false, "no space contains the addr");
2693   return last_space_id;
2694 }
2695 
2696 void PSParallelCompact::update_deferred_objects(ParCompactionManager* cm,
2697                                                 SpaceId id) {
2698   assert(id < last_space_id, "bad space id");
2699 
2700   ParallelCompactData& sd = summary_data();
2701   const SpaceInfo* const space_info = _space_info + id;
2702   ObjectStartArray* const start_array = space_info->start_array();
2703 
2704   const MutableSpace* const space = space_info->space();
2705   assert(space_info->dense_prefix() >= space->bottom(), "dense_prefix not set");
2706   HeapWord* const beg_addr = space_info->dense_prefix();
2707   HeapWord* const end_addr = sd.region_align_up(space_info->new_top());
2708 
2709   const RegionData* const beg_region = sd.addr_to_region_ptr(beg_addr);
2710   const RegionData* const end_region = sd.addr_to_region_ptr(end_addr);
2711   const RegionData* cur_region;
2712   for (cur_region = beg_region; cur_region < end_region; ++cur_region) {
2713     HeapWord* const addr = cur_region->deferred_obj_addr();
2714     if (addr != NULL) {
2715       if (start_array != NULL) {
2716         start_array->allocate_block(addr);
2717       }
2718       cm->update_contents(cast_to_oop(addr));
2719       assert(oopDesc::is_oop_or_null(cast_to_oop(addr)), "Expected an oop or NULL at " PTR_FORMAT, p2i(cast_to_oop(addr)));
2720     }
2721   }
2722 }
2723 
2724 // Skip over count live words starting from beg, and return the address of the
2725 // next live word.  Unless marked, the word corresponding to beg is assumed to
2726 // be dead.  Callers must either ensure beg does not correspond to the middle of
2727 // an object, or account for those live words in some other way.  Callers must
2728 // also ensure that there are enough live words in the range [beg, end) to skip.
2729 HeapWord*
2730 PSParallelCompact::skip_live_words(HeapWord* beg, HeapWord* end, size_t count)
2731 {
2732   assert(count > 0, "sanity");
2733 
2734   ParMarkBitMap* m = mark_bitmap();
2735   idx_t bits_to_skip = m->words_to_bits(count);
2736   idx_t cur_beg = m->addr_to_bit(beg);
2737   const idx_t search_end = m->align_range_end(m->addr_to_bit(end));
2738 
2739   do {
2740     cur_beg = m->find_obj_beg(cur_beg, search_end);
2741     idx_t cur_end = m->find_obj_end(cur_beg, search_end);
2742     const size_t obj_bits = cur_end - cur_beg + 1;
2743     if (obj_bits > bits_to_skip) {
2744       return m->bit_to_addr(cur_beg + bits_to_skip);
2745     }
2746     bits_to_skip -= obj_bits;
2747     cur_beg = cur_end + 1;
2748   } while (bits_to_skip > 0);
2749 
2750   // Skipping the desired number of words landed just past the end of an object.
2751   // Find the start of the next object.
2752   cur_beg = m->find_obj_beg(cur_beg, search_end);
2753   assert(cur_beg < m->addr_to_bit(end), "not enough live words to skip");
2754   return m->bit_to_addr(cur_beg);
2755 }
2756 
2757 HeapWord* PSParallelCompact::first_src_addr(HeapWord* const dest_addr,
2758                                             SpaceId src_space_id,
2759                                             size_t src_region_idx)
2760 {
2761   assert(summary_data().is_region_aligned(dest_addr), "not aligned");
2762 
2763   const SplitInfo& split_info = _space_info[src_space_id].split_info();
2764   if (split_info.dest_region_addr() == dest_addr) {
2765     // The partial object ending at the split point contains the first word to
2766     // be copied to dest_addr.
2767     return split_info.first_src_addr();
2768   }
2769 
2770   const ParallelCompactData& sd = summary_data();
2771   ParMarkBitMap* const bitmap = mark_bitmap();
2772   const size_t RegionSize = ParallelCompactData::RegionSize;
2773 
2774   assert(sd.is_region_aligned(dest_addr), "not aligned");
2775   const RegionData* const src_region_ptr = sd.region(src_region_idx);
2776   const size_t partial_obj_size = src_region_ptr->partial_obj_size();
2777   HeapWord* const src_region_destination = src_region_ptr->destination();
2778 
2779   assert(dest_addr >= src_region_destination, "wrong src region");
2780   assert(src_region_ptr->data_size() > 0, "src region cannot be empty");
2781 
2782   HeapWord* const src_region_beg = sd.region_to_addr(src_region_idx);
2783   HeapWord* const src_region_end = src_region_beg + RegionSize;
2784 
2785   HeapWord* addr = src_region_beg;
2786   if (dest_addr == src_region_destination) {
2787     // Return the first live word in the source region.
2788     if (partial_obj_size == 0) {
2789       addr = bitmap->find_obj_beg(addr, src_region_end);
2790       assert(addr < src_region_end, "no objects start in src region");
2791     }
2792     return addr;
2793   }
2794 
2795   // Must skip some live data.
2796   size_t words_to_skip = dest_addr - src_region_destination;
2797   assert(src_region_ptr->data_size() > words_to_skip, "wrong src region");
2798 
2799   if (partial_obj_size >= words_to_skip) {
2800     // All the live words to skip are part of the partial object.
2801     addr += words_to_skip;
2802     if (partial_obj_size == words_to_skip) {
2803       // Find the first live word past the partial object.
2804       addr = bitmap->find_obj_beg(addr, src_region_end);
2805       assert(addr < src_region_end, "wrong src region");
2806     }
2807     return addr;
2808   }
2809 
2810   // Skip over the partial object (if any).
2811   if (partial_obj_size != 0) {
2812     words_to_skip -= partial_obj_size;
2813     addr += partial_obj_size;
2814   }
2815 
2816   // Skip over live words due to objects that start in the region.
2817   addr = skip_live_words(addr, src_region_end, words_to_skip);
2818   assert(addr < src_region_end, "wrong src region");
2819   return addr;
2820 }
2821 
2822 void PSParallelCompact::decrement_destination_counts(ParCompactionManager* cm,
2823                                                      SpaceId src_space_id,
2824                                                      size_t beg_region,
2825                                                      HeapWord* end_addr)
2826 {
2827   ParallelCompactData& sd = summary_data();
2828 
2829 #ifdef ASSERT
2830   MutableSpace* const src_space = _space_info[src_space_id].space();
2831   HeapWord* const beg_addr = sd.region_to_addr(beg_region);
2832   assert(src_space->contains(beg_addr) || beg_addr == src_space->end(),
2833          "src_space_id does not match beg_addr");
2834   assert(src_space->contains(end_addr) || end_addr == src_space->end(),
2835          "src_space_id does not match end_addr");
2836 #endif // #ifdef ASSERT
2837 
2838   RegionData* const beg = sd.region(beg_region);
2839   RegionData* const end = sd.addr_to_region_ptr(sd.region_align_up(end_addr));
2840 
2841   // Regions up to new_top() are enqueued if they become available.
2842   HeapWord* const new_top = _space_info[src_space_id].new_top();
2843   RegionData* const enqueue_end =
2844     sd.addr_to_region_ptr(sd.region_align_up(new_top));
2845 
2846   for (RegionData* cur = beg; cur < end; ++cur) {
2847     assert(cur->data_size() > 0, "region must have live data");
2848     cur->decrement_destination_count();
2849     if (cur < enqueue_end && cur->available() && cur->claim()) {
2850       if (cur->mark_normal()) {
2851         cm->push_region(sd.region(cur));
2852       } else if (cur->mark_copied()) {
2853         // Try to copy the content of the shadow region back to its corresponding
2854         // heap region if the shadow region is filled. Otherwise, the GC thread
2855         // fills the shadow region will copy the data back (see
2856         // MoveAndUpdateShadowClosure::complete_region).
2857         copy_back(sd.region_to_addr(cur->shadow_region()), sd.region_to_addr(cur));
2858         ParCompactionManager::push_shadow_region_mt_safe(cur->shadow_region());
2859         cur->set_completed();
2860       }
2861     }
2862   }
2863 }
2864 
2865 size_t PSParallelCompact::next_src_region(MoveAndUpdateClosure& closure,
2866                                           SpaceId& src_space_id,
2867                                           HeapWord*& src_space_top,
2868                                           HeapWord* end_addr)
2869 {
2870   typedef ParallelCompactData::RegionData RegionData;
2871 
2872   ParallelCompactData& sd = PSParallelCompact::summary_data();
2873   const size_t region_size = ParallelCompactData::RegionSize;
2874 
2875   size_t src_region_idx = 0;
2876 
2877   // Skip empty regions (if any) up to the top of the space.
2878   HeapWord* const src_aligned_up = sd.region_align_up(end_addr);
2879   RegionData* src_region_ptr = sd.addr_to_region_ptr(src_aligned_up);
2880   HeapWord* const top_aligned_up = sd.region_align_up(src_space_top);
2881   const RegionData* const top_region_ptr =
2882     sd.addr_to_region_ptr(top_aligned_up);
2883   while (src_region_ptr < top_region_ptr && src_region_ptr->data_size() == 0) {
2884     ++src_region_ptr;
2885   }
2886 
2887   if (src_region_ptr < top_region_ptr) {
2888     // The next source region is in the current space.  Update src_region_idx
2889     // and the source address to match src_region_ptr.
2890     src_region_idx = sd.region(src_region_ptr);
2891     HeapWord* const src_region_addr = sd.region_to_addr(src_region_idx);
2892     if (src_region_addr > closure.source()) {
2893       closure.set_source(src_region_addr);
2894     }
2895     return src_region_idx;
2896   }
2897 
2898   // Switch to a new source space and find the first non-empty region.
2899   unsigned int space_id = src_space_id + 1;
2900   assert(space_id < last_space_id, "not enough spaces");
2901 
2902   HeapWord* const destination = closure.destination();
2903 
2904   do {
2905     MutableSpace* space = _space_info[space_id].space();
2906     HeapWord* const bottom = space->bottom();
2907     const RegionData* const bottom_cp = sd.addr_to_region_ptr(bottom);
2908 
2909     // Iterate over the spaces that do not compact into themselves.
2910     if (bottom_cp->destination() != bottom) {
2911       HeapWord* const top_aligned_up = sd.region_align_up(space->top());
2912       const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up);
2913 
2914       for (const RegionData* src_cp = bottom_cp; src_cp < top_cp; ++src_cp) {
2915         if (src_cp->live_obj_size() > 0) {
2916           // Found it.
2917           assert(src_cp->destination() == destination,
2918                  "first live obj in the space must match the destination");
2919           assert(src_cp->partial_obj_size() == 0,
2920                  "a space cannot begin with a partial obj");
2921 
2922           src_space_id = SpaceId(space_id);
2923           src_space_top = space->top();
2924           const size_t src_region_idx = sd.region(src_cp);
2925           closure.set_source(sd.region_to_addr(src_region_idx));
2926           return src_region_idx;
2927         } else {
2928           assert(src_cp->data_size() == 0, "sanity");
2929         }
2930       }
2931     }
2932   } while (++space_id < last_space_id);
2933 
2934   assert(false, "no source region was found");
2935   return 0;
2936 }
2937 
2938 void PSParallelCompact::fill_region(ParCompactionManager* cm, MoveAndUpdateClosure& closure, size_t region_idx)
2939 {
2940   typedef ParMarkBitMap::IterationStatus IterationStatus;
2941   ParMarkBitMap* const bitmap = mark_bitmap();
2942   ParallelCompactData& sd = summary_data();
2943   RegionData* const region_ptr = sd.region(region_idx);
2944 
2945   // Get the source region and related info.
2946   size_t src_region_idx = region_ptr->source_region();
2947   SpaceId src_space_id = space_id(sd.region_to_addr(src_region_idx));
2948   HeapWord* src_space_top = _space_info[src_space_id].space()->top();
2949   HeapWord* dest_addr = sd.region_to_addr(region_idx);
2950 
2951   closure.set_source(first_src_addr(dest_addr, src_space_id, src_region_idx));
2952 
2953   // Adjust src_region_idx to prepare for decrementing destination counts (the
2954   // destination count is not decremented when a region is copied to itself).
2955   if (src_region_idx == region_idx) {
2956     src_region_idx += 1;
2957   }
2958 
2959   if (bitmap->is_unmarked(closure.source())) {
2960     // The first source word is in the middle of an object; copy the remainder
2961     // of the object or as much as will fit.  The fact that pointer updates were
2962     // deferred will be noted when the object header is processed.
2963     HeapWord* const old_src_addr = closure.source();
2964     closure.copy_partial_obj();
2965     if (closure.is_full()) {
2966       decrement_destination_counts(cm, src_space_id, src_region_idx,
2967                                    closure.source());
2968       region_ptr->set_deferred_obj_addr(NULL);
2969       closure.complete_region(cm, dest_addr, region_ptr);
2970       return;
2971     }
2972 
2973     HeapWord* const end_addr = sd.region_align_down(closure.source());
2974     if (sd.region_align_down(old_src_addr) != end_addr) {
2975       // The partial object was copied from more than one source region.
2976       decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr);
2977 
2978       // Move to the next source region, possibly switching spaces as well.  All
2979       // args except end_addr may be modified.
2980       src_region_idx = next_src_region(closure, src_space_id, src_space_top,
2981                                        end_addr);
2982     }
2983   }
2984 
2985   do {
2986     HeapWord* const cur_addr = closure.source();
2987     HeapWord* const end_addr = MIN2(sd.region_align_up(cur_addr + 1),
2988                                     src_space_top);
2989     IterationStatus status = bitmap->iterate(&closure, cur_addr, end_addr);
2990 
2991     if (status == ParMarkBitMap::incomplete) {
2992       // The last obj that starts in the source region does not end in the
2993       // region.
2994       assert(closure.source() < end_addr, "sanity");
2995       HeapWord* const obj_beg = closure.source();
2996       HeapWord* const range_end = MIN2(obj_beg + closure.words_remaining(),
2997                                        src_space_top);
2998       HeapWord* const obj_end = bitmap->find_obj_end(obj_beg, range_end);
2999       if (obj_end < range_end) {
3000         // The end was found; the entire object will fit.
3001         status = closure.do_addr(obj_beg, bitmap->obj_size(obj_beg, obj_end));
3002         assert(status != ParMarkBitMap::would_overflow, "sanity");
3003       } else {
3004         // The end was not found; the object will not fit.
3005         assert(range_end < src_space_top, "obj cannot cross space boundary");
3006         status = ParMarkBitMap::would_overflow;
3007       }
3008     }
3009 
3010     if (status == ParMarkBitMap::would_overflow) {
3011       // The last object did not fit.  Note that interior oop updates were
3012       // deferred, then copy enough of the object to fill the region.
3013       region_ptr->set_deferred_obj_addr(closure.destination());
3014       status = closure.copy_until_full(); // copies from closure.source()
3015 
3016       decrement_destination_counts(cm, src_space_id, src_region_idx,
3017                                    closure.source());
3018       closure.complete_region(cm, dest_addr, region_ptr);
3019       return;
3020     }
3021 
3022     if (status == ParMarkBitMap::full) {
3023       decrement_destination_counts(cm, src_space_id, src_region_idx,
3024                                    closure.source());
3025       region_ptr->set_deferred_obj_addr(NULL);
3026       closure.complete_region(cm, dest_addr, region_ptr);
3027       return;
3028     }
3029 
3030     decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr);
3031 
3032     // Move to the next source region, possibly switching spaces as well.  All
3033     // args except end_addr may be modified.
3034     src_region_idx = next_src_region(closure, src_space_id, src_space_top,
3035                                      end_addr);
3036   } while (true);
3037 }
3038 
3039 void PSParallelCompact::fill_and_update_region(ParCompactionManager* cm, size_t region_idx)
3040 {
3041   MoveAndUpdateClosure cl(mark_bitmap(), cm, region_idx);
3042   fill_region(cm, cl, region_idx);
3043 }
3044 
3045 void PSParallelCompact::fill_and_update_shadow_region(ParCompactionManager* cm, size_t region_idx)
3046 {
3047   // Get a shadow region first
3048   ParallelCompactData& sd = summary_data();
3049   RegionData* const region_ptr = sd.region(region_idx);
3050   size_t shadow_region = ParCompactionManager::pop_shadow_region_mt_safe(region_ptr);
3051   // The InvalidShadow return value indicates the corresponding heap region is available,
3052   // so use MoveAndUpdateClosure to fill the normal region. Otherwise, use
3053   // MoveAndUpdateShadowClosure to fill the acquired shadow region.
3054   if (shadow_region == ParCompactionManager::InvalidShadow) {
3055     MoveAndUpdateClosure cl(mark_bitmap(), cm, region_idx);
3056     region_ptr->shadow_to_normal();
3057     return fill_region(cm, cl, region_idx);
3058   } else {
3059     MoveAndUpdateShadowClosure cl(mark_bitmap(), cm, region_idx, shadow_region);
3060     return fill_region(cm, cl, region_idx);
3061   }
3062 }
3063 
3064 void PSParallelCompact::copy_back(HeapWord *shadow_addr, HeapWord *region_addr)
3065 {
3066   Copy::aligned_conjoint_words(shadow_addr, region_addr, _summary_data.RegionSize);
3067 }
3068 
3069 bool PSParallelCompact::steal_unavailable_region(ParCompactionManager* cm, size_t &region_idx)
3070 {
3071   size_t next = cm->next_shadow_region();
3072   ParallelCompactData& sd = summary_data();
3073   size_t old_new_top = sd.addr_to_region_idx(_space_info[old_space_id].new_top());
3074   uint active_gc_threads = ParallelScavengeHeap::heap()->workers().active_workers();
3075 
3076   while (next < old_new_top) {
3077     if (sd.region(next)->mark_shadow()) {
3078       region_idx = next;
3079       return true;
3080     }
3081     next = cm->move_next_shadow_region_by(active_gc_threads);
3082   }
3083 
3084   return false;
3085 }
3086 
3087 // The shadow region is an optimization to address region dependencies in full GC. The basic
3088 // idea is making more regions available by temporally storing their live objects in empty
3089 // shadow regions to resolve dependencies between them and the destination regions. Therefore,
3090 // GC threads need not wait destination regions to be available before processing sources.
3091 //
3092 // A typical workflow would be:
3093 // After draining its own stack and failing to steal from others, a GC worker would pick an
3094 // unavailable region (destination count > 0) and get a shadow region. Then the worker fills
3095 // the shadow region by copying live objects from source regions of the unavailable one. Once
3096 // the unavailable region becomes available, the data in the shadow region will be copied back.
3097 // Shadow regions are empty regions in the to-space and regions between top and end of other spaces.
3098 void PSParallelCompact::initialize_shadow_regions(uint parallel_gc_threads)
3099 {
3100   const ParallelCompactData& sd = PSParallelCompact::summary_data();
3101 
3102   for (unsigned int id = old_space_id; id < last_space_id; ++id) {
3103     SpaceInfo* const space_info = _space_info + id;
3104     MutableSpace* const space = space_info->space();
3105 
3106     const size_t beg_region =
3107       sd.addr_to_region_idx(sd.region_align_up(MAX2(space_info->new_top(), space->top())));
3108     const size_t end_region =
3109       sd.addr_to_region_idx(sd.region_align_down(space->end()));
3110 
3111     for (size_t cur = beg_region; cur < end_region; ++cur) {
3112       ParCompactionManager::push_shadow_region(cur);
3113     }
3114   }
3115 
3116   size_t beg_region = sd.addr_to_region_idx(_space_info[old_space_id].dense_prefix());
3117   for (uint i = 0; i < parallel_gc_threads; i++) {
3118     ParCompactionManager *cm = ParCompactionManager::gc_thread_compaction_manager(i);
3119     cm->set_next_shadow_region(beg_region + i);
3120   }
3121 }
3122 
3123 void PSParallelCompact::fill_blocks(size_t region_idx)
3124 {
3125   // Fill in the block table elements for the specified region.  Each block
3126   // table element holds the number of live words in the region that are to the
3127   // left of the first object that starts in the block.  Thus only blocks in
3128   // which an object starts need to be filled.
3129   //
3130   // The algorithm scans the section of the bitmap that corresponds to the
3131   // region, keeping a running total of the live words.  When an object start is
3132   // found, if it's the first to start in the block that contains it, the
3133   // current total is written to the block table element.
3134   const size_t Log2BlockSize = ParallelCompactData::Log2BlockSize;
3135   const size_t Log2RegionSize = ParallelCompactData::Log2RegionSize;
3136   const size_t RegionSize = ParallelCompactData::RegionSize;
3137 
3138   ParallelCompactData& sd = summary_data();
3139   const size_t partial_obj_size = sd.region(region_idx)->partial_obj_size();
3140   if (partial_obj_size >= RegionSize) {
3141     return; // No objects start in this region.
3142   }
3143 
3144   // Ensure the first loop iteration decides that the block has changed.
3145   size_t cur_block = sd.block_count();
3146 
3147   const ParMarkBitMap* const bitmap = mark_bitmap();
3148 
3149   const size_t Log2BitsPerBlock = Log2BlockSize - LogMinObjAlignment;
3150   assert((size_t)1 << Log2BitsPerBlock ==
3151          bitmap->words_to_bits(ParallelCompactData::BlockSize), "sanity");
3152 
3153   size_t beg_bit = bitmap->words_to_bits(region_idx << Log2RegionSize);
3154   const size_t range_end = beg_bit + bitmap->words_to_bits(RegionSize);
3155   size_t live_bits = bitmap->words_to_bits(partial_obj_size);
3156   beg_bit = bitmap->find_obj_beg(beg_bit + live_bits, range_end);
3157   while (beg_bit < range_end) {
3158     const size_t new_block = beg_bit >> Log2BitsPerBlock;
3159     if (new_block != cur_block) {
3160       cur_block = new_block;
3161       sd.block(cur_block)->set_offset(bitmap->bits_to_words(live_bits));
3162     }
3163 
3164     const size_t end_bit = bitmap->find_obj_end(beg_bit, range_end);
3165     if (end_bit < range_end - 1) {
3166       live_bits += end_bit - beg_bit + 1;
3167       beg_bit = bitmap->find_obj_beg(end_bit + 1, range_end);
3168     } else {
3169       return;
3170     }
3171   }
3172 }
3173 
3174 ParMarkBitMap::IterationStatus MoveAndUpdateClosure::copy_until_full()
3175 {
3176   if (source() != copy_destination()) {
3177     DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3178     Copy::aligned_conjoint_words(source(), copy_destination(), words_remaining());
3179   }
3180   update_state(words_remaining());
3181   assert(is_full(), "sanity");
3182   return ParMarkBitMap::full;
3183 }
3184 
3185 void MoveAndUpdateClosure::copy_partial_obj()
3186 {
3187   size_t words = words_remaining();
3188 
3189   HeapWord* const range_end = MIN2(source() + words, bitmap()->region_end());
3190   HeapWord* const end_addr = bitmap()->find_obj_end(source(), range_end);
3191   if (end_addr < range_end) {
3192     words = bitmap()->obj_size(source(), end_addr);
3193   }
3194 
3195   // This test is necessary; if omitted, the pointer updates to a partial object
3196   // that crosses the dense prefix boundary could be overwritten.
3197   if (source() != copy_destination()) {
3198     DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3199     Copy::aligned_conjoint_words(source(), copy_destination(), words);
3200   }
3201   update_state(words);
3202 }
3203 
3204 void MoveAndUpdateClosure::complete_region(ParCompactionManager *cm, HeapWord *dest_addr,
3205                                            PSParallelCompact::RegionData *region_ptr) {
3206   assert(region_ptr->shadow_state() == ParallelCompactData::RegionData::NormalRegion, "Region should be finished");
3207   region_ptr->set_completed();
3208 }
3209 
3210 ParMarkBitMapClosure::IterationStatus
3211 MoveAndUpdateClosure::do_addr(HeapWord* addr, size_t words) {
3212   assert(destination() != NULL, "sanity");
3213   assert(bitmap()->obj_size(addr) == words, "bad size");
3214 
3215   _source = addr;
3216   assert(PSParallelCompact::summary_data().calc_new_pointer(source(), compaction_manager()) ==
3217          destination(), "wrong destination");
3218 
3219   if (words > words_remaining()) {
3220     return ParMarkBitMap::would_overflow;
3221   }
3222 
3223   // The start_array must be updated even if the object is not moving.
3224   if (_start_array != NULL) {
3225     _start_array->allocate_block(destination());
3226   }
3227 
3228   if (copy_destination() != source()) {
3229     DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3230     Copy::aligned_conjoint_words(source(), copy_destination(), words);
3231   }
3232 
3233   oop moved_oop = cast_to_oop(copy_destination());
3234   compaction_manager()->update_contents(moved_oop);
3235   assert(oopDesc::is_oop_or_null(moved_oop), "Expected an oop or NULL at " PTR_FORMAT, p2i(moved_oop));
3236 
3237   update_state(words);
3238   assert(copy_destination() == cast_from_oop<HeapWord*>(moved_oop) + moved_oop->size(), "sanity");
3239   return is_full() ? ParMarkBitMap::full : ParMarkBitMap::incomplete;
3240 }
3241 
3242 void MoveAndUpdateShadowClosure::complete_region(ParCompactionManager *cm, HeapWord *dest_addr,
3243                                                  PSParallelCompact::RegionData *region_ptr) {
3244   assert(region_ptr->shadow_state() == ParallelCompactData::RegionData::ShadowRegion, "Region should be shadow");
3245   // Record the shadow region index
3246   region_ptr->set_shadow_region(_shadow);
3247   // Mark the shadow region as filled to indicate the data is ready to be
3248   // copied back
3249   region_ptr->mark_filled();
3250   // Try to copy the content of the shadow region back to its corresponding
3251   // heap region if available; the GC thread that decreases the destination
3252   // count to zero will do the copying otherwise (see
3253   // PSParallelCompact::decrement_destination_counts).
3254   if (((region_ptr->available() && region_ptr->claim()) || region_ptr->claimed()) && region_ptr->mark_copied()) {
3255     region_ptr->set_completed();
3256     PSParallelCompact::copy_back(PSParallelCompact::summary_data().region_to_addr(_shadow), dest_addr);
3257     ParCompactionManager::push_shadow_region_mt_safe(_shadow);
3258   }
3259 }
3260 
3261 UpdateOnlyClosure::UpdateOnlyClosure(ParMarkBitMap* mbm,
3262                                      ParCompactionManager* cm,
3263                                      PSParallelCompact::SpaceId space_id) :
3264   ParMarkBitMapClosure(mbm, cm),
3265   _space_id(space_id),
3266   _start_array(PSParallelCompact::start_array(space_id))
3267 {
3268 }
3269 
3270 // Updates the references in the object to their new values.
3271 ParMarkBitMapClosure::IterationStatus
3272 UpdateOnlyClosure::do_addr(HeapWord* addr, size_t words) {
3273   do_addr(addr);
3274   return ParMarkBitMap::incomplete;
3275 }
3276 
3277 FillClosure::FillClosure(ParCompactionManager* cm, PSParallelCompact::SpaceId space_id) :
3278   ParMarkBitMapClosure(PSParallelCompact::mark_bitmap(), cm),
3279   _start_array(PSParallelCompact::start_array(space_id))
3280 {
3281   assert(space_id == PSParallelCompact::old_space_id,
3282          "cannot use FillClosure in the young gen");
3283 }
3284 
3285 ParMarkBitMapClosure::IterationStatus
3286 FillClosure::do_addr(HeapWord* addr, size_t size) {
3287   CollectedHeap::fill_with_objects(addr, size);
3288   HeapWord* const end = addr + size;
3289   do {
3290     _start_array->allocate_block(addr);
3291     addr += cast_to_oop(addr)->size();
3292   } while (addr < end);
3293   return ParMarkBitMap::incomplete;
3294 }