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