1 /*
   2  * Copyright (c) 2005, 2013, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #include "precompiled.hpp"
  26 #include "classfile/symbolTable.hpp"
  27 #include "classfile/systemDictionary.hpp"
  28 #include "code/codeCache.hpp"
  29 #include "gc_implementation/parallelScavenge/gcTaskManager.hpp"
  30 #include "gc_implementation/parallelScavenge/parallelScavengeHeap.inline.hpp"
  31 #include "gc_implementation/parallelScavenge/pcTasks.hpp"
  32 #include "gc_implementation/parallelScavenge/psAdaptiveSizePolicy.hpp"
  33 #include "gc_implementation/parallelScavenge/psCompactionManager.inline.hpp"
  34 #include "gc_implementation/parallelScavenge/psMarkSweep.hpp"
  35 #include "gc_implementation/parallelScavenge/psMarkSweepDecorator.hpp"
  36 #include "gc_implementation/parallelScavenge/psOldGen.hpp"
  37 #include "gc_implementation/parallelScavenge/psParallelCompact.hpp"
  38 #include "gc_implementation/parallelScavenge/psPromotionManager.inline.hpp"
  39 #include "gc_implementation/parallelScavenge/psScavenge.hpp"
  40 #include "gc_implementation/parallelScavenge/psYoungGen.hpp"
  41 #include "gc_implementation/shared/gcHeapSummary.hpp"
  42 #include "gc_implementation/shared/gcTimer.hpp"
  43 #include "gc_implementation/shared/gcTrace.hpp"
  44 #include "gc_implementation/shared/gcTraceTime.hpp"
  45 #include "gc_implementation/shared/isGCActiveMark.hpp"
  46 #include "gc_interface/gcCause.hpp"
  47 #include "memory/gcLocker.inline.hpp"
  48 #include "memory/referencePolicy.hpp"
  49 #include "memory/referenceProcessor.hpp"
  50 #include "oops/methodData.hpp"
  51 #include "oops/oop.inline.hpp"
  52 #include "oops/oop.pcgc.inline.hpp"
  53 #include "runtime/fprofiler.hpp"
  54 #include "runtime/safepoint.hpp"
  55 #include "runtime/vmThread.hpp"
  56 #include "services/management.hpp"
  57 #include "services/memoryService.hpp"
  58 #include "services/memTracker.hpp"
  59 #include "utilities/events.hpp"
  60 #include "utilities/stack.inline.hpp"
  61 
  62 #include <math.h>
  63 
  64 // All sizes are in HeapWords.
  65 const size_t ParallelCompactData::Log2RegionSize  = 16; // 64K words
  66 const size_t ParallelCompactData::RegionSize      = (size_t)1 << Log2RegionSize;
  67 const size_t ParallelCompactData::RegionSizeBytes =
  68   RegionSize << LogHeapWordSize;
  69 const size_t ParallelCompactData::RegionSizeOffsetMask = RegionSize - 1;
  70 const size_t ParallelCompactData::RegionAddrOffsetMask = RegionSizeBytes - 1;
  71 const size_t ParallelCompactData::RegionAddrMask       = ~RegionAddrOffsetMask;
  72 
  73 const size_t ParallelCompactData::Log2BlockSize   = 7; // 128 words
  74 const size_t ParallelCompactData::BlockSize       = (size_t)1 << Log2BlockSize;
  75 const size_t ParallelCompactData::BlockSizeBytes  =
  76   BlockSize << LogHeapWordSize;
  77 const size_t ParallelCompactData::BlockSizeOffsetMask = BlockSize - 1;
  78 const size_t ParallelCompactData::BlockAddrOffsetMask = BlockSizeBytes - 1;
  79 const size_t ParallelCompactData::BlockAddrMask       = ~BlockAddrOffsetMask;
  80 
  81 const size_t ParallelCompactData::BlocksPerRegion = RegionSize / BlockSize;
  82 const size_t ParallelCompactData::Log2BlocksPerRegion =
  83   Log2RegionSize - Log2BlockSize;
  84 
  85 const ParallelCompactData::RegionData::region_sz_t
  86 ParallelCompactData::RegionData::dc_shift = 27;
  87 
  88 const ParallelCompactData::RegionData::region_sz_t
  89 ParallelCompactData::RegionData::dc_mask = ~0U << dc_shift;
  90 
  91 const ParallelCompactData::RegionData::region_sz_t
  92 ParallelCompactData::RegionData::dc_one = 0x1U << dc_shift;
  93 
  94 const ParallelCompactData::RegionData::region_sz_t
  95 ParallelCompactData::RegionData::los_mask = ~dc_mask;
  96 
  97 const ParallelCompactData::RegionData::region_sz_t
  98 ParallelCompactData::RegionData::dc_claimed = 0x8U << dc_shift;
  99 
 100 const ParallelCompactData::RegionData::region_sz_t
 101 ParallelCompactData::RegionData::dc_completed = 0xcU << dc_shift;
 102 
 103 SpaceInfo PSParallelCompact::_space_info[PSParallelCompact::last_space_id];
 104 bool      PSParallelCompact::_print_phases = false;
 105 
 106 ReferenceProcessor* PSParallelCompact::_ref_processor = NULL;
 107 Klass*              PSParallelCompact::_updated_int_array_klass_obj = NULL;
 108 
 109 double PSParallelCompact::_dwl_mean;
 110 double PSParallelCompact::_dwl_std_dev;
 111 double PSParallelCompact::_dwl_first_term;
 112 double PSParallelCompact::_dwl_adjustment;
 113 #ifdef  ASSERT
 114 bool   PSParallelCompact::_dwl_initialized = false;
 115 #endif  // #ifdef ASSERT
 116 
 117 void SplitInfo::record(size_t src_region_idx, size_t partial_obj_size,
 118                        HeapWord* destination)
 119 {
 120   assert(src_region_idx != 0, "invalid src_region_idx");
 121   assert(partial_obj_size != 0, "invalid partial_obj_size argument");
 122   assert(destination != NULL, "invalid destination argument");
 123 
 124   _src_region_idx = src_region_idx;
 125   _partial_obj_size = partial_obj_size;
 126   _destination = destination;
 127 
 128   // These fields may not be updated below, so make sure they're clear.
 129   assert(_dest_region_addr == NULL, "should have been cleared");
 130   assert(_first_src_addr == NULL, "should have been cleared");
 131 
 132   // Determine the number of destination regions for the partial object.
 133   HeapWord* const last_word = destination + partial_obj_size - 1;
 134   const ParallelCompactData& sd = PSParallelCompact::summary_data();
 135   HeapWord* const beg_region_addr = sd.region_align_down(destination);
 136   HeapWord* const end_region_addr = sd.region_align_down(last_word);
 137 
 138   if (beg_region_addr == end_region_addr) {
 139     // One destination region.
 140     _destination_count = 1;
 141     if (end_region_addr == destination) {
 142       // The destination falls on a region boundary, thus the first word of the
 143       // partial object will be the first word copied to the destination region.
 144       _dest_region_addr = end_region_addr;
 145       _first_src_addr = sd.region_to_addr(src_region_idx);
 146     }
 147   } else {
 148     // Two destination regions.  When copied, the partial object will cross a
 149     // destination region boundary, so a word somewhere within the partial
 150     // object will be the first word copied to the second destination region.
 151     _destination_count = 2;
 152     _dest_region_addr = end_region_addr;
 153     const size_t ofs = pointer_delta(end_region_addr, destination);
 154     assert(ofs < _partial_obj_size, "sanity");
 155     _first_src_addr = sd.region_to_addr(src_region_idx) + ofs;
 156   }
 157 }
 158 
 159 void SplitInfo::clear()
 160 {
 161   _src_region_idx = 0;
 162   _partial_obj_size = 0;
 163   _destination = NULL;
 164   _destination_count = 0;
 165   _dest_region_addr = NULL;
 166   _first_src_addr = NULL;
 167   assert(!is_valid(), "sanity");
 168 }
 169 
 170 #ifdef  ASSERT
 171 void SplitInfo::verify_clear()
 172 {
 173   assert(_src_region_idx == 0, "not clear");
 174   assert(_partial_obj_size == 0, "not clear");
 175   assert(_destination == NULL, "not clear");
 176   assert(_destination_count == 0, "not clear");
 177   assert(_dest_region_addr == NULL, "not clear");
 178   assert(_first_src_addr == NULL, "not clear");
 179 }
 180 #endif  // #ifdef ASSERT
 181 
 182 
 183 void PSParallelCompact::print_on_error(outputStream* st) {
 184   _mark_bitmap.print_on_error(st);
 185 }
 186 
 187 #ifndef PRODUCT
 188 const char* PSParallelCompact::space_names[] = {
 189   "old ", "eden", "from", "to  "
 190 };
 191 
 192 void PSParallelCompact::print_region_ranges()
 193 {
 194   tty->print_cr("space  bottom     top        end        new_top");
 195   tty->print_cr("------ ---------- ---------- ---------- ----------");
 196 
 197   for (unsigned int id = 0; id < last_space_id; ++id) {
 198     const MutableSpace* space = _space_info[id].space();
 199     tty->print_cr("%u %s "
 200                   SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " "
 201                   SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " ",
 202                   id, space_names[id],
 203                   summary_data().addr_to_region_idx(space->bottom()),
 204                   summary_data().addr_to_region_idx(space->top()),
 205                   summary_data().addr_to_region_idx(space->end()),
 206                   summary_data().addr_to_region_idx(_space_info[id].new_top()));
 207   }
 208 }
 209 
 210 void
 211 print_generic_summary_region(size_t i, const ParallelCompactData::RegionData* c)
 212 {
 213 #define REGION_IDX_FORMAT        SIZE_FORMAT_W(7)
 214 #define REGION_DATA_FORMAT       SIZE_FORMAT_W(5)
 215 
 216   ParallelCompactData& sd = PSParallelCompact::summary_data();
 217   size_t dci = c->destination() ? sd.addr_to_region_idx(c->destination()) : 0;
 218   tty->print_cr(REGION_IDX_FORMAT " " PTR_FORMAT " "
 219                 REGION_IDX_FORMAT " " PTR_FORMAT " "
 220                 REGION_DATA_FORMAT " " REGION_DATA_FORMAT " "
 221                 REGION_DATA_FORMAT " " REGION_IDX_FORMAT " %d",
 222                 i, c->data_location(), dci, c->destination(),
 223                 c->partial_obj_size(), c->live_obj_size(),
 224                 c->data_size(), c->source_region(), c->destination_count());
 225 
 226 #undef  REGION_IDX_FORMAT
 227 #undef  REGION_DATA_FORMAT
 228 }
 229 
 230 void
 231 print_generic_summary_data(ParallelCompactData& summary_data,
 232                            HeapWord* const beg_addr,
 233                            HeapWord* const end_addr)
 234 {
 235   size_t total_words = 0;
 236   size_t i = summary_data.addr_to_region_idx(beg_addr);
 237   const size_t last = summary_data.addr_to_region_idx(end_addr);
 238   HeapWord* pdest = 0;
 239 
 240   while (i <= last) {
 241     ParallelCompactData::RegionData* c = summary_data.region(i);
 242     if (c->data_size() != 0 || c->destination() != pdest) {
 243       print_generic_summary_region(i, c);
 244       total_words += c->data_size();
 245       pdest = c->destination();
 246     }
 247     ++i;
 248   }
 249 
 250   tty->print_cr("summary_data_bytes=" SIZE_FORMAT, total_words * HeapWordSize);
 251 }
 252 
 253 void
 254 print_generic_summary_data(ParallelCompactData& summary_data,
 255                            SpaceInfo* space_info)
 256 {
 257   for (unsigned int id = 0; id < PSParallelCompact::last_space_id; ++id) {
 258     const MutableSpace* space = space_info[id].space();
 259     print_generic_summary_data(summary_data, space->bottom(),
 260                                MAX2(space->top(), space_info[id].new_top()));
 261   }
 262 }
 263 
 264 void
 265 print_initial_summary_region(size_t i,
 266                              const ParallelCompactData::RegionData* c,
 267                              bool newline = true)
 268 {
 269   tty->print(SIZE_FORMAT_W(5) " " PTR_FORMAT " "
 270              SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " "
 271              SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d",
 272              i, c->destination(),
 273              c->partial_obj_size(), c->live_obj_size(),
 274              c->data_size(), c->source_region(), c->destination_count());
 275   if (newline) tty->cr();
 276 }
 277 
 278 void
 279 print_initial_summary_data(ParallelCompactData& summary_data,
 280                            const MutableSpace* space) {
 281   if (space->top() == space->bottom()) {
 282     return;
 283   }
 284 
 285   const size_t region_size = ParallelCompactData::RegionSize;
 286   typedef ParallelCompactData::RegionData RegionData;
 287   HeapWord* const top_aligned_up = summary_data.region_align_up(space->top());
 288   const size_t end_region = summary_data.addr_to_region_idx(top_aligned_up);
 289   const RegionData* c = summary_data.region(end_region - 1);
 290   HeapWord* end_addr = c->destination() + c->data_size();
 291   const size_t live_in_space = pointer_delta(end_addr, space->bottom());
 292 
 293   // Print (and count) the full regions at the beginning of the space.
 294   size_t full_region_count = 0;
 295   size_t i = summary_data.addr_to_region_idx(space->bottom());
 296   while (i < end_region && summary_data.region(i)->data_size() == region_size) {
 297     print_initial_summary_region(i, summary_data.region(i));
 298     ++full_region_count;
 299     ++i;
 300   }
 301 
 302   size_t live_to_right = live_in_space - full_region_count * region_size;
 303 
 304   double max_reclaimed_ratio = 0.0;
 305   size_t max_reclaimed_ratio_region = 0;
 306   size_t max_dead_to_right = 0;
 307   size_t max_live_to_right = 0;
 308 
 309   // Print the 'reclaimed ratio' for regions while there is something live in
 310   // the region or to the right of it.  The remaining regions are empty (and
 311   // uninteresting), and computing the ratio will result in division by 0.
 312   while (i < end_region && live_to_right > 0) {
 313     c = summary_data.region(i);
 314     HeapWord* const region_addr = summary_data.region_to_addr(i);
 315     const size_t used_to_right = pointer_delta(space->top(), region_addr);
 316     const size_t dead_to_right = used_to_right - live_to_right;
 317     const double reclaimed_ratio = double(dead_to_right) / live_to_right;
 318 
 319     if (reclaimed_ratio > max_reclaimed_ratio) {
 320             max_reclaimed_ratio = reclaimed_ratio;
 321             max_reclaimed_ratio_region = i;
 322             max_dead_to_right = dead_to_right;
 323             max_live_to_right = live_to_right;
 324     }
 325 
 326     print_initial_summary_region(i, c, false);
 327     tty->print_cr(" %12.10f " SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10),
 328                   reclaimed_ratio, dead_to_right, live_to_right);
 329 
 330     live_to_right -= c->data_size();
 331     ++i;
 332   }
 333 
 334   // Any remaining regions are empty.  Print one more if there is one.
 335   if (i < end_region) {
 336     print_initial_summary_region(i, summary_data.region(i));
 337   }
 338 
 339   tty->print_cr("max:  " SIZE_FORMAT_W(4) " d2r=" SIZE_FORMAT_W(10) " "
 340                 "l2r=" SIZE_FORMAT_W(10) " max_ratio=%14.12f",
 341                 max_reclaimed_ratio_region, max_dead_to_right,
 342                 max_live_to_right, max_reclaimed_ratio);
 343 }
 344 
 345 void
 346 print_initial_summary_data(ParallelCompactData& summary_data,
 347                            SpaceInfo* space_info) {
 348   unsigned int id = PSParallelCompact::old_space_id;
 349   const MutableSpace* space;
 350   do {
 351     space = space_info[id].space();
 352     print_initial_summary_data(summary_data, space);
 353   } while (++id < PSParallelCompact::eden_space_id);
 354 
 355   do {
 356     space = space_info[id].space();
 357     print_generic_summary_data(summary_data, space->bottom(), space->top());
 358   } while (++id < PSParallelCompact::last_space_id);
 359 }
 360 #endif  // #ifndef PRODUCT
 361 
 362 #ifdef  ASSERT
 363 size_t add_obj_count;
 364 size_t add_obj_size;
 365 size_t mark_bitmap_count;
 366 size_t mark_bitmap_size;
 367 #endif  // #ifdef ASSERT
 368 
 369 ParallelCompactData::ParallelCompactData()
 370 {
 371   _region_start = 0;
 372 
 373   _region_vspace = 0;
 374   _reserved_byte_size = 0;
 375   _region_data = 0;
 376   _region_count = 0;
 377 
 378   _block_vspace = 0;
 379   _block_data = 0;
 380   _block_count = 0;
 381 }
 382 
 383 bool ParallelCompactData::initialize(MemRegion covered_region)
 384 {
 385   _region_start = covered_region.start();
 386   const size_t region_size = covered_region.word_size();
 387   DEBUG_ONLY(_region_end = _region_start + region_size;)
 388 
 389   assert(region_align_down(_region_start) == _region_start,
 390          "region start not aligned");
 391   assert((region_size & RegionSizeOffsetMask) == 0,
 392          "region size not a multiple of RegionSize");
 393 
 394   bool result = initialize_region_data(region_size) && initialize_block_data();
 395   return result;
 396 }
 397 
 398 PSVirtualSpace*
 399 ParallelCompactData::create_vspace(size_t count, size_t element_size)
 400 {
 401   const size_t raw_bytes = count * element_size;
 402   const size_t page_sz = os::page_size_for_region(raw_bytes, raw_bytes, 10);
 403   const size_t granularity = os::vm_allocation_granularity();
 404   _reserved_byte_size = align_size_up(raw_bytes, MAX2(page_sz, granularity));
 405 
 406   const size_t rs_align = page_sz == (size_t) os::vm_page_size() ? 0 :
 407     MAX2(page_sz, granularity);
 408   ReservedSpace rs(_reserved_byte_size, rs_align, rs_align > 0);
 409   os::trace_page_sizes("par compact", raw_bytes, raw_bytes, page_sz, rs.base(),
 410                        rs.size());
 411 
 412   MemTracker::record_virtual_memory_type((address)rs.base(), mtGC);
 413 
 414   PSVirtualSpace* vspace = new PSVirtualSpace(rs, page_sz);
 415   if (vspace != 0) {
 416     if (vspace->expand_by(_reserved_byte_size)) {
 417       return vspace;
 418     }
 419     delete vspace;
 420     // Release memory reserved in the space.
 421     rs.release();
 422   }
 423 
 424   return 0;
 425 }
 426 
 427 bool ParallelCompactData::initialize_region_data(size_t region_size)
 428 {
 429   const size_t count = (region_size + RegionSizeOffsetMask) >> Log2RegionSize;
 430   _region_vspace = create_vspace(count, sizeof(RegionData));
 431   if (_region_vspace != 0) {
 432     _region_data = (RegionData*)_region_vspace->reserved_low_addr();
 433     _region_count = count;
 434     return true;
 435   }
 436   return false;
 437 }
 438 
 439 bool ParallelCompactData::initialize_block_data()
 440 {
 441   assert(_region_count != 0, "region data must be initialized first");
 442   const size_t count = _region_count << Log2BlocksPerRegion;
 443   _block_vspace = create_vspace(count, sizeof(BlockData));
 444   if (_block_vspace != 0) {
 445     _block_data = (BlockData*)_block_vspace->reserved_low_addr();
 446     _block_count = count;
 447     return true;
 448   }
 449   return false;
 450 }
 451 
 452 void ParallelCompactData::clear()
 453 {
 454   memset(_region_data, 0, _region_vspace->committed_size());
 455   memset(_block_data, 0, _block_vspace->committed_size());
 456 }
 457 
 458 void ParallelCompactData::clear_range(size_t beg_region, size_t end_region) {
 459   assert(beg_region <= _region_count, "beg_region out of range");
 460   assert(end_region <= _region_count, "end_region out of range");
 461   assert(RegionSize % BlockSize == 0, "RegionSize not a multiple of BlockSize");
 462 
 463   const size_t region_cnt = end_region - beg_region;
 464   memset(_region_data + beg_region, 0, region_cnt * sizeof(RegionData));
 465 
 466   const size_t beg_block = beg_region * BlocksPerRegion;
 467   const size_t block_cnt = region_cnt * BlocksPerRegion;
 468   memset(_block_data + beg_block, 0, block_cnt * sizeof(BlockData));
 469 }
 470 
 471 HeapWord* ParallelCompactData::partial_obj_end(size_t region_idx) const
 472 {
 473   const RegionData* cur_cp = region(region_idx);
 474   const RegionData* const end_cp = region(region_count() - 1);
 475 
 476   HeapWord* result = region_to_addr(region_idx);
 477   if (cur_cp < end_cp) {
 478     do {
 479       result += cur_cp->partial_obj_size();
 480     } while (cur_cp->partial_obj_size() == RegionSize && ++cur_cp < end_cp);
 481   }
 482   return result;
 483 }
 484 
 485 void ParallelCompactData::add_obj(HeapWord* addr, size_t len)
 486 {
 487   const size_t obj_ofs = pointer_delta(addr, _region_start);
 488   const size_t beg_region = obj_ofs >> Log2RegionSize;
 489   const size_t end_region = (obj_ofs + len - 1) >> Log2RegionSize;
 490 
 491   DEBUG_ONLY(Atomic::inc_ptr(&add_obj_count);)
 492   DEBUG_ONLY(Atomic::add_ptr(len, &add_obj_size);)
 493 
 494   if (beg_region == end_region) {
 495     // All in one region.
 496     _region_data[beg_region].add_live_obj(len);
 497     return;
 498   }
 499 
 500   // First region.
 501   const size_t beg_ofs = region_offset(addr);
 502   _region_data[beg_region].add_live_obj(RegionSize - beg_ofs);
 503 
 504   Klass* klass = ((oop)addr)->klass();
 505   // Middle regions--completely spanned by this object.
 506   for (size_t region = beg_region + 1; region < end_region; ++region) {
 507     _region_data[region].set_partial_obj_size(RegionSize);
 508     _region_data[region].set_partial_obj_addr(addr);
 509   }
 510 
 511   // Last region.
 512   const size_t end_ofs = region_offset(addr + len - 1);
 513   _region_data[end_region].set_partial_obj_size(end_ofs + 1);
 514   _region_data[end_region].set_partial_obj_addr(addr);
 515 }
 516 
 517 void
 518 ParallelCompactData::summarize_dense_prefix(HeapWord* beg, HeapWord* end)
 519 {
 520   assert(region_offset(beg) == 0, "not RegionSize aligned");
 521   assert(region_offset(end) == 0, "not RegionSize aligned");
 522 
 523   size_t cur_region = addr_to_region_idx(beg);
 524   const size_t end_region = addr_to_region_idx(end);
 525   HeapWord* addr = beg;
 526   while (cur_region < end_region) {
 527     _region_data[cur_region].set_destination(addr);
 528     _region_data[cur_region].set_destination_count(0);
 529     _region_data[cur_region].set_source_region(cur_region);
 530     _region_data[cur_region].set_data_location(addr);
 531 
 532     // Update live_obj_size so the region appears completely full.
 533     size_t live_size = RegionSize - _region_data[cur_region].partial_obj_size();
 534     _region_data[cur_region].set_live_obj_size(live_size);
 535 
 536     ++cur_region;
 537     addr += RegionSize;
 538   }
 539 }
 540 
 541 // Find the point at which a space can be split and, if necessary, record the
 542 // split point.
 543 //
 544 // If the current src region (which overflowed the destination space) doesn't
 545 // have a partial object, the split point is at the beginning of the current src
 546 // region (an "easy" split, no extra bookkeeping required).
 547 //
 548 // If the current src region has a partial object, the split point is in the
 549 // region where that partial object starts (call it the split_region).  If
 550 // split_region has a partial object, then the split point is just after that
 551 // partial object (a "hard" split where we have to record the split data and
 552 // zero the partial_obj_size field).  With a "hard" split, we know that the
 553 // partial_obj ends within split_region because the partial object that caused
 554 // the overflow starts in split_region.  If split_region doesn't have a partial
 555 // obj, then the split is at the beginning of split_region (another "easy"
 556 // split).
 557 HeapWord*
 558 ParallelCompactData::summarize_split_space(size_t src_region,
 559                                            SplitInfo& split_info,
 560                                            HeapWord* destination,
 561                                            HeapWord* target_end,
 562                                            HeapWord** target_next)
 563 {
 564   assert(destination <= target_end, "sanity");
 565   assert(destination + _region_data[src_region].data_size() > target_end,
 566     "region should not fit into target space");
 567   assert(is_region_aligned(target_end), "sanity");
 568 
 569   size_t split_region = src_region;
 570   HeapWord* split_destination = destination;
 571   size_t partial_obj_size = _region_data[src_region].partial_obj_size();
 572 
 573   if (destination + partial_obj_size > target_end) {
 574     // The split point is just after the partial object (if any) in the
 575     // src_region that contains the start of the object that overflowed the
 576     // destination space.
 577     //
 578     // Find the start of the "overflow" object and set split_region to the
 579     // region containing it.
 580     HeapWord* const overflow_obj = _region_data[src_region].partial_obj_addr();
 581     split_region = addr_to_region_idx(overflow_obj);
 582 
 583     // Clear the source_region field of all destination regions whose first word
 584     // came from data after the split point (a non-null source_region field
 585     // implies a region must be filled).
 586     //
 587     // An alternative to the simple loop below:  clear during post_compact(),
 588     // which uses memcpy instead of individual stores, and is easy to
 589     // parallelize.  (The downside is that it clears the entire RegionData
 590     // object as opposed to just one field.)
 591     //
 592     // post_compact() would have to clear the summary data up to the highest
 593     // address that was written during the summary phase, which would be
 594     //
 595     //         max(top, max(new_top, clear_top))
 596     //
 597     // where clear_top is a new field in SpaceInfo.  Would have to set clear_top
 598     // to target_end.
 599     const RegionData* const sr = region(split_region);
 600     const size_t beg_idx =
 601       addr_to_region_idx(region_align_up(sr->destination() +
 602                                          sr->partial_obj_size()));
 603     const size_t end_idx = addr_to_region_idx(target_end);
 604 
 605     if (TraceParallelOldGCSummaryPhase) {
 606         gclog_or_tty->print_cr("split:  clearing source_region field in ["
 607                                SIZE_FORMAT ", " SIZE_FORMAT ")",
 608                                beg_idx, end_idx);
 609     }
 610     for (size_t idx = beg_idx; idx < end_idx; ++idx) {
 611       _region_data[idx].set_source_region(0);
 612     }
 613 
 614     // Set split_destination and partial_obj_size to reflect the split region.
 615     split_destination = sr->destination();
 616     partial_obj_size = sr->partial_obj_size();
 617   }
 618 
 619   // The split is recorded only if a partial object extends onto the region.
 620   if (partial_obj_size != 0) {
 621     _region_data[split_region].set_partial_obj_size(0);
 622     split_info.record(split_region, partial_obj_size, split_destination);
 623   }
 624 
 625   // Setup the continuation addresses.
 626   *target_next = split_destination + partial_obj_size;
 627   HeapWord* const source_next = region_to_addr(split_region) + partial_obj_size;
 628 
 629   if (TraceParallelOldGCSummaryPhase) {
 630     const char * split_type = partial_obj_size == 0 ? "easy" : "hard";
 631     gclog_or_tty->print_cr("%s split:  src=" PTR_FORMAT " src_c=" SIZE_FORMAT
 632                            " pos=" SIZE_FORMAT,
 633                            split_type, source_next, split_region,
 634                            partial_obj_size);
 635     gclog_or_tty->print_cr("%s split:  dst=" PTR_FORMAT " dst_c=" SIZE_FORMAT
 636                            " tn=" PTR_FORMAT,
 637                            split_type, split_destination,
 638                            addr_to_region_idx(split_destination),
 639                            *target_next);
 640 
 641     if (partial_obj_size != 0) {
 642       HeapWord* const po_beg = split_info.destination();
 643       HeapWord* const po_end = po_beg + split_info.partial_obj_size();
 644       gclog_or_tty->print_cr("%s split:  "
 645                              "po_beg=" PTR_FORMAT " " SIZE_FORMAT " "
 646                              "po_end=" PTR_FORMAT " " SIZE_FORMAT,
 647                              split_type,
 648                              po_beg, addr_to_region_idx(po_beg),
 649                              po_end, addr_to_region_idx(po_end));
 650     }
 651   }
 652 
 653   return source_next;
 654 }
 655 
 656 bool ParallelCompactData::summarize(SplitInfo& split_info,
 657                                     HeapWord* source_beg, HeapWord* source_end,
 658                                     HeapWord** source_next,
 659                                     HeapWord* target_beg, HeapWord* target_end,
 660                                     HeapWord** target_next)
 661 {
 662   if (TraceParallelOldGCSummaryPhase) {
 663     HeapWord* const source_next_val = source_next == NULL ? NULL : *source_next;
 664     tty->print_cr("sb=" PTR_FORMAT " se=" PTR_FORMAT " sn=" PTR_FORMAT
 665                   "tb=" PTR_FORMAT " te=" PTR_FORMAT " tn=" PTR_FORMAT,
 666                   source_beg, source_end, source_next_val,
 667                   target_beg, target_end, *target_next);
 668   }
 669 
 670   size_t cur_region = addr_to_region_idx(source_beg);
 671   const size_t end_region = addr_to_region_idx(region_align_up(source_end));
 672 
 673   HeapWord *dest_addr = target_beg;
 674   while (cur_region < end_region) {
 675     // The destination must be set even if the region has no data.
 676     _region_data[cur_region].set_destination(dest_addr);
 677 
 678     size_t words = _region_data[cur_region].data_size();
 679     if (words > 0) {
 680       // If cur_region does not fit entirely into the target space, find a point
 681       // at which the source space can be 'split' so that part is copied to the
 682       // target space and the rest is copied elsewhere.
 683       if (dest_addr + words > target_end) {
 684         assert(source_next != NULL, "source_next is NULL when splitting");
 685         *source_next = summarize_split_space(cur_region, split_info, dest_addr,
 686                                              target_end, target_next);
 687         return false;
 688       }
 689 
 690       // Compute the destination_count for cur_region, and if necessary, update
 691       // source_region for a destination region.  The source_region field is
 692       // updated if cur_region is the first (left-most) region to be copied to a
 693       // destination region.
 694       //
 695       // The destination_count calculation is a bit subtle.  A region that has
 696       // data that compacts into itself does not count itself as a destination.
 697       // This maintains the invariant that a zero count means the region is
 698       // available and can be claimed and then filled.
 699       uint destination_count = 0;
 700       if (split_info.is_split(cur_region)) {
 701         // The current region has been split:  the partial object will be copied
 702         // to one destination space and the remaining data will be copied to
 703         // another destination space.  Adjust the initial destination_count and,
 704         // if necessary, set the source_region field if the partial object will
 705         // cross a destination region boundary.
 706         destination_count = split_info.destination_count();
 707         if (destination_count == 2) {
 708           size_t dest_idx = addr_to_region_idx(split_info.dest_region_addr());
 709           _region_data[dest_idx].set_source_region(cur_region);
 710         }
 711       }
 712 
 713       HeapWord* const last_addr = dest_addr + words - 1;
 714       const size_t dest_region_1 = addr_to_region_idx(dest_addr);
 715       const size_t dest_region_2 = addr_to_region_idx(last_addr);
 716 
 717       // Initially assume that the destination regions will be the same and
 718       // adjust the value below if necessary.  Under this assumption, if
 719       // cur_region == dest_region_2, then cur_region will be compacted
 720       // completely into itself.
 721       destination_count += cur_region == dest_region_2 ? 0 : 1;
 722       if (dest_region_1 != dest_region_2) {
 723         // Destination regions differ; adjust destination_count.
 724         destination_count += 1;
 725         // Data from cur_region will be copied to the start of dest_region_2.
 726         _region_data[dest_region_2].set_source_region(cur_region);
 727       } else if (region_offset(dest_addr) == 0) {
 728         // Data from cur_region will be copied to the start of the destination
 729         // region.
 730         _region_data[dest_region_1].set_source_region(cur_region);
 731       }
 732 
 733       _region_data[cur_region].set_destination_count(destination_count);
 734       _region_data[cur_region].set_data_location(region_to_addr(cur_region));
 735       dest_addr += words;
 736     }
 737 
 738     ++cur_region;
 739   }
 740 
 741   *target_next = dest_addr;
 742   return true;
 743 }
 744 
 745 HeapWord* ParallelCompactData::calc_new_pointer(HeapWord* addr) {
 746   assert(addr != NULL, "Should detect NULL oop earlier");
 747   assert(PSParallelCompact::gc_heap()->is_in(addr), "not in heap");
 748   assert(PSParallelCompact::mark_bitmap()->is_marked(addr), "not marked");
 749 
 750   // Region covering the object.
 751   RegionData* const region_ptr = addr_to_region_ptr(addr);
 752   HeapWord* result = region_ptr->destination();
 753 
 754   // If the entire Region is live, the new location is region->destination + the
 755   // offset of the object within in the Region.
 756 
 757   // Run some performance tests to determine if this special case pays off.  It
 758   // is worth it for pointers into the dense prefix.  If the optimization to
 759   // avoid pointer updates in regions that only point to the dense prefix is
 760   // ever implemented, this should be revisited.
 761   if (region_ptr->data_size() == RegionSize) {
 762     result += region_offset(addr);
 763     return result;
 764   }
 765 
 766   // Otherwise, the new location is region->destination + block offset + the
 767   // number of live words in the Block that are (a) to the left of addr and (b)
 768   // due to objects that start in the Block.
 769 
 770   // Fill in the block table if necessary.  This is unsynchronized, so multiple
 771   // threads may fill the block table for a region (harmless, since it is
 772   // idempotent).
 773   if (!region_ptr->blocks_filled()) {
 774     PSParallelCompact::fill_blocks(addr_to_region_idx(addr));
 775     region_ptr->set_blocks_filled();
 776   }
 777 
 778   HeapWord* const search_start = block_align_down(addr);
 779   const size_t block_offset = addr_to_block_ptr(addr)->offset();
 780 
 781   const ParMarkBitMap* bitmap = PSParallelCompact::mark_bitmap();
 782   const size_t live = bitmap->live_words_in_range(search_start, oop(addr));
 783   result += block_offset + live;
 784   DEBUG_ONLY(PSParallelCompact::check_new_location(addr, result));
 785   return result;
 786 }
 787 
 788 #ifdef ASSERT
 789 void ParallelCompactData::verify_clear(const PSVirtualSpace* vspace)
 790 {
 791   const size_t* const beg = (const size_t*)vspace->committed_low_addr();
 792   const size_t* const end = (const size_t*)vspace->committed_high_addr();
 793   for (const size_t* p = beg; p < end; ++p) {
 794     assert(*p == 0, "not zero");
 795   }
 796 }
 797 
 798 void ParallelCompactData::verify_clear()
 799 {
 800   verify_clear(_region_vspace);
 801   verify_clear(_block_vspace);
 802 }
 803 #endif  // #ifdef ASSERT
 804 
 805 STWGCTimer          PSParallelCompact::_gc_timer;
 806 ParallelOldTracer   PSParallelCompact::_gc_tracer;
 807 elapsedTimer        PSParallelCompact::_accumulated_time;
 808 unsigned int        PSParallelCompact::_total_invocations = 0;
 809 unsigned int        PSParallelCompact::_maximum_compaction_gc_num = 0;
 810 jlong               PSParallelCompact::_time_of_last_gc = 0;
 811 CollectorCounters*  PSParallelCompact::_counters = NULL;
 812 ParMarkBitMap       PSParallelCompact::_mark_bitmap;
 813 ParallelCompactData PSParallelCompact::_summary_data;
 814 
 815 PSParallelCompact::IsAliveClosure PSParallelCompact::_is_alive_closure;
 816 
 817 bool PSParallelCompact::IsAliveClosure::do_object_b(oop p) { return mark_bitmap()->is_marked(p); }
 818 
 819 void PSParallelCompact::KeepAliveClosure::do_oop(oop* p)       { PSParallelCompact::KeepAliveClosure::do_oop_work(p); }
 820 void PSParallelCompact::KeepAliveClosure::do_oop(narrowOop* p) { PSParallelCompact::KeepAliveClosure::do_oop_work(p); }
 821 
 822 PSParallelCompact::AdjustPointerClosure PSParallelCompact::_adjust_pointer_closure;
 823 PSParallelCompact::AdjustKlassClosure PSParallelCompact::_adjust_klass_closure;
 824 
 825 void PSParallelCompact::AdjustPointerClosure::do_oop(oop* p)       { adjust_pointer(p); }
 826 void PSParallelCompact::AdjustPointerClosure::do_oop(narrowOop* p) { adjust_pointer(p); }
 827 
 828 void PSParallelCompact::FollowStackClosure::do_void() { _compaction_manager->follow_marking_stacks(); }
 829 
 830 void PSParallelCompact::MarkAndPushClosure::do_oop(oop* p)       {
 831   mark_and_push(_compaction_manager, p);
 832 }
 833 void PSParallelCompact::MarkAndPushClosure::do_oop(narrowOop* p) { mark_and_push(_compaction_manager, p); }
 834 
 835 void PSParallelCompact::FollowKlassClosure::do_klass(Klass* klass) {
 836   klass->oops_do(_mark_and_push_closure);
 837 }
 838 void PSParallelCompact::AdjustKlassClosure::do_klass(Klass* klass) {
 839   klass->oops_do(&PSParallelCompact::_adjust_pointer_closure);
 840 }
 841 
 842 void PSParallelCompact::post_initialize() {
 843   ParallelScavengeHeap* heap = gc_heap();
 844   assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity");
 845 
 846   MemRegion mr = heap->reserved_region();
 847   _ref_processor =
 848     new ReferenceProcessor(mr,            // span
 849                            ParallelRefProcEnabled && (ParallelGCThreads > 1), // mt processing
 850                            (int) ParallelGCThreads, // mt processing degree
 851                            true,          // mt discovery
 852                            (int) ParallelGCThreads, // mt discovery degree
 853                            true,          // atomic_discovery
 854                            &_is_alive_closure, // non-header is alive closure
 855                            false);        // write barrier for next field updates
 856   _counters = new CollectorCounters("PSParallelCompact", 1);
 857 
 858   // Initialize static fields in ParCompactionManager.
 859   ParCompactionManager::initialize(mark_bitmap());
 860 }
 861 
 862 bool PSParallelCompact::initialize() {
 863   ParallelScavengeHeap* heap = gc_heap();
 864   assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity");
 865   MemRegion mr = heap->reserved_region();
 866 
 867   // Was the old gen get allocated successfully?
 868   if (!heap->old_gen()->is_allocated()) {
 869     return false;
 870   }
 871 
 872   initialize_space_info();
 873   initialize_dead_wood_limiter();
 874 
 875   if (!_mark_bitmap.initialize(mr)) {
 876     vm_shutdown_during_initialization(
 877       err_msg("Unable to allocate " SIZE_FORMAT "KB bitmaps for parallel "
 878       "garbage collection for the requested " SIZE_FORMAT "KB heap.",
 879       _mark_bitmap.reserved_byte_size()/K, mr.byte_size()/K));
 880     return false;
 881   }
 882 
 883   if (!_summary_data.initialize(mr)) {
 884     vm_shutdown_during_initialization(
 885       err_msg("Unable to allocate " SIZE_FORMAT "KB card tables for parallel "
 886       "garbage collection for the requested " SIZE_FORMAT "KB heap.",
 887       _summary_data.reserved_byte_size()/K, mr.byte_size()/K));
 888     return false;
 889   }
 890 
 891   return true;
 892 }
 893 
 894 void PSParallelCompact::initialize_space_info()
 895 {
 896   memset(&_space_info, 0, sizeof(_space_info));
 897 
 898   ParallelScavengeHeap* heap = gc_heap();
 899   PSYoungGen* young_gen = heap->young_gen();
 900 
 901   _space_info[old_space_id].set_space(heap->old_gen()->object_space());
 902   _space_info[eden_space_id].set_space(young_gen->eden_space());
 903   _space_info[from_space_id].set_space(young_gen->from_space());
 904   _space_info[to_space_id].set_space(young_gen->to_space());
 905 
 906   _space_info[old_space_id].set_start_array(heap->old_gen()->start_array());
 907 }
 908 
 909 void PSParallelCompact::initialize_dead_wood_limiter()
 910 {
 911   const size_t max = 100;
 912   _dwl_mean = double(MIN2(ParallelOldDeadWoodLimiterMean, max)) / 100.0;
 913   _dwl_std_dev = double(MIN2(ParallelOldDeadWoodLimiterStdDev, max)) / 100.0;
 914   _dwl_first_term = 1.0 / (sqrt(2.0 * M_PI) * _dwl_std_dev);
 915   DEBUG_ONLY(_dwl_initialized = true;)
 916   _dwl_adjustment = normal_distribution(1.0);
 917 }
 918 
 919 // Simple class for storing info about the heap at the start of GC, to be used
 920 // after GC for comparison/printing.
 921 class PreGCValues {
 922 public:
 923   PreGCValues() { }
 924   PreGCValues(ParallelScavengeHeap* heap) { fill(heap); }
 925 
 926   void fill(ParallelScavengeHeap* heap) {
 927     _heap_used      = heap->used();
 928     _young_gen_used = heap->young_gen()->used_in_bytes();
 929     _old_gen_used   = heap->old_gen()->used_in_bytes();
 930     _metadata_used  = MetaspaceAux::allocated_used_bytes();
 931   };
 932 
 933   size_t heap_used() const      { return _heap_used; }
 934   size_t young_gen_used() const { return _young_gen_used; }
 935   size_t old_gen_used() const   { return _old_gen_used; }
 936   size_t metadata_used() const  { return _metadata_used; }
 937 
 938 private:
 939   size_t _heap_used;
 940   size_t _young_gen_used;
 941   size_t _old_gen_used;
 942   size_t _metadata_used;
 943 };
 944 
 945 void
 946 PSParallelCompact::clear_data_covering_space(SpaceId id)
 947 {
 948   // At this point, top is the value before GC, new_top() is the value that will
 949   // be set at the end of GC.  The marking bitmap is cleared to top; nothing
 950   // should be marked above top.  The summary data is cleared to the larger of
 951   // top & new_top.
 952   MutableSpace* const space = _space_info[id].space();
 953   HeapWord* const bot = space->bottom();
 954   HeapWord* const top = space->top();
 955   HeapWord* const max_top = MAX2(top, _space_info[id].new_top());
 956 
 957   const idx_t beg_bit = _mark_bitmap.addr_to_bit(bot);
 958   const idx_t end_bit = BitMap::word_align_up(_mark_bitmap.addr_to_bit(top));
 959   _mark_bitmap.clear_range(beg_bit, end_bit);
 960 
 961   const size_t beg_region = _summary_data.addr_to_region_idx(bot);
 962   const size_t end_region =
 963     _summary_data.addr_to_region_idx(_summary_data.region_align_up(max_top));
 964   _summary_data.clear_range(beg_region, end_region);
 965 
 966   // Clear the data used to 'split' regions.
 967   SplitInfo& split_info = _space_info[id].split_info();
 968   if (split_info.is_valid()) {
 969     split_info.clear();
 970   }
 971   DEBUG_ONLY(split_info.verify_clear();)
 972 }
 973 
 974 void PSParallelCompact::pre_compact(PreGCValues* pre_gc_values)
 975 {
 976   // Update the from & to space pointers in space_info, since they are swapped
 977   // at each young gen gc.  Do the update unconditionally (even though a
 978   // promotion failure does not swap spaces) because an unknown number of minor
 979   // collections will have swapped the spaces an unknown number of times.
 980   GCTraceTime tm("pre compact", print_phases(), true, &_gc_timer);
 981   ParallelScavengeHeap* heap = gc_heap();
 982   _space_info[from_space_id].set_space(heap->young_gen()->from_space());
 983   _space_info[to_space_id].set_space(heap->young_gen()->to_space());
 984 
 985   pre_gc_values->fill(heap);
 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->accumulate_statistics_all_tlabs();
1001   heap->ensure_parsability(true);  // retire TLABs
1002 
1003   if (VerifyBeforeGC && heap->total_collections() >= VerifyGCStartAt) {
1004     HandleMark hm;  // Discard invalid handles created during verification
1005     Universe::verify(" VerifyBeforeGC:");
1006   }
1007 
1008   // Verify object start arrays
1009   if (VerifyObjectStartArray &&
1010       VerifyBeforeGC) {
1011     heap->old_gen()->verify_object_start_array();
1012   }
1013 
1014   DEBUG_ONLY(mark_bitmap()->verify_clear();)
1015   DEBUG_ONLY(summary_data().verify_clear();)
1016 
1017   // Have worker threads release resources the next time they run a task.
1018   gc_task_manager()->release_all_resources();
1019 }
1020 
1021 void PSParallelCompact::post_compact()
1022 {
1023   GCTraceTime tm("post compact", print_phases(), true, &_gc_timer);
1024 
1025   for (unsigned int id = old_space_id; id < last_space_id; ++id) {
1026     // Clear the marking bitmap, summary data and split info.
1027     clear_data_covering_space(SpaceId(id));
1028     // Update top().  Must be done after clearing the bitmap and summary data.
1029     _space_info[id].publish_new_top();
1030   }
1031 
1032   MutableSpace* const eden_space = _space_info[eden_space_id].space();
1033   MutableSpace* const from_space = _space_info[from_space_id].space();
1034   MutableSpace* const to_space   = _space_info[to_space_id].space();
1035 
1036   ParallelScavengeHeap* heap = gc_heap();
1037   bool eden_empty = eden_space->is_empty();
1038   if (!eden_empty) {
1039     eden_empty = absorb_live_data_from_eden(heap->size_policy(),
1040                                             heap->young_gen(), heap->old_gen());
1041   }
1042 
1043   // Update heap occupancy information which is used as input to the soft ref
1044   // clearing policy at the next gc.
1045   Universe::update_heap_info_at_gc();
1046 
1047   bool young_gen_empty = eden_empty && from_space->is_empty() &&
1048     to_space->is_empty();
1049 
1050   BarrierSet* bs = heap->barrier_set();
1051   if (bs->is_a(BarrierSet::ModRef)) {
1052     ModRefBarrierSet* modBS = (ModRefBarrierSet*)bs;
1053     MemRegion old_mr = heap->old_gen()->reserved();
1054 
1055     if (young_gen_empty) {
1056       modBS->clear(MemRegion(old_mr.start(), old_mr.end()));
1057     } else {
1058       modBS->invalidate(MemRegion(old_mr.start(), old_mr.end()));
1059     }
1060   }
1061 
1062   // Delete metaspaces for unloaded class loaders and clean up loader_data graph
1063   ClassLoaderDataGraph::purge();
1064   MetaspaceAux::verify_metrics();
1065 
1066   Threads::gc_epilogue();
1067   CodeCache::gc_epilogue();
1068   JvmtiExport::gc_epilogue();
1069 
1070   COMPILER2_PRESENT(DerivedPointerTable::update_pointers());
1071 
1072   ref_processor()->enqueue_discovered_references(NULL);
1073 
1074   if (ZapUnusedHeapArea) {
1075     heap->gen_mangle_unused_area();
1076   }
1077 
1078   // Update time of last GC
1079   reset_millis_since_last_gc();
1080 }
1081 
1082 HeapWord*
1083 PSParallelCompact::compute_dense_prefix_via_density(const SpaceId id,
1084                                                     bool maximum_compaction)
1085 {
1086   const size_t region_size = ParallelCompactData::RegionSize;
1087   const ParallelCompactData& sd = summary_data();
1088 
1089   const MutableSpace* const space = _space_info[id].space();
1090   HeapWord* const top_aligned_up = sd.region_align_up(space->top());
1091   const RegionData* const beg_cp = sd.addr_to_region_ptr(space->bottom());
1092   const RegionData* const end_cp = sd.addr_to_region_ptr(top_aligned_up);
1093 
1094   // Skip full regions at the beginning of the space--they are necessarily part
1095   // of the dense prefix.
1096   size_t full_count = 0;
1097   const RegionData* cp;
1098   for (cp = beg_cp; cp < end_cp && cp->data_size() == region_size; ++cp) {
1099     ++full_count;
1100   }
1101 
1102   assert(total_invocations() >= _maximum_compaction_gc_num, "sanity");
1103   const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num;
1104   const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval;
1105   if (maximum_compaction || cp == end_cp || interval_ended) {
1106     _maximum_compaction_gc_num = total_invocations();
1107     return sd.region_to_addr(cp);
1108   }
1109 
1110   HeapWord* const new_top = _space_info[id].new_top();
1111   const size_t space_live = pointer_delta(new_top, space->bottom());
1112   const size_t space_used = space->used_in_words();
1113   const size_t space_capacity = space->capacity_in_words();
1114 
1115   const double cur_density = double(space_live) / space_capacity;
1116   const double deadwood_density =
1117     (1.0 - cur_density) * (1.0 - cur_density) * cur_density * cur_density;
1118   const size_t deadwood_goal = size_t(space_capacity * deadwood_density);
1119 
1120   if (TraceParallelOldGCDensePrefix) {
1121     tty->print_cr("cur_dens=%5.3f dw_dens=%5.3f dw_goal=" SIZE_FORMAT,
1122                   cur_density, deadwood_density, deadwood_goal);
1123     tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " "
1124                   "space_cap=" SIZE_FORMAT,
1125                   space_live, space_used,
1126                   space_capacity);
1127   }
1128 
1129   // XXX - Use binary search?
1130   HeapWord* dense_prefix = sd.region_to_addr(cp);
1131   const RegionData* full_cp = cp;
1132   const RegionData* const top_cp = sd.addr_to_region_ptr(space->top() - 1);
1133   while (cp < end_cp) {
1134     HeapWord* region_destination = cp->destination();
1135     const size_t cur_deadwood = pointer_delta(dense_prefix, region_destination);
1136     if (TraceParallelOldGCDensePrefix && Verbose) {
1137       tty->print_cr("c#=" SIZE_FORMAT_W(4) " dst=" PTR_FORMAT " "
1138                     "dp=" SIZE_FORMAT_W(8) " " "cdw=" SIZE_FORMAT_W(8),
1139                     sd.region(cp), region_destination,
1140                     dense_prefix, cur_deadwood);
1141     }
1142 
1143     if (cur_deadwood >= deadwood_goal) {
1144       // Found the region that has the correct amount of deadwood to the left.
1145       // This typically occurs after crossing a fairly sparse set of regions, so
1146       // iterate backwards over those sparse regions, looking for the region
1147       // that has the lowest density of live objects 'to the right.'
1148       size_t space_to_left = sd.region(cp) * region_size;
1149       size_t live_to_left = space_to_left - cur_deadwood;
1150       size_t space_to_right = space_capacity - space_to_left;
1151       size_t live_to_right = space_live - live_to_left;
1152       double density_to_right = double(live_to_right) / space_to_right;
1153       while (cp > full_cp) {
1154         --cp;
1155         const size_t prev_region_live_to_right = live_to_right -
1156           cp->data_size();
1157         const size_t prev_region_space_to_right = space_to_right + region_size;
1158         double prev_region_density_to_right =
1159           double(prev_region_live_to_right) / prev_region_space_to_right;
1160         if (density_to_right <= prev_region_density_to_right) {
1161           return dense_prefix;
1162         }
1163         if (TraceParallelOldGCDensePrefix && Verbose) {
1164           tty->print_cr("backing up from c=" SIZE_FORMAT_W(4) " d2r=%10.8f "
1165                         "pc_d2r=%10.8f", sd.region(cp), density_to_right,
1166                         prev_region_density_to_right);
1167         }
1168         dense_prefix -= region_size;
1169         live_to_right = prev_region_live_to_right;
1170         space_to_right = prev_region_space_to_right;
1171         density_to_right = prev_region_density_to_right;
1172       }
1173       return dense_prefix;
1174     }
1175 
1176     dense_prefix += region_size;
1177     ++cp;
1178   }
1179 
1180   return dense_prefix;
1181 }
1182 
1183 #ifndef PRODUCT
1184 void PSParallelCompact::print_dense_prefix_stats(const char* const algorithm,
1185                                                  const SpaceId id,
1186                                                  const bool maximum_compaction,
1187                                                  HeapWord* const addr)
1188 {
1189   const size_t region_idx = summary_data().addr_to_region_idx(addr);
1190   RegionData* const cp = summary_data().region(region_idx);
1191   const MutableSpace* const space = _space_info[id].space();
1192   HeapWord* const new_top = _space_info[id].new_top();
1193 
1194   const size_t space_live = pointer_delta(new_top, space->bottom());
1195   const size_t dead_to_left = pointer_delta(addr, cp->destination());
1196   const size_t space_cap = space->capacity_in_words();
1197   const double dead_to_left_pct = double(dead_to_left) / space_cap;
1198   const size_t live_to_right = new_top - cp->destination();
1199   const size_t dead_to_right = space->top() - addr - live_to_right;
1200 
1201   tty->print_cr("%s=" PTR_FORMAT " dpc=" SIZE_FORMAT_W(5) " "
1202                 "spl=" SIZE_FORMAT " "
1203                 "d2l=" SIZE_FORMAT " d2l%%=%6.4f "
1204                 "d2r=" SIZE_FORMAT " l2r=" SIZE_FORMAT
1205                 " ratio=%10.8f",
1206                 algorithm, addr, region_idx,
1207                 space_live,
1208                 dead_to_left, dead_to_left_pct,
1209                 dead_to_right, live_to_right,
1210                 double(dead_to_right) / live_to_right);
1211 }
1212 #endif  // #ifndef PRODUCT
1213 
1214 // Return a fraction indicating how much of the generation can be treated as
1215 // "dead wood" (i.e., not reclaimed).  The function uses a normal distribution
1216 // based on the density of live objects in the generation to determine a limit,
1217 // which is then adjusted so the return value is min_percent when the density is
1218 // 1.
1219 //
1220 // The following table shows some return values for a different values of the
1221 // standard deviation (ParallelOldDeadWoodLimiterStdDev); the mean is 0.5 and
1222 // min_percent is 1.
1223 //
1224 //                          fraction allowed as dead wood
1225 //         -----------------------------------------------------------------
1226 // density std_dev=70 std_dev=75 std_dev=80 std_dev=85 std_dev=90 std_dev=95
1227 // ------- ---------- ---------- ---------- ---------- ---------- ----------
1228 // 0.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000
1229 // 0.05000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941
1230 // 0.10000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272
1231 // 0.15000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066
1232 // 0.20000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975
1233 // 0.25000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313
1234 // 0.30000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132
1235 // 0.35000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289
1236 // 0.40000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500
1237 // 0.45000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386
1238 // 0.50000 0.13832410 0.11599237 0.09847664 0.08456518 0.07338887 0.06431510
1239 // 0.55000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386
1240 // 0.60000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500
1241 // 0.65000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289
1242 // 0.70000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132
1243 // 0.75000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313
1244 // 0.80000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975
1245 // 0.85000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066
1246 // 0.90000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272
1247 // 0.95000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941
1248 // 1.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000
1249 
1250 double PSParallelCompact::dead_wood_limiter(double density, size_t min_percent)
1251 {
1252   assert(_dwl_initialized, "uninitialized");
1253 
1254   // The raw limit is the value of the normal distribution at x = density.
1255   const double raw_limit = normal_distribution(density);
1256 
1257   // Adjust the raw limit so it becomes the minimum when the density is 1.
1258   //
1259   // First subtract the adjustment value (which is simply the precomputed value
1260   // normal_distribution(1.0)); this yields a value of 0 when the density is 1.
1261   // Then add the minimum value, so the minimum is returned when the density is
1262   // 1.  Finally, prevent negative values, which occur when the mean is not 0.5.
1263   const double min = double(min_percent) / 100.0;
1264   const double limit = raw_limit - _dwl_adjustment + min;
1265   return MAX2(limit, 0.0);
1266 }
1267 
1268 ParallelCompactData::RegionData*
1269 PSParallelCompact::first_dead_space_region(const RegionData* beg,
1270                                            const RegionData* end)
1271 {
1272   const size_t region_size = ParallelCompactData::RegionSize;
1273   ParallelCompactData& sd = summary_data();
1274   size_t left = sd.region(beg);
1275   size_t right = end > beg ? sd.region(end) - 1 : left;
1276 
1277   // Binary search.
1278   while (left < right) {
1279     // Equivalent to (left + right) / 2, but does not overflow.
1280     const size_t middle = left + (right - left) / 2;
1281     RegionData* const middle_ptr = sd.region(middle);
1282     HeapWord* const dest = middle_ptr->destination();
1283     HeapWord* const addr = sd.region_to_addr(middle);
1284     assert(dest != NULL, "sanity");
1285     assert(dest <= addr, "must move left");
1286 
1287     if (middle > left && dest < addr) {
1288       right = middle - 1;
1289     } else if (middle < right && middle_ptr->data_size() == region_size) {
1290       left = middle + 1;
1291     } else {
1292       return middle_ptr;
1293     }
1294   }
1295   return sd.region(left);
1296 }
1297 
1298 ParallelCompactData::RegionData*
1299 PSParallelCompact::dead_wood_limit_region(const RegionData* beg,
1300                                           const RegionData* end,
1301                                           size_t dead_words)
1302 {
1303   ParallelCompactData& sd = summary_data();
1304   size_t left = sd.region(beg);
1305   size_t right = end > beg ? sd.region(end) - 1 : left;
1306 
1307   // Binary search.
1308   while (left < right) {
1309     // Equivalent to (left + right) / 2, but does not overflow.
1310     const size_t middle = left + (right - left) / 2;
1311     RegionData* const middle_ptr = sd.region(middle);
1312     HeapWord* const dest = middle_ptr->destination();
1313     HeapWord* const addr = sd.region_to_addr(middle);
1314     assert(dest != NULL, "sanity");
1315     assert(dest <= addr, "must move left");
1316 
1317     const size_t dead_to_left = pointer_delta(addr, dest);
1318     if (middle > left && dead_to_left > dead_words) {
1319       right = middle - 1;
1320     } else if (middle < right && dead_to_left < dead_words) {
1321       left = middle + 1;
1322     } else {
1323       return middle_ptr;
1324     }
1325   }
1326   return sd.region(left);
1327 }
1328 
1329 // The result is valid during the summary phase, after the initial summarization
1330 // of each space into itself, and before final summarization.
1331 inline double
1332 PSParallelCompact::reclaimed_ratio(const RegionData* const cp,
1333                                    HeapWord* const bottom,
1334                                    HeapWord* const top,
1335                                    HeapWord* const new_top)
1336 {
1337   ParallelCompactData& sd = summary_data();
1338 
1339   assert(cp != NULL, "sanity");
1340   assert(bottom != NULL, "sanity");
1341   assert(top != NULL, "sanity");
1342   assert(new_top != NULL, "sanity");
1343   assert(top >= new_top, "summary data problem?");
1344   assert(new_top > bottom, "space is empty; should not be here");
1345   assert(new_top >= cp->destination(), "sanity");
1346   assert(top >= sd.region_to_addr(cp), "sanity");
1347 
1348   HeapWord* const destination = cp->destination();
1349   const size_t dense_prefix_live  = pointer_delta(destination, bottom);
1350   const size_t compacted_region_live = pointer_delta(new_top, destination);
1351   const size_t compacted_region_used = pointer_delta(top,
1352                                                      sd.region_to_addr(cp));
1353   const size_t reclaimable = compacted_region_used - compacted_region_live;
1354 
1355   const double divisor = dense_prefix_live + 1.25 * compacted_region_live;
1356   return double(reclaimable) / divisor;
1357 }
1358 
1359 // Return the address of the end of the dense prefix, a.k.a. the start of the
1360 // compacted region.  The address is always on a region boundary.
1361 //
1362 // Completely full regions at the left are skipped, since no compaction can
1363 // occur in those regions.  Then the maximum amount of dead wood to allow is
1364 // computed, based on the density (amount live / capacity) of the generation;
1365 // the region with approximately that amount of dead space to the left is
1366 // identified as the limit region.  Regions between the last completely full
1367 // region and the limit region are scanned and the one that has the best
1368 // (maximum) reclaimed_ratio() is selected.
1369 HeapWord*
1370 PSParallelCompact::compute_dense_prefix(const SpaceId id,
1371                                         bool maximum_compaction)
1372 {
1373   if (ParallelOldGCSplitALot) {
1374     if (_space_info[id].dense_prefix() != _space_info[id].space()->bottom()) {
1375       // The value was chosen to provoke splitting a young gen space; use it.
1376       return _space_info[id].dense_prefix();
1377     }
1378   }
1379 
1380   const size_t region_size = ParallelCompactData::RegionSize;
1381   const ParallelCompactData& sd = summary_data();
1382 
1383   const MutableSpace* const space = _space_info[id].space();
1384   HeapWord* const top = space->top();
1385   HeapWord* const top_aligned_up = sd.region_align_up(top);
1386   HeapWord* const new_top = _space_info[id].new_top();
1387   HeapWord* const new_top_aligned_up = sd.region_align_up(new_top);
1388   HeapWord* const bottom = space->bottom();
1389   const RegionData* const beg_cp = sd.addr_to_region_ptr(bottom);
1390   const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up);
1391   const RegionData* const new_top_cp =
1392     sd.addr_to_region_ptr(new_top_aligned_up);
1393 
1394   // Skip full regions at the beginning of the space--they are necessarily part
1395   // of the dense prefix.
1396   const RegionData* const full_cp = first_dead_space_region(beg_cp, new_top_cp);
1397   assert(full_cp->destination() == sd.region_to_addr(full_cp) ||
1398          space->is_empty(), "no dead space allowed to the left");
1399   assert(full_cp->data_size() < region_size || full_cp == new_top_cp - 1,
1400          "region must have dead space");
1401 
1402   // The gc number is saved whenever a maximum compaction is done, and used to
1403   // determine when the maximum compaction interval has expired.  This avoids
1404   // successive max compactions for different reasons.
1405   assert(total_invocations() >= _maximum_compaction_gc_num, "sanity");
1406   const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num;
1407   const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval ||
1408     total_invocations() == HeapFirstMaximumCompactionCount;
1409   if (maximum_compaction || full_cp == top_cp || interval_ended) {
1410     _maximum_compaction_gc_num = total_invocations();
1411     return sd.region_to_addr(full_cp);
1412   }
1413 
1414   const size_t space_live = pointer_delta(new_top, bottom);
1415   const size_t space_used = space->used_in_words();
1416   const size_t space_capacity = space->capacity_in_words();
1417 
1418   const double density = double(space_live) / double(space_capacity);
1419   const size_t min_percent_free = MarkSweepDeadRatio;
1420   const double limiter = dead_wood_limiter(density, min_percent_free);
1421   const size_t dead_wood_max = space_used - space_live;
1422   const size_t dead_wood_limit = MIN2(size_t(space_capacity * limiter),
1423                                       dead_wood_max);
1424 
1425   if (TraceParallelOldGCDensePrefix) {
1426     tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " "
1427                   "space_cap=" SIZE_FORMAT,
1428                   space_live, space_used,
1429                   space_capacity);
1430     tty->print_cr("dead_wood_limiter(%6.4f, %d)=%6.4f "
1431                   "dead_wood_max=" SIZE_FORMAT " dead_wood_limit=" SIZE_FORMAT,
1432                   density, min_percent_free, limiter,
1433                   dead_wood_max, dead_wood_limit);
1434   }
1435 
1436   // Locate the region with the desired amount of dead space to the left.
1437   const RegionData* const limit_cp =
1438     dead_wood_limit_region(full_cp, top_cp, dead_wood_limit);
1439 
1440   // Scan from the first region with dead space to the limit region and find the
1441   // one with the best (largest) reclaimed ratio.
1442   double best_ratio = 0.0;
1443   const RegionData* best_cp = full_cp;
1444   for (const RegionData* cp = full_cp; cp < limit_cp; ++cp) {
1445     double tmp_ratio = reclaimed_ratio(cp, bottom, top, new_top);
1446     if (tmp_ratio > best_ratio) {
1447       best_cp = cp;
1448       best_ratio = tmp_ratio;
1449     }
1450   }
1451 
1452 #if     0
1453   // Something to consider:  if the region with the best ratio is 'close to' the
1454   // first region w/free space, choose the first region with free space
1455   // ("first-free").  The first-free region is usually near the start of the
1456   // heap, which means we are copying most of the heap already, so copy a bit
1457   // more to get complete compaction.
1458   if (pointer_delta(best_cp, full_cp, sizeof(RegionData)) < 4) {
1459     _maximum_compaction_gc_num = total_invocations();
1460     best_cp = full_cp;
1461   }
1462 #endif  // #if 0
1463 
1464   return sd.region_to_addr(best_cp);
1465 }
1466 
1467 #ifndef PRODUCT
1468 void
1469 PSParallelCompact::fill_with_live_objects(SpaceId id, HeapWord* const start,
1470                                           size_t words)
1471 {
1472   if (TraceParallelOldGCSummaryPhase) {
1473     tty->print_cr("fill_with_live_objects [" PTR_FORMAT " " PTR_FORMAT ") "
1474                   SIZE_FORMAT, start, start + words, words);
1475   }
1476 
1477   ObjectStartArray* const start_array = _space_info[id].start_array();
1478   CollectedHeap::fill_with_objects(start, words);
1479   for (HeapWord* p = start; p < start + words; p += oop(p)->size()) {
1480     _mark_bitmap.mark_obj(p, words);
1481     _summary_data.add_obj(p, words);
1482     start_array->allocate_block(p);
1483   }
1484 }
1485 
1486 void
1487 PSParallelCompact::summarize_new_objects(SpaceId id, HeapWord* start)
1488 {
1489   ParallelCompactData& sd = summary_data();
1490   MutableSpace* space = _space_info[id].space();
1491 
1492   // Find the source and destination start addresses.
1493   HeapWord* const src_addr = sd.region_align_down(start);
1494   HeapWord* dst_addr;
1495   if (src_addr < start) {
1496     dst_addr = sd.addr_to_region_ptr(src_addr)->destination();
1497   } else if (src_addr > space->bottom()) {
1498     // The start (the original top() value) is aligned to a region boundary so
1499     // the associated region does not have a destination.  Compute the
1500     // destination from the previous region.
1501     RegionData* const cp = sd.addr_to_region_ptr(src_addr) - 1;
1502     dst_addr = cp->destination() + cp->data_size();
1503   } else {
1504     // Filling the entire space.
1505     dst_addr = space->bottom();
1506   }
1507   assert(dst_addr != NULL, "sanity");
1508 
1509   // Update the summary data.
1510   bool result = _summary_data.summarize(_space_info[id].split_info(),
1511                                         src_addr, space->top(), NULL,
1512                                         dst_addr, space->end(),
1513                                         _space_info[id].new_top_addr());
1514   assert(result, "should not fail:  bad filler object size");
1515 }
1516 
1517 void
1518 PSParallelCompact::provoke_split_fill_survivor(SpaceId id)
1519 {
1520   if (total_invocations() % (ParallelOldGCSplitInterval * 3) != 0) {
1521     return;
1522   }
1523 
1524   MutableSpace* const space = _space_info[id].space();
1525   if (space->is_empty()) {
1526     HeapWord* b = space->bottom();
1527     HeapWord* t = b + space->capacity_in_words() / 2;
1528     space->set_top(t);
1529     if (ZapUnusedHeapArea) {
1530       space->set_top_for_allocations();
1531     }
1532 
1533     size_t min_size = CollectedHeap::min_fill_size();
1534     size_t obj_len = min_size;
1535     while (b + obj_len <= t) {
1536       CollectedHeap::fill_with_object(b, obj_len);
1537       mark_bitmap()->mark_obj(b, obj_len);
1538       summary_data().add_obj(b, obj_len);
1539       b += obj_len;
1540       obj_len = (obj_len & (min_size*3)) + min_size; // 8 16 24 32 8 16 24 32 ...
1541     }
1542     if (b < t) {
1543       // The loop didn't completely fill to t (top); adjust top downward.
1544       space->set_top(b);
1545       if (ZapUnusedHeapArea) {
1546         space->set_top_for_allocations();
1547       }
1548     }
1549 
1550     HeapWord** nta = _space_info[id].new_top_addr();
1551     bool result = summary_data().summarize(_space_info[id].split_info(),
1552                                            space->bottom(), space->top(), NULL,
1553                                            space->bottom(), space->end(), nta);
1554     assert(result, "space must fit into itself");
1555   }
1556 }
1557 
1558 void
1559 PSParallelCompact::provoke_split(bool & max_compaction)
1560 {
1561   if (total_invocations() % ParallelOldGCSplitInterval != 0) {
1562     return;
1563   }
1564 
1565   const size_t region_size = ParallelCompactData::RegionSize;
1566   ParallelCompactData& sd = summary_data();
1567 
1568   MutableSpace* const eden_space = _space_info[eden_space_id].space();
1569   MutableSpace* const from_space = _space_info[from_space_id].space();
1570   const size_t eden_live = pointer_delta(eden_space->top(),
1571                                          _space_info[eden_space_id].new_top());
1572   const size_t from_live = pointer_delta(from_space->top(),
1573                                          _space_info[from_space_id].new_top());
1574 
1575   const size_t min_fill_size = CollectedHeap::min_fill_size();
1576   const size_t eden_free = pointer_delta(eden_space->end(), eden_space->top());
1577   const size_t eden_fillable = eden_free >= min_fill_size ? eden_free : 0;
1578   const size_t from_free = pointer_delta(from_space->end(), from_space->top());
1579   const size_t from_fillable = from_free >= min_fill_size ? from_free : 0;
1580 
1581   // Choose the space to split; need at least 2 regions live (or fillable).
1582   SpaceId id;
1583   MutableSpace* space;
1584   size_t live_words;
1585   size_t fill_words;
1586   if (eden_live + eden_fillable >= region_size * 2) {
1587     id = eden_space_id;
1588     space = eden_space;
1589     live_words = eden_live;
1590     fill_words = eden_fillable;
1591   } else if (from_live + from_fillable >= region_size * 2) {
1592     id = from_space_id;
1593     space = from_space;
1594     live_words = from_live;
1595     fill_words = from_fillable;
1596   } else {
1597     return; // Give up.
1598   }
1599   assert(fill_words == 0 || fill_words >= min_fill_size, "sanity");
1600 
1601   if (live_words < region_size * 2) {
1602     // Fill from top() to end() w/live objects of mixed sizes.
1603     HeapWord* const fill_start = space->top();
1604     live_words += fill_words;
1605 
1606     space->set_top(fill_start + fill_words);
1607     if (ZapUnusedHeapArea) {
1608       space->set_top_for_allocations();
1609     }
1610 
1611     HeapWord* cur_addr = fill_start;
1612     while (fill_words > 0) {
1613       const size_t r = (size_t)os::random() % (region_size / 2) + min_fill_size;
1614       size_t cur_size = MIN2(align_object_size_(r), fill_words);
1615       if (fill_words - cur_size < min_fill_size) {
1616         cur_size = fill_words; // Avoid leaving a fragment too small to fill.
1617       }
1618 
1619       CollectedHeap::fill_with_object(cur_addr, cur_size);
1620       mark_bitmap()->mark_obj(cur_addr, cur_size);
1621       sd.add_obj(cur_addr, cur_size);
1622 
1623       cur_addr += cur_size;
1624       fill_words -= cur_size;
1625     }
1626 
1627     summarize_new_objects(id, fill_start);
1628   }
1629 
1630   max_compaction = false;
1631 
1632   // Manipulate the old gen so that it has room for about half of the live data
1633   // in the target young gen space (live_words / 2).
1634   id = old_space_id;
1635   space = _space_info[id].space();
1636   const size_t free_at_end = space->free_in_words();
1637   const size_t free_target = align_object_size(live_words / 2);
1638   const size_t dead = pointer_delta(space->top(), _space_info[id].new_top());
1639 
1640   if (free_at_end >= free_target + min_fill_size) {
1641     // Fill space above top() and set the dense prefix so everything survives.
1642     HeapWord* const fill_start = space->top();
1643     const size_t fill_size = free_at_end - free_target;
1644     space->set_top(space->top() + fill_size);
1645     if (ZapUnusedHeapArea) {
1646       space->set_top_for_allocations();
1647     }
1648     fill_with_live_objects(id, fill_start, fill_size);
1649     summarize_new_objects(id, fill_start);
1650     _space_info[id].set_dense_prefix(sd.region_align_down(space->top()));
1651   } else if (dead + free_at_end > free_target) {
1652     // Find a dense prefix that makes the right amount of space available.
1653     HeapWord* cur = sd.region_align_down(space->top());
1654     HeapWord* cur_destination = sd.addr_to_region_ptr(cur)->destination();
1655     size_t dead_to_right = pointer_delta(space->end(), cur_destination);
1656     while (dead_to_right < free_target) {
1657       cur -= region_size;
1658       cur_destination = sd.addr_to_region_ptr(cur)->destination();
1659       dead_to_right = pointer_delta(space->end(), cur_destination);
1660     }
1661     _space_info[id].set_dense_prefix(cur);
1662   }
1663 }
1664 #endif // #ifndef PRODUCT
1665 
1666 void PSParallelCompact::summarize_spaces_quick()
1667 {
1668   for (unsigned int i = 0; i < last_space_id; ++i) {
1669     const MutableSpace* space = _space_info[i].space();
1670     HeapWord** nta = _space_info[i].new_top_addr();
1671     bool result = _summary_data.summarize(_space_info[i].split_info(),
1672                                           space->bottom(), space->top(), NULL,
1673                                           space->bottom(), space->end(), nta);
1674     assert(result, "space must fit into itself");
1675     _space_info[i].set_dense_prefix(space->bottom());
1676   }
1677 
1678 #ifndef PRODUCT
1679   if (ParallelOldGCSplitALot) {
1680     provoke_split_fill_survivor(to_space_id);
1681   }
1682 #endif // #ifndef PRODUCT
1683 }
1684 
1685 void PSParallelCompact::fill_dense_prefix_end(SpaceId id)
1686 {
1687   HeapWord* const dense_prefix_end = dense_prefix(id);
1688   const RegionData* region = _summary_data.addr_to_region_ptr(dense_prefix_end);
1689   const idx_t dense_prefix_bit = _mark_bitmap.addr_to_bit(dense_prefix_end);
1690   if (dead_space_crosses_boundary(region, dense_prefix_bit)) {
1691     // Only enough dead space is filled so that any remaining dead space to the
1692     // left is larger than the minimum filler object.  (The remainder is filled
1693     // during the copy/update phase.)
1694     //
1695     // The size of the dead space to the right of the boundary is not a
1696     // concern, since compaction will be able to use whatever space is
1697     // available.
1698     //
1699     // Here '||' is the boundary, 'x' represents a don't care bit and a box
1700     // surrounds the space to be filled with an object.
1701     //
1702     // In the 32-bit VM, each bit represents two 32-bit words:
1703     //                              +---+
1704     // a) beg_bits:  ...  x   x   x | 0 | ||   0   x  x  ...
1705     //    end_bits:  ...  x   x   x | 0 | ||   0   x  x  ...
1706     //                              +---+
1707     //
1708     // In the 64-bit VM, each bit represents one 64-bit word:
1709     //                              +------------+
1710     // b) beg_bits:  ...  x   x   x | 0   ||   0 | x  x  ...
1711     //    end_bits:  ...  x   x   1 | 0   ||   0 | x  x  ...
1712     //                              +------------+
1713     //                          +-------+
1714     // c) beg_bits:  ...  x   x | 0   0 | ||   0   x  x  ...
1715     //    end_bits:  ...  x   1 | 0   0 | ||   0   x  x  ...
1716     //                          +-------+
1717     //                      +-----------+
1718     // d) beg_bits:  ...  x | 0   0   0 | ||   0   x  x  ...
1719     //    end_bits:  ...  1 | 0   0   0 | ||   0   x  x  ...
1720     //                      +-----------+
1721     //                          +-------+
1722     // e) beg_bits:  ...  0   0 | 0   0 | ||   0   x  x  ...
1723     //    end_bits:  ...  0   0 | 0   0 | ||   0   x  x  ...
1724     //                          +-------+
1725 
1726     // Initially assume case a, c or e will apply.
1727     size_t obj_len = CollectedHeap::min_fill_size();
1728     HeapWord* obj_beg = dense_prefix_end - obj_len;
1729 
1730 #ifdef  _LP64
1731     if (MinObjAlignment > 1) { // object alignment > heap word size
1732       // Cases a, c or e.
1733     } else if (_mark_bitmap.is_obj_end(dense_prefix_bit - 2)) {
1734       // Case b above.
1735       obj_beg = dense_prefix_end - 1;
1736     } else if (!_mark_bitmap.is_obj_end(dense_prefix_bit - 3) &&
1737                _mark_bitmap.is_obj_end(dense_prefix_bit - 4)) {
1738       // Case d above.
1739       obj_beg = dense_prefix_end - 3;
1740       obj_len = 3;
1741     }
1742 #endif  // #ifdef _LP64
1743 
1744     CollectedHeap::fill_with_object(obj_beg, obj_len);
1745     _mark_bitmap.mark_obj(obj_beg, obj_len);
1746     _summary_data.add_obj(obj_beg, obj_len);
1747     assert(start_array(id) != NULL, "sanity");
1748     start_array(id)->allocate_block(obj_beg);
1749   }
1750 }
1751 
1752 void
1753 PSParallelCompact::clear_source_region(HeapWord* beg_addr, HeapWord* end_addr)
1754 {
1755   RegionData* const beg_ptr = _summary_data.addr_to_region_ptr(beg_addr);
1756   HeapWord* const end_aligned_up = _summary_data.region_align_up(end_addr);
1757   RegionData* const end_ptr = _summary_data.addr_to_region_ptr(end_aligned_up);
1758   for (RegionData* cur = beg_ptr; cur < end_ptr; ++cur) {
1759     cur->set_source_region(0);
1760   }
1761 }
1762 
1763 void
1764 PSParallelCompact::summarize_space(SpaceId id, bool maximum_compaction)
1765 {
1766   assert(id < last_space_id, "id out of range");
1767   assert(_space_info[id].dense_prefix() == _space_info[id].space()->bottom() ||
1768          ParallelOldGCSplitALot && id == old_space_id,
1769          "should have been reset in summarize_spaces_quick()");
1770 
1771   const MutableSpace* space = _space_info[id].space();
1772   if (_space_info[id].new_top() != space->bottom()) {
1773     HeapWord* dense_prefix_end = compute_dense_prefix(id, maximum_compaction);
1774     _space_info[id].set_dense_prefix(dense_prefix_end);
1775 
1776 #ifndef PRODUCT
1777     if (TraceParallelOldGCDensePrefix) {
1778       print_dense_prefix_stats("ratio", id, maximum_compaction,
1779                                dense_prefix_end);
1780       HeapWord* addr = compute_dense_prefix_via_density(id, maximum_compaction);
1781       print_dense_prefix_stats("density", id, maximum_compaction, addr);
1782     }
1783 #endif  // #ifndef PRODUCT
1784 
1785     // Recompute the summary data, taking into account the dense prefix.  If
1786     // every last byte will be reclaimed, then the existing summary data which
1787     // compacts everything can be left in place.
1788     if (!maximum_compaction && dense_prefix_end != space->bottom()) {
1789       // If dead space crosses the dense prefix boundary, it is (at least
1790       // partially) filled with a dummy object, marked live and added to the
1791       // summary data.  This simplifies the copy/update phase and must be done
1792       // before the final locations of objects are determined, to prevent
1793       // leaving a fragment of dead space that is too small to fill.
1794       fill_dense_prefix_end(id);
1795 
1796       // Compute the destination of each Region, and thus each object.
1797       _summary_data.summarize_dense_prefix(space->bottom(), dense_prefix_end);
1798       _summary_data.summarize(_space_info[id].split_info(),
1799                               dense_prefix_end, space->top(), NULL,
1800                               dense_prefix_end, space->end(),
1801                               _space_info[id].new_top_addr());
1802     }
1803   }
1804 
1805   if (TraceParallelOldGCSummaryPhase) {
1806     const size_t region_size = ParallelCompactData::RegionSize;
1807     HeapWord* const dense_prefix_end = _space_info[id].dense_prefix();
1808     const size_t dp_region = _summary_data.addr_to_region_idx(dense_prefix_end);
1809     const size_t dp_words = pointer_delta(dense_prefix_end, space->bottom());
1810     HeapWord* const new_top = _space_info[id].new_top();
1811     const HeapWord* nt_aligned_up = _summary_data.region_align_up(new_top);
1812     const size_t cr_words = pointer_delta(nt_aligned_up, dense_prefix_end);
1813     tty->print_cr("id=%d cap=" SIZE_FORMAT " dp=" PTR_FORMAT " "
1814                   "dp_region=" SIZE_FORMAT " " "dp_count=" SIZE_FORMAT " "
1815                   "cr_count=" SIZE_FORMAT " " "nt=" PTR_FORMAT,
1816                   id, space->capacity_in_words(), dense_prefix_end,
1817                   dp_region, dp_words / region_size,
1818                   cr_words / region_size, new_top);
1819   }
1820 }
1821 
1822 #ifndef PRODUCT
1823 void PSParallelCompact::summary_phase_msg(SpaceId dst_space_id,
1824                                           HeapWord* dst_beg, HeapWord* dst_end,
1825                                           SpaceId src_space_id,
1826                                           HeapWord* src_beg, HeapWord* src_end)
1827 {
1828   if (TraceParallelOldGCSummaryPhase) {
1829     tty->print_cr("summarizing %d [%s] into %d [%s]:  "
1830                   "src=" PTR_FORMAT "-" PTR_FORMAT " "
1831                   SIZE_FORMAT "-" SIZE_FORMAT " "
1832                   "dst=" PTR_FORMAT "-" PTR_FORMAT " "
1833                   SIZE_FORMAT "-" SIZE_FORMAT,
1834                   src_space_id, space_names[src_space_id],
1835                   dst_space_id, space_names[dst_space_id],
1836                   src_beg, src_end,
1837                   _summary_data.addr_to_region_idx(src_beg),
1838                   _summary_data.addr_to_region_idx(src_end),
1839                   dst_beg, dst_end,
1840                   _summary_data.addr_to_region_idx(dst_beg),
1841                   _summary_data.addr_to_region_idx(dst_end));
1842   }
1843 }
1844 #endif  // #ifndef PRODUCT
1845 
1846 void PSParallelCompact::summary_phase(ParCompactionManager* cm,
1847                                       bool maximum_compaction)
1848 {
1849   GCTraceTime tm("summary phase", print_phases(), true, &_gc_timer);
1850   // trace("2");
1851 
1852 #ifdef  ASSERT
1853   if (TraceParallelOldGCMarkingPhase) {
1854     tty->print_cr("add_obj_count=" SIZE_FORMAT " "
1855                   "add_obj_bytes=" SIZE_FORMAT,
1856                   add_obj_count, add_obj_size * HeapWordSize);
1857     tty->print_cr("mark_bitmap_count=" SIZE_FORMAT " "
1858                   "mark_bitmap_bytes=" SIZE_FORMAT,
1859                   mark_bitmap_count, mark_bitmap_size * HeapWordSize);
1860   }
1861 #endif  // #ifdef ASSERT
1862 
1863   // Quick summarization of each space into itself, to see how much is live.
1864   summarize_spaces_quick();
1865 
1866   if (TraceParallelOldGCSummaryPhase) {
1867     tty->print_cr("summary_phase:  after summarizing each space to self");
1868     Universe::print();
1869     NOT_PRODUCT(print_region_ranges());
1870     if (Verbose) {
1871       NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info));
1872     }
1873   }
1874 
1875   // The amount of live data that will end up in old space (assuming it fits).
1876   size_t old_space_total_live = 0;
1877   for (unsigned int id = old_space_id; id < last_space_id; ++id) {
1878     old_space_total_live += pointer_delta(_space_info[id].new_top(),
1879                                           _space_info[id].space()->bottom());
1880   }
1881 
1882   MutableSpace* const old_space = _space_info[old_space_id].space();
1883   const size_t old_capacity = old_space->capacity_in_words();
1884   if (old_space_total_live > old_capacity) {
1885     // XXX - should also try to expand
1886     maximum_compaction = true;
1887   }
1888 #ifndef PRODUCT
1889   if (ParallelOldGCSplitALot && old_space_total_live < old_capacity) {
1890     provoke_split(maximum_compaction);
1891   }
1892 #endif // #ifndef PRODUCT
1893 
1894   // Old generations.
1895   summarize_space(old_space_id, maximum_compaction);
1896 
1897   // Summarize the remaining spaces in the young gen.  The initial target space
1898   // is the old gen.  If a space does not fit entirely into the target, then the
1899   // remainder is compacted into the space itself and that space becomes the new
1900   // target.
1901   SpaceId dst_space_id = old_space_id;
1902   HeapWord* dst_space_end = old_space->end();
1903   HeapWord** new_top_addr = _space_info[dst_space_id].new_top_addr();
1904   for (unsigned int id = eden_space_id; id < last_space_id; ++id) {
1905     const MutableSpace* space = _space_info[id].space();
1906     const size_t live = pointer_delta(_space_info[id].new_top(),
1907                                       space->bottom());
1908     const size_t available = pointer_delta(dst_space_end, *new_top_addr);
1909 
1910     NOT_PRODUCT(summary_phase_msg(dst_space_id, *new_top_addr, dst_space_end,
1911                                   SpaceId(id), space->bottom(), space->top());)
1912     if (live > 0 && live <= available) {
1913       // All the live data will fit.
1914       bool done = _summary_data.summarize(_space_info[id].split_info(),
1915                                           space->bottom(), space->top(),
1916                                           NULL,
1917                                           *new_top_addr, dst_space_end,
1918                                           new_top_addr);
1919       assert(done, "space must fit into old gen");
1920 
1921       // Reset the new_top value for the space.
1922       _space_info[id].set_new_top(space->bottom());
1923     } else if (live > 0) {
1924       // Attempt to fit part of the source space into the target space.
1925       HeapWord* next_src_addr = NULL;
1926       bool done = _summary_data.summarize(_space_info[id].split_info(),
1927                                           space->bottom(), space->top(),
1928                                           &next_src_addr,
1929                                           *new_top_addr, dst_space_end,
1930                                           new_top_addr);
1931       assert(!done, "space should not fit into old gen");
1932       assert(next_src_addr != NULL, "sanity");
1933 
1934       // The source space becomes the new target, so the remainder is compacted
1935       // within the space itself.
1936       dst_space_id = SpaceId(id);
1937       dst_space_end = space->end();
1938       new_top_addr = _space_info[id].new_top_addr();
1939       NOT_PRODUCT(summary_phase_msg(dst_space_id,
1940                                     space->bottom(), dst_space_end,
1941                                     SpaceId(id), next_src_addr, space->top());)
1942       done = _summary_data.summarize(_space_info[id].split_info(),
1943                                      next_src_addr, space->top(),
1944                                      NULL,
1945                                      space->bottom(), dst_space_end,
1946                                      new_top_addr);
1947       assert(done, "space must fit when compacted into itself");
1948       assert(*new_top_addr <= space->top(), "usage should not grow");
1949     }
1950   }
1951 
1952   if (TraceParallelOldGCSummaryPhase) {
1953     tty->print_cr("summary_phase:  after final summarization");
1954     Universe::print();
1955     NOT_PRODUCT(print_region_ranges());
1956     if (Verbose) {
1957       NOT_PRODUCT(print_generic_summary_data(_summary_data, _space_info));
1958     }
1959   }
1960 }
1961 
1962 // This method should contain all heap-specific policy for invoking a full
1963 // collection.  invoke_no_policy() will only attempt to compact the heap; it
1964 // will do nothing further.  If we need to bail out for policy reasons, scavenge
1965 // before full gc, or any other specialized behavior, it needs to be added here.
1966 //
1967 // Note that this method should only be called from the vm_thread while at a
1968 // safepoint.
1969 //
1970 // Note that the all_soft_refs_clear flag in the collector policy
1971 // may be true because this method can be called without intervening
1972 // activity.  For example when the heap space is tight and full measure
1973 // are being taken to free space.
1974 void PSParallelCompact::invoke(bool maximum_heap_compaction) {
1975   assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
1976   assert(Thread::current() == (Thread*)VMThread::vm_thread(),
1977          "should be in vm thread");
1978 
1979   ParallelScavengeHeap* heap = gc_heap();
1980   GCCause::Cause gc_cause = heap->gc_cause();
1981   assert(!heap->is_gc_active(), "not reentrant");
1982 
1983   PSAdaptiveSizePolicy* policy = heap->size_policy();
1984   IsGCActiveMark mark;
1985 
1986   if (ScavengeBeforeFullGC) {
1987     PSScavenge::invoke_no_policy();
1988   }
1989 
1990   const bool clear_all_soft_refs =
1991     heap->collector_policy()->should_clear_all_soft_refs();
1992 
1993   PSParallelCompact::invoke_no_policy(clear_all_soft_refs ||
1994                                       maximum_heap_compaction);
1995 }
1996 
1997 // This method contains no policy. You should probably
1998 // be calling invoke() instead.
1999 bool PSParallelCompact::invoke_no_policy(bool maximum_heap_compaction) {
2000   assert(SafepointSynchronize::is_at_safepoint(), "must be at a safepoint");
2001   assert(ref_processor() != NULL, "Sanity");
2002 
2003   if (GC_locker::check_active_before_gc()) {
2004     return false;
2005   }
2006 
2007   ParallelScavengeHeap* heap = gc_heap();
2008 
2009   _gc_timer.register_gc_start(os::elapsed_counter());
2010   _gc_tracer.report_gc_start(heap->gc_cause(), _gc_timer.gc_start());
2011 
2012   TimeStamp marking_start;
2013   TimeStamp compaction_start;
2014   TimeStamp collection_exit;
2015 
2016   GCCause::Cause gc_cause = heap->gc_cause();
2017   PSYoungGen* young_gen = heap->young_gen();
2018   PSOldGen* old_gen = heap->old_gen();
2019   PSAdaptiveSizePolicy* size_policy = heap->size_policy();
2020 
2021   // The scope of casr should end after code that can change
2022   // CollectorPolicy::_should_clear_all_soft_refs.
2023   ClearedAllSoftRefs casr(maximum_heap_compaction,
2024                           heap->collector_policy());
2025 
2026   if (ZapUnusedHeapArea) {
2027     // Save information needed to minimize mangling
2028     heap->record_gen_tops_before_GC();
2029   }
2030 
2031   heap->pre_full_gc_dump(&_gc_timer);
2032 
2033   _print_phases = PrintGCDetails && PrintParallelOldGCPhaseTimes;
2034 
2035   // Make sure data structures are sane, make the heap parsable, and do other
2036   // miscellaneous bookkeeping.
2037   PreGCValues pre_gc_values;
2038   pre_compact(&pre_gc_values);
2039 
2040   // Get the compaction manager reserved for the VM thread.
2041   ParCompactionManager* const vmthread_cm =
2042     ParCompactionManager::manager_array(gc_task_manager()->workers());
2043 
2044   // Place after pre_compact() where the number of invocations is incremented.
2045   AdaptiveSizePolicyOutput(size_policy, heap->total_collections());
2046 
2047   {
2048     ResourceMark rm;
2049     HandleMark hm;
2050 
2051     // Set the number of GC threads to be used in this collection
2052     gc_task_manager()->set_active_gang();
2053     gc_task_manager()->task_idle_workers();
2054     heap->set_par_threads(gc_task_manager()->active_workers());
2055 
2056     gclog_or_tty->date_stamp(PrintGC && PrintGCDateStamps);
2057     TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
2058     GCTraceTime t1(GCCauseString("Full GC", gc_cause), PrintGC, !PrintGCDetails, NULL);
2059     TraceCollectorStats tcs(counters());
2060     TraceMemoryManagerStats tms(true /* Full GC */,gc_cause);
2061 
2062     if (TraceGen1Time) accumulated_time()->start();
2063 
2064     // Let the size policy know we're starting
2065     size_policy->major_collection_begin();
2066 
2067     CodeCache::gc_prologue();
2068     Threads::gc_prologue();
2069 
2070     COMPILER2_PRESENT(DerivedPointerTable::clear());
2071 
2072     ref_processor()->enable_discovery(true /*verify_disabled*/, true /*verify_no_refs*/);
2073     ref_processor()->setup_policy(maximum_heap_compaction);
2074 
2075     bool marked_for_unloading = false;
2076 
2077     marking_start.update();
2078     marking_phase(vmthread_cm, maximum_heap_compaction, &_gc_tracer);
2079 
2080     bool max_on_system_gc = UseMaximumCompactionOnSystemGC
2081       && gc_cause == GCCause::_java_lang_system_gc;
2082     summary_phase(vmthread_cm, maximum_heap_compaction || max_on_system_gc);
2083 
2084     COMPILER2_PRESENT(assert(DerivedPointerTable::is_active(), "Sanity"));
2085     COMPILER2_PRESENT(DerivedPointerTable::set_active(false));
2086 
2087     // adjust_roots() updates Universe::_intArrayKlassObj which is
2088     // needed by the compaction for filling holes in the dense prefix.
2089     adjust_roots();
2090 
2091     compaction_start.update();
2092     compact();
2093 
2094     // Reset the mark bitmap, summary data, and do other bookkeeping.  Must be
2095     // done before resizing.
2096     post_compact();
2097 
2098     // Let the size policy know we're done
2099     size_policy->major_collection_end(old_gen->used_in_bytes(), gc_cause);
2100 
2101     if (UseAdaptiveSizePolicy) {
2102       if (PrintAdaptiveSizePolicy) {
2103         gclog_or_tty->print("AdaptiveSizeStart: ");
2104         gclog_or_tty->stamp();
2105         gclog_or_tty->print_cr(" collection: %d ",
2106                        heap->total_collections());
2107         if (Verbose) {
2108           gclog_or_tty->print("old_gen_capacity: %d young_gen_capacity: %d",
2109             old_gen->capacity_in_bytes(), young_gen->capacity_in_bytes());
2110         }
2111       }
2112 
2113       // Don't check if the size_policy is ready here.  Let
2114       // the size_policy check that internally.
2115       if (UseAdaptiveGenerationSizePolicyAtMajorCollection &&
2116           ((gc_cause != GCCause::_java_lang_system_gc) ||
2117             UseAdaptiveSizePolicyWithSystemGC)) {
2118         // Calculate optimal free space amounts
2119         assert(young_gen->max_size() >
2120           young_gen->from_space()->capacity_in_bytes() +
2121           young_gen->to_space()->capacity_in_bytes(),
2122           "Sizes of space in young gen are out-of-bounds");
2123 
2124         size_t young_live = young_gen->used_in_bytes();
2125         size_t eden_live = young_gen->eden_space()->used_in_bytes();
2126         size_t old_live = old_gen->used_in_bytes();
2127         size_t cur_eden = young_gen->eden_space()->capacity_in_bytes();
2128         size_t max_old_gen_size = old_gen->max_gen_size();
2129         size_t max_eden_size = young_gen->max_size() -
2130           young_gen->from_space()->capacity_in_bytes() -
2131           young_gen->to_space()->capacity_in_bytes();
2132 
2133         // Used for diagnostics
2134         size_policy->clear_generation_free_space_flags();
2135 
2136         size_policy->compute_generations_free_space(young_live,
2137                                                     eden_live,
2138                                                     old_live,
2139                                                     cur_eden,
2140                                                     max_old_gen_size,
2141                                                     max_eden_size,
2142                                                     true /* full gc*/);
2143 
2144         size_policy->check_gc_overhead_limit(young_live,
2145                                              eden_live,
2146                                              max_old_gen_size,
2147                                              max_eden_size,
2148                                              true /* full gc*/,
2149                                              gc_cause,
2150                                              heap->collector_policy());
2151 
2152         size_policy->decay_supplemental_growth(true /* full gc*/);
2153 
2154         heap->resize_old_gen(
2155           size_policy->calculated_old_free_size_in_bytes());
2156 
2157         // Don't resize the young generation at an major collection.  A
2158         // desired young generation size may have been calculated but
2159         // resizing the young generation complicates the code because the
2160         // resizing of the old generation may have moved the boundary
2161         // between the young generation and the old generation.  Let the
2162         // young generation resizing happen at the minor collections.
2163       }
2164       if (PrintAdaptiveSizePolicy) {
2165         gclog_or_tty->print_cr("AdaptiveSizeStop: collection: %d ",
2166                        heap->total_collections());
2167       }
2168     }
2169 
2170     if (UsePerfData) {
2171       PSGCAdaptivePolicyCounters* const counters = heap->gc_policy_counters();
2172       counters->update_counters();
2173       counters->update_old_capacity(old_gen->capacity_in_bytes());
2174       counters->update_young_capacity(young_gen->capacity_in_bytes());
2175     }
2176 
2177     heap->resize_all_tlabs();
2178 
2179     // Resize the metaspace capactiy after a collection
2180     MetaspaceGC::compute_new_size();
2181 
2182     if (TraceGen1Time) accumulated_time()->stop();
2183 
2184     if (PrintGC) {
2185       if (PrintGCDetails) {
2186         // No GC timestamp here.  This is after GC so it would be confusing.
2187         young_gen->print_used_change(pre_gc_values.young_gen_used());
2188         old_gen->print_used_change(pre_gc_values.old_gen_used());
2189         heap->print_heap_change(pre_gc_values.heap_used());
2190         MetaspaceAux::print_metaspace_change(pre_gc_values.metadata_used());
2191       } else {
2192         heap->print_heap_change(pre_gc_values.heap_used());
2193       }
2194     }
2195 
2196     // Track memory usage and detect low memory
2197     MemoryService::track_memory_usage();
2198     heap->update_counters();
2199     gc_task_manager()->release_idle_workers();
2200   }
2201 
2202 #ifdef ASSERT
2203   for (size_t i = 0; i < ParallelGCThreads + 1; ++i) {
2204     ParCompactionManager* const cm =
2205       ParCompactionManager::manager_array(int(i));
2206     assert(cm->marking_stack()->is_empty(),       "should be empty");
2207     assert(ParCompactionManager::region_list(int(i))->is_empty(), "should be empty");
2208   }
2209 #endif // ASSERT
2210 
2211   if (VerifyAfterGC && heap->total_collections() >= VerifyGCStartAt) {
2212     HandleMark hm;  // Discard invalid handles created during verification
2213     Universe::verify(" VerifyAfterGC:");
2214   }
2215 
2216   // Re-verify object start arrays
2217   if (VerifyObjectStartArray &&
2218       VerifyAfterGC) {
2219     old_gen->verify_object_start_array();
2220   }
2221 
2222   if (ZapUnusedHeapArea) {
2223     old_gen->object_space()->check_mangled_unused_area_complete();
2224   }
2225 
2226   NOT_PRODUCT(ref_processor()->verify_no_references_recorded());
2227 
2228   collection_exit.update();
2229 
2230   heap->print_heap_after_gc();
2231   heap->trace_heap_after_gc(&_gc_tracer);
2232 
2233   if (PrintGCTaskTimeStamps) {
2234     gclog_or_tty->print_cr("VM-Thread " INT64_FORMAT " " INT64_FORMAT " "
2235                            INT64_FORMAT,
2236                            marking_start.ticks(), compaction_start.ticks(),
2237                            collection_exit.ticks());
2238     gc_task_manager()->print_task_time_stamps();
2239   }
2240 
2241   heap->post_full_gc_dump(&_gc_timer);
2242 
2243 #ifdef TRACESPINNING
2244   ParallelTaskTerminator::print_termination_counts();
2245 #endif
2246 
2247   _gc_timer.register_gc_end(os::elapsed_counter());
2248 
2249   _gc_tracer.report_dense_prefix(dense_prefix(old_space_id));
2250   _gc_tracer.report_gc_end(_gc_timer.gc_end(), _gc_timer.time_partitions());
2251 
2252   return true;
2253 }
2254 
2255 bool PSParallelCompact::absorb_live_data_from_eden(PSAdaptiveSizePolicy* size_policy,
2256                                              PSYoungGen* young_gen,
2257                                              PSOldGen* old_gen) {
2258   MutableSpace* const eden_space = young_gen->eden_space();
2259   assert(!eden_space->is_empty(), "eden must be non-empty");
2260   assert(young_gen->virtual_space()->alignment() ==
2261          old_gen->virtual_space()->alignment(), "alignments do not match");
2262 
2263   if (!(UseAdaptiveSizePolicy && UseAdaptiveGCBoundary)) {
2264     return false;
2265   }
2266 
2267   // Both generations must be completely committed.
2268   if (young_gen->virtual_space()->uncommitted_size() != 0) {
2269     return false;
2270   }
2271   if (old_gen->virtual_space()->uncommitted_size() != 0) {
2272     return false;
2273   }
2274 
2275   // Figure out how much to take from eden.  Include the average amount promoted
2276   // in the total; otherwise the next young gen GC will simply bail out to a
2277   // full GC.
2278   const size_t alignment = old_gen->virtual_space()->alignment();
2279   const size_t eden_used = eden_space->used_in_bytes();
2280   const size_t promoted = (size_t)size_policy->avg_promoted()->padded_average();
2281   const size_t absorb_size = align_size_up(eden_used + promoted, alignment);
2282   const size_t eden_capacity = eden_space->capacity_in_bytes();
2283 
2284   if (absorb_size >= eden_capacity) {
2285     return false; // Must leave some space in eden.
2286   }
2287 
2288   const size_t new_young_size = young_gen->capacity_in_bytes() - absorb_size;
2289   if (new_young_size < young_gen->min_gen_size()) {
2290     return false; // Respect young gen minimum size.
2291   }
2292 
2293   if (TraceAdaptiveGCBoundary && Verbose) {
2294     gclog_or_tty->print(" absorbing " SIZE_FORMAT "K:  "
2295                         "eden " SIZE_FORMAT "K->" SIZE_FORMAT "K "
2296                         "from " SIZE_FORMAT "K, to " SIZE_FORMAT "K "
2297                         "young_gen " SIZE_FORMAT "K->" SIZE_FORMAT "K ",
2298                         absorb_size / K,
2299                         eden_capacity / K, (eden_capacity - absorb_size) / K,
2300                         young_gen->from_space()->used_in_bytes() / K,
2301                         young_gen->to_space()->used_in_bytes() / K,
2302                         young_gen->capacity_in_bytes() / K, new_young_size / K);
2303   }
2304 
2305   // Fill the unused part of the old gen.
2306   MutableSpace* const old_space = old_gen->object_space();
2307   HeapWord* const unused_start = old_space->top();
2308   size_t const unused_words = pointer_delta(old_space->end(), unused_start);
2309 
2310   if (unused_words > 0) {
2311     if (unused_words < CollectedHeap::min_fill_size()) {
2312       return false;  // If the old gen cannot be filled, must give up.
2313     }
2314     CollectedHeap::fill_with_objects(unused_start, unused_words);
2315   }
2316 
2317   // Take the live data from eden and set both top and end in the old gen to
2318   // eden top.  (Need to set end because reset_after_change() mangles the region
2319   // from end to virtual_space->high() in debug builds).
2320   HeapWord* const new_top = eden_space->top();
2321   old_gen->virtual_space()->expand_into(young_gen->virtual_space(),
2322                                         absorb_size);
2323   young_gen->reset_after_change();
2324   old_space->set_top(new_top);
2325   old_space->set_end(new_top);
2326   old_gen->reset_after_change();
2327 
2328   // Update the object start array for the filler object and the data from eden.
2329   ObjectStartArray* const start_array = old_gen->start_array();
2330   for (HeapWord* p = unused_start; p < new_top; p += oop(p)->size()) {
2331     start_array->allocate_block(p);
2332   }
2333 
2334   // Could update the promoted average here, but it is not typically updated at
2335   // full GCs and the value to use is unclear.  Something like
2336   //
2337   // cur_promoted_avg + absorb_size / number_of_scavenges_since_last_full_gc.
2338 
2339   size_policy->set_bytes_absorbed_from_eden(absorb_size);
2340   return true;
2341 }
2342 
2343 GCTaskManager* const PSParallelCompact::gc_task_manager() {
2344   assert(ParallelScavengeHeap::gc_task_manager() != NULL,
2345     "shouldn't return NULL");
2346   return ParallelScavengeHeap::gc_task_manager();
2347 }
2348 
2349 void PSParallelCompact::marking_phase(ParCompactionManager* cm,
2350                                       bool maximum_heap_compaction,
2351                                       ParallelOldTracer *gc_tracer) {
2352   // Recursively traverse all live objects and mark them
2353   GCTraceTime tm("marking phase", print_phases(), true, &_gc_timer);
2354 
2355   ParallelScavengeHeap* heap = gc_heap();
2356   uint parallel_gc_threads = heap->gc_task_manager()->workers();
2357   uint active_gc_threads = heap->gc_task_manager()->active_workers();
2358   TaskQueueSetSuper* qset = ParCompactionManager::region_array();
2359   ParallelTaskTerminator terminator(active_gc_threads, qset);
2360 
2361   PSParallelCompact::MarkAndPushClosure mark_and_push_closure(cm);
2362   PSParallelCompact::FollowStackClosure follow_stack_closure(cm);
2363 
2364   // Need new claim bits before marking starts.
2365   ClassLoaderDataGraph::clear_claimed_marks();
2366 
2367   {
2368     GCTraceTime tm_m("par mark", print_phases(), true, &_gc_timer);
2369 
2370     ParallelScavengeHeap::ParStrongRootsScope psrs;
2371 
2372     GCTaskQueue* q = GCTaskQueue::create();
2373 
2374     q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::universe));
2375     q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jni_handles));
2376     // We scan the thread roots in parallel
2377     Threads::create_thread_roots_marking_tasks(q);
2378     q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::object_synchronizer));
2379     q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::flat_profiler));
2380     q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::management));
2381     q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::system_dictionary));
2382     q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::class_loader_data));
2383     q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jvmti));
2384     q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::code_cache));
2385 
2386     if (active_gc_threads > 1) {
2387       for (uint j = 0; j < active_gc_threads; j++) {
2388         q->enqueue(new StealMarkingTask(&terminator));
2389       }
2390     }
2391 
2392     gc_task_manager()->execute_and_wait(q);
2393   }
2394 
2395   // Process reference objects found during marking
2396   {
2397     GCTraceTime tm_r("reference processing", print_phases(), true, &_gc_timer);
2398 
2399     ReferenceProcessorStats stats;
2400     if (ref_processor()->processing_is_mt()) {
2401       RefProcTaskExecutor task_executor;
2402       stats = ref_processor()->process_discovered_references(
2403         is_alive_closure(), &mark_and_push_closure, &follow_stack_closure,
2404         &task_executor, &_gc_timer);
2405     } else {
2406       stats = ref_processor()->process_discovered_references(
2407         is_alive_closure(), &mark_and_push_closure, &follow_stack_closure, NULL,
2408         &_gc_timer);
2409     }
2410 
2411     gc_tracer->report_gc_reference_stats(stats);
2412   }
2413 
2414   GCTraceTime tm_c("class unloading", print_phases(), true, &_gc_timer);
2415 
2416   // This is the point where the entire marking should have completed.
2417   assert(cm->marking_stacks_empty(), "Marking should have completed");
2418 
2419   // Follow system dictionary roots and unload classes.
2420   bool purged_class = SystemDictionary::do_unloading(is_alive_closure());
2421 
2422   // Unload nmethods.
2423   CodeCache::do_unloading(is_alive_closure(), purged_class);
2424 
2425   // Prune dead klasses from subklass/sibling/implementor lists.
2426   Klass::clean_weak_klass_links(is_alive_closure());
2427 
2428   // Delete entries for dead interned strings.
2429   StringTable::unlink(is_alive_closure());
2430 
2431   // Clean up unreferenced symbols in symbol table.
2432   SymbolTable::unlink();
2433   _gc_tracer.report_object_count_after_gc(is_alive_closure());
2434 }
2435 
2436 void PSParallelCompact::follow_klass(ParCompactionManager* cm, Klass* klass) {
2437   ClassLoaderData* cld = klass->class_loader_data();
2438   // The actual processing of the klass is done when we
2439   // traverse the list of Klasses in the class loader data.
2440   PSParallelCompact::follow_class_loader(cm, cld);
2441 }
2442 
2443 void PSParallelCompact::adjust_klass(ParCompactionManager* cm, Klass* klass) {
2444   ClassLoaderData* cld = klass->class_loader_data();
2445   // The actual processing of the klass is done when we
2446   // traverse the list of Klasses in the class loader data.
2447   PSParallelCompact::adjust_class_loader(cm, cld);
2448 }
2449 
2450 void PSParallelCompact::follow_class_loader(ParCompactionManager* cm,
2451                                             ClassLoaderData* cld) {
2452   PSParallelCompact::MarkAndPushClosure mark_and_push_closure(cm);
2453   PSParallelCompact::FollowKlassClosure follow_klass_closure(&mark_and_push_closure);
2454 
2455   cld->oops_do(&mark_and_push_closure, &follow_klass_closure, true);
2456 }
2457 
2458 void PSParallelCompact::adjust_class_loader(ParCompactionManager* cm,
2459                                             ClassLoaderData* cld) {
2460   cld->oops_do(PSParallelCompact::adjust_pointer_closure(),
2461                PSParallelCompact::adjust_klass_closure(),
2462                true);
2463 }
2464 
2465 // This should be moved to the shared markSweep code!
2466 class PSAlwaysTrueClosure: public BoolObjectClosure {
2467 public:
2468   bool do_object_b(oop p) { return true; }
2469 };
2470 static PSAlwaysTrueClosure always_true;
2471 
2472 void PSParallelCompact::adjust_roots() {
2473   // Adjust the pointers to reflect the new locations
2474   GCTraceTime tm("adjust roots", print_phases(), true, &_gc_timer);
2475 
2476   // Need new claim bits when tracing through and adjusting pointers.
2477   ClassLoaderDataGraph::clear_claimed_marks();
2478 
2479   // General strong roots.
2480   Universe::oops_do(adjust_pointer_closure());
2481   JNIHandles::oops_do(adjust_pointer_closure());   // Global (strong) JNI handles
2482   CLDToOopClosure adjust_from_cld(adjust_pointer_closure());
2483   Threads::oops_do(adjust_pointer_closure(), &adjust_from_cld, NULL);
2484   ObjectSynchronizer::oops_do(adjust_pointer_closure());
2485   FlatProfiler::oops_do(adjust_pointer_closure());
2486   Management::oops_do(adjust_pointer_closure());
2487   JvmtiExport::oops_do(adjust_pointer_closure());
2488   // SO_AllClasses
2489   SystemDictionary::oops_do(adjust_pointer_closure());
2490   ClassLoaderDataGraph::oops_do(adjust_pointer_closure(), adjust_klass_closure(), true);
2491 
2492   // Now adjust pointers in remaining weak roots.  (All of which should
2493   // have been cleared if they pointed to non-surviving objects.)
2494   // Global (weak) JNI handles
2495   JNIHandles::weak_oops_do(&always_true, adjust_pointer_closure());
2496 
2497   CodeCache::oops_do(adjust_pointer_closure());
2498   StringTable::oops_do(adjust_pointer_closure());
2499   ref_processor()->weak_oops_do(adjust_pointer_closure());
2500   // Roots were visited so references into the young gen in roots
2501   // may have been scanned.  Process them also.
2502   // Should the reference processor have a span that excludes
2503   // young gen objects?
2504   PSScavenge::reference_processor()->weak_oops_do(adjust_pointer_closure());
2505 }
2506 
2507 void PSParallelCompact::enqueue_region_draining_tasks(GCTaskQueue* q,
2508                                                       uint parallel_gc_threads)
2509 {
2510   GCTraceTime tm("drain task setup", print_phases(), true, &_gc_timer);
2511 
2512   // Find the threads that are active
2513   unsigned int which = 0;
2514 
2515   const uint task_count = MAX2(parallel_gc_threads, 1U);
2516   for (uint j = 0; j < task_count; j++) {
2517     q->enqueue(new DrainStacksCompactionTask(j));
2518     ParCompactionManager::verify_region_list_empty(j);
2519     // Set the region stacks variables to "no" region stack values
2520     // so that they will be recognized and needing a region stack
2521     // in the stealing tasks if they do not get one by executing
2522     // a draining stack.
2523     ParCompactionManager* cm = ParCompactionManager::manager_array(j);
2524     cm->set_region_stack(NULL);
2525     cm->set_region_stack_index((uint)max_uintx);
2526   }
2527   ParCompactionManager::reset_recycled_stack_index();
2528 
2529   // Find all regions that are available (can be filled immediately) and
2530   // distribute them to the thread stacks.  The iteration is done in reverse
2531   // order (high to low) so the regions will be removed in ascending order.
2532 
2533   const ParallelCompactData& sd = PSParallelCompact::summary_data();
2534 
2535   size_t fillable_regions = 0;   // A count for diagnostic purposes.
2536   // A region index which corresponds to the tasks created above.
2537   // "which" must be 0 <= which < task_count
2538 
2539   which = 0;
2540   // id + 1 is used to test termination so unsigned  can
2541   // be used with an old_space_id == 0.
2542   for (unsigned int id = to_space_id; id + 1 > old_space_id; --id) {
2543     SpaceInfo* const space_info = _space_info + id;
2544     MutableSpace* const space = space_info->space();
2545     HeapWord* const new_top = space_info->new_top();
2546 
2547     const size_t beg_region = sd.addr_to_region_idx(space_info->dense_prefix());
2548     const size_t end_region =
2549       sd.addr_to_region_idx(sd.region_align_up(new_top));
2550 
2551     for (size_t cur = end_region - 1; cur + 1 > beg_region; --cur) {
2552       if (sd.region(cur)->claim_unsafe()) {
2553         ParCompactionManager::region_list_push(which, cur);
2554 
2555         if (TraceParallelOldGCCompactionPhase && Verbose) {
2556           const size_t count_mod_8 = fillable_regions & 7;
2557           if (count_mod_8 == 0) gclog_or_tty->print("fillable: ");
2558           gclog_or_tty->print(" " SIZE_FORMAT_W(7), cur);
2559           if (count_mod_8 == 7) gclog_or_tty->cr();
2560         }
2561 
2562         NOT_PRODUCT(++fillable_regions;)
2563 
2564         // Assign regions to tasks in round-robin fashion.
2565         if (++which == task_count) {
2566           assert(which <= parallel_gc_threads,
2567             "Inconsistent number of workers");
2568           which = 0;
2569         }
2570       }
2571     }
2572   }
2573 
2574   if (TraceParallelOldGCCompactionPhase) {
2575     if (Verbose && (fillable_regions & 7) != 0) gclog_or_tty->cr();
2576     gclog_or_tty->print_cr("%u initially fillable regions", fillable_regions);
2577   }
2578 }
2579 
2580 #define PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING 4
2581 
2582 void PSParallelCompact::enqueue_dense_prefix_tasks(GCTaskQueue* q,
2583                                                     uint parallel_gc_threads) {
2584   GCTraceTime tm("dense prefix task setup", print_phases(), true, &_gc_timer);
2585 
2586   ParallelCompactData& sd = PSParallelCompact::summary_data();
2587 
2588   // Iterate over all the spaces adding tasks for updating
2589   // regions in the dense prefix.  Assume that 1 gc thread
2590   // will work on opening the gaps and the remaining gc threads
2591   // will work on the dense prefix.
2592   unsigned int space_id;
2593   for (space_id = old_space_id; space_id < last_space_id; ++ space_id) {
2594     HeapWord* const dense_prefix_end = _space_info[space_id].dense_prefix();
2595     const MutableSpace* const space = _space_info[space_id].space();
2596 
2597     if (dense_prefix_end == space->bottom()) {
2598       // There is no dense prefix for this space.
2599       continue;
2600     }
2601 
2602     // The dense prefix is before this region.
2603     size_t region_index_end_dense_prefix =
2604         sd.addr_to_region_idx(dense_prefix_end);
2605     RegionData* const dense_prefix_cp =
2606       sd.region(region_index_end_dense_prefix);
2607     assert(dense_prefix_end == space->end() ||
2608            dense_prefix_cp->available() ||
2609            dense_prefix_cp->claimed(),
2610            "The region after the dense prefix should always be ready to fill");
2611 
2612     size_t region_index_start = sd.addr_to_region_idx(space->bottom());
2613 
2614     // Is there dense prefix work?
2615     size_t total_dense_prefix_regions =
2616       region_index_end_dense_prefix - region_index_start;
2617     // How many regions of the dense prefix should be given to
2618     // each thread?
2619     if (total_dense_prefix_regions > 0) {
2620       uint tasks_for_dense_prefix = 1;
2621       if (total_dense_prefix_regions <=
2622           (parallel_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING)) {
2623         // Don't over partition.  This assumes that
2624         // PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING is a small integer value
2625         // so there are not many regions to process.
2626         tasks_for_dense_prefix = parallel_gc_threads;
2627       } else {
2628         // Over partition
2629         tasks_for_dense_prefix = parallel_gc_threads *
2630           PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING;
2631       }
2632       size_t regions_per_thread = total_dense_prefix_regions /
2633         tasks_for_dense_prefix;
2634       // Give each thread at least 1 region.
2635       if (regions_per_thread == 0) {
2636         regions_per_thread = 1;
2637       }
2638 
2639       for (uint k = 0; k < tasks_for_dense_prefix; k++) {
2640         if (region_index_start >= region_index_end_dense_prefix) {
2641           break;
2642         }
2643         // region_index_end is not processed
2644         size_t region_index_end = MIN2(region_index_start + regions_per_thread,
2645                                        region_index_end_dense_prefix);
2646         q->enqueue(new UpdateDensePrefixTask(SpaceId(space_id),
2647                                              region_index_start,
2648                                              region_index_end));
2649         region_index_start = region_index_end;
2650       }
2651     }
2652     // This gets any part of the dense prefix that did not
2653     // fit evenly.
2654     if (region_index_start < region_index_end_dense_prefix) {
2655       q->enqueue(new UpdateDensePrefixTask(SpaceId(space_id),
2656                                            region_index_start,
2657                                            region_index_end_dense_prefix));
2658     }
2659   }
2660 }
2661 
2662 void PSParallelCompact::enqueue_region_stealing_tasks(
2663                                      GCTaskQueue* q,
2664                                      ParallelTaskTerminator* terminator_ptr,
2665                                      uint parallel_gc_threads) {
2666   GCTraceTime tm("steal task setup", print_phases(), true, &_gc_timer);
2667 
2668   // Once a thread has drained it's stack, it should try to steal regions from
2669   // other threads.
2670   if (parallel_gc_threads > 1) {
2671     for (uint j = 0; j < parallel_gc_threads; j++) {
2672       q->enqueue(new StealRegionCompactionTask(terminator_ptr));
2673     }
2674   }
2675 }
2676 
2677 #ifdef ASSERT
2678 // Write a histogram of the number of times the block table was filled for a
2679 // region.
2680 void PSParallelCompact::write_block_fill_histogram(outputStream* const out)
2681 {
2682   if (!TraceParallelOldGCCompactionPhase) return;
2683 
2684   typedef ParallelCompactData::RegionData rd_t;
2685   ParallelCompactData& sd = summary_data();
2686 
2687   for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2688     MutableSpace* const spc = _space_info[id].space();
2689     if (spc->bottom() != spc->top()) {
2690       const rd_t* const beg = sd.addr_to_region_ptr(spc->bottom());
2691       HeapWord* const top_aligned_up = sd.region_align_up(spc->top());
2692       const rd_t* const end = sd.addr_to_region_ptr(top_aligned_up);
2693 
2694       size_t histo[5] = { 0, 0, 0, 0, 0 };
2695       const size_t histo_len = sizeof(histo) / sizeof(size_t);
2696       const size_t region_cnt = pointer_delta(end, beg, sizeof(rd_t));
2697 
2698       for (const rd_t* cur = beg; cur < end; ++cur) {
2699         ++histo[MIN2(cur->blocks_filled_count(), histo_len - 1)];
2700       }
2701       out->print("%u %-4s" SIZE_FORMAT_W(5), id, space_names[id], region_cnt);
2702       for (size_t i = 0; i < histo_len; ++i) {
2703         out->print(" " SIZE_FORMAT_W(5) " %5.1f%%",
2704                    histo[i], 100.0 * histo[i] / region_cnt);
2705       }
2706       out->cr();
2707     }
2708   }
2709 }
2710 #endif // #ifdef ASSERT
2711 
2712 void PSParallelCompact::compact() {
2713   // trace("5");
2714   GCTraceTime tm("compaction phase", print_phases(), true, &_gc_timer);
2715 
2716   ParallelScavengeHeap* heap = (ParallelScavengeHeap*)Universe::heap();
2717   assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity");
2718   PSOldGen* old_gen = heap->old_gen();
2719   old_gen->start_array()->reset();
2720   uint parallel_gc_threads = heap->gc_task_manager()->workers();
2721   uint active_gc_threads = heap->gc_task_manager()->active_workers();
2722   TaskQueueSetSuper* qset = ParCompactionManager::region_array();
2723   ParallelTaskTerminator terminator(active_gc_threads, qset);
2724 
2725   GCTaskQueue* q = GCTaskQueue::create();
2726   enqueue_region_draining_tasks(q, active_gc_threads);
2727   enqueue_dense_prefix_tasks(q, active_gc_threads);
2728   enqueue_region_stealing_tasks(q, &terminator, active_gc_threads);
2729 
2730   {
2731     GCTraceTime tm_pc("par compact", print_phases(), true, &_gc_timer);
2732 
2733     gc_task_manager()->execute_and_wait(q);
2734 
2735 #ifdef  ASSERT
2736     // Verify that all regions have been processed before the deferred updates.
2737     for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2738       verify_complete(SpaceId(id));
2739     }
2740 #endif
2741   }
2742 
2743   {
2744     // Update the deferred objects, if any.  Any compaction manager can be used.
2745     GCTraceTime tm_du("deferred updates", print_phases(), true, &_gc_timer);
2746     ParCompactionManager* cm = ParCompactionManager::manager_array(0);
2747     for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2748       update_deferred_objects(cm, SpaceId(id));
2749     }
2750   }
2751 
2752   DEBUG_ONLY(write_block_fill_histogram(gclog_or_tty));
2753 }
2754 
2755 #ifdef  ASSERT
2756 void PSParallelCompact::verify_complete(SpaceId space_id) {
2757   // All Regions between space bottom() to new_top() should be marked as filled
2758   // and all Regions between new_top() and top() should be available (i.e.,
2759   // should have been emptied).
2760   ParallelCompactData& sd = summary_data();
2761   SpaceInfo si = _space_info[space_id];
2762   HeapWord* new_top_addr = sd.region_align_up(si.new_top());
2763   HeapWord* old_top_addr = sd.region_align_up(si.space()->top());
2764   const size_t beg_region = sd.addr_to_region_idx(si.space()->bottom());
2765   const size_t new_top_region = sd.addr_to_region_idx(new_top_addr);
2766   const size_t old_top_region = sd.addr_to_region_idx(old_top_addr);
2767 
2768   bool issued_a_warning = false;
2769 
2770   size_t cur_region;
2771   for (cur_region = beg_region; cur_region < new_top_region; ++cur_region) {
2772     const RegionData* const c = sd.region(cur_region);
2773     if (!c->completed()) {
2774       warning("region " SIZE_FORMAT " not filled:  "
2775               "destination_count=" SIZE_FORMAT,
2776               cur_region, c->destination_count());
2777       issued_a_warning = true;
2778     }
2779   }
2780 
2781   for (cur_region = new_top_region; cur_region < old_top_region; ++cur_region) {
2782     const RegionData* const c = sd.region(cur_region);
2783     if (!c->available()) {
2784       warning("region " SIZE_FORMAT " not empty:   "
2785               "destination_count=" SIZE_FORMAT,
2786               cur_region, c->destination_count());
2787       issued_a_warning = true;
2788     }
2789   }
2790 
2791   if (issued_a_warning) {
2792     print_region_ranges();
2793   }
2794 }
2795 #endif  // #ifdef ASSERT
2796 
2797 // Update interior oops in the ranges of regions [beg_region, end_region).
2798 void
2799 PSParallelCompact::update_and_deadwood_in_dense_prefix(ParCompactionManager* cm,
2800                                                        SpaceId space_id,
2801                                                        size_t beg_region,
2802                                                        size_t end_region) {
2803   ParallelCompactData& sd = summary_data();
2804   ParMarkBitMap* const mbm = mark_bitmap();
2805 
2806   HeapWord* beg_addr = sd.region_to_addr(beg_region);
2807   HeapWord* const end_addr = sd.region_to_addr(end_region);
2808   assert(beg_region <= end_region, "bad region range");
2809   assert(end_addr <= dense_prefix(space_id), "not in the dense prefix");
2810 
2811 #ifdef  ASSERT
2812   // Claim the regions to avoid triggering an assert when they are marked as
2813   // filled.
2814   for (size_t claim_region = beg_region; claim_region < end_region; ++claim_region) {
2815     assert(sd.region(claim_region)->claim_unsafe(), "claim() failed");
2816   }
2817 #endif  // #ifdef ASSERT
2818 
2819   if (beg_addr != space(space_id)->bottom()) {
2820     // Find the first live object or block of dead space that *starts* in this
2821     // range of regions.  If a partial object crosses onto the region, skip it;
2822     // it will be marked for 'deferred update' when the object head is
2823     // processed.  If dead space crosses onto the region, it is also skipped; it
2824     // will be filled when the prior region is processed.  If neither of those
2825     // apply, the first word in the region is the start of a live object or dead
2826     // space.
2827     assert(beg_addr > space(space_id)->bottom(), "sanity");
2828     const RegionData* const cp = sd.region(beg_region);
2829     if (cp->partial_obj_size() != 0) {
2830       beg_addr = sd.partial_obj_end(beg_region);
2831     } else if (dead_space_crosses_boundary(cp, mbm->addr_to_bit(beg_addr))) {
2832       beg_addr = mbm->find_obj_beg(beg_addr, end_addr);
2833     }
2834   }
2835 
2836   if (beg_addr < end_addr) {
2837     // A live object or block of dead space starts in this range of Regions.
2838      HeapWord* const dense_prefix_end = dense_prefix(space_id);
2839 
2840     // Create closures and iterate.
2841     UpdateOnlyClosure update_closure(mbm, cm, space_id);
2842     FillClosure fill_closure(cm, space_id);
2843     ParMarkBitMap::IterationStatus status;
2844     status = mbm->iterate(&update_closure, &fill_closure, beg_addr, end_addr,
2845                           dense_prefix_end);
2846     if (status == ParMarkBitMap::incomplete) {
2847       update_closure.do_addr(update_closure.source());
2848     }
2849   }
2850 
2851   // Mark the regions as filled.
2852   RegionData* const beg_cp = sd.region(beg_region);
2853   RegionData* const end_cp = sd.region(end_region);
2854   for (RegionData* cp = beg_cp; cp < end_cp; ++cp) {
2855     cp->set_completed();
2856   }
2857 }
2858 
2859 // Return the SpaceId for the space containing addr.  If addr is not in the
2860 // heap, last_space_id is returned.  In debug mode it expects the address to be
2861 // in the heap and asserts such.
2862 PSParallelCompact::SpaceId PSParallelCompact::space_id(HeapWord* addr) {
2863   assert(Universe::heap()->is_in_reserved(addr), "addr not in the heap");
2864 
2865   for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2866     if (_space_info[id].space()->contains(addr)) {
2867       return SpaceId(id);
2868     }
2869   }
2870 
2871   assert(false, "no space contains the addr");
2872   return last_space_id;
2873 }
2874 
2875 void PSParallelCompact::update_deferred_objects(ParCompactionManager* cm,
2876                                                 SpaceId id) {
2877   assert(id < last_space_id, "bad space id");
2878 
2879   ParallelCompactData& sd = summary_data();
2880   const SpaceInfo* const space_info = _space_info + id;
2881   ObjectStartArray* const start_array = space_info->start_array();
2882 
2883   const MutableSpace* const space = space_info->space();
2884   assert(space_info->dense_prefix() >= space->bottom(), "dense_prefix not set");
2885   HeapWord* const beg_addr = space_info->dense_prefix();
2886   HeapWord* const end_addr = sd.region_align_up(space_info->new_top());
2887 
2888   const RegionData* const beg_region = sd.addr_to_region_ptr(beg_addr);
2889   const RegionData* const end_region = sd.addr_to_region_ptr(end_addr);
2890   const RegionData* cur_region;
2891   for (cur_region = beg_region; cur_region < end_region; ++cur_region) {
2892     HeapWord* const addr = cur_region->deferred_obj_addr();
2893     if (addr != NULL) {
2894       if (start_array != NULL) {
2895         start_array->allocate_block(addr);
2896       }
2897       oop(addr)->update_contents(cm);
2898       assert(oop(addr)->is_oop_or_null(), "should be an oop now");
2899     }
2900   }
2901 }
2902 
2903 // Skip over count live words starting from beg, and return the address of the
2904 // next live word.  Unless marked, the word corresponding to beg is assumed to
2905 // be dead.  Callers must either ensure beg does not correspond to the middle of
2906 // an object, or account for those live words in some other way.  Callers must
2907 // also ensure that there are enough live words in the range [beg, end) to skip.
2908 HeapWord*
2909 PSParallelCompact::skip_live_words(HeapWord* beg, HeapWord* end, size_t count)
2910 {
2911   assert(count > 0, "sanity");
2912 
2913   ParMarkBitMap* m = mark_bitmap();
2914   idx_t bits_to_skip = m->words_to_bits(count);
2915   idx_t cur_beg = m->addr_to_bit(beg);
2916   const idx_t search_end = BitMap::word_align_up(m->addr_to_bit(end));
2917 
2918   do {
2919     cur_beg = m->find_obj_beg(cur_beg, search_end);
2920     idx_t cur_end = m->find_obj_end(cur_beg, search_end);
2921     const size_t obj_bits = cur_end - cur_beg + 1;
2922     if (obj_bits > bits_to_skip) {
2923       return m->bit_to_addr(cur_beg + bits_to_skip);
2924     }
2925     bits_to_skip -= obj_bits;
2926     cur_beg = cur_end + 1;
2927   } while (bits_to_skip > 0);
2928 
2929   // Skipping the desired number of words landed just past the end of an object.
2930   // Find the start of the next object.
2931   cur_beg = m->find_obj_beg(cur_beg, search_end);
2932   assert(cur_beg < m->addr_to_bit(end), "not enough live words to skip");
2933   return m->bit_to_addr(cur_beg);
2934 }
2935 
2936 HeapWord* PSParallelCompact::first_src_addr(HeapWord* const dest_addr,
2937                                             SpaceId src_space_id,
2938                                             size_t src_region_idx)
2939 {
2940   assert(summary_data().is_region_aligned(dest_addr), "not aligned");
2941 
2942   const SplitInfo& split_info = _space_info[src_space_id].split_info();
2943   if (split_info.dest_region_addr() == dest_addr) {
2944     // The partial object ending at the split point contains the first word to
2945     // be copied to dest_addr.
2946     return split_info.first_src_addr();
2947   }
2948 
2949   const ParallelCompactData& sd = summary_data();
2950   ParMarkBitMap* const bitmap = mark_bitmap();
2951   const size_t RegionSize = ParallelCompactData::RegionSize;
2952 
2953   assert(sd.is_region_aligned(dest_addr), "not aligned");
2954   const RegionData* const src_region_ptr = sd.region(src_region_idx);
2955   const size_t partial_obj_size = src_region_ptr->partial_obj_size();
2956   HeapWord* const src_region_destination = src_region_ptr->destination();
2957 
2958   assert(dest_addr >= src_region_destination, "wrong src region");
2959   assert(src_region_ptr->data_size() > 0, "src region cannot be empty");
2960 
2961   HeapWord* const src_region_beg = sd.region_to_addr(src_region_idx);
2962   HeapWord* const src_region_end = src_region_beg + RegionSize;
2963 
2964   HeapWord* addr = src_region_beg;
2965   if (dest_addr == src_region_destination) {
2966     // Return the first live word in the source region.
2967     if (partial_obj_size == 0) {
2968       addr = bitmap->find_obj_beg(addr, src_region_end);
2969       assert(addr < src_region_end, "no objects start in src region");
2970     }
2971     return addr;
2972   }
2973 
2974   // Must skip some live data.
2975   size_t words_to_skip = dest_addr - src_region_destination;
2976   assert(src_region_ptr->data_size() > words_to_skip, "wrong src region");
2977 
2978   if (partial_obj_size >= words_to_skip) {
2979     // All the live words to skip are part of the partial object.
2980     addr += words_to_skip;
2981     if (partial_obj_size == words_to_skip) {
2982       // Find the first live word past the partial object.
2983       addr = bitmap->find_obj_beg(addr, src_region_end);
2984       assert(addr < src_region_end, "wrong src region");
2985     }
2986     return addr;
2987   }
2988 
2989   // Skip over the partial object (if any).
2990   if (partial_obj_size != 0) {
2991     words_to_skip -= partial_obj_size;
2992     addr += partial_obj_size;
2993   }
2994 
2995   // Skip over live words due to objects that start in the region.
2996   addr = skip_live_words(addr, src_region_end, words_to_skip);
2997   assert(addr < src_region_end, "wrong src region");
2998   return addr;
2999 }
3000 
3001 void PSParallelCompact::decrement_destination_counts(ParCompactionManager* cm,
3002                                                      SpaceId src_space_id,
3003                                                      size_t beg_region,
3004                                                      HeapWord* end_addr)
3005 {
3006   ParallelCompactData& sd = summary_data();
3007 
3008 #ifdef ASSERT
3009   MutableSpace* const src_space = _space_info[src_space_id].space();
3010   HeapWord* const beg_addr = sd.region_to_addr(beg_region);
3011   assert(src_space->contains(beg_addr) || beg_addr == src_space->end(),
3012          "src_space_id does not match beg_addr");
3013   assert(src_space->contains(end_addr) || end_addr == src_space->end(),
3014          "src_space_id does not match end_addr");
3015 #endif // #ifdef ASSERT
3016 
3017   RegionData* const beg = sd.region(beg_region);
3018   RegionData* const end = sd.addr_to_region_ptr(sd.region_align_up(end_addr));
3019 
3020   // Regions up to new_top() are enqueued if they become available.
3021   HeapWord* const new_top = _space_info[src_space_id].new_top();
3022   RegionData* const enqueue_end =
3023     sd.addr_to_region_ptr(sd.region_align_up(new_top));
3024 
3025   for (RegionData* cur = beg; cur < end; ++cur) {
3026     assert(cur->data_size() > 0, "region must have live data");
3027     cur->decrement_destination_count();
3028     if (cur < enqueue_end && cur->available() && cur->claim()) {
3029       cm->push_region(sd.region(cur));
3030     }
3031   }
3032 }
3033 
3034 size_t PSParallelCompact::next_src_region(MoveAndUpdateClosure& closure,
3035                                           SpaceId& src_space_id,
3036                                           HeapWord*& src_space_top,
3037                                           HeapWord* end_addr)
3038 {
3039   typedef ParallelCompactData::RegionData RegionData;
3040 
3041   ParallelCompactData& sd = PSParallelCompact::summary_data();
3042   const size_t region_size = ParallelCompactData::RegionSize;
3043 
3044   size_t src_region_idx = 0;
3045 
3046   // Skip empty regions (if any) up to the top of the space.
3047   HeapWord* const src_aligned_up = sd.region_align_up(end_addr);
3048   RegionData* src_region_ptr = sd.addr_to_region_ptr(src_aligned_up);
3049   HeapWord* const top_aligned_up = sd.region_align_up(src_space_top);
3050   const RegionData* const top_region_ptr =
3051     sd.addr_to_region_ptr(top_aligned_up);
3052   while (src_region_ptr < top_region_ptr && src_region_ptr->data_size() == 0) {
3053     ++src_region_ptr;
3054   }
3055 
3056   if (src_region_ptr < top_region_ptr) {
3057     // The next source region is in the current space.  Update src_region_idx
3058     // and the source address to match src_region_ptr.
3059     src_region_idx = sd.region(src_region_ptr);
3060     HeapWord* const src_region_addr = sd.region_to_addr(src_region_idx);
3061     if (src_region_addr > closure.source()) {
3062       closure.set_source(src_region_addr);
3063     }
3064     return src_region_idx;
3065   }
3066 
3067   // Switch to a new source space and find the first non-empty region.
3068   unsigned int space_id = src_space_id + 1;
3069   assert(space_id < last_space_id, "not enough spaces");
3070 
3071   HeapWord* const destination = closure.destination();
3072 
3073   do {
3074     MutableSpace* space = _space_info[space_id].space();
3075     HeapWord* const bottom = space->bottom();
3076     const RegionData* const bottom_cp = sd.addr_to_region_ptr(bottom);
3077 
3078     // Iterate over the spaces that do not compact into themselves.
3079     if (bottom_cp->destination() != bottom) {
3080       HeapWord* const top_aligned_up = sd.region_align_up(space->top());
3081       const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up);
3082 
3083       for (const RegionData* src_cp = bottom_cp; src_cp < top_cp; ++src_cp) {
3084         if (src_cp->live_obj_size() > 0) {
3085           // Found it.
3086           assert(src_cp->destination() == destination,
3087                  "first live obj in the space must match the destination");
3088           assert(src_cp->partial_obj_size() == 0,
3089                  "a space cannot begin with a partial obj");
3090 
3091           src_space_id = SpaceId(space_id);
3092           src_space_top = space->top();
3093           const size_t src_region_idx = sd.region(src_cp);
3094           closure.set_source(sd.region_to_addr(src_region_idx));
3095           return src_region_idx;
3096         } else {
3097           assert(src_cp->data_size() == 0, "sanity");
3098         }
3099       }
3100     }
3101   } while (++space_id < last_space_id);
3102 
3103   assert(false, "no source region was found");
3104   return 0;
3105 }
3106 
3107 void PSParallelCompact::fill_region(ParCompactionManager* cm, size_t region_idx)
3108 {
3109   typedef ParMarkBitMap::IterationStatus IterationStatus;
3110   const size_t RegionSize = ParallelCompactData::RegionSize;
3111   ParMarkBitMap* const bitmap = mark_bitmap();
3112   ParallelCompactData& sd = summary_data();
3113   RegionData* const region_ptr = sd.region(region_idx);
3114 
3115   // Get the items needed to construct the closure.
3116   HeapWord* dest_addr = sd.region_to_addr(region_idx);
3117   SpaceId dest_space_id = space_id(dest_addr);
3118   ObjectStartArray* start_array = _space_info[dest_space_id].start_array();
3119   HeapWord* new_top = _space_info[dest_space_id].new_top();
3120   assert(dest_addr < new_top, "sanity");
3121   const size_t words = MIN2(pointer_delta(new_top, dest_addr), RegionSize);
3122 
3123   // Get the source region and related info.
3124   size_t src_region_idx = region_ptr->source_region();
3125   SpaceId src_space_id = space_id(sd.region_to_addr(src_region_idx));
3126   HeapWord* src_space_top = _space_info[src_space_id].space()->top();
3127 
3128   MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words);
3129   closure.set_source(first_src_addr(dest_addr, src_space_id, src_region_idx));
3130 
3131   // Adjust src_region_idx to prepare for decrementing destination counts (the
3132   // destination count is not decremented when a region is copied to itself).
3133   if (src_region_idx == region_idx) {
3134     src_region_idx += 1;
3135   }
3136 
3137   if (bitmap->is_unmarked(closure.source())) {
3138     // The first source word is in the middle of an object; copy the remainder
3139     // of the object or as much as will fit.  The fact that pointer updates were
3140     // deferred will be noted when the object header is processed.
3141     HeapWord* const old_src_addr = closure.source();
3142     closure.copy_partial_obj();
3143     if (closure.is_full()) {
3144       decrement_destination_counts(cm, src_space_id, src_region_idx,
3145                                    closure.source());
3146       region_ptr->set_deferred_obj_addr(NULL);
3147       region_ptr->set_completed();
3148       return;
3149     }
3150 
3151     HeapWord* const end_addr = sd.region_align_down(closure.source());
3152     if (sd.region_align_down(old_src_addr) != end_addr) {
3153       // The partial object was copied from more than one source region.
3154       decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr);
3155 
3156       // Move to the next source region, possibly switching spaces as well.  All
3157       // args except end_addr may be modified.
3158       src_region_idx = next_src_region(closure, src_space_id, src_space_top,
3159                                        end_addr);
3160     }
3161   }
3162 
3163   do {
3164     HeapWord* const cur_addr = closure.source();
3165     HeapWord* const end_addr = MIN2(sd.region_align_up(cur_addr + 1),
3166                                     src_space_top);
3167     IterationStatus status = bitmap->iterate(&closure, cur_addr, end_addr);
3168 
3169     if (status == ParMarkBitMap::incomplete) {
3170       // The last obj that starts in the source region does not end in the
3171       // region.
3172       assert(closure.source() < end_addr, "sanity");
3173       HeapWord* const obj_beg = closure.source();
3174       HeapWord* const range_end = MIN2(obj_beg + closure.words_remaining(),
3175                                        src_space_top);
3176       HeapWord* const obj_end = bitmap->find_obj_end(obj_beg, range_end);
3177       if (obj_end < range_end) {
3178         // The end was found; the entire object will fit.
3179         status = closure.do_addr(obj_beg, bitmap->obj_size(obj_beg, obj_end));
3180         assert(status != ParMarkBitMap::would_overflow, "sanity");
3181       } else {
3182         // The end was not found; the object will not fit.
3183         assert(range_end < src_space_top, "obj cannot cross space boundary");
3184         status = ParMarkBitMap::would_overflow;
3185       }
3186     }
3187 
3188     if (status == ParMarkBitMap::would_overflow) {
3189       // The last object did not fit.  Note that interior oop updates were
3190       // deferred, then copy enough of the object to fill the region.
3191       region_ptr->set_deferred_obj_addr(closure.destination());
3192       status = closure.copy_until_full(); // copies from closure.source()
3193 
3194       decrement_destination_counts(cm, src_space_id, src_region_idx,
3195                                    closure.source());
3196       region_ptr->set_completed();
3197       return;
3198     }
3199 
3200     if (status == ParMarkBitMap::full) {
3201       decrement_destination_counts(cm, src_space_id, src_region_idx,
3202                                    closure.source());
3203       region_ptr->set_deferred_obj_addr(NULL);
3204       region_ptr->set_completed();
3205       return;
3206     }
3207 
3208     decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr);
3209 
3210     // Move to the next source region, possibly switching spaces as well.  All
3211     // args except end_addr may be modified.
3212     src_region_idx = next_src_region(closure, src_space_id, src_space_top,
3213                                      end_addr);
3214   } while (true);
3215 }
3216 
3217 void PSParallelCompact::fill_blocks(size_t region_idx)
3218 {
3219   // Fill in the block table elements for the specified region.  Each block
3220   // table element holds the number of live words in the region that are to the
3221   // left of the first object that starts in the block.  Thus only blocks in
3222   // which an object starts need to be filled.
3223   //
3224   // The algorithm scans the section of the bitmap that corresponds to the
3225   // region, keeping a running total of the live words.  When an object start is
3226   // found, if it's the first to start in the block that contains it, the
3227   // current total is written to the block table element.
3228   const size_t Log2BlockSize = ParallelCompactData::Log2BlockSize;
3229   const size_t Log2RegionSize = ParallelCompactData::Log2RegionSize;
3230   const size_t RegionSize = ParallelCompactData::RegionSize;
3231 
3232   ParallelCompactData& sd = summary_data();
3233   const size_t partial_obj_size = sd.region(region_idx)->partial_obj_size();
3234   if (partial_obj_size >= RegionSize) {
3235     return; // No objects start in this region.
3236   }
3237 
3238   // Ensure the first loop iteration decides that the block has changed.
3239   size_t cur_block = sd.block_count();
3240 
3241   const ParMarkBitMap* const bitmap = mark_bitmap();
3242 
3243   const size_t Log2BitsPerBlock = Log2BlockSize - LogMinObjAlignment;
3244   assert((size_t)1 << Log2BitsPerBlock ==
3245          bitmap->words_to_bits(ParallelCompactData::BlockSize), "sanity");
3246 
3247   size_t beg_bit = bitmap->words_to_bits(region_idx << Log2RegionSize);
3248   const size_t range_end = beg_bit + bitmap->words_to_bits(RegionSize);
3249   size_t live_bits = bitmap->words_to_bits(partial_obj_size);
3250   beg_bit = bitmap->find_obj_beg(beg_bit + live_bits, range_end);
3251   while (beg_bit < range_end) {
3252     const size_t new_block = beg_bit >> Log2BitsPerBlock;
3253     if (new_block != cur_block) {
3254       cur_block = new_block;
3255       sd.block(cur_block)->set_offset(bitmap->bits_to_words(live_bits));
3256     }
3257 
3258     const size_t end_bit = bitmap->find_obj_end(beg_bit, range_end);
3259     if (end_bit < range_end - 1) {
3260       live_bits += end_bit - beg_bit + 1;
3261       beg_bit = bitmap->find_obj_beg(end_bit + 1, range_end);
3262     } else {
3263       return;
3264     }
3265   }
3266 }
3267 
3268 void
3269 PSParallelCompact::move_and_update(ParCompactionManager* cm, SpaceId space_id) {
3270   const MutableSpace* sp = space(space_id);
3271   if (sp->is_empty()) {
3272     return;
3273   }
3274 
3275   ParallelCompactData& sd = PSParallelCompact::summary_data();
3276   ParMarkBitMap* const bitmap = mark_bitmap();
3277   HeapWord* const dp_addr = dense_prefix(space_id);
3278   HeapWord* beg_addr = sp->bottom();
3279   HeapWord* end_addr = sp->top();
3280 
3281   assert(beg_addr <= dp_addr && dp_addr <= end_addr, "bad dense prefix");
3282 
3283   const size_t beg_region = sd.addr_to_region_idx(beg_addr);
3284   const size_t dp_region = sd.addr_to_region_idx(dp_addr);
3285   if (beg_region < dp_region) {
3286     update_and_deadwood_in_dense_prefix(cm, space_id, beg_region, dp_region);
3287   }
3288 
3289   // The destination of the first live object that starts in the region is one
3290   // past the end of the partial object entering the region (if any).
3291   HeapWord* const dest_addr = sd.partial_obj_end(dp_region);
3292   HeapWord* const new_top = _space_info[space_id].new_top();
3293   assert(new_top >= dest_addr, "bad new_top value");
3294   const size_t words = pointer_delta(new_top, dest_addr);
3295 
3296   if (words > 0) {
3297     ObjectStartArray* start_array = _space_info[space_id].start_array();
3298     MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words);
3299 
3300     ParMarkBitMap::IterationStatus status;
3301     status = bitmap->iterate(&closure, dest_addr, end_addr);
3302     assert(status == ParMarkBitMap::full, "iteration not complete");
3303     assert(bitmap->find_obj_beg(closure.source(), end_addr) == end_addr,
3304            "live objects skipped because closure is full");
3305   }
3306 }
3307 
3308 jlong PSParallelCompact::millis_since_last_gc() {
3309   // We need a monotonically non-deccreasing time in ms but
3310   // os::javaTimeMillis() does not guarantee monotonicity.
3311   jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
3312   jlong ret_val = now - _time_of_last_gc;
3313   // XXX See note in genCollectedHeap::millis_since_last_gc().
3314   if (ret_val < 0) {
3315     NOT_PRODUCT(warning("time warp: "INT64_FORMAT, ret_val);)
3316     return 0;
3317   }
3318   return ret_val;
3319 }
3320 
3321 void PSParallelCompact::reset_millis_since_last_gc() {
3322   // We need a monotonically non-deccreasing time in ms but
3323   // os::javaTimeMillis() does not guarantee monotonicity.
3324   _time_of_last_gc = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
3325 }
3326 
3327 ParMarkBitMap::IterationStatus MoveAndUpdateClosure::copy_until_full()
3328 {
3329   if (source() != destination()) {
3330     DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3331     Copy::aligned_conjoint_words(source(), destination(), words_remaining());
3332   }
3333   update_state(words_remaining());
3334   assert(is_full(), "sanity");
3335   return ParMarkBitMap::full;
3336 }
3337 
3338 void MoveAndUpdateClosure::copy_partial_obj()
3339 {
3340   size_t words = words_remaining();
3341 
3342   HeapWord* const range_end = MIN2(source() + words, bitmap()->region_end());
3343   HeapWord* const end_addr = bitmap()->find_obj_end(source(), range_end);
3344   if (end_addr < range_end) {
3345     words = bitmap()->obj_size(source(), end_addr);
3346   }
3347 
3348   // This test is necessary; if omitted, the pointer updates to a partial object
3349   // that crosses the dense prefix boundary could be overwritten.
3350   if (source() != destination()) {
3351     DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3352     Copy::aligned_conjoint_words(source(), destination(), words);
3353   }
3354   update_state(words);
3355 }
3356 
3357 ParMarkBitMapClosure::IterationStatus
3358 MoveAndUpdateClosure::do_addr(HeapWord* addr, size_t words) {
3359   assert(destination() != NULL, "sanity");
3360   assert(bitmap()->obj_size(addr) == words, "bad size");
3361 
3362   _source = addr;
3363   assert(PSParallelCompact::summary_data().calc_new_pointer(source()) ==
3364          destination(), "wrong destination");
3365 
3366   if (words > words_remaining()) {
3367     return ParMarkBitMap::would_overflow;
3368   }
3369 
3370   // The start_array must be updated even if the object is not moving.
3371   if (_start_array != NULL) {
3372     _start_array->allocate_block(destination());
3373   }
3374 
3375   if (destination() != source()) {
3376     DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3377     Copy::aligned_conjoint_words(source(), destination(), words);
3378   }
3379 
3380   oop moved_oop = (oop) destination();
3381   moved_oop->update_contents(compaction_manager());
3382   assert(moved_oop->is_oop_or_null(), "Object should be whole at this point");
3383 
3384   update_state(words);
3385   assert(destination() == (HeapWord*)moved_oop + moved_oop->size(), "sanity");
3386   return is_full() ? ParMarkBitMap::full : ParMarkBitMap::incomplete;
3387 }
3388 
3389 UpdateOnlyClosure::UpdateOnlyClosure(ParMarkBitMap* mbm,
3390                                      ParCompactionManager* cm,
3391                                      PSParallelCompact::SpaceId space_id) :
3392   ParMarkBitMapClosure(mbm, cm),
3393   _space_id(space_id),
3394   _start_array(PSParallelCompact::start_array(space_id))
3395 {
3396 }
3397 
3398 // Updates the references in the object to their new values.
3399 ParMarkBitMapClosure::IterationStatus
3400 UpdateOnlyClosure::do_addr(HeapWord* addr, size_t words) {
3401   do_addr(addr);
3402   return ParMarkBitMap::incomplete;
3403 }