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_implementation/shared/spaceDecorator.hpp"
  47 #include "gc_interface/gcCause.hpp"
  48 #include "memory/gcLocker.inline.hpp"
  49 #include "memory/referencePolicy.hpp"
  50 #include "memory/referenceProcessor.hpp"
  51 #include "oops/methodData.hpp"
  52 #include "oops/oop.inline.hpp"
  53 #include "oops/oop.pcgc.inline.hpp"
  54 #include "runtime/fprofiler.hpp"
  55 #include "runtime/safepoint.hpp"
  56 #include "runtime/vmThread.hpp"
  57 #include "services/management.hpp"
  58 #include "services/memoryService.hpp"
  59 #include "services/memTracker.hpp"
  60 #include "utilities/events.hpp"
  61 #include "utilities/stack.inline.hpp"
  62 
  63 #include <math.h>
  64 
  65 // All sizes are in HeapWords.
  66 const size_t ParallelCompactData::Log2RegionSize  = 16; // 64K words
  67 const size_t ParallelCompactData::RegionSize      = (size_t)1 << Log2RegionSize;
  68 const size_t ParallelCompactData::RegionSizeBytes =
  69   RegionSize << LogHeapWordSize;
  70 const size_t ParallelCompactData::RegionSizeOffsetMask = RegionSize - 1;
  71 const size_t ParallelCompactData::RegionAddrOffsetMask = RegionSizeBytes - 1;
  72 const size_t ParallelCompactData::RegionAddrMask       = ~RegionAddrOffsetMask;
  73 
  74 const size_t ParallelCompactData::Log2BlockSize   = 7; // 128 words
  75 const size_t ParallelCompactData::BlockSize       = (size_t)1 << Log2BlockSize;
  76 const size_t ParallelCompactData::BlockSizeBytes  =
  77   BlockSize << LogHeapWordSize;
  78 const size_t ParallelCompactData::BlockSizeOffsetMask = BlockSize - 1;
  79 const size_t ParallelCompactData::BlockAddrOffsetMask = BlockSizeBytes - 1;
  80 const size_t ParallelCompactData::BlockAddrMask       = ~BlockAddrOffsetMask;
  81 
  82 const size_t ParallelCompactData::BlocksPerRegion = RegionSize / BlockSize;
  83 const size_t ParallelCompactData::Log2BlocksPerRegion =
  84   Log2RegionSize - Log2BlockSize;
  85 
  86 const ParallelCompactData::RegionData::region_sz_t
  87 ParallelCompactData::RegionData::dc_shift = 27;
  88 
  89 const ParallelCompactData::RegionData::region_sz_t
  90 ParallelCompactData::RegionData::dc_mask = ~0U << dc_shift;
  91 
  92 const ParallelCompactData::RegionData::region_sz_t
  93 ParallelCompactData::RegionData::dc_one = 0x1U << dc_shift;
  94 
  95 const ParallelCompactData::RegionData::region_sz_t
  96 ParallelCompactData::RegionData::los_mask = ~dc_mask;
  97 
  98 const ParallelCompactData::RegionData::region_sz_t
  99 ParallelCompactData::RegionData::dc_claimed = 0x8U << dc_shift;
 100 
 101 const ParallelCompactData::RegionData::region_sz_t
 102 ParallelCompactData::RegionData::dc_completed = 0xcU << dc_shift;
 103 
 104 SpaceInfo PSParallelCompact::_space_info[PSParallelCompact::last_space_id];
 105 bool      PSParallelCompact::_print_phases = false;
 106 
 107 ReferenceProcessor* PSParallelCompact::_ref_processor = NULL;
 108 Klass*              PSParallelCompact::_updated_int_array_klass_obj = NULL;
 109 
 110 double PSParallelCompact::_dwl_mean;
 111 double PSParallelCompact::_dwl_std_dev;
 112 double PSParallelCompact::_dwl_first_term;
 113 double PSParallelCompact::_dwl_adjustment;
 114 #ifdef  ASSERT
 115 bool   PSParallelCompact::_dwl_initialized = false;
 116 #endif  // #ifdef ASSERT
 117 
 118 void SplitInfo::record(size_t src_region_idx, size_t partial_obj_size,
 119                        HeapWord* destination)
 120 {
 121   assert(src_region_idx != 0, "invalid src_region_idx");
 122   assert(partial_obj_size != 0, "invalid partial_obj_size argument");
 123   assert(destination != NULL, "invalid destination argument");
 124 
 125   _src_region_idx = src_region_idx;
 126   _partial_obj_size = partial_obj_size;
 127   _destination = destination;
 128 
 129   // These fields may not be updated below, so make sure they're clear.
 130   assert(_dest_region_addr == NULL, "should have been cleared");
 131   assert(_first_src_addr == NULL, "should have been cleared");
 132 
 133   // Determine the number of destination regions for the partial object.
 134   HeapWord* const last_word = destination + partial_obj_size - 1;
 135   const ParallelCompactData& sd = PSParallelCompact::summary_data();
 136   HeapWord* const beg_region_addr = sd.region_align_down(destination);
 137   HeapWord* const end_region_addr = sd.region_align_down(last_word);
 138 
 139   if (beg_region_addr == end_region_addr) {
 140     // One destination region.
 141     _destination_count = 1;
 142     if (end_region_addr == destination) {
 143       // The destination falls on a region boundary, thus the first word of the
 144       // partial object will be the first word copied to the destination region.
 145       _dest_region_addr = end_region_addr;
 146       _first_src_addr = sd.region_to_addr(src_region_idx);
 147     }
 148   } else {
 149     // Two destination regions.  When copied, the partial object will cross a
 150     // destination region boundary, so a word somewhere within the partial
 151     // object will be the first word copied to the second destination region.
 152     _destination_count = 2;
 153     _dest_region_addr = end_region_addr;
 154     const size_t ofs = pointer_delta(end_region_addr, destination);
 155     assert(ofs < _partial_obj_size, "sanity");
 156     _first_src_addr = sd.region_to_addr(src_region_idx) + ofs;
 157   }
 158 }
 159 
 160 void SplitInfo::clear()
 161 {
 162   _src_region_idx = 0;
 163   _partial_obj_size = 0;
 164   _destination = NULL;
 165   _destination_count = 0;
 166   _dest_region_addr = NULL;
 167   _first_src_addr = NULL;
 168   assert(!is_valid(), "sanity");
 169 }
 170 
 171 #ifdef  ASSERT
 172 void SplitInfo::verify_clear()
 173 {
 174   assert(_src_region_idx == 0, "not clear");
 175   assert(_partial_obj_size == 0, "not clear");
 176   assert(_destination == NULL, "not clear");
 177   assert(_destination_count == 0, "not clear");
 178   assert(_dest_region_addr == NULL, "not clear");
 179   assert(_first_src_addr == NULL, "not clear");
 180 }
 181 #endif  // #ifdef ASSERT
 182 
 183 
 184 void PSParallelCompact::print_on_error(outputStream* st) {
 185   _mark_bitmap.print_on_error(st);
 186 }
 187 
 188 #ifndef PRODUCT
 189 const char* PSParallelCompact::space_names[] = {
 190   "old ", "eden", "from", "to  "
 191 };
 192 
 193 void PSParallelCompact::print_region_ranges()
 194 {
 195   tty->print_cr("space  bottom     top        end        new_top");
 196   tty->print_cr("------ ---------- ---------- ---------- ----------");
 197 
 198   for (unsigned int id = 0; id < last_space_id; ++id) {
 199     const MutableSpace* space = _space_info[id].space();
 200     tty->print_cr("%u %s "
 201                   SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " "
 202                   SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " ",
 203                   id, space_names[id],
 204                   summary_data().addr_to_region_idx(space->bottom()),
 205                   summary_data().addr_to_region_idx(space->top()),
 206                   summary_data().addr_to_region_idx(space->end()),
 207                   summary_data().addr_to_region_idx(_space_info[id].new_top()));
 208   }
 209 }
 210 
 211 void
 212 print_generic_summary_region(size_t i, const ParallelCompactData::RegionData* c)
 213 {
 214 #define REGION_IDX_FORMAT        SIZE_FORMAT_W(7)
 215 #define REGION_DATA_FORMAT       SIZE_FORMAT_W(5)
 216 
 217   ParallelCompactData& sd = PSParallelCompact::summary_data();
 218   size_t dci = c->destination() ? sd.addr_to_region_idx(c->destination()) : 0;
 219   tty->print_cr(REGION_IDX_FORMAT " " PTR_FORMAT " "
 220                 REGION_IDX_FORMAT " " PTR_FORMAT " "
 221                 REGION_DATA_FORMAT " " REGION_DATA_FORMAT " "
 222                 REGION_DATA_FORMAT " " REGION_IDX_FORMAT " %d",
 223                 i, c->data_location(), dci, c->destination(),
 224                 c->partial_obj_size(), c->live_obj_size(),
 225                 c->data_size(), c->source_region(), c->destination_count());
 226 
 227 #undef  REGION_IDX_FORMAT
 228 #undef  REGION_DATA_FORMAT
 229 }
 230 
 231 void
 232 print_generic_summary_data(ParallelCompactData& summary_data,
 233                            HeapWord* const beg_addr,
 234                            HeapWord* const end_addr)
 235 {
 236   size_t total_words = 0;
 237   size_t i = summary_data.addr_to_region_idx(beg_addr);
 238   const size_t last = summary_data.addr_to_region_idx(end_addr);
 239   HeapWord* pdest = 0;
 240 
 241   while (i <= last) {
 242     ParallelCompactData::RegionData* c = summary_data.region(i);
 243     if (c->data_size() != 0 || c->destination() != pdest) {
 244       print_generic_summary_region(i, c);
 245       total_words += c->data_size();
 246       pdest = c->destination();
 247     }
 248     ++i;
 249   }
 250 
 251   tty->print_cr("summary_data_bytes=" SIZE_FORMAT, total_words * HeapWordSize);
 252 }
 253 
 254 void
 255 print_generic_summary_data(ParallelCompactData& summary_data,
 256                            SpaceInfo* space_info)
 257 {
 258   for (unsigned int id = 0; id < PSParallelCompact::last_space_id; ++id) {
 259     const MutableSpace* space = space_info[id].space();
 260     print_generic_summary_data(summary_data, space->bottom(),
 261                                MAX2(space->top(), space_info[id].new_top()));
 262   }
 263 }
 264 
 265 void
 266 print_initial_summary_region(size_t i,
 267                              const ParallelCompactData::RegionData* c,
 268                              bool newline = true)
 269 {
 270   tty->print(SIZE_FORMAT_W(5) " " PTR_FORMAT " "
 271              SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " "
 272              SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d",
 273              i, c->destination(),
 274              c->partial_obj_size(), c->live_obj_size(),
 275              c->data_size(), c->source_region(), c->destination_count());
 276   if (newline) tty->cr();
 277 }
 278 
 279 void
 280 print_initial_summary_data(ParallelCompactData& summary_data,
 281                            const MutableSpace* space) {
 282   if (space->top() == space->bottom()) {
 283     return;
 284   }
 285 
 286   const size_t region_size = ParallelCompactData::RegionSize;
 287   typedef ParallelCompactData::RegionData RegionData;
 288   HeapWord* const top_aligned_up = summary_data.region_align_up(space->top());
 289   const size_t end_region = summary_data.addr_to_region_idx(top_aligned_up);
 290   const RegionData* c = summary_data.region(end_region - 1);
 291   HeapWord* end_addr = c->destination() + c->data_size();
 292   const size_t live_in_space = pointer_delta(end_addr, space->bottom());
 293 
 294   // Print (and count) the full regions at the beginning of the space.
 295   size_t full_region_count = 0;
 296   size_t i = summary_data.addr_to_region_idx(space->bottom());
 297   while (i < end_region && summary_data.region(i)->data_size() == region_size) {
 298     print_initial_summary_region(i, summary_data.region(i));
 299     ++full_region_count;
 300     ++i;
 301   }
 302 
 303   size_t live_to_right = live_in_space - full_region_count * region_size;
 304 
 305   double max_reclaimed_ratio = 0.0;
 306   size_t max_reclaimed_ratio_region = 0;
 307   size_t max_dead_to_right = 0;
 308   size_t max_live_to_right = 0;
 309 
 310   // Print the 'reclaimed ratio' for regions while there is something live in
 311   // the region or to the right of it.  The remaining regions are empty (and
 312   // uninteresting), and computing the ratio will result in division by 0.
 313   while (i < end_region && live_to_right > 0) {
 314     c = summary_data.region(i);
 315     HeapWord* const region_addr = summary_data.region_to_addr(i);
 316     const size_t used_to_right = pointer_delta(space->top(), region_addr);
 317     const size_t dead_to_right = used_to_right - live_to_right;
 318     const double reclaimed_ratio = double(dead_to_right) / live_to_right;
 319 
 320     if (reclaimed_ratio > max_reclaimed_ratio) {
 321             max_reclaimed_ratio = reclaimed_ratio;
 322             max_reclaimed_ratio_region = i;
 323             max_dead_to_right = dead_to_right;
 324             max_live_to_right = live_to_right;
 325     }
 326 
 327     print_initial_summary_region(i, c, false);
 328     tty->print_cr(" %12.10f " SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10),
 329                   reclaimed_ratio, dead_to_right, live_to_right);
 330 
 331     live_to_right -= c->data_size();
 332     ++i;
 333   }
 334 
 335   // Any remaining regions are empty.  Print one more if there is one.
 336   if (i < end_region) {
 337     print_initial_summary_region(i, summary_data.region(i));
 338   }
 339 
 340   tty->print_cr("max:  " SIZE_FORMAT_W(4) " d2r=" SIZE_FORMAT_W(10) " "
 341                 "l2r=" SIZE_FORMAT_W(10) " max_ratio=%14.12f",
 342                 max_reclaimed_ratio_region, max_dead_to_right,
 343                 max_live_to_right, max_reclaimed_ratio);
 344 }
 345 
 346 void
 347 print_initial_summary_data(ParallelCompactData& summary_data,
 348                            SpaceInfo* space_info) {
 349   unsigned int id = PSParallelCompact::old_space_id;
 350   const MutableSpace* space;
 351   do {
 352     space = space_info[id].space();
 353     print_initial_summary_data(summary_data, space);
 354   } while (++id < PSParallelCompact::eden_space_id);
 355 
 356   do {
 357     space = space_info[id].space();
 358     print_generic_summary_data(summary_data, space->bottom(), space->top());
 359   } while (++id < PSParallelCompact::last_space_id);
 360 }
 361 #endif  // #ifndef PRODUCT
 362 
 363 #ifdef  ASSERT
 364 size_t add_obj_count;
 365 size_t add_obj_size;
 366 size_t mark_bitmap_count;
 367 size_t mark_bitmap_size;
 368 #endif  // #ifdef ASSERT
 369 
 370 ParallelCompactData::ParallelCompactData()
 371 {
 372   _region_start = 0;
 373 
 374   _region_vspace = 0;
 375   _reserved_byte_size = 0;
 376   _region_data = 0;
 377   _region_count = 0;
 378 
 379   _block_vspace = 0;
 380   _block_data = 0;
 381   _block_count = 0;
 382 }
 383 
 384 bool ParallelCompactData::initialize(MemRegion covered_region)
 385 {
 386   _region_start = covered_region.start();
 387   const size_t region_size = covered_region.word_size();
 388   DEBUG_ONLY(_region_end = _region_start + region_size;)
 389 
 390   assert(region_align_down(_region_start) == _region_start,
 391          "region start not aligned");
 392   assert((region_size & RegionSizeOffsetMask) == 0,
 393          "region size not a multiple of RegionSize");
 394 
 395   bool result = initialize_region_data(region_size) && initialize_block_data();
 396   return result;
 397 }
 398 
 399 PSVirtualSpace*
 400 ParallelCompactData::create_vspace(size_t count, size_t element_size)
 401 {
 402   const size_t raw_bytes = count * element_size;
 403   const size_t page_sz = os::page_size_for_region(raw_bytes, raw_bytes, 10);
 404   const size_t granularity = os::vm_allocation_granularity();
 405   _reserved_byte_size = align_size_up(raw_bytes, MAX2(page_sz, granularity));
 406 
 407   const size_t rs_align = page_sz == (size_t) os::vm_page_size() ? 0 :
 408     MAX2(page_sz, granularity);
 409   ReservedSpace rs(_reserved_byte_size, rs_align, rs_align > 0);
 410   os::trace_page_sizes("par compact", raw_bytes, raw_bytes, page_sz, rs.base(),
 411                        rs.size());
 412 
 413   MemTracker::record_virtual_memory_type((address)rs.base(), mtGC);
 414 
 415   PSVirtualSpace* vspace = new PSVirtualSpace(rs, page_sz);
 416   if (vspace != 0) {
 417     if (vspace->expand_by(_reserved_byte_size)) {
 418       return vspace;
 419     }
 420     delete vspace;
 421     // Release memory reserved in the space.
 422     rs.release();
 423   }
 424 
 425   return 0;
 426 }
 427 
 428 bool ParallelCompactData::initialize_region_data(size_t region_size)
 429 {
 430   const size_t count = (region_size + RegionSizeOffsetMask) >> Log2RegionSize;
 431   _region_vspace = create_vspace(count, sizeof(RegionData));
 432   if (_region_vspace != 0) {
 433     _region_data = (RegionData*)_region_vspace->reserved_low_addr();
 434     _region_count = count;
 435     return true;
 436   }
 437   return false;
 438 }
 439 
 440 bool ParallelCompactData::initialize_block_data()
 441 {
 442   assert(_region_count != 0, "region data must be initialized first");
 443   const size_t count = _region_count << Log2BlocksPerRegion;
 444   _block_vspace = create_vspace(count, sizeof(BlockData));
 445   if (_block_vspace != 0) {
 446     _block_data = (BlockData*)_block_vspace->reserved_low_addr();
 447     _block_count = count;
 448     return true;
 449   }
 450   return false;
 451 }
 452 
 453 void ParallelCompactData::clear()
 454 {
 455   memset(_region_data, 0, _region_vspace->committed_size());
 456   memset(_block_data, 0, _block_vspace->committed_size());
 457 }
 458 
 459 void ParallelCompactData::clear_range(size_t beg_region, size_t end_region) {
 460   assert(beg_region <= _region_count, "beg_region out of range");
 461   assert(end_region <= _region_count, "end_region out of range");
 462   assert(RegionSize % BlockSize == 0, "RegionSize not a multiple of BlockSize");
 463 
 464   const size_t region_cnt = end_region - beg_region;
 465   memset(_region_data + beg_region, 0, region_cnt * sizeof(RegionData));
 466 
 467   const size_t beg_block = beg_region * BlocksPerRegion;
 468   const size_t block_cnt = region_cnt * BlocksPerRegion;
 469   memset(_block_data + beg_block, 0, block_cnt * sizeof(BlockData));
 470 }
 471 
 472 HeapWord* ParallelCompactData::partial_obj_end(size_t region_idx) const
 473 {
 474   const RegionData* cur_cp = region(region_idx);
 475   const RegionData* const end_cp = region(region_count() - 1);
 476 
 477   HeapWord* result = region_to_addr(region_idx);
 478   if (cur_cp < end_cp) {
 479     do {
 480       result += cur_cp->partial_obj_size();
 481     } while (cur_cp->partial_obj_size() == RegionSize && ++cur_cp < end_cp);
 482   }
 483   return result;
 484 }
 485 
 486 void ParallelCompactData::add_obj(HeapWord* addr, size_t len)
 487 {
 488   const size_t obj_ofs = pointer_delta(addr, _region_start);
 489   const size_t beg_region = obj_ofs >> Log2RegionSize;
 490   const size_t end_region = (obj_ofs + len - 1) >> Log2RegionSize;
 491 
 492   DEBUG_ONLY(Atomic::inc_ptr(&add_obj_count);)
 493   DEBUG_ONLY(Atomic::add_ptr(len, &add_obj_size);)
 494 
 495   if (beg_region == end_region) {
 496     // All in one region.
 497     _region_data[beg_region].add_live_obj(len);
 498     return;
 499   }
 500 
 501   // First region.
 502   const size_t beg_ofs = region_offset(addr);
 503   _region_data[beg_region].add_live_obj(RegionSize - beg_ofs);
 504 
 505   Klass* klass = ((oop)addr)->klass();
 506   // Middle regions--completely spanned by this object.
 507   for (size_t region = beg_region + 1; region < end_region; ++region) {
 508     _region_data[region].set_partial_obj_size(RegionSize);
 509     _region_data[region].set_partial_obj_addr(addr);
 510   }
 511 
 512   // Last region.
 513   const size_t end_ofs = region_offset(addr + len - 1);
 514   _region_data[end_region].set_partial_obj_size(end_ofs + 1);
 515   _region_data[end_region].set_partial_obj_addr(addr);
 516 }
 517 
 518 void
 519 ParallelCompactData::summarize_dense_prefix(HeapWord* beg, HeapWord* end)
 520 {
 521   assert(region_offset(beg) == 0, "not RegionSize aligned");
 522   assert(region_offset(end) == 0, "not RegionSize aligned");
 523 
 524   size_t cur_region = addr_to_region_idx(beg);
 525   const size_t end_region = addr_to_region_idx(end);
 526   HeapWord* addr = beg;
 527   while (cur_region < end_region) {
 528     _region_data[cur_region].set_destination(addr);
 529     _region_data[cur_region].set_destination_count(0);
 530     _region_data[cur_region].set_source_region(cur_region);
 531     _region_data[cur_region].set_data_location(addr);
 532 
 533     // Update live_obj_size so the region appears completely full.
 534     size_t live_size = RegionSize - _region_data[cur_region].partial_obj_size();
 535     _region_data[cur_region].set_live_obj_size(live_size);
 536 
 537     ++cur_region;
 538     addr += RegionSize;
 539   }
 540 }
 541 
 542 // Find the point at which a space can be split and, if necessary, record the
 543 // split point.
 544 //
 545 // If the current src region (which overflowed the destination space) doesn't
 546 // have a partial object, the split point is at the beginning of the current src
 547 // region (an "easy" split, no extra bookkeeping required).
 548 //
 549 // If the current src region has a partial object, the split point is in the
 550 // region where that partial object starts (call it the split_region).  If
 551 // split_region has a partial object, then the split point is just after that
 552 // partial object (a "hard" split where we have to record the split data and
 553 // zero the partial_obj_size field).  With a "hard" split, we know that the
 554 // partial_obj ends within split_region because the partial object that caused
 555 // the overflow starts in split_region.  If split_region doesn't have a partial
 556 // obj, then the split is at the beginning of split_region (another "easy"
 557 // split).
 558 HeapWord*
 559 ParallelCompactData::summarize_split_space(size_t src_region,
 560                                            SplitInfo& split_info,
 561                                            HeapWord* destination,
 562                                            HeapWord* target_end,
 563                                            HeapWord** target_next)
 564 {
 565   assert(destination <= target_end, "sanity");
 566   assert(destination + _region_data[src_region].data_size() > target_end,
 567     "region should not fit into target space");
 568   assert(is_region_aligned(target_end), "sanity");
 569 
 570   size_t split_region = src_region;
 571   HeapWord* split_destination = destination;
 572   size_t partial_obj_size = _region_data[src_region].partial_obj_size();
 573 
 574   if (destination + partial_obj_size > target_end) {
 575     // The split point is just after the partial object (if any) in the
 576     // src_region that contains the start of the object that overflowed the
 577     // destination space.
 578     //
 579     // Find the start of the "overflow" object and set split_region to the
 580     // region containing it.
 581     HeapWord* const overflow_obj = _region_data[src_region].partial_obj_addr();
 582     split_region = addr_to_region_idx(overflow_obj);
 583 
 584     // Clear the source_region field of all destination regions whose first word
 585     // came from data after the split point (a non-null source_region field
 586     // implies a region must be filled).
 587     //
 588     // An alternative to the simple loop below:  clear during post_compact(),
 589     // which uses memcpy instead of individual stores, and is easy to
 590     // parallelize.  (The downside is that it clears the entire RegionData
 591     // object as opposed to just one field.)
 592     //
 593     // post_compact() would have to clear the summary data up to the highest
 594     // address that was written during the summary phase, which would be
 595     //
 596     //         max(top, max(new_top, clear_top))
 597     //
 598     // where clear_top is a new field in SpaceInfo.  Would have to set clear_top
 599     // to target_end.
 600     const RegionData* const sr = region(split_region);
 601     const size_t beg_idx =
 602       addr_to_region_idx(region_align_up(sr->destination() +
 603                                          sr->partial_obj_size()));
 604     const size_t end_idx = addr_to_region_idx(target_end);
 605 
 606     if (TraceParallelOldGCSummaryPhase) {
 607         gclog_or_tty->print_cr("split:  clearing source_region field in ["
 608                                SIZE_FORMAT ", " SIZE_FORMAT ")",
 609                                beg_idx, end_idx);
 610     }
 611     for (size_t idx = beg_idx; idx < end_idx; ++idx) {
 612       _region_data[idx].set_source_region(0);
 613     }
 614 
 615     // Set split_destination and partial_obj_size to reflect the split region.
 616     split_destination = sr->destination();
 617     partial_obj_size = sr->partial_obj_size();
 618   }
 619 
 620   // The split is recorded only if a partial object extends onto the region.
 621   if (partial_obj_size != 0) {
 622     _region_data[split_region].set_partial_obj_size(0);
 623     split_info.record(split_region, partial_obj_size, split_destination);
 624   }
 625 
 626   // Setup the continuation addresses.
 627   *target_next = split_destination + partial_obj_size;
 628   HeapWord* const source_next = region_to_addr(split_region) + partial_obj_size;
 629 
 630   if (TraceParallelOldGCSummaryPhase) {
 631     const char * split_type = partial_obj_size == 0 ? "easy" : "hard";
 632     gclog_or_tty->print_cr("%s split:  src=" PTR_FORMAT " src_c=" SIZE_FORMAT
 633                            " pos=" SIZE_FORMAT,
 634                            split_type, source_next, split_region,
 635                            partial_obj_size);
 636     gclog_or_tty->print_cr("%s split:  dst=" PTR_FORMAT " dst_c=" SIZE_FORMAT
 637                            " tn=" PTR_FORMAT,
 638                            split_type, split_destination,
 639                            addr_to_region_idx(split_destination),
 640                            *target_next);
 641 
 642     if (partial_obj_size != 0) {
 643       HeapWord* const po_beg = split_info.destination();
 644       HeapWord* const po_end = po_beg + split_info.partial_obj_size();
 645       gclog_or_tty->print_cr("%s split:  "
 646                              "po_beg=" PTR_FORMAT " " SIZE_FORMAT " "
 647                              "po_end=" PTR_FORMAT " " SIZE_FORMAT,
 648                              split_type,
 649                              po_beg, addr_to_region_idx(po_beg),
 650                              po_end, addr_to_region_idx(po_end));
 651     }
 652   }
 653 
 654   return source_next;
 655 }
 656 
 657 bool ParallelCompactData::summarize(SplitInfo& split_info,
 658                                     HeapWord* source_beg, HeapWord* source_end,
 659                                     HeapWord** source_next,
 660                                     HeapWord* target_beg, HeapWord* target_end,
 661                                     HeapWord** target_next)
 662 {
 663   if (TraceParallelOldGCSummaryPhase) {
 664     HeapWord* const source_next_val = source_next == NULL ? NULL : *source_next;
 665     tty->print_cr("sb=" PTR_FORMAT " se=" PTR_FORMAT " sn=" PTR_FORMAT
 666                   "tb=" PTR_FORMAT " te=" PTR_FORMAT " tn=" PTR_FORMAT,
 667                   source_beg, source_end, source_next_val,
 668                   target_beg, target_end, *target_next);
 669   }
 670 
 671   size_t cur_region = addr_to_region_idx(source_beg);
 672   const size_t end_region = addr_to_region_idx(region_align_up(source_end));
 673 
 674   HeapWord *dest_addr = target_beg;
 675   while (cur_region < end_region) {
 676     // The destination must be set even if the region has no data.
 677     _region_data[cur_region].set_destination(dest_addr);
 678 
 679     size_t words = _region_data[cur_region].data_size();
 680     if (words > 0) {
 681       // If cur_region does not fit entirely into the target space, find a point
 682       // at which the source space can be 'split' so that part is copied to the
 683       // target space and the rest is copied elsewhere.
 684       if (dest_addr + words > target_end) {
 685         assert(source_next != NULL, "source_next is NULL when splitting");
 686         *source_next = summarize_split_space(cur_region, split_info, dest_addr,
 687                                              target_end, target_next);
 688         return false;
 689       }
 690 
 691       // Compute the destination_count for cur_region, and if necessary, update
 692       // source_region for a destination region.  The source_region field is
 693       // updated if cur_region is the first (left-most) region to be copied to a
 694       // destination region.
 695       //
 696       // The destination_count calculation is a bit subtle.  A region that has
 697       // data that compacts into itself does not count itself as a destination.
 698       // This maintains the invariant that a zero count means the region is
 699       // available and can be claimed and then filled.
 700       uint destination_count = 0;
 701       if (split_info.is_split(cur_region)) {
 702         // The current region has been split:  the partial object will be copied
 703         // to one destination space and the remaining data will be copied to
 704         // another destination space.  Adjust the initial destination_count and,
 705         // if necessary, set the source_region field if the partial object will
 706         // cross a destination region boundary.
 707         destination_count = split_info.destination_count();
 708         if (destination_count == 2) {
 709           size_t dest_idx = addr_to_region_idx(split_info.dest_region_addr());
 710           _region_data[dest_idx].set_source_region(cur_region);
 711         }
 712       }
 713 
 714       HeapWord* const last_addr = dest_addr + words - 1;
 715       const size_t dest_region_1 = addr_to_region_idx(dest_addr);
 716       const size_t dest_region_2 = addr_to_region_idx(last_addr);
 717 
 718       // Initially assume that the destination regions will be the same and
 719       // adjust the value below if necessary.  Under this assumption, if
 720       // cur_region == dest_region_2, then cur_region will be compacted
 721       // completely into itself.
 722       destination_count += cur_region == dest_region_2 ? 0 : 1;
 723       if (dest_region_1 != dest_region_2) {
 724         // Destination regions differ; adjust destination_count.
 725         destination_count += 1;
 726         // Data from cur_region will be copied to the start of dest_region_2.
 727         _region_data[dest_region_2].set_source_region(cur_region);
 728       } else if (region_offset(dest_addr) == 0) {
 729         // Data from cur_region will be copied to the start of the destination
 730         // region.
 731         _region_data[dest_region_1].set_source_region(cur_region);
 732       }
 733 
 734       _region_data[cur_region].set_destination_count(destination_count);
 735       _region_data[cur_region].set_data_location(region_to_addr(cur_region));
 736       dest_addr += words;
 737     }
 738 
 739     ++cur_region;
 740   }
 741 
 742   *target_next = dest_addr;
 743   return true;
 744 }
 745 
 746 HeapWord* ParallelCompactData::calc_new_pointer(HeapWord* addr) {
 747   assert(addr != NULL, "Should detect NULL oop earlier");
 748   assert(PSParallelCompact::gc_heap()->is_in(addr), "not in heap");
 749   assert(PSParallelCompact::mark_bitmap()->is_marked(addr), "not marked");
 750 
 751   // Region covering the object.
 752   RegionData* const region_ptr = addr_to_region_ptr(addr);
 753   HeapWord* result = region_ptr->destination();
 754 
 755   // If the entire Region is live, the new location is region->destination + the
 756   // offset of the object within in the Region.
 757 
 758   // Run some performance tests to determine if this special case pays off.  It
 759   // is worth it for pointers into the dense prefix.  If the optimization to
 760   // avoid pointer updates in regions that only point to the dense prefix is
 761   // ever implemented, this should be revisited.
 762   if (region_ptr->data_size() == RegionSize) {
 763     result += region_offset(addr);
 764     return result;
 765   }
 766 
 767   // Otherwise, the new location is region->destination + block offset + the
 768   // number of live words in the Block that are (a) to the left of addr and (b)
 769   // due to objects that start in the Block.
 770 
 771   // Fill in the block table if necessary.  This is unsynchronized, so multiple
 772   // threads may fill the block table for a region (harmless, since it is
 773   // idempotent).
 774   if (!region_ptr->blocks_filled()) {
 775     PSParallelCompact::fill_blocks(addr_to_region_idx(addr));
 776     region_ptr->set_blocks_filled();
 777   }
 778 
 779   HeapWord* const search_start = block_align_down(addr);
 780   const size_t block_offset = addr_to_block_ptr(addr)->offset();
 781 
 782   const ParMarkBitMap* bitmap = PSParallelCompact::mark_bitmap();
 783   const size_t live = bitmap->live_words_in_range(search_start, oop(addr));
 784   result += block_offset + live;
 785   DEBUG_ONLY(PSParallelCompact::check_new_location(addr, result));
 786   return result;
 787 }
 788 
 789 #ifdef ASSERT
 790 void ParallelCompactData::verify_clear(const PSVirtualSpace* vspace)
 791 {
 792   const size_t* const beg = (const size_t*)vspace->committed_low_addr();
 793   const size_t* const end = (const size_t*)vspace->committed_high_addr();
 794   for (const size_t* p = beg; p < end; ++p) {
 795     assert(*p == 0, "not zero");
 796   }
 797 }
 798 
 799 void ParallelCompactData::verify_clear()
 800 {
 801   verify_clear(_region_vspace);
 802   verify_clear(_block_vspace);
 803 }
 804 #endif  // #ifdef ASSERT
 805 
 806 STWGCTimer          PSParallelCompact::_gc_timer;
 807 ParallelOldTracer   PSParallelCompact::_gc_tracer;
 808 elapsedTimer        PSParallelCompact::_accumulated_time;
 809 unsigned int        PSParallelCompact::_total_invocations = 0;
 810 unsigned int        PSParallelCompact::_maximum_compaction_gc_num = 0;
 811 jlong               PSParallelCompact::_time_of_last_gc = 0;
 812 CollectorCounters*  PSParallelCompact::_counters = NULL;
 813 ParMarkBitMap       PSParallelCompact::_mark_bitmap;
 814 ParallelCompactData PSParallelCompact::_summary_data;
 815 
 816 PSParallelCompact::IsAliveClosure PSParallelCompact::_is_alive_closure;
 817 
 818 bool PSParallelCompact::IsAliveClosure::do_object_b(oop p) { return mark_bitmap()->is_marked(p); }
 819 
 820 void PSParallelCompact::KeepAliveClosure::do_oop(oop* p)       { PSParallelCompact::KeepAliveClosure::do_oop_work(p); }
 821 void PSParallelCompact::KeepAliveClosure::do_oop(narrowOop* p) { PSParallelCompact::KeepAliveClosure::do_oop_work(p); }
 822 
 823 PSParallelCompact::AdjustPointerClosure PSParallelCompact::_adjust_pointer_closure;
 824 PSParallelCompact::AdjustKlassClosure PSParallelCompact::_adjust_klass_closure;
 825 
 826 void PSParallelCompact::AdjustPointerClosure::do_oop(oop* p)       { adjust_pointer(p); }
 827 void PSParallelCompact::AdjustPointerClosure::do_oop(narrowOop* p) { adjust_pointer(p); }
 828 
 829 void PSParallelCompact::FollowStackClosure::do_void() { _compaction_manager->follow_marking_stacks(); }
 830 
 831 void PSParallelCompact::MarkAndPushClosure::do_oop(oop* p)       {
 832   mark_and_push(_compaction_manager, p);
 833 }
 834 void PSParallelCompact::MarkAndPushClosure::do_oop(narrowOop* p) { mark_and_push(_compaction_manager, p); }
 835 
 836 void PSParallelCompact::FollowKlassClosure::do_klass(Klass* klass) {
 837   klass->oops_do(_mark_and_push_closure);
 838 }
 839 void PSParallelCompact::AdjustKlassClosure::do_klass(Klass* klass) {
 840   klass->oops_do(&PSParallelCompact::_adjust_pointer_closure);
 841 }
 842 
 843 void PSParallelCompact::post_initialize() {
 844   ParallelScavengeHeap* heap = gc_heap();
 845   assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity");
 846 
 847   MemRegion mr = heap->reserved_region();
 848   _ref_processor =
 849     new ReferenceProcessor(mr,            // span
 850                            ParallelRefProcEnabled && (ParallelGCThreads > 1), // mt processing
 851                            (int) ParallelGCThreads, // mt processing degree
 852                            true,          // mt discovery
 853                            (int) ParallelGCThreads, // mt discovery degree
 854                            true,          // atomic_discovery
 855                            &_is_alive_closure, // non-header is alive closure
 856                            false);        // write barrier for next field updates
 857   _counters = new CollectorCounters("PSParallelCompact", 1);
 858 
 859   // Initialize static fields in ParCompactionManager.
 860   ParCompactionManager::initialize(mark_bitmap());
 861 }
 862 
 863 bool PSParallelCompact::initialize() {
 864   ParallelScavengeHeap* heap = gc_heap();
 865   assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity");
 866   MemRegion mr = heap->reserved_region();
 867 
 868   // Was the old gen get allocated successfully?
 869   if (!heap->old_gen()->is_allocated()) {
 870     return false;
 871   }
 872 
 873   initialize_space_info();
 874   initialize_dead_wood_limiter();
 875 
 876   if (!_mark_bitmap.initialize(mr)) {
 877     vm_shutdown_during_initialization(
 878       err_msg("Unable to allocate " SIZE_FORMAT "KB bitmaps for parallel "
 879       "garbage collection for the requested " SIZE_FORMAT "KB heap.",
 880       _mark_bitmap.reserved_byte_size()/K, mr.byte_size()/K));
 881     return false;
 882   }
 883 
 884   if (!_summary_data.initialize(mr)) {
 885     vm_shutdown_during_initialization(
 886       err_msg("Unable to allocate " SIZE_FORMAT "KB card tables for parallel "
 887       "garbage collection for the requested " SIZE_FORMAT "KB heap.",
 888       _summary_data.reserved_byte_size()/K, mr.byte_size()/K));
 889     return false;
 890   }
 891 
 892   return true;
 893 }
 894 
 895 void PSParallelCompact::initialize_space_info()
 896 {
 897   memset(&_space_info, 0, sizeof(_space_info));
 898 
 899   ParallelScavengeHeap* heap = gc_heap();
 900   PSYoungGen* young_gen = heap->young_gen();
 901 
 902   _space_info[old_space_id].set_space(heap->old_gen()->object_space());
 903   _space_info[eden_space_id].set_space(young_gen->eden_space());
 904   _space_info[from_space_id].set_space(young_gen->from_space());
 905   _space_info[to_space_id].set_space(young_gen->to_space());
 906 
 907   _space_info[old_space_id].set_start_array(heap->old_gen()->start_array());
 908 }
 909 
 910 void PSParallelCompact::initialize_dead_wood_limiter()
 911 {
 912   const size_t max = 100;
 913   _dwl_mean = double(MIN2(ParallelOldDeadWoodLimiterMean, max)) / 100.0;
 914   _dwl_std_dev = double(MIN2(ParallelOldDeadWoodLimiterStdDev, max)) / 100.0;
 915   _dwl_first_term = 1.0 / (sqrt(2.0 * M_PI) * _dwl_std_dev);
 916   DEBUG_ONLY(_dwl_initialized = true;)
 917   _dwl_adjustment = normal_distribution(1.0);
 918 }
 919 
 920 // Simple class for storing info about the heap at the start of GC, to be used
 921 // after GC for comparison/printing.
 922 class PreGCValues {
 923 public:
 924   PreGCValues() { }
 925   PreGCValues(ParallelScavengeHeap* heap) { fill(heap); }
 926 
 927   void fill(ParallelScavengeHeap* heap) {
 928     _heap_used      = heap->used();
 929     _young_gen_used = heap->young_gen()->used_in_bytes();
 930     _old_gen_used   = heap->old_gen()->used_in_bytes();
 931     _metadata_used  = MetaspaceAux::allocated_used_bytes();
 932   };
 933 
 934   size_t heap_used() const      { return _heap_used; }
 935   size_t young_gen_used() const { return _young_gen_used; }
 936   size_t old_gen_used() const   { return _old_gen_used; }
 937   size_t metadata_used() const  { return _metadata_used; }
 938 
 939 private:
 940   size_t _heap_used;
 941   size_t _young_gen_used;
 942   size_t _old_gen_used;
 943   size_t _metadata_used;
 944 };
 945 
 946 void
 947 PSParallelCompact::clear_data_covering_space(SpaceId id)
 948 {
 949   // At this point, top is the value before GC, new_top() is the value that will
 950   // be set at the end of GC.  The marking bitmap is cleared to top; nothing
 951   // should be marked above top.  The summary data is cleared to the larger of
 952   // top & new_top.
 953   MutableSpace* const space = _space_info[id].space();
 954   HeapWord* const bot = space->bottom();
 955   HeapWord* const top = space->top();
 956   HeapWord* const max_top = MAX2(top, _space_info[id].new_top());
 957 
 958   const idx_t beg_bit = _mark_bitmap.addr_to_bit(bot);
 959   const idx_t end_bit = BitMap::word_align_up(_mark_bitmap.addr_to_bit(top));
 960   _mark_bitmap.clear_range(beg_bit, end_bit);
 961 
 962   const size_t beg_region = _summary_data.addr_to_region_idx(bot);
 963   const size_t end_region =
 964     _summary_data.addr_to_region_idx(_summary_data.region_align_up(max_top));
 965   _summary_data.clear_range(beg_region, end_region);
 966 
 967   // Clear the data used to 'split' regions.
 968   SplitInfo& split_info = _space_info[id].split_info();
 969   if (split_info.is_valid()) {
 970     split_info.clear();
 971   }
 972   DEBUG_ONLY(split_info.verify_clear();)
 973 }
 974 
 975 void PSParallelCompact::pre_compact(PreGCValues* pre_gc_values)
 976 {
 977   // Update the from & to space pointers in space_info, since they are swapped
 978   // at each young gen gc.  Do the update unconditionally (even though a
 979   // promotion failure does not swap spaces) because an unknown number of minor
 980   // collections will have swapped the spaces an unknown number of times.
 981   GCTraceTime tm("pre compact", print_phases(), true, &_gc_timer);
 982   ParallelScavengeHeap* heap = gc_heap();
 983   _space_info[from_space_id].set_space(heap->young_gen()->from_space());
 984   _space_info[to_space_id].set_space(heap->young_gen()->to_space());
 985 
 986   pre_gc_values->fill(heap);
 987 
 988   DEBUG_ONLY(add_obj_count = add_obj_size = 0;)
 989   DEBUG_ONLY(mark_bitmap_count = mark_bitmap_size = 0;)
 990 
 991   // Increment the invocation count
 992   heap->increment_total_collections(true);
 993 
 994   // We need to track unique mark sweep invocations as well.
 995   _total_invocations++;
 996 
 997   heap->print_heap_before_gc();
 998   heap->trace_heap_before_gc(&_gc_tracer);
 999 
1000   // Fill in TLABs
1001   heap->accumulate_statistics_all_tlabs();
1002   heap->ensure_parsability(true);  // retire TLABs
1003 
1004   if (VerifyBeforeGC && heap->total_collections() >= VerifyGCStartAt) {
1005     HandleMark hm;  // Discard invalid handles created during verification
1006     Universe::verify(" VerifyBeforeGC:");
1007   }
1008 
1009   // Verify object start arrays
1010   if (VerifyObjectStartArray &&
1011       VerifyBeforeGC) {
1012     heap->old_gen()->verify_object_start_array();
1013   }
1014 
1015   DEBUG_ONLY(mark_bitmap()->verify_clear();)
1016   DEBUG_ONLY(summary_data().verify_clear();)
1017 
1018   // Have worker threads release resources the next time they run a task.
1019   gc_task_manager()->release_all_resources();
1020 }
1021 
1022 void PSParallelCompact::post_compact()
1023 {
1024   GCTraceTime tm("post compact", print_phases(), true, &_gc_timer);
1025 
1026   for (unsigned int id = old_space_id; id < last_space_id; ++id) {
1027     // Clear the marking bitmap, summary data and split info.
1028     clear_data_covering_space(SpaceId(id));
1029     // Update top().  Must be done after clearing the bitmap and summary data.
1030     _space_info[id].publish_new_top();
1031   }
1032 
1033   MutableSpace* const eden_space = _space_info[eden_space_id].space();
1034   MutableSpace* const from_space = _space_info[from_space_id].space();
1035   MutableSpace* const to_space   = _space_info[to_space_id].space();
1036 
1037   ParallelScavengeHeap* heap = gc_heap();
1038   bool eden_empty = eden_space->is_empty();
1039   if (!eden_empty) {
1040     eden_empty = absorb_live_data_from_eden(heap->size_policy(),
1041                                             heap->young_gen(), heap->old_gen());
1042   }
1043 
1044   // Update heap occupancy information which is used as input to the soft ref
1045   // clearing policy at the next gc.
1046   Universe::update_heap_info_at_gc();
1047 
1048   bool young_gen_empty = eden_empty && from_space->is_empty() &&
1049     to_space->is_empty();
1050 
1051   BarrierSet* bs = heap->barrier_set();
1052   if (bs->is_a(BarrierSet::ModRef)) {
1053     ModRefBarrierSet* modBS = (ModRefBarrierSet*)bs;
1054     MemRegion old_mr = heap->old_gen()->reserved();
1055 
1056     if (young_gen_empty) {
1057       modBS->clear(MemRegion(old_mr.start(), old_mr.end()));
1058     } else {
1059       modBS->invalidate(MemRegion(old_mr.start(), old_mr.end()));
1060     }
1061   }
1062 
1063   // Delete metaspaces for unloaded class loaders and clean up loader_data graph
1064   ClassLoaderDataGraph::purge();
1065   MetaspaceAux::verify_metrics();
1066 
1067   Threads::gc_epilogue();
1068   CodeCache::gc_epilogue();
1069   JvmtiExport::gc_epilogue();
1070 
1071   COMPILER2_PRESENT(DerivedPointerTable::update_pointers());
1072 
1073   ref_processor()->enqueue_discovered_references(NULL);
1074 
1075   if (ZapUnusedHeapArea) {
1076     heap->gen_mangle_unused_area();
1077   }
1078 
1079   // Update time of last GC
1080   reset_millis_since_last_gc();
1081 }
1082 
1083 HeapWord*
1084 PSParallelCompact::compute_dense_prefix_via_density(const SpaceId id,
1085                                                     bool maximum_compaction)
1086 {
1087   const size_t region_size = ParallelCompactData::RegionSize;
1088   const ParallelCompactData& sd = summary_data();
1089 
1090   const MutableSpace* const space = _space_info[id].space();
1091   HeapWord* const top_aligned_up = sd.region_align_up(space->top());
1092   const RegionData* const beg_cp = sd.addr_to_region_ptr(space->bottom());
1093   const RegionData* const end_cp = sd.addr_to_region_ptr(top_aligned_up);
1094 
1095   // Skip full regions at the beginning of the space--they are necessarily part
1096   // of the dense prefix.
1097   size_t full_count = 0;
1098   const RegionData* cp;
1099   for (cp = beg_cp; cp < end_cp && cp->data_size() == region_size; ++cp) {
1100     ++full_count;
1101   }
1102 
1103   assert(total_invocations() >= _maximum_compaction_gc_num, "sanity");
1104   const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num;
1105   const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval;
1106   if (maximum_compaction || cp == end_cp || interval_ended) {
1107     _maximum_compaction_gc_num = total_invocations();
1108     return sd.region_to_addr(cp);
1109   }
1110 
1111   HeapWord* const new_top = _space_info[id].new_top();
1112   const size_t space_live = pointer_delta(new_top, space->bottom());
1113   const size_t space_used = space->used_in_words();
1114   const size_t space_capacity = space->capacity_in_words();
1115 
1116   const double cur_density = double(space_live) / space_capacity;
1117   const double deadwood_density =
1118     (1.0 - cur_density) * (1.0 - cur_density) * cur_density * cur_density;
1119   const size_t deadwood_goal = size_t(space_capacity * deadwood_density);
1120 
1121   if (TraceParallelOldGCDensePrefix) {
1122     tty->print_cr("cur_dens=%5.3f dw_dens=%5.3f dw_goal=" SIZE_FORMAT,
1123                   cur_density, deadwood_density, deadwood_goal);
1124     tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " "
1125                   "space_cap=" SIZE_FORMAT,
1126                   space_live, space_used,
1127                   space_capacity);
1128   }
1129 
1130   // XXX - Use binary search?
1131   HeapWord* dense_prefix = sd.region_to_addr(cp);
1132   const RegionData* full_cp = cp;
1133   const RegionData* const top_cp = sd.addr_to_region_ptr(space->top() - 1);
1134   while (cp < end_cp) {
1135     HeapWord* region_destination = cp->destination();
1136     const size_t cur_deadwood = pointer_delta(dense_prefix, region_destination);
1137     if (TraceParallelOldGCDensePrefix && Verbose) {
1138       tty->print_cr("c#=" SIZE_FORMAT_W(4) " dst=" PTR_FORMAT " "
1139                     "dp=" SIZE_FORMAT_W(8) " " "cdw=" SIZE_FORMAT_W(8),
1140                     sd.region(cp), region_destination,
1141                     dense_prefix, cur_deadwood);
1142     }
1143 
1144     if (cur_deadwood >= deadwood_goal) {
1145       // Found the region that has the correct amount of deadwood to the left.
1146       // This typically occurs after crossing a fairly sparse set of regions, so
1147       // iterate backwards over those sparse regions, looking for the region
1148       // that has the lowest density of live objects 'to the right.'
1149       size_t space_to_left = sd.region(cp) * region_size;
1150       size_t live_to_left = space_to_left - cur_deadwood;
1151       size_t space_to_right = space_capacity - space_to_left;
1152       size_t live_to_right = space_live - live_to_left;
1153       double density_to_right = double(live_to_right) / space_to_right;
1154       while (cp > full_cp) {
1155         --cp;
1156         const size_t prev_region_live_to_right = live_to_right -
1157           cp->data_size();
1158         const size_t prev_region_space_to_right = space_to_right + region_size;
1159         double prev_region_density_to_right =
1160           double(prev_region_live_to_right) / prev_region_space_to_right;
1161         if (density_to_right <= prev_region_density_to_right) {
1162           return dense_prefix;
1163         }
1164         if (TraceParallelOldGCDensePrefix && Verbose) {
1165           tty->print_cr("backing up from c=" SIZE_FORMAT_W(4) " d2r=%10.8f "
1166                         "pc_d2r=%10.8f", sd.region(cp), density_to_right,
1167                         prev_region_density_to_right);
1168         }
1169         dense_prefix -= region_size;
1170         live_to_right = prev_region_live_to_right;
1171         space_to_right = prev_region_space_to_right;
1172         density_to_right = prev_region_density_to_right;
1173       }
1174       return dense_prefix;
1175     }
1176 
1177     dense_prefix += region_size;
1178     ++cp;
1179   }
1180 
1181   return dense_prefix;
1182 }
1183 
1184 #ifndef PRODUCT
1185 void PSParallelCompact::print_dense_prefix_stats(const char* const algorithm,
1186                                                  const SpaceId id,
1187                                                  const bool maximum_compaction,
1188                                                  HeapWord* const addr)
1189 {
1190   const size_t region_idx = summary_data().addr_to_region_idx(addr);
1191   RegionData* const cp = summary_data().region(region_idx);
1192   const MutableSpace* const space = _space_info[id].space();
1193   HeapWord* const new_top = _space_info[id].new_top();
1194 
1195   const size_t space_live = pointer_delta(new_top, space->bottom());
1196   const size_t dead_to_left = pointer_delta(addr, cp->destination());
1197   const size_t space_cap = space->capacity_in_words();
1198   const double dead_to_left_pct = double(dead_to_left) / space_cap;
1199   const size_t live_to_right = new_top - cp->destination();
1200   const size_t dead_to_right = space->top() - addr - live_to_right;
1201 
1202   tty->print_cr("%s=" PTR_FORMAT " dpc=" SIZE_FORMAT_W(5) " "
1203                 "spl=" SIZE_FORMAT " "
1204                 "d2l=" SIZE_FORMAT " d2l%%=%6.4f "
1205                 "d2r=" SIZE_FORMAT " l2r=" SIZE_FORMAT
1206                 " ratio=%10.8f",
1207                 algorithm, addr, region_idx,
1208                 space_live,
1209                 dead_to_left, dead_to_left_pct,
1210                 dead_to_right, live_to_right,
1211                 double(dead_to_right) / live_to_right);
1212 }
1213 #endif  // #ifndef PRODUCT
1214 
1215 // Return a fraction indicating how much of the generation can be treated as
1216 // "dead wood" (i.e., not reclaimed).  The function uses a normal distribution
1217 // based on the density of live objects in the generation to determine a limit,
1218 // which is then adjusted so the return value is min_percent when the density is
1219 // 1.
1220 //
1221 // The following table shows some return values for a different values of the
1222 // standard deviation (ParallelOldDeadWoodLimiterStdDev); the mean is 0.5 and
1223 // min_percent is 1.
1224 //
1225 //                          fraction allowed as dead wood
1226 //         -----------------------------------------------------------------
1227 // density std_dev=70 std_dev=75 std_dev=80 std_dev=85 std_dev=90 std_dev=95
1228 // ------- ---------- ---------- ---------- ---------- ---------- ----------
1229 // 0.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000
1230 // 0.05000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941
1231 // 0.10000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272
1232 // 0.15000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066
1233 // 0.20000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975
1234 // 0.25000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313
1235 // 0.30000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132
1236 // 0.35000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289
1237 // 0.40000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500
1238 // 0.45000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386
1239 // 0.50000 0.13832410 0.11599237 0.09847664 0.08456518 0.07338887 0.06431510
1240 // 0.55000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386
1241 // 0.60000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500
1242 // 0.65000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289
1243 // 0.70000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132
1244 // 0.75000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313
1245 // 0.80000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975
1246 // 0.85000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066
1247 // 0.90000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272
1248 // 0.95000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941
1249 // 1.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000
1250 
1251 double PSParallelCompact::dead_wood_limiter(double density, size_t min_percent)
1252 {
1253   assert(_dwl_initialized, "uninitialized");
1254 
1255   // The raw limit is the value of the normal distribution at x = density.
1256   const double raw_limit = normal_distribution(density);
1257 
1258   // Adjust the raw limit so it becomes the minimum when the density is 1.
1259   //
1260   // First subtract the adjustment value (which is simply the precomputed value
1261   // normal_distribution(1.0)); this yields a value of 0 when the density is 1.
1262   // Then add the minimum value, so the minimum is returned when the density is
1263   // 1.  Finally, prevent negative values, which occur when the mean is not 0.5.
1264   const double min = double(min_percent) / 100.0;
1265   const double limit = raw_limit - _dwl_adjustment + min;
1266   return MAX2(limit, 0.0);
1267 }
1268 
1269 ParallelCompactData::RegionData*
1270 PSParallelCompact::first_dead_space_region(const RegionData* beg,
1271                                            const RegionData* end)
1272 {
1273   const size_t region_size = ParallelCompactData::RegionSize;
1274   ParallelCompactData& sd = summary_data();
1275   size_t left = sd.region(beg);
1276   size_t right = end > beg ? sd.region(end) - 1 : left;
1277 
1278   // Binary search.
1279   while (left < right) {
1280     // Equivalent to (left + right) / 2, but does not overflow.
1281     const size_t middle = left + (right - left) / 2;
1282     RegionData* const middle_ptr = sd.region(middle);
1283     HeapWord* const dest = middle_ptr->destination();
1284     HeapWord* const addr = sd.region_to_addr(middle);
1285     assert(dest != NULL, "sanity");
1286     assert(dest <= addr, "must move left");
1287 
1288     if (middle > left && dest < addr) {
1289       right = middle - 1;
1290     } else if (middle < right && middle_ptr->data_size() == region_size) {
1291       left = middle + 1;
1292     } else {
1293       return middle_ptr;
1294     }
1295   }
1296   return sd.region(left);
1297 }
1298 
1299 ParallelCompactData::RegionData*
1300 PSParallelCompact::dead_wood_limit_region(const RegionData* beg,
1301                                           const RegionData* end,
1302                                           size_t dead_words)
1303 {
1304   ParallelCompactData& sd = summary_data();
1305   size_t left = sd.region(beg);
1306   size_t right = end > beg ? sd.region(end) - 1 : left;
1307 
1308   // Binary search.
1309   while (left < right) {
1310     // Equivalent to (left + right) / 2, but does not overflow.
1311     const size_t middle = left + (right - left) / 2;
1312     RegionData* const middle_ptr = sd.region(middle);
1313     HeapWord* const dest = middle_ptr->destination();
1314     HeapWord* const addr = sd.region_to_addr(middle);
1315     assert(dest != NULL, "sanity");
1316     assert(dest <= addr, "must move left");
1317 
1318     const size_t dead_to_left = pointer_delta(addr, dest);
1319     if (middle > left && dead_to_left > dead_words) {
1320       right = middle - 1;
1321     } else if (middle < right && dead_to_left < dead_words) {
1322       left = middle + 1;
1323     } else {
1324       return middle_ptr;
1325     }
1326   }
1327   return sd.region(left);
1328 }
1329 
1330 // The result is valid during the summary phase, after the initial summarization
1331 // of each space into itself, and before final summarization.
1332 inline double
1333 PSParallelCompact::reclaimed_ratio(const RegionData* const cp,
1334                                    HeapWord* const bottom,
1335                                    HeapWord* const top,
1336                                    HeapWord* const new_top)
1337 {
1338   ParallelCompactData& sd = summary_data();
1339 
1340   assert(cp != NULL, "sanity");
1341   assert(bottom != NULL, "sanity");
1342   assert(top != NULL, "sanity");
1343   assert(new_top != NULL, "sanity");
1344   assert(top >= new_top, "summary data problem?");
1345   assert(new_top > bottom, "space is empty; should not be here");
1346   assert(new_top >= cp->destination(), "sanity");
1347   assert(top >= sd.region_to_addr(cp), "sanity");
1348 
1349   HeapWord* const destination = cp->destination();
1350   const size_t dense_prefix_live  = pointer_delta(destination, bottom);
1351   const size_t compacted_region_live = pointer_delta(new_top, destination);
1352   const size_t compacted_region_used = pointer_delta(top,
1353                                                      sd.region_to_addr(cp));
1354   const size_t reclaimable = compacted_region_used - compacted_region_live;
1355 
1356   const double divisor = dense_prefix_live + 1.25 * compacted_region_live;
1357   return double(reclaimable) / divisor;
1358 }
1359 
1360 // Return the address of the end of the dense prefix, a.k.a. the start of the
1361 // compacted region.  The address is always on a region boundary.
1362 //
1363 // Completely full regions at the left are skipped, since no compaction can
1364 // occur in those regions.  Then the maximum amount of dead wood to allow is
1365 // computed, based on the density (amount live / capacity) of the generation;
1366 // the region with approximately that amount of dead space to the left is
1367 // identified as the limit region.  Regions between the last completely full
1368 // region and the limit region are scanned and the one that has the best
1369 // (maximum) reclaimed_ratio() is selected.
1370 HeapWord*
1371 PSParallelCompact::compute_dense_prefix(const SpaceId id,
1372                                         bool maximum_compaction)
1373 {
1374   if (ParallelOldGCSplitALot) {
1375     if (_space_info[id].dense_prefix() != _space_info[id].space()->bottom()) {
1376       // The value was chosen to provoke splitting a young gen space; use it.
1377       return _space_info[id].dense_prefix();
1378     }
1379   }
1380 
1381   const size_t region_size = ParallelCompactData::RegionSize;
1382   const ParallelCompactData& sd = summary_data();
1383 
1384   const MutableSpace* const space = _space_info[id].space();
1385   HeapWord* const top = space->top();
1386   HeapWord* const top_aligned_up = sd.region_align_up(top);
1387   HeapWord* const new_top = _space_info[id].new_top();
1388   HeapWord* const new_top_aligned_up = sd.region_align_up(new_top);
1389   HeapWord* const bottom = space->bottom();
1390   const RegionData* const beg_cp = sd.addr_to_region_ptr(bottom);
1391   const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up);
1392   const RegionData* const new_top_cp =
1393     sd.addr_to_region_ptr(new_top_aligned_up);
1394 
1395   // Skip full regions at the beginning of the space--they are necessarily part
1396   // of the dense prefix.
1397   const RegionData* const full_cp = first_dead_space_region(beg_cp, new_top_cp);
1398   assert(full_cp->destination() == sd.region_to_addr(full_cp) ||
1399          space->is_empty(), "no dead space allowed to the left");
1400   assert(full_cp->data_size() < region_size || full_cp == new_top_cp - 1,
1401          "region must have dead space");
1402 
1403   // The gc number is saved whenever a maximum compaction is done, and used to
1404   // determine when the maximum compaction interval has expired.  This avoids
1405   // successive max compactions for different reasons.
1406   assert(total_invocations() >= _maximum_compaction_gc_num, "sanity");
1407   const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num;
1408   const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval ||
1409     total_invocations() == HeapFirstMaximumCompactionCount;
1410   if (maximum_compaction || full_cp == top_cp || interval_ended) {
1411     _maximum_compaction_gc_num = total_invocations();
1412     return sd.region_to_addr(full_cp);
1413   }
1414 
1415   const size_t space_live = pointer_delta(new_top, bottom);
1416   const size_t space_used = space->used_in_words();
1417   const size_t space_capacity = space->capacity_in_words();
1418 
1419   const double density = double(space_live) / double(space_capacity);
1420   const size_t min_percent_free = MarkSweepDeadRatio;
1421   const double limiter = dead_wood_limiter(density, min_percent_free);
1422   const size_t dead_wood_max = space_used - space_live;
1423   const size_t dead_wood_limit = MIN2(size_t(space_capacity * limiter),
1424                                       dead_wood_max);
1425 
1426   if (TraceParallelOldGCDensePrefix) {
1427     tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " "
1428                   "space_cap=" SIZE_FORMAT,
1429                   space_live, space_used,
1430                   space_capacity);
1431     tty->print_cr("dead_wood_limiter(%6.4f, %d)=%6.4f "
1432                   "dead_wood_max=" SIZE_FORMAT " dead_wood_limit=" SIZE_FORMAT,
1433                   density, min_percent_free, limiter,
1434                   dead_wood_max, dead_wood_limit);
1435   }
1436 
1437   // Locate the region with the desired amount of dead space to the left.
1438   const RegionData* const limit_cp =
1439     dead_wood_limit_region(full_cp, top_cp, dead_wood_limit);
1440 
1441   // Scan from the first region with dead space to the limit region and find the
1442   // one with the best (largest) reclaimed ratio.
1443   double best_ratio = 0.0;
1444   const RegionData* best_cp = full_cp;
1445   for (const RegionData* cp = full_cp; cp < limit_cp; ++cp) {
1446     double tmp_ratio = reclaimed_ratio(cp, bottom, top, new_top);
1447     if (tmp_ratio > best_ratio) {
1448       best_cp = cp;
1449       best_ratio = tmp_ratio;
1450     }
1451   }
1452 
1453 #if     0
1454   // Something to consider:  if the region with the best ratio is 'close to' the
1455   // first region w/free space, choose the first region with free space
1456   // ("first-free").  The first-free region is usually near the start of the
1457   // heap, which means we are copying most of the heap already, so copy a bit
1458   // more to get complete compaction.
1459   if (pointer_delta(best_cp, full_cp, sizeof(RegionData)) < 4) {
1460     _maximum_compaction_gc_num = total_invocations();
1461     best_cp = full_cp;
1462   }
1463 #endif  // #if 0
1464 
1465   return sd.region_to_addr(best_cp);
1466 }
1467 
1468 #ifndef PRODUCT
1469 void
1470 PSParallelCompact::fill_with_live_objects(SpaceId id, HeapWord* const start,
1471                                           size_t words)
1472 {
1473   if (TraceParallelOldGCSummaryPhase) {
1474     tty->print_cr("fill_with_live_objects [" PTR_FORMAT " " PTR_FORMAT ") "
1475                   SIZE_FORMAT, start, start + words, words);
1476   }
1477 
1478   ObjectStartArray* const start_array = _space_info[id].start_array();
1479   CollectedHeap::fill_with_objects(start, words);
1480   for (HeapWord* p = start; p < start + words; p += oop(p)->size()) {
1481     _mark_bitmap.mark_obj(p, words);
1482     _summary_data.add_obj(p, words);
1483     start_array->allocate_block(p);
1484   }
1485 }
1486 
1487 void
1488 PSParallelCompact::summarize_new_objects(SpaceId id, HeapWord* start)
1489 {
1490   ParallelCompactData& sd = summary_data();
1491   MutableSpace* space = _space_info[id].space();
1492 
1493   // Find the source and destination start addresses.
1494   HeapWord* const src_addr = sd.region_align_down(start);
1495   HeapWord* dst_addr;
1496   if (src_addr < start) {
1497     dst_addr = sd.addr_to_region_ptr(src_addr)->destination();
1498   } else if (src_addr > space->bottom()) {
1499     // The start (the original top() value) is aligned to a region boundary so
1500     // the associated region does not have a destination.  Compute the
1501     // destination from the previous region.
1502     RegionData* const cp = sd.addr_to_region_ptr(src_addr) - 1;
1503     dst_addr = cp->destination() + cp->data_size();
1504   } else {
1505     // Filling the entire space.
1506     dst_addr = space->bottom();
1507   }
1508   assert(dst_addr != NULL, "sanity");
1509 
1510   // Update the summary data.
1511   bool result = _summary_data.summarize(_space_info[id].split_info(),
1512                                         src_addr, space->top(), NULL,
1513                                         dst_addr, space->end(),
1514                                         _space_info[id].new_top_addr());
1515   assert(result, "should not fail:  bad filler object size");
1516 }
1517 
1518 void
1519 PSParallelCompact::provoke_split_fill_survivor(SpaceId id)
1520 {
1521   if (total_invocations() % (ParallelOldGCSplitInterval * 3) != 0) {
1522     return;
1523   }
1524 
1525   MutableSpace* const space = _space_info[id].space();
1526   if (space->is_empty()) {
1527     HeapWord* b = space->bottom();
1528     HeapWord* t = b + space->capacity_in_words() / 2;
1529     space->set_top(t);
1530     if (ZapUnusedHeapArea) {
1531       space->set_top_for_allocations();
1532     }
1533 
1534     size_t min_size = CollectedHeap::min_fill_size();
1535     size_t obj_len = min_size;
1536     while (b + obj_len <= t) {
1537       CollectedHeap::fill_with_object(b, obj_len);
1538       mark_bitmap()->mark_obj(b, obj_len);
1539       summary_data().add_obj(b, obj_len);
1540       b += obj_len;
1541       obj_len = (obj_len & (min_size*3)) + min_size; // 8 16 24 32 8 16 24 32 ...
1542     }
1543     if (b < t) {
1544       // The loop didn't completely fill to t (top); adjust top downward.
1545       space->set_top(b);
1546       if (ZapUnusedHeapArea) {
1547         space->set_top_for_allocations();
1548       }
1549     }
1550 
1551     HeapWord** nta = _space_info[id].new_top_addr();
1552     bool result = summary_data().summarize(_space_info[id].split_info(),
1553                                            space->bottom(), space->top(), NULL,
1554                                            space->bottom(), space->end(), nta);
1555     assert(result, "space must fit into itself");
1556   }
1557 }
1558 
1559 void
1560 PSParallelCompact::provoke_split(bool & max_compaction)
1561 {
1562   if (total_invocations() % ParallelOldGCSplitInterval != 0) {
1563     return;
1564   }
1565 
1566   const size_t region_size = ParallelCompactData::RegionSize;
1567   ParallelCompactData& sd = summary_data();
1568 
1569   MutableSpace* const eden_space = _space_info[eden_space_id].space();
1570   MutableSpace* const from_space = _space_info[from_space_id].space();
1571   const size_t eden_live = pointer_delta(eden_space->top(),
1572                                          _space_info[eden_space_id].new_top());
1573   const size_t from_live = pointer_delta(from_space->top(),
1574                                          _space_info[from_space_id].new_top());
1575 
1576   const size_t min_fill_size = CollectedHeap::min_fill_size();
1577   const size_t eden_free = pointer_delta(eden_space->end(), eden_space->top());
1578   const size_t eden_fillable = eden_free >= min_fill_size ? eden_free : 0;
1579   const size_t from_free = pointer_delta(from_space->end(), from_space->top());
1580   const size_t from_fillable = from_free >= min_fill_size ? from_free : 0;
1581 
1582   // Choose the space to split; need at least 2 regions live (or fillable).
1583   SpaceId id;
1584   MutableSpace* space;
1585   size_t live_words;
1586   size_t fill_words;
1587   if (eden_live + eden_fillable >= region_size * 2) {
1588     id = eden_space_id;
1589     space = eden_space;
1590     live_words = eden_live;
1591     fill_words = eden_fillable;
1592   } else if (from_live + from_fillable >= region_size * 2) {
1593     id = from_space_id;
1594     space = from_space;
1595     live_words = from_live;
1596     fill_words = from_fillable;
1597   } else {
1598     return; // Give up.
1599   }
1600   assert(fill_words == 0 || fill_words >= min_fill_size, "sanity");
1601 
1602   if (live_words < region_size * 2) {
1603     // Fill from top() to end() w/live objects of mixed sizes.
1604     HeapWord* const fill_start = space->top();
1605     live_words += fill_words;
1606 
1607     space->set_top(fill_start + fill_words);
1608     if (ZapUnusedHeapArea) {
1609       space->set_top_for_allocations();
1610     }
1611 
1612     HeapWord* cur_addr = fill_start;
1613     while (fill_words > 0) {
1614       const size_t r = (size_t)os::random() % (region_size / 2) + min_fill_size;
1615       size_t cur_size = MIN2(align_object_size_(r), fill_words);
1616       if (fill_words - cur_size < min_fill_size) {
1617         cur_size = fill_words; // Avoid leaving a fragment too small to fill.
1618       }
1619 
1620       CollectedHeap::fill_with_object(cur_addr, cur_size);
1621       mark_bitmap()->mark_obj(cur_addr, cur_size);
1622       sd.add_obj(cur_addr, cur_size);
1623 
1624       cur_addr += cur_size;
1625       fill_words -= cur_size;
1626     }
1627 
1628     summarize_new_objects(id, fill_start);
1629   }
1630 
1631   max_compaction = false;
1632 
1633   // Manipulate the old gen so that it has room for about half of the live data
1634   // in the target young gen space (live_words / 2).
1635   id = old_space_id;
1636   space = _space_info[id].space();
1637   const size_t free_at_end = space->free_in_words();
1638   const size_t free_target = align_object_size(live_words / 2);
1639   const size_t dead = pointer_delta(space->top(), _space_info[id].new_top());
1640 
1641   if (free_at_end >= free_target + min_fill_size) {
1642     // Fill space above top() and set the dense prefix so everything survives.
1643     HeapWord* const fill_start = space->top();
1644     const size_t fill_size = free_at_end - free_target;
1645     space->set_top(space->top() + fill_size);
1646     if (ZapUnusedHeapArea) {
1647       space->set_top_for_allocations();
1648     }
1649     fill_with_live_objects(id, fill_start, fill_size);
1650     summarize_new_objects(id, fill_start);
1651     _space_info[id].set_dense_prefix(sd.region_align_down(space->top()));
1652   } else if (dead + free_at_end > free_target) {
1653     // Find a dense prefix that makes the right amount of space available.
1654     HeapWord* cur = sd.region_align_down(space->top());
1655     HeapWord* cur_destination = sd.addr_to_region_ptr(cur)->destination();
1656     size_t dead_to_right = pointer_delta(space->end(), cur_destination);
1657     while (dead_to_right < free_target) {
1658       cur -= region_size;
1659       cur_destination = sd.addr_to_region_ptr(cur)->destination();
1660       dead_to_right = pointer_delta(space->end(), cur_destination);
1661     }
1662     _space_info[id].set_dense_prefix(cur);
1663   }
1664 }
1665 #endif // #ifndef PRODUCT
1666 
1667 void PSParallelCompact::summarize_spaces_quick()
1668 {
1669   for (unsigned int i = 0; i < last_space_id; ++i) {
1670     const MutableSpace* space = _space_info[i].space();
1671     HeapWord** nta = _space_info[i].new_top_addr();
1672     bool result = _summary_data.summarize(_space_info[i].split_info(),
1673                                           space->bottom(), space->top(), NULL,
1674                                           space->bottom(), space->end(), nta);
1675     assert(result, "space must fit into itself");
1676     _space_info[i].set_dense_prefix(space->bottom());
1677   }
1678 
1679 #ifndef PRODUCT
1680   if (ParallelOldGCSplitALot) {
1681     provoke_split_fill_survivor(to_space_id);
1682   }
1683 #endif // #ifndef PRODUCT
1684 }
1685 
1686 void PSParallelCompact::fill_dense_prefix_end(SpaceId id)
1687 {
1688   HeapWord* const dense_prefix_end = dense_prefix(id);
1689   const RegionData* region = _summary_data.addr_to_region_ptr(dense_prefix_end);
1690   const idx_t dense_prefix_bit = _mark_bitmap.addr_to_bit(dense_prefix_end);
1691   if (dead_space_crosses_boundary(region, dense_prefix_bit)) {
1692     // Only enough dead space is filled so that any remaining dead space to the
1693     // left is larger than the minimum filler object.  (The remainder is filled
1694     // during the copy/update phase.)
1695     //
1696     // The size of the dead space to the right of the boundary is not a
1697     // concern, since compaction will be able to use whatever space is
1698     // available.
1699     //
1700     // Here '||' is the boundary, 'x' represents a don't care bit and a box
1701     // surrounds the space to be filled with an object.
1702     //
1703     // In the 32-bit VM, each bit represents two 32-bit words:
1704     //                              +---+
1705     // a) beg_bits:  ...  x   x   x | 0 | ||   0   x  x  ...
1706     //    end_bits:  ...  x   x   x | 0 | ||   0   x  x  ...
1707     //                              +---+
1708     //
1709     // In the 64-bit VM, each bit represents one 64-bit word:
1710     //                              +------------+
1711     // b) beg_bits:  ...  x   x   x | 0   ||   0 | x  x  ...
1712     //    end_bits:  ...  x   x   1 | 0   ||   0 | x  x  ...
1713     //                              +------------+
1714     //                          +-------+
1715     // c) beg_bits:  ...  x   x | 0   0 | ||   0   x  x  ...
1716     //    end_bits:  ...  x   1 | 0   0 | ||   0   x  x  ...
1717     //                          +-------+
1718     //                      +-----------+
1719     // d) beg_bits:  ...  x | 0   0   0 | ||   0   x  x  ...
1720     //    end_bits:  ...  1 | 0   0   0 | ||   0   x  x  ...
1721     //                      +-----------+
1722     //                          +-------+
1723     // e) beg_bits:  ...  0   0 | 0   0 | ||   0   x  x  ...
1724     //    end_bits:  ...  0   0 | 0   0 | ||   0   x  x  ...
1725     //                          +-------+
1726 
1727     // Initially assume case a, c or e will apply.
1728     size_t obj_len = CollectedHeap::min_fill_size();
1729     HeapWord* obj_beg = dense_prefix_end - obj_len;
1730 
1731 #ifdef  _LP64
1732     if (MinObjAlignment > 1) { // object alignment > heap word size
1733       // Cases a, c or e.
1734     } else if (_mark_bitmap.is_obj_end(dense_prefix_bit - 2)) {
1735       // Case b above.
1736       obj_beg = dense_prefix_end - 1;
1737     } else if (!_mark_bitmap.is_obj_end(dense_prefix_bit - 3) &&
1738                _mark_bitmap.is_obj_end(dense_prefix_bit - 4)) {
1739       // Case d above.
1740       obj_beg = dense_prefix_end - 3;
1741       obj_len = 3;
1742     }
1743 #endif  // #ifdef _LP64
1744 
1745     CollectedHeap::fill_with_object(obj_beg, obj_len);
1746     _mark_bitmap.mark_obj(obj_beg, obj_len);
1747     _summary_data.add_obj(obj_beg, obj_len);
1748     assert(start_array(id) != NULL, "sanity");
1749     start_array(id)->allocate_block(obj_beg);
1750   }
1751 }
1752 
1753 void
1754 PSParallelCompact::clear_source_region(HeapWord* beg_addr, HeapWord* end_addr)
1755 {
1756   RegionData* const beg_ptr = _summary_data.addr_to_region_ptr(beg_addr);
1757   HeapWord* const end_aligned_up = _summary_data.region_align_up(end_addr);
1758   RegionData* const end_ptr = _summary_data.addr_to_region_ptr(end_aligned_up);
1759   for (RegionData* cur = beg_ptr; cur < end_ptr; ++cur) {
1760     cur->set_source_region(0);
1761   }
1762 }
1763 
1764 void
1765 PSParallelCompact::summarize_space(SpaceId id, bool maximum_compaction)
1766 {
1767   assert(id < last_space_id, "id out of range");
1768   assert(_space_info[id].dense_prefix() == _space_info[id].space()->bottom() ||
1769          ParallelOldGCSplitALot && id == old_space_id,
1770          "should have been reset in summarize_spaces_quick()");
1771 
1772   const MutableSpace* space = _space_info[id].space();
1773   if (_space_info[id].new_top() != space->bottom()) {
1774     HeapWord* dense_prefix_end = compute_dense_prefix(id, maximum_compaction);
1775     _space_info[id].set_dense_prefix(dense_prefix_end);
1776 
1777 #ifndef PRODUCT
1778     if (TraceParallelOldGCDensePrefix) {
1779       print_dense_prefix_stats("ratio", id, maximum_compaction,
1780                                dense_prefix_end);
1781       HeapWord* addr = compute_dense_prefix_via_density(id, maximum_compaction);
1782       print_dense_prefix_stats("density", id, maximum_compaction, addr);
1783     }
1784 #endif  // #ifndef PRODUCT
1785 
1786     // Recompute the summary data, taking into account the dense prefix.  If
1787     // every last byte will be reclaimed, then the existing summary data which
1788     // compacts everything can be left in place.
1789     if (!maximum_compaction && dense_prefix_end != space->bottom()) {
1790       // If dead space crosses the dense prefix boundary, it is (at least
1791       // partially) filled with a dummy object, marked live and added to the
1792       // summary data.  This simplifies the copy/update phase and must be done
1793       // before the final locations of objects are determined, to prevent
1794       // leaving a fragment of dead space that is too small to fill.
1795       fill_dense_prefix_end(id);
1796 
1797       // Compute the destination of each Region, and thus each object.
1798       _summary_data.summarize_dense_prefix(space->bottom(), dense_prefix_end);
1799       _summary_data.summarize(_space_info[id].split_info(),
1800                               dense_prefix_end, space->top(), NULL,
1801                               dense_prefix_end, space->end(),
1802                               _space_info[id].new_top_addr());
1803     }
1804   }
1805 
1806   if (TraceParallelOldGCSummaryPhase) {
1807     const size_t region_size = ParallelCompactData::RegionSize;
1808     HeapWord* const dense_prefix_end = _space_info[id].dense_prefix();
1809     const size_t dp_region = _summary_data.addr_to_region_idx(dense_prefix_end);
1810     const size_t dp_words = pointer_delta(dense_prefix_end, space->bottom());
1811     HeapWord* const new_top = _space_info[id].new_top();
1812     const HeapWord* nt_aligned_up = _summary_data.region_align_up(new_top);
1813     const size_t cr_words = pointer_delta(nt_aligned_up, dense_prefix_end);
1814     tty->print_cr("id=%d cap=" SIZE_FORMAT " dp=" PTR_FORMAT " "
1815                   "dp_region=" SIZE_FORMAT " " "dp_count=" SIZE_FORMAT " "
1816                   "cr_count=" SIZE_FORMAT " " "nt=" PTR_FORMAT,
1817                   id, space->capacity_in_words(), dense_prefix_end,
1818                   dp_region, dp_words / region_size,
1819                   cr_words / region_size, new_top);
1820   }
1821 }
1822 
1823 #ifndef PRODUCT
1824 void PSParallelCompact::summary_phase_msg(SpaceId dst_space_id,
1825                                           HeapWord* dst_beg, HeapWord* dst_end,
1826                                           SpaceId src_space_id,
1827                                           HeapWord* src_beg, HeapWord* src_end)
1828 {
1829   if (TraceParallelOldGCSummaryPhase) {
1830     tty->print_cr("summarizing %d [%s] into %d [%s]:  "
1831                   "src=" PTR_FORMAT "-" PTR_FORMAT " "
1832                   SIZE_FORMAT "-" SIZE_FORMAT " "
1833                   "dst=" PTR_FORMAT "-" PTR_FORMAT " "
1834                   SIZE_FORMAT "-" SIZE_FORMAT,
1835                   src_space_id, space_names[src_space_id],
1836                   dst_space_id, space_names[dst_space_id],
1837                   src_beg, src_end,
1838                   _summary_data.addr_to_region_idx(src_beg),
1839                   _summary_data.addr_to_region_idx(src_end),
1840                   dst_beg, dst_end,
1841                   _summary_data.addr_to_region_idx(dst_beg),
1842                   _summary_data.addr_to_region_idx(dst_end));
1843   }
1844 }
1845 #endif  // #ifndef PRODUCT
1846 
1847 void PSParallelCompact::summary_phase(ParCompactionManager* cm,
1848                                       bool maximum_compaction)
1849 {
1850   GCTraceTime tm("summary phase", print_phases(), true, &_gc_timer);
1851   // trace("2");
1852 
1853 #ifdef  ASSERT
1854   if (TraceParallelOldGCMarkingPhase) {
1855     tty->print_cr("add_obj_count=" SIZE_FORMAT " "
1856                   "add_obj_bytes=" SIZE_FORMAT,
1857                   add_obj_count, add_obj_size * HeapWordSize);
1858     tty->print_cr("mark_bitmap_count=" SIZE_FORMAT " "
1859                   "mark_bitmap_bytes=" SIZE_FORMAT,
1860                   mark_bitmap_count, mark_bitmap_size * HeapWordSize);
1861   }
1862 #endif  // #ifdef ASSERT
1863 
1864   // Quick summarization of each space into itself, to see how much is live.
1865   summarize_spaces_quick();
1866 
1867   if (TraceParallelOldGCSummaryPhase) {
1868     tty->print_cr("summary_phase:  after summarizing each space to self");
1869     Universe::print();
1870     NOT_PRODUCT(print_region_ranges());
1871     if (Verbose) {
1872       NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info));
1873     }
1874   }
1875 
1876   // The amount of live data that will end up in old space (assuming it fits).
1877   size_t old_space_total_live = 0;
1878   for (unsigned int id = old_space_id; id < last_space_id; ++id) {
1879     old_space_total_live += pointer_delta(_space_info[id].new_top(),
1880                                           _space_info[id].space()->bottom());
1881   }
1882 
1883   MutableSpace* const old_space = _space_info[old_space_id].space();
1884   const size_t old_capacity = old_space->capacity_in_words();
1885   if (old_space_total_live > old_capacity) {
1886     // XXX - should also try to expand
1887     maximum_compaction = true;
1888   }
1889 #ifndef PRODUCT
1890   if (ParallelOldGCSplitALot && old_space_total_live < old_capacity) {
1891     provoke_split(maximum_compaction);
1892   }
1893 #endif // #ifndef PRODUCT
1894 
1895   // Old generations.
1896   summarize_space(old_space_id, maximum_compaction);
1897 
1898   // Summarize the remaining spaces in the young gen.  The initial target space
1899   // is the old gen.  If a space does not fit entirely into the target, then the
1900   // remainder is compacted into the space itself and that space becomes the new
1901   // target.
1902   SpaceId dst_space_id = old_space_id;
1903   HeapWord* dst_space_end = old_space->end();
1904   HeapWord** new_top_addr = _space_info[dst_space_id].new_top_addr();
1905   for (unsigned int id = eden_space_id; id < last_space_id; ++id) {
1906     const MutableSpace* space = _space_info[id].space();
1907     const size_t live = pointer_delta(_space_info[id].new_top(),
1908                                       space->bottom());
1909     const size_t available = pointer_delta(dst_space_end, *new_top_addr);
1910 
1911     NOT_PRODUCT(summary_phase_msg(dst_space_id, *new_top_addr, dst_space_end,
1912                                   SpaceId(id), space->bottom(), space->top());)
1913     if (live > 0 && live <= available) {
1914       // All the live data will fit.
1915       bool done = _summary_data.summarize(_space_info[id].split_info(),
1916                                           space->bottom(), space->top(),
1917                                           NULL,
1918                                           *new_top_addr, dst_space_end,
1919                                           new_top_addr);
1920       assert(done, "space must fit into old gen");
1921 
1922       // Reset the new_top value for the space.
1923       _space_info[id].set_new_top(space->bottom());
1924     } else if (live > 0) {
1925       // Attempt to fit part of the source space into the target space.
1926       HeapWord* next_src_addr = NULL;
1927       bool done = _summary_data.summarize(_space_info[id].split_info(),
1928                                           space->bottom(), space->top(),
1929                                           &next_src_addr,
1930                                           *new_top_addr, dst_space_end,
1931                                           new_top_addr);
1932       assert(!done, "space should not fit into old gen");
1933       assert(next_src_addr != NULL, "sanity");
1934 
1935       // The source space becomes the new target, so the remainder is compacted
1936       // within the space itself.
1937       dst_space_id = SpaceId(id);
1938       dst_space_end = space->end();
1939       new_top_addr = _space_info[id].new_top_addr();
1940       NOT_PRODUCT(summary_phase_msg(dst_space_id,
1941                                     space->bottom(), dst_space_end,
1942                                     SpaceId(id), next_src_addr, space->top());)
1943       done = _summary_data.summarize(_space_info[id].split_info(),
1944                                      next_src_addr, space->top(),
1945                                      NULL,
1946                                      space->bottom(), dst_space_end,
1947                                      new_top_addr);
1948       assert(done, "space must fit when compacted into itself");
1949       assert(*new_top_addr <= space->top(), "usage should not grow");
1950     }
1951   }
1952 
1953   if (TraceParallelOldGCSummaryPhase) {
1954     tty->print_cr("summary_phase:  after final summarization");
1955     Universe::print();
1956     NOT_PRODUCT(print_region_ranges());
1957     if (Verbose) {
1958       NOT_PRODUCT(print_generic_summary_data(_summary_data, _space_info));
1959     }
1960   }
1961 }
1962 
1963 // This method should contain all heap-specific policy for invoking a full
1964 // collection.  invoke_no_policy() will only attempt to compact the heap; it
1965 // will do nothing further.  If we need to bail out for policy reasons, scavenge
1966 // before full gc, or any other specialized behavior, it needs to be added here.
1967 //
1968 // Note that this method should only be called from the vm_thread while at a
1969 // safepoint.
1970 //
1971 // Note that the all_soft_refs_clear flag in the collector policy
1972 // may be true because this method can be called without intervening
1973 // activity.  For example when the heap space is tight and full measure
1974 // are being taken to free space.
1975 void PSParallelCompact::invoke(bool maximum_heap_compaction) {
1976   assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
1977   assert(Thread::current() == (Thread*)VMThread::vm_thread(),
1978          "should be in vm thread");
1979 
1980   ParallelScavengeHeap* heap = gc_heap();
1981   GCCause::Cause gc_cause = heap->gc_cause();
1982   assert(!heap->is_gc_active(), "not reentrant");
1983 
1984   PSAdaptiveSizePolicy* policy = heap->size_policy();
1985   IsGCActiveMark mark;
1986 
1987   if (ScavengeBeforeFullGC) {
1988     PSScavenge::invoke_no_policy();
1989   }
1990 
1991   const bool clear_all_soft_refs =
1992     heap->collector_policy()->should_clear_all_soft_refs();
1993 
1994   PSParallelCompact::invoke_no_policy(clear_all_soft_refs ||
1995                                       maximum_heap_compaction);
1996 }
1997 
1998 // This method contains no policy. You should probably
1999 // be calling invoke() instead.
2000 bool PSParallelCompact::invoke_no_policy(bool maximum_heap_compaction) {
2001   assert(SafepointSynchronize::is_at_safepoint(), "must be at a safepoint");
2002   assert(ref_processor() != NULL, "Sanity");
2003 
2004   if (GC_locker::check_active_before_gc()) {
2005     return false;
2006   }
2007 
2008   ParallelScavengeHeap* heap = gc_heap();
2009 
2010   _gc_timer.register_gc_start(os::elapsed_counter());
2011   _gc_tracer.report_gc_start(heap->gc_cause(), _gc_timer.gc_start());
2012 
2013   TimeStamp marking_start;
2014   TimeStamp compaction_start;
2015   TimeStamp collection_exit;
2016 
2017   GCCause::Cause gc_cause = heap->gc_cause();
2018   PSYoungGen* young_gen = heap->young_gen();
2019   PSOldGen* old_gen = heap->old_gen();
2020   PSAdaptiveSizePolicy* size_policy = heap->size_policy();
2021 
2022   // The scope of casr should end after code that can change
2023   // CollectorPolicy::_should_clear_all_soft_refs.
2024   ClearedAllSoftRefs casr(maximum_heap_compaction,
2025                           heap->collector_policy());
2026 
2027   if (ZapUnusedHeapArea) {
2028     // Save information needed to minimize mangling
2029     heap->record_gen_tops_before_GC();
2030   }
2031 
2032   heap->pre_full_gc_dump(&_gc_timer);
2033 
2034   _print_phases = PrintGCDetails && PrintParallelOldGCPhaseTimes;
2035 
2036   // Make sure data structures are sane, make the heap parsable, and do other
2037   // miscellaneous bookkeeping.
2038   PreGCValues pre_gc_values;
2039   pre_compact(&pre_gc_values);
2040 
2041   // Get the compaction manager reserved for the VM thread.
2042   ParCompactionManager* const vmthread_cm =
2043     ParCompactionManager::manager_array(gc_task_manager()->workers());
2044 
2045   // Place after pre_compact() where the number of invocations is incremented.
2046   AdaptiveSizePolicyOutput(size_policy, heap->total_collections());
2047 
2048   {
2049     ResourceMark rm;
2050     HandleMark hm;
2051 
2052     // Set the number of GC threads to be used in this collection
2053     gc_task_manager()->set_active_gang();
2054     gc_task_manager()->task_idle_workers();
2055     heap->set_par_threads(gc_task_manager()->active_workers());
2056 
2057     gclog_or_tty->date_stamp(PrintGC && PrintGCDateStamps);
2058     TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
2059     GCTraceTime t1(GCCauseString("Full GC", gc_cause), PrintGC, !PrintGCDetails, NULL);
2060     TraceCollectorStats tcs(counters());
2061     TraceMemoryManagerStats tms(true /* Full GC */,gc_cause);
2062 
2063     if (TraceGen1Time) accumulated_time()->start();
2064 
2065     // Let the size policy know we're starting
2066     size_policy->major_collection_begin();
2067 
2068     CodeCache::gc_prologue();
2069     Threads::gc_prologue();
2070 
2071     COMPILER2_PRESENT(DerivedPointerTable::clear());
2072 
2073     ref_processor()->enable_discovery(true /*verify_disabled*/, true /*verify_no_refs*/);
2074     ref_processor()->setup_policy(maximum_heap_compaction);
2075 
2076     bool marked_for_unloading = false;
2077 
2078     marking_start.update();
2079     marking_phase(vmthread_cm, maximum_heap_compaction, &_gc_tracer);
2080 
2081     bool max_on_system_gc = UseMaximumCompactionOnSystemGC
2082       && gc_cause == GCCause::_java_lang_system_gc;
2083     summary_phase(vmthread_cm, maximum_heap_compaction || max_on_system_gc);
2084 
2085     COMPILER2_PRESENT(assert(DerivedPointerTable::is_active(), "Sanity"));
2086     COMPILER2_PRESENT(DerivedPointerTable::set_active(false));
2087 
2088     // adjust_roots() updates Universe::_intArrayKlassObj which is
2089     // needed by the compaction for filling holes in the dense prefix.
2090     adjust_roots();
2091 
2092     compaction_start.update();
2093     compact();
2094 
2095     // Reset the mark bitmap, summary data, and do other bookkeeping.  Must be
2096     // done before resizing.
2097     post_compact();
2098 
2099     // Let the size policy know we're done
2100     size_policy->major_collection_end(old_gen->used_in_bytes(), gc_cause);
2101 
2102     if (UseAdaptiveSizePolicy) {
2103       if (PrintAdaptiveSizePolicy) {
2104         gclog_or_tty->print("AdaptiveSizeStart: ");
2105         gclog_or_tty->stamp();
2106         gclog_or_tty->print_cr(" collection: %d ",
2107                        heap->total_collections());
2108         if (Verbose) {
2109           gclog_or_tty->print("old_gen_capacity: %d young_gen_capacity: %d",
2110             old_gen->capacity_in_bytes(), young_gen->capacity_in_bytes());
2111         }
2112       }
2113 
2114       // Don't check if the size_policy is ready here.  Let
2115       // the size_policy check that internally.
2116       if (UseAdaptiveGenerationSizePolicyAtMajorCollection &&
2117           ((gc_cause != GCCause::_java_lang_system_gc) ||
2118             UseAdaptiveSizePolicyWithSystemGC)) {
2119         // Swap the survivor spaces if from_space is empty
2120         if (UseAdaptiveSizePolicyResizeYoungGenAtMajorCollection &&
2121             young_gen->from_space()->is_empty()) {
2122           young_gen->from_space()->clear(SpaceDecorator::Mangle);
2123           young_gen->swap_spaces();
2124         }
2125 
2126         // Calculate optimal free space amounts
2127         assert(young_gen->max_size() >
2128           young_gen->from_space()->capacity_in_bytes() +
2129           young_gen->to_space()->capacity_in_bytes(),
2130           "Sizes of space in young gen are out-of-bounds");
2131 
2132         size_t young_live = young_gen->used_in_bytes();
2133         size_t eden_live = young_gen->eden_space()->used_in_bytes();
2134         size_t old_live = old_gen->used_in_bytes();
2135         size_t cur_eden = young_gen->eden_space()->capacity_in_bytes();
2136         size_t max_old_gen_size = old_gen->max_gen_size();
2137         size_t max_eden_size = young_gen->max_size() -
2138           young_gen->from_space()->capacity_in_bytes() -
2139           young_gen->to_space()->capacity_in_bytes();
2140 
2141         // Used for diagnostics
2142         size_policy->clear_generation_free_space_flags();
2143 
2144         size_policy->compute_generations_free_space(young_live,
2145                                                     eden_live,
2146                                                     old_live,
2147                                                     cur_eden,
2148                                                     max_old_gen_size,
2149                                                     max_eden_size,
2150                                                     true /* full gc*/);
2151 
2152         size_policy->check_gc_overhead_limit(young_live,
2153                                              eden_live,
2154                                              max_old_gen_size,
2155                                              max_eden_size,
2156                                              true /* full gc*/,
2157                                              gc_cause,
2158                                              heap->collector_policy());
2159 
2160         size_policy->decay_supplemental_growth(true /* full gc*/);
2161 
2162         heap->resize_old_gen(
2163           size_policy->calculated_old_free_size_in_bytes());
2164 
2165         if (UseAdaptiveSizePolicyResizeYoungGenAtMajorCollection) {
2166           heap->resize_young_gen(size_policy->calculated_eden_size_in_bytes(),
2167                                  size_policy->calculated_survivor_size_in_bytes());
2168         }
2169       }
2170       if (PrintAdaptiveSizePolicy) {
2171         gclog_or_tty->print_cr("AdaptiveSizeStop: collection: %d ",
2172                        heap->total_collections());
2173       }
2174     }
2175 
2176     if (UsePerfData) {
2177       PSGCAdaptivePolicyCounters* const counters = heap->gc_policy_counters();
2178       counters->update_counters();
2179       counters->update_old_capacity(old_gen->capacity_in_bytes());
2180       counters->update_young_capacity(young_gen->capacity_in_bytes());
2181     }
2182 
2183     heap->resize_all_tlabs();
2184 
2185     // Resize the metaspace capactiy after a collection
2186     MetaspaceGC::compute_new_size();
2187 
2188     if (TraceGen1Time) accumulated_time()->stop();
2189 
2190     if (PrintGC) {
2191       if (PrintGCDetails) {
2192         // No GC timestamp here.  This is after GC so it would be confusing.
2193         young_gen->print_used_change(pre_gc_values.young_gen_used());
2194         old_gen->print_used_change(pre_gc_values.old_gen_used());
2195         heap->print_heap_change(pre_gc_values.heap_used());
2196         MetaspaceAux::print_metaspace_change(pre_gc_values.metadata_used());
2197       } else {
2198         heap->print_heap_change(pre_gc_values.heap_used());
2199       }
2200     }
2201 
2202     // Track memory usage and detect low memory
2203     MemoryService::track_memory_usage();
2204     heap->update_counters();
2205     gc_task_manager()->release_idle_workers();
2206   }
2207 
2208 #ifdef ASSERT
2209   for (size_t i = 0; i < ParallelGCThreads + 1; ++i) {
2210     ParCompactionManager* const cm =
2211       ParCompactionManager::manager_array(int(i));
2212     assert(cm->marking_stack()->is_empty(),       "should be empty");
2213     assert(ParCompactionManager::region_list(int(i))->is_empty(), "should be empty");
2214   }
2215 #endif // ASSERT
2216 
2217   if (VerifyAfterGC && heap->total_collections() >= VerifyGCStartAt) {
2218     HandleMark hm;  // Discard invalid handles created during verification
2219     Universe::verify(" VerifyAfterGC:");
2220   }
2221 
2222   // Re-verify object start arrays
2223   if (VerifyObjectStartArray &&
2224       VerifyAfterGC) {
2225     old_gen->verify_object_start_array();
2226   }
2227 
2228   if (ZapUnusedHeapArea) {
2229     old_gen->object_space()->check_mangled_unused_area_complete();
2230   }
2231 
2232   NOT_PRODUCT(ref_processor()->verify_no_references_recorded());
2233 
2234   collection_exit.update();
2235 
2236   heap->print_heap_after_gc();
2237   heap->trace_heap_after_gc(&_gc_tracer);
2238 
2239   if (PrintGCTaskTimeStamps) {
2240     gclog_or_tty->print_cr("VM-Thread " INT64_FORMAT " " INT64_FORMAT " "
2241                            INT64_FORMAT,
2242                            marking_start.ticks(), compaction_start.ticks(),
2243                            collection_exit.ticks());
2244     gc_task_manager()->print_task_time_stamps();
2245   }
2246 
2247   heap->post_full_gc_dump(&_gc_timer);
2248 
2249 #ifdef TRACESPINNING
2250   ParallelTaskTerminator::print_termination_counts();
2251 #endif
2252 
2253   _gc_timer.register_gc_end(os::elapsed_counter());
2254 
2255   _gc_tracer.report_dense_prefix(dense_prefix(old_space_id));
2256   _gc_tracer.report_gc_end(_gc_timer.gc_end(), _gc_timer.time_partitions());
2257 
2258   return true;
2259 }
2260 
2261 bool PSParallelCompact::absorb_live_data_from_eden(PSAdaptiveSizePolicy* size_policy,
2262                                              PSYoungGen* young_gen,
2263                                              PSOldGen* old_gen) {
2264   MutableSpace* const eden_space = young_gen->eden_space();
2265   assert(!eden_space->is_empty(), "eden must be non-empty");
2266   assert(young_gen->virtual_space()->alignment() ==
2267          old_gen->virtual_space()->alignment(), "alignments do not match");
2268 
2269   if (!(UseAdaptiveSizePolicy && UseAdaptiveGCBoundary)) {
2270     return false;
2271   }
2272 
2273   // Both generations must be completely committed.
2274   if (young_gen->virtual_space()->uncommitted_size() != 0) {
2275     return false;
2276   }
2277   if (old_gen->virtual_space()->uncommitted_size() != 0) {
2278     return false;
2279   }
2280 
2281   // Figure out how much to take from eden.  Include the average amount promoted
2282   // in the total; otherwise the next young gen GC will simply bail out to a
2283   // full GC.
2284   const size_t alignment = old_gen->virtual_space()->alignment();
2285   const size_t eden_used = eden_space->used_in_bytes();
2286   const size_t promoted = (size_t)size_policy->avg_promoted()->padded_average();
2287   const size_t absorb_size = align_size_up(eden_used + promoted, alignment);
2288   const size_t eden_capacity = eden_space->capacity_in_bytes();
2289 
2290   if (absorb_size >= eden_capacity) {
2291     return false; // Must leave some space in eden.
2292   }
2293 
2294   const size_t new_young_size = young_gen->capacity_in_bytes() - absorb_size;
2295   if (new_young_size < young_gen->min_gen_size()) {
2296     return false; // Respect young gen minimum size.
2297   }
2298 
2299   if (TraceAdaptiveGCBoundary && Verbose) {
2300     gclog_or_tty->print(" absorbing " SIZE_FORMAT "K:  "
2301                         "eden " SIZE_FORMAT "K->" SIZE_FORMAT "K "
2302                         "from " SIZE_FORMAT "K, to " SIZE_FORMAT "K "
2303                         "young_gen " SIZE_FORMAT "K->" SIZE_FORMAT "K ",
2304                         absorb_size / K,
2305                         eden_capacity / K, (eden_capacity - absorb_size) / K,
2306                         young_gen->from_space()->used_in_bytes() / K,
2307                         young_gen->to_space()->used_in_bytes() / K,
2308                         young_gen->capacity_in_bytes() / K, new_young_size / K);
2309   }
2310 
2311   // Fill the unused part of the old gen.
2312   MutableSpace* const old_space = old_gen->object_space();
2313   HeapWord* const unused_start = old_space->top();
2314   size_t const unused_words = pointer_delta(old_space->end(), unused_start);
2315 
2316   if (unused_words > 0) {
2317     if (unused_words < CollectedHeap::min_fill_size()) {
2318       return false;  // If the old gen cannot be filled, must give up.
2319     }
2320     CollectedHeap::fill_with_objects(unused_start, unused_words);
2321   }
2322 
2323   // Take the live data from eden and set both top and end in the old gen to
2324   // eden top.  (Need to set end because reset_after_change() mangles the region
2325   // from end to virtual_space->high() in debug builds).
2326   HeapWord* const new_top = eden_space->top();
2327   old_gen->virtual_space()->expand_into(young_gen->virtual_space(),
2328                                         absorb_size);
2329   young_gen->reset_after_change();
2330   old_space->set_top(new_top);
2331   old_space->set_end(new_top);
2332   old_gen->reset_after_change();
2333 
2334   // Update the object start array for the filler object and the data from eden.
2335   ObjectStartArray* const start_array = old_gen->start_array();
2336   for (HeapWord* p = unused_start; p < new_top; p += oop(p)->size()) {
2337     start_array->allocate_block(p);
2338   }
2339 
2340   // Could update the promoted average here, but it is not typically updated at
2341   // full GCs and the value to use is unclear.  Something like
2342   //
2343   // cur_promoted_avg + absorb_size / number_of_scavenges_since_last_full_gc.
2344 
2345   size_policy->set_bytes_absorbed_from_eden(absorb_size);
2346   return true;
2347 }
2348 
2349 GCTaskManager* const PSParallelCompact::gc_task_manager() {
2350   assert(ParallelScavengeHeap::gc_task_manager() != NULL,
2351     "shouldn't return NULL");
2352   return ParallelScavengeHeap::gc_task_manager();
2353 }
2354 
2355 void PSParallelCompact::marking_phase(ParCompactionManager* cm,
2356                                       bool maximum_heap_compaction,
2357                                       ParallelOldTracer *gc_tracer) {
2358   // Recursively traverse all live objects and mark them
2359   GCTraceTime tm("marking phase", print_phases(), true, &_gc_timer);
2360 
2361   ParallelScavengeHeap* heap = gc_heap();
2362   uint parallel_gc_threads = heap->gc_task_manager()->workers();
2363   uint active_gc_threads = heap->gc_task_manager()->active_workers();
2364   TaskQueueSetSuper* qset = ParCompactionManager::region_array();
2365   ParallelTaskTerminator terminator(active_gc_threads, qset);
2366 
2367   PSParallelCompact::MarkAndPushClosure mark_and_push_closure(cm);
2368   PSParallelCompact::FollowStackClosure follow_stack_closure(cm);
2369 
2370   // Need new claim bits before marking starts.
2371   ClassLoaderDataGraph::clear_claimed_marks();
2372 
2373   {
2374     GCTraceTime tm_m("par mark", print_phases(), true, &_gc_timer);
2375 
2376     ParallelScavengeHeap::ParStrongRootsScope psrs;
2377 
2378     GCTaskQueue* q = GCTaskQueue::create();
2379 
2380     q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::universe));
2381     q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jni_handles));
2382     // We scan the thread roots in parallel
2383     Threads::create_thread_roots_marking_tasks(q);
2384     q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::object_synchronizer));
2385     q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::flat_profiler));
2386     q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::management));
2387     q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::system_dictionary));
2388     q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::class_loader_data));
2389     q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jvmti));
2390     q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::code_cache));
2391 
2392     if (active_gc_threads > 1) {
2393       for (uint j = 0; j < active_gc_threads; j++) {
2394         q->enqueue(new StealMarkingTask(&terminator));
2395       }
2396     }
2397 
2398     gc_task_manager()->execute_and_wait(q);
2399   }
2400 
2401   // Process reference objects found during marking
2402   {
2403     GCTraceTime tm_r("reference processing", print_phases(), true, &_gc_timer);
2404 
2405     ReferenceProcessorStats stats;
2406     if (ref_processor()->processing_is_mt()) {
2407       RefProcTaskExecutor task_executor;
2408       stats = ref_processor()->process_discovered_references(
2409         is_alive_closure(), &mark_and_push_closure, &follow_stack_closure,
2410         &task_executor, &_gc_timer);
2411     } else {
2412       stats = ref_processor()->process_discovered_references(
2413         is_alive_closure(), &mark_and_push_closure, &follow_stack_closure, NULL,
2414         &_gc_timer);
2415     }
2416 
2417     gc_tracer->report_gc_reference_stats(stats);
2418   }
2419 
2420   GCTraceTime tm_c("class unloading", print_phases(), true, &_gc_timer);
2421 
2422   // This is the point where the entire marking should have completed.
2423   assert(cm->marking_stacks_empty(), "Marking should have completed");
2424 
2425   // Follow system dictionary roots and unload classes.
2426   bool purged_class = SystemDictionary::do_unloading(is_alive_closure());
2427 
2428   // Unload nmethods.
2429   CodeCache::do_unloading(is_alive_closure(), purged_class);
2430 
2431   // Prune dead klasses from subklass/sibling/implementor lists.
2432   Klass::clean_weak_klass_links(is_alive_closure());
2433 
2434   // Delete entries for dead interned strings.
2435   StringTable::unlink(is_alive_closure());
2436 
2437   // Clean up unreferenced symbols in symbol table.
2438   SymbolTable::unlink();
2439   _gc_tracer.report_object_count_after_gc(is_alive_closure());
2440 }
2441 
2442 void PSParallelCompact::follow_klass(ParCompactionManager* cm, Klass* klass) {
2443   ClassLoaderData* cld = klass->class_loader_data();
2444   // The actual processing of the klass is done when we
2445   // traverse the list of Klasses in the class loader data.
2446   PSParallelCompact::follow_class_loader(cm, cld);
2447 }
2448 
2449 void PSParallelCompact::adjust_klass(ParCompactionManager* cm, Klass* klass) {
2450   ClassLoaderData* cld = klass->class_loader_data();
2451   // The actual processing of the klass is done when we
2452   // traverse the list of Klasses in the class loader data.
2453   PSParallelCompact::adjust_class_loader(cm, cld);
2454 }
2455 
2456 void PSParallelCompact::follow_class_loader(ParCompactionManager* cm,
2457                                             ClassLoaderData* cld) {
2458   PSParallelCompact::MarkAndPushClosure mark_and_push_closure(cm);
2459   PSParallelCompact::FollowKlassClosure follow_klass_closure(&mark_and_push_closure);
2460 
2461   cld->oops_do(&mark_and_push_closure, &follow_klass_closure, true);
2462 }
2463 
2464 void PSParallelCompact::adjust_class_loader(ParCompactionManager* cm,
2465                                             ClassLoaderData* cld) {
2466   cld->oops_do(PSParallelCompact::adjust_pointer_closure(),
2467                PSParallelCompact::adjust_klass_closure(),
2468                true);
2469 }
2470 
2471 // This should be moved to the shared markSweep code!
2472 class PSAlwaysTrueClosure: public BoolObjectClosure {
2473 public:
2474   bool do_object_b(oop p) { return true; }
2475 };
2476 static PSAlwaysTrueClosure always_true;
2477 
2478 void PSParallelCompact::adjust_roots() {
2479   // Adjust the pointers to reflect the new locations
2480   GCTraceTime tm("adjust roots", print_phases(), true, &_gc_timer);
2481 
2482   // Need new claim bits when tracing through and adjusting pointers.
2483   ClassLoaderDataGraph::clear_claimed_marks();
2484 
2485   // General strong roots.
2486   Universe::oops_do(adjust_pointer_closure());
2487   JNIHandles::oops_do(adjust_pointer_closure());   // Global (strong) JNI handles
2488   CLDToOopClosure adjust_from_cld(adjust_pointer_closure());
2489   Threads::oops_do(adjust_pointer_closure(), &adjust_from_cld, NULL);
2490   ObjectSynchronizer::oops_do(adjust_pointer_closure());
2491   FlatProfiler::oops_do(adjust_pointer_closure());
2492   Management::oops_do(adjust_pointer_closure());
2493   JvmtiExport::oops_do(adjust_pointer_closure());
2494   // SO_AllClasses
2495   SystemDictionary::oops_do(adjust_pointer_closure());
2496   ClassLoaderDataGraph::oops_do(adjust_pointer_closure(), adjust_klass_closure(), true);
2497 
2498   // Now adjust pointers in remaining weak roots.  (All of which should
2499   // have been cleared if they pointed to non-surviving objects.)
2500   // Global (weak) JNI handles
2501   JNIHandles::weak_oops_do(&always_true, adjust_pointer_closure());
2502 
2503   CodeCache::oops_do(adjust_pointer_closure());
2504   StringTable::oops_do(adjust_pointer_closure());
2505   ref_processor()->weak_oops_do(adjust_pointer_closure());
2506   // Roots were visited so references into the young gen in roots
2507   // may have been scanned.  Process them also.
2508   // Should the reference processor have a span that excludes
2509   // young gen objects?
2510   PSScavenge::reference_processor()->weak_oops_do(adjust_pointer_closure());
2511 }
2512 
2513 void PSParallelCompact::enqueue_region_draining_tasks(GCTaskQueue* q,
2514                                                       uint parallel_gc_threads)
2515 {
2516   GCTraceTime tm("drain task setup", print_phases(), true, &_gc_timer);
2517 
2518   // Find the threads that are active
2519   unsigned int which = 0;
2520 
2521   const uint task_count = MAX2(parallel_gc_threads, 1U);
2522   for (uint j = 0; j < task_count; j++) {
2523     q->enqueue(new DrainStacksCompactionTask(j));
2524     ParCompactionManager::verify_region_list_empty(j);
2525     // Set the region stacks variables to "no" region stack values
2526     // so that they will be recognized and needing a region stack
2527     // in the stealing tasks if they do not get one by executing
2528     // a draining stack.
2529     ParCompactionManager* cm = ParCompactionManager::manager_array(j);
2530     cm->set_region_stack(NULL);
2531     cm->set_region_stack_index((uint)max_uintx);
2532   }
2533   ParCompactionManager::reset_recycled_stack_index();
2534 
2535   // Find all regions that are available (can be filled immediately) and
2536   // distribute them to the thread stacks.  The iteration is done in reverse
2537   // order (high to low) so the regions will be removed in ascending order.
2538 
2539   const ParallelCompactData& sd = PSParallelCompact::summary_data();
2540 
2541   size_t fillable_regions = 0;   // A count for diagnostic purposes.
2542   // A region index which corresponds to the tasks created above.
2543   // "which" must be 0 <= which < task_count
2544 
2545   which = 0;
2546   // id + 1 is used to test termination so unsigned  can
2547   // be used with an old_space_id == 0.
2548   for (unsigned int id = to_space_id; id + 1 > old_space_id; --id) {
2549     SpaceInfo* const space_info = _space_info + id;
2550     MutableSpace* const space = space_info->space();
2551     HeapWord* const new_top = space_info->new_top();
2552 
2553     const size_t beg_region = sd.addr_to_region_idx(space_info->dense_prefix());
2554     const size_t end_region =
2555       sd.addr_to_region_idx(sd.region_align_up(new_top));
2556 
2557     for (size_t cur = end_region - 1; cur + 1 > beg_region; --cur) {
2558       if (sd.region(cur)->claim_unsafe()) {
2559         ParCompactionManager::region_list_push(which, cur);
2560 
2561         if (TraceParallelOldGCCompactionPhase && Verbose) {
2562           const size_t count_mod_8 = fillable_regions & 7;
2563           if (count_mod_8 == 0) gclog_or_tty->print("fillable: ");
2564           gclog_or_tty->print(" " SIZE_FORMAT_W(7), cur);
2565           if (count_mod_8 == 7) gclog_or_tty->cr();
2566         }
2567 
2568         NOT_PRODUCT(++fillable_regions;)
2569 
2570         // Assign regions to tasks in round-robin fashion.
2571         if (++which == task_count) {
2572           assert(which <= parallel_gc_threads,
2573             "Inconsistent number of workers");
2574           which = 0;
2575         }
2576       }
2577     }
2578   }
2579 
2580   if (TraceParallelOldGCCompactionPhase) {
2581     if (Verbose && (fillable_regions & 7) != 0) gclog_or_tty->cr();
2582     gclog_or_tty->print_cr("%u initially fillable regions", fillable_regions);
2583   }
2584 }
2585 
2586 #define PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING 4
2587 
2588 void PSParallelCompact::enqueue_dense_prefix_tasks(GCTaskQueue* q,
2589                                                     uint parallel_gc_threads) {
2590   GCTraceTime tm("dense prefix task setup", print_phases(), true, &_gc_timer);
2591 
2592   ParallelCompactData& sd = PSParallelCompact::summary_data();
2593 
2594   // Iterate over all the spaces adding tasks for updating
2595   // regions in the dense prefix.  Assume that 1 gc thread
2596   // will work on opening the gaps and the remaining gc threads
2597   // will work on the dense prefix.
2598   unsigned int space_id;
2599   for (space_id = old_space_id; space_id < last_space_id; ++ space_id) {
2600     HeapWord* const dense_prefix_end = _space_info[space_id].dense_prefix();
2601     const MutableSpace* const space = _space_info[space_id].space();
2602 
2603     if (dense_prefix_end == space->bottom()) {
2604       // There is no dense prefix for this space.
2605       continue;
2606     }
2607 
2608     // The dense prefix is before this region.
2609     size_t region_index_end_dense_prefix =
2610         sd.addr_to_region_idx(dense_prefix_end);
2611     RegionData* const dense_prefix_cp =
2612       sd.region(region_index_end_dense_prefix);
2613     assert(dense_prefix_end == space->end() ||
2614            dense_prefix_cp->available() ||
2615            dense_prefix_cp->claimed(),
2616            "The region after the dense prefix should always be ready to fill");
2617 
2618     size_t region_index_start = sd.addr_to_region_idx(space->bottom());
2619 
2620     // Is there dense prefix work?
2621     size_t total_dense_prefix_regions =
2622       region_index_end_dense_prefix - region_index_start;
2623     // How many regions of the dense prefix should be given to
2624     // each thread?
2625     if (total_dense_prefix_regions > 0) {
2626       uint tasks_for_dense_prefix = 1;
2627       if (total_dense_prefix_regions <=
2628           (parallel_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING)) {
2629         // Don't over partition.  This assumes that
2630         // PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING is a small integer value
2631         // so there are not many regions to process.
2632         tasks_for_dense_prefix = parallel_gc_threads;
2633       } else {
2634         // Over partition
2635         tasks_for_dense_prefix = parallel_gc_threads *
2636           PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING;
2637       }
2638       size_t regions_per_thread = total_dense_prefix_regions /
2639         tasks_for_dense_prefix;
2640       // Give each thread at least 1 region.
2641       if (regions_per_thread == 0) {
2642         regions_per_thread = 1;
2643       }
2644 
2645       for (uint k = 0; k < tasks_for_dense_prefix; k++) {
2646         if (region_index_start >= region_index_end_dense_prefix) {
2647           break;
2648         }
2649         // region_index_end is not processed
2650         size_t region_index_end = MIN2(region_index_start + regions_per_thread,
2651                                        region_index_end_dense_prefix);
2652         q->enqueue(new UpdateDensePrefixTask(SpaceId(space_id),
2653                                              region_index_start,
2654                                              region_index_end));
2655         region_index_start = region_index_end;
2656       }
2657     }
2658     // This gets any part of the dense prefix that did not
2659     // fit evenly.
2660     if (region_index_start < region_index_end_dense_prefix) {
2661       q->enqueue(new UpdateDensePrefixTask(SpaceId(space_id),
2662                                            region_index_start,
2663                                            region_index_end_dense_prefix));
2664     }
2665   }
2666 }
2667 
2668 void PSParallelCompact::enqueue_region_stealing_tasks(
2669                                      GCTaskQueue* q,
2670                                      ParallelTaskTerminator* terminator_ptr,
2671                                      uint parallel_gc_threads) {
2672   GCTraceTime tm("steal task setup", print_phases(), true, &_gc_timer);
2673 
2674   // Once a thread has drained it's stack, it should try to steal regions from
2675   // other threads.
2676   if (parallel_gc_threads > 1) {
2677     for (uint j = 0; j < parallel_gc_threads; j++) {
2678       q->enqueue(new StealRegionCompactionTask(terminator_ptr));
2679     }
2680   }
2681 }
2682 
2683 #ifdef ASSERT
2684 // Write a histogram of the number of times the block table was filled for a
2685 // region.
2686 void PSParallelCompact::write_block_fill_histogram(outputStream* const out)
2687 {
2688   if (!TraceParallelOldGCCompactionPhase) return;
2689 
2690   typedef ParallelCompactData::RegionData rd_t;
2691   ParallelCompactData& sd = summary_data();
2692 
2693   for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2694     MutableSpace* const spc = _space_info[id].space();
2695     if (spc->bottom() != spc->top()) {
2696       const rd_t* const beg = sd.addr_to_region_ptr(spc->bottom());
2697       HeapWord* const top_aligned_up = sd.region_align_up(spc->top());
2698       const rd_t* const end = sd.addr_to_region_ptr(top_aligned_up);
2699 
2700       size_t histo[5] = { 0, 0, 0, 0, 0 };
2701       const size_t histo_len = sizeof(histo) / sizeof(size_t);
2702       const size_t region_cnt = pointer_delta(end, beg, sizeof(rd_t));
2703 
2704       for (const rd_t* cur = beg; cur < end; ++cur) {
2705         ++histo[MIN2(cur->blocks_filled_count(), histo_len - 1)];
2706       }
2707       out->print("%u %-4s" SIZE_FORMAT_W(5), id, space_names[id], region_cnt);
2708       for (size_t i = 0; i < histo_len; ++i) {
2709         out->print(" " SIZE_FORMAT_W(5) " %5.1f%%",
2710                    histo[i], 100.0 * histo[i] / region_cnt);
2711       }
2712       out->cr();
2713     }
2714   }
2715 }
2716 #endif // #ifdef ASSERT
2717 
2718 void PSParallelCompact::compact() {
2719   // trace("5");
2720   GCTraceTime tm("compaction phase", print_phases(), true, &_gc_timer);
2721 
2722   ParallelScavengeHeap* heap = (ParallelScavengeHeap*)Universe::heap();
2723   assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity");
2724   PSOldGen* old_gen = heap->old_gen();
2725   old_gen->start_array()->reset();
2726   uint parallel_gc_threads = heap->gc_task_manager()->workers();
2727   uint active_gc_threads = heap->gc_task_manager()->active_workers();
2728   TaskQueueSetSuper* qset = ParCompactionManager::region_array();
2729   ParallelTaskTerminator terminator(active_gc_threads, qset);
2730 
2731   GCTaskQueue* q = GCTaskQueue::create();
2732   enqueue_region_draining_tasks(q, active_gc_threads);
2733   enqueue_dense_prefix_tasks(q, active_gc_threads);
2734   enqueue_region_stealing_tasks(q, &terminator, active_gc_threads);
2735 
2736   {
2737     GCTraceTime tm_pc("par compact", print_phases(), true, &_gc_timer);
2738 
2739     gc_task_manager()->execute_and_wait(q);
2740 
2741 #ifdef  ASSERT
2742     // Verify that all regions have been processed before the deferred updates.
2743     for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2744       verify_complete(SpaceId(id));
2745     }
2746 #endif
2747   }
2748 
2749   {
2750     // Update the deferred objects, if any.  Any compaction manager can be used.
2751     GCTraceTime tm_du("deferred updates", print_phases(), true, &_gc_timer);
2752     ParCompactionManager* cm = ParCompactionManager::manager_array(0);
2753     for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2754       update_deferred_objects(cm, SpaceId(id));
2755     }
2756   }
2757 
2758   DEBUG_ONLY(write_block_fill_histogram(gclog_or_tty));
2759 }
2760 
2761 #ifdef  ASSERT
2762 void PSParallelCompact::verify_complete(SpaceId space_id) {
2763   // All Regions between space bottom() to new_top() should be marked as filled
2764   // and all Regions between new_top() and top() should be available (i.e.,
2765   // should have been emptied).
2766   ParallelCompactData& sd = summary_data();
2767   SpaceInfo si = _space_info[space_id];
2768   HeapWord* new_top_addr = sd.region_align_up(si.new_top());
2769   HeapWord* old_top_addr = sd.region_align_up(si.space()->top());
2770   const size_t beg_region = sd.addr_to_region_idx(si.space()->bottom());
2771   const size_t new_top_region = sd.addr_to_region_idx(new_top_addr);
2772   const size_t old_top_region = sd.addr_to_region_idx(old_top_addr);
2773 
2774   bool issued_a_warning = false;
2775 
2776   size_t cur_region;
2777   for (cur_region = beg_region; cur_region < new_top_region; ++cur_region) {
2778     const RegionData* const c = sd.region(cur_region);
2779     if (!c->completed()) {
2780       warning("region " SIZE_FORMAT " not filled:  "
2781               "destination_count=" SIZE_FORMAT,
2782               cur_region, c->destination_count());
2783       issued_a_warning = true;
2784     }
2785   }
2786 
2787   for (cur_region = new_top_region; cur_region < old_top_region; ++cur_region) {
2788     const RegionData* const c = sd.region(cur_region);
2789     if (!c->available()) {
2790       warning("region " SIZE_FORMAT " not empty:   "
2791               "destination_count=" SIZE_FORMAT,
2792               cur_region, c->destination_count());
2793       issued_a_warning = true;
2794     }
2795   }
2796 
2797   if (issued_a_warning) {
2798     print_region_ranges();
2799   }
2800 }
2801 #endif  // #ifdef ASSERT
2802 
2803 // Update interior oops in the ranges of regions [beg_region, end_region).
2804 void
2805 PSParallelCompact::update_and_deadwood_in_dense_prefix(ParCompactionManager* cm,
2806                                                        SpaceId space_id,
2807                                                        size_t beg_region,
2808                                                        size_t end_region) {
2809   ParallelCompactData& sd = summary_data();
2810   ParMarkBitMap* const mbm = mark_bitmap();
2811 
2812   HeapWord* beg_addr = sd.region_to_addr(beg_region);
2813   HeapWord* const end_addr = sd.region_to_addr(end_region);
2814   assert(beg_region <= end_region, "bad region range");
2815   assert(end_addr <= dense_prefix(space_id), "not in the dense prefix");
2816 
2817 #ifdef  ASSERT
2818   // Claim the regions to avoid triggering an assert when they are marked as
2819   // filled.
2820   for (size_t claim_region = beg_region; claim_region < end_region; ++claim_region) {
2821     assert(sd.region(claim_region)->claim_unsafe(), "claim() failed");
2822   }
2823 #endif  // #ifdef ASSERT
2824 
2825   if (beg_addr != space(space_id)->bottom()) {
2826     // Find the first live object or block of dead space that *starts* in this
2827     // range of regions.  If a partial object crosses onto the region, skip it;
2828     // it will be marked for 'deferred update' when the object head is
2829     // processed.  If dead space crosses onto the region, it is also skipped; it
2830     // will be filled when the prior region is processed.  If neither of those
2831     // apply, the first word in the region is the start of a live object or dead
2832     // space.
2833     assert(beg_addr > space(space_id)->bottom(), "sanity");
2834     const RegionData* const cp = sd.region(beg_region);
2835     if (cp->partial_obj_size() != 0) {
2836       beg_addr = sd.partial_obj_end(beg_region);
2837     } else if (dead_space_crosses_boundary(cp, mbm->addr_to_bit(beg_addr))) {
2838       beg_addr = mbm->find_obj_beg(beg_addr, end_addr);
2839     }
2840   }
2841 
2842   if (beg_addr < end_addr) {
2843     // A live object or block of dead space starts in this range of Regions.
2844      HeapWord* const dense_prefix_end = dense_prefix(space_id);
2845 
2846     // Create closures and iterate.
2847     UpdateOnlyClosure update_closure(mbm, cm, space_id);
2848     FillClosure fill_closure(cm, space_id);
2849     ParMarkBitMap::IterationStatus status;
2850     status = mbm->iterate(&update_closure, &fill_closure, beg_addr, end_addr,
2851                           dense_prefix_end);
2852     if (status == ParMarkBitMap::incomplete) {
2853       update_closure.do_addr(update_closure.source());
2854     }
2855   }
2856 
2857   // Mark the regions as filled.
2858   RegionData* const beg_cp = sd.region(beg_region);
2859   RegionData* const end_cp = sd.region(end_region);
2860   for (RegionData* cp = beg_cp; cp < end_cp; ++cp) {
2861     cp->set_completed();
2862   }
2863 }
2864 
2865 // Return the SpaceId for the space containing addr.  If addr is not in the
2866 // heap, last_space_id is returned.  In debug mode it expects the address to be
2867 // in the heap and asserts such.
2868 PSParallelCompact::SpaceId PSParallelCompact::space_id(HeapWord* addr) {
2869   assert(Universe::heap()->is_in_reserved(addr), "addr not in the heap");
2870 
2871   for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2872     if (_space_info[id].space()->contains(addr)) {
2873       return SpaceId(id);
2874     }
2875   }
2876 
2877   assert(false, "no space contains the addr");
2878   return last_space_id;
2879 }
2880 
2881 void PSParallelCompact::update_deferred_objects(ParCompactionManager* cm,
2882                                                 SpaceId id) {
2883   assert(id < last_space_id, "bad space id");
2884 
2885   ParallelCompactData& sd = summary_data();
2886   const SpaceInfo* const space_info = _space_info + id;
2887   ObjectStartArray* const start_array = space_info->start_array();
2888 
2889   const MutableSpace* const space = space_info->space();
2890   assert(space_info->dense_prefix() >= space->bottom(), "dense_prefix not set");
2891   HeapWord* const beg_addr = space_info->dense_prefix();
2892   HeapWord* const end_addr = sd.region_align_up(space_info->new_top());
2893 
2894   const RegionData* const beg_region = sd.addr_to_region_ptr(beg_addr);
2895   const RegionData* const end_region = sd.addr_to_region_ptr(end_addr);
2896   const RegionData* cur_region;
2897   for (cur_region = beg_region; cur_region < end_region; ++cur_region) {
2898     HeapWord* const addr = cur_region->deferred_obj_addr();
2899     if (addr != NULL) {
2900       if (start_array != NULL) {
2901         start_array->allocate_block(addr);
2902       }
2903       oop(addr)->update_contents(cm);
2904       assert(oop(addr)->is_oop_or_null(), "should be an oop now");
2905     }
2906   }
2907 }
2908 
2909 // Skip over count live words starting from beg, and return the address of the
2910 // next live word.  Unless marked, the word corresponding to beg is assumed to
2911 // be dead.  Callers must either ensure beg does not correspond to the middle of
2912 // an object, or account for those live words in some other way.  Callers must
2913 // also ensure that there are enough live words in the range [beg, end) to skip.
2914 HeapWord*
2915 PSParallelCompact::skip_live_words(HeapWord* beg, HeapWord* end, size_t count)
2916 {
2917   assert(count > 0, "sanity");
2918 
2919   ParMarkBitMap* m = mark_bitmap();
2920   idx_t bits_to_skip = m->words_to_bits(count);
2921   idx_t cur_beg = m->addr_to_bit(beg);
2922   const idx_t search_end = BitMap::word_align_up(m->addr_to_bit(end));
2923 
2924   do {
2925     cur_beg = m->find_obj_beg(cur_beg, search_end);
2926     idx_t cur_end = m->find_obj_end(cur_beg, search_end);
2927     const size_t obj_bits = cur_end - cur_beg + 1;
2928     if (obj_bits > bits_to_skip) {
2929       return m->bit_to_addr(cur_beg + bits_to_skip);
2930     }
2931     bits_to_skip -= obj_bits;
2932     cur_beg = cur_end + 1;
2933   } while (bits_to_skip > 0);
2934 
2935   // Skipping the desired number of words landed just past the end of an object.
2936   // Find the start of the next object.
2937   cur_beg = m->find_obj_beg(cur_beg, search_end);
2938   assert(cur_beg < m->addr_to_bit(end), "not enough live words to skip");
2939   return m->bit_to_addr(cur_beg);
2940 }
2941 
2942 HeapWord* PSParallelCompact::first_src_addr(HeapWord* const dest_addr,
2943                                             SpaceId src_space_id,
2944                                             size_t src_region_idx)
2945 {
2946   assert(summary_data().is_region_aligned(dest_addr), "not aligned");
2947 
2948   const SplitInfo& split_info = _space_info[src_space_id].split_info();
2949   if (split_info.dest_region_addr() == dest_addr) {
2950     // The partial object ending at the split point contains the first word to
2951     // be copied to dest_addr.
2952     return split_info.first_src_addr();
2953   }
2954 
2955   const ParallelCompactData& sd = summary_data();
2956   ParMarkBitMap* const bitmap = mark_bitmap();
2957   const size_t RegionSize = ParallelCompactData::RegionSize;
2958 
2959   assert(sd.is_region_aligned(dest_addr), "not aligned");
2960   const RegionData* const src_region_ptr = sd.region(src_region_idx);
2961   const size_t partial_obj_size = src_region_ptr->partial_obj_size();
2962   HeapWord* const src_region_destination = src_region_ptr->destination();
2963 
2964   assert(dest_addr >= src_region_destination, "wrong src region");
2965   assert(src_region_ptr->data_size() > 0, "src region cannot be empty");
2966 
2967   HeapWord* const src_region_beg = sd.region_to_addr(src_region_idx);
2968   HeapWord* const src_region_end = src_region_beg + RegionSize;
2969 
2970   HeapWord* addr = src_region_beg;
2971   if (dest_addr == src_region_destination) {
2972     // Return the first live word in the source region.
2973     if (partial_obj_size == 0) {
2974       addr = bitmap->find_obj_beg(addr, src_region_end);
2975       assert(addr < src_region_end, "no objects start in src region");
2976     }
2977     return addr;
2978   }
2979 
2980   // Must skip some live data.
2981   size_t words_to_skip = dest_addr - src_region_destination;
2982   assert(src_region_ptr->data_size() > words_to_skip, "wrong src region");
2983 
2984   if (partial_obj_size >= words_to_skip) {
2985     // All the live words to skip are part of the partial object.
2986     addr += words_to_skip;
2987     if (partial_obj_size == words_to_skip) {
2988       // Find the first live word past the partial object.
2989       addr = bitmap->find_obj_beg(addr, src_region_end);
2990       assert(addr < src_region_end, "wrong src region");
2991     }
2992     return addr;
2993   }
2994 
2995   // Skip over the partial object (if any).
2996   if (partial_obj_size != 0) {
2997     words_to_skip -= partial_obj_size;
2998     addr += partial_obj_size;
2999   }
3000 
3001   // Skip over live words due to objects that start in the region.
3002   addr = skip_live_words(addr, src_region_end, words_to_skip);
3003   assert(addr < src_region_end, "wrong src region");
3004   return addr;
3005 }
3006 
3007 void PSParallelCompact::decrement_destination_counts(ParCompactionManager* cm,
3008                                                      SpaceId src_space_id,
3009                                                      size_t beg_region,
3010                                                      HeapWord* end_addr)
3011 {
3012   ParallelCompactData& sd = summary_data();
3013 
3014 #ifdef ASSERT
3015   MutableSpace* const src_space = _space_info[src_space_id].space();
3016   HeapWord* const beg_addr = sd.region_to_addr(beg_region);
3017   assert(src_space->contains(beg_addr) || beg_addr == src_space->end(),
3018          "src_space_id does not match beg_addr");
3019   assert(src_space->contains(end_addr) || end_addr == src_space->end(),
3020          "src_space_id does not match end_addr");
3021 #endif // #ifdef ASSERT
3022 
3023   RegionData* const beg = sd.region(beg_region);
3024   RegionData* const end = sd.addr_to_region_ptr(sd.region_align_up(end_addr));
3025 
3026   // Regions up to new_top() are enqueued if they become available.
3027   HeapWord* const new_top = _space_info[src_space_id].new_top();
3028   RegionData* const enqueue_end =
3029     sd.addr_to_region_ptr(sd.region_align_up(new_top));
3030 
3031   for (RegionData* cur = beg; cur < end; ++cur) {
3032     assert(cur->data_size() > 0, "region must have live data");
3033     cur->decrement_destination_count();
3034     if (cur < enqueue_end && cur->available() && cur->claim()) {
3035       cm->push_region(sd.region(cur));
3036     }
3037   }
3038 }
3039 
3040 size_t PSParallelCompact::next_src_region(MoveAndUpdateClosure& closure,
3041                                           SpaceId& src_space_id,
3042                                           HeapWord*& src_space_top,
3043                                           HeapWord* end_addr)
3044 {
3045   typedef ParallelCompactData::RegionData RegionData;
3046 
3047   ParallelCompactData& sd = PSParallelCompact::summary_data();
3048   const size_t region_size = ParallelCompactData::RegionSize;
3049 
3050   size_t src_region_idx = 0;
3051 
3052   // Skip empty regions (if any) up to the top of the space.
3053   HeapWord* const src_aligned_up = sd.region_align_up(end_addr);
3054   RegionData* src_region_ptr = sd.addr_to_region_ptr(src_aligned_up);
3055   HeapWord* const top_aligned_up = sd.region_align_up(src_space_top);
3056   const RegionData* const top_region_ptr =
3057     sd.addr_to_region_ptr(top_aligned_up);
3058   while (src_region_ptr < top_region_ptr && src_region_ptr->data_size() == 0) {
3059     ++src_region_ptr;
3060   }
3061 
3062   if (src_region_ptr < top_region_ptr) {
3063     // The next source region is in the current space.  Update src_region_idx
3064     // and the source address to match src_region_ptr.
3065     src_region_idx = sd.region(src_region_ptr);
3066     HeapWord* const src_region_addr = sd.region_to_addr(src_region_idx);
3067     if (src_region_addr > closure.source()) {
3068       closure.set_source(src_region_addr);
3069     }
3070     return src_region_idx;
3071   }
3072 
3073   // Switch to a new source space and find the first non-empty region.
3074   unsigned int space_id = src_space_id + 1;
3075   assert(space_id < last_space_id, "not enough spaces");
3076 
3077   HeapWord* const destination = closure.destination();
3078 
3079   do {
3080     MutableSpace* space = _space_info[space_id].space();
3081     HeapWord* const bottom = space->bottom();
3082     const RegionData* const bottom_cp = sd.addr_to_region_ptr(bottom);
3083 
3084     // Iterate over the spaces that do not compact into themselves.
3085     if (bottom_cp->destination() != bottom) {
3086       HeapWord* const top_aligned_up = sd.region_align_up(space->top());
3087       const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up);
3088 
3089       for (const RegionData* src_cp = bottom_cp; src_cp < top_cp; ++src_cp) {
3090         if (src_cp->live_obj_size() > 0) {
3091           // Found it.
3092           assert(src_cp->destination() == destination,
3093                  "first live obj in the space must match the destination");
3094           assert(src_cp->partial_obj_size() == 0,
3095                  "a space cannot begin with a partial obj");
3096 
3097           src_space_id = SpaceId(space_id);
3098           src_space_top = space->top();
3099           const size_t src_region_idx = sd.region(src_cp);
3100           closure.set_source(sd.region_to_addr(src_region_idx));
3101           return src_region_idx;
3102         } else {
3103           assert(src_cp->data_size() == 0, "sanity");
3104         }
3105       }
3106     }
3107   } while (++space_id < last_space_id);
3108 
3109   assert(false, "no source region was found");
3110   return 0;
3111 }
3112 
3113 void PSParallelCompact::fill_region(ParCompactionManager* cm, size_t region_idx)
3114 {
3115   typedef ParMarkBitMap::IterationStatus IterationStatus;
3116   const size_t RegionSize = ParallelCompactData::RegionSize;
3117   ParMarkBitMap* const bitmap = mark_bitmap();
3118   ParallelCompactData& sd = summary_data();
3119   RegionData* const region_ptr = sd.region(region_idx);
3120 
3121   // Get the items needed to construct the closure.
3122   HeapWord* dest_addr = sd.region_to_addr(region_idx);
3123   SpaceId dest_space_id = space_id(dest_addr);
3124   ObjectStartArray* start_array = _space_info[dest_space_id].start_array();
3125   HeapWord* new_top = _space_info[dest_space_id].new_top();
3126   assert(dest_addr < new_top, "sanity");
3127   const size_t words = MIN2(pointer_delta(new_top, dest_addr), RegionSize);
3128 
3129   // Get the source region and related info.
3130   size_t src_region_idx = region_ptr->source_region();
3131   SpaceId src_space_id = space_id(sd.region_to_addr(src_region_idx));
3132   HeapWord* src_space_top = _space_info[src_space_id].space()->top();
3133 
3134   MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words);
3135   closure.set_source(first_src_addr(dest_addr, src_space_id, src_region_idx));
3136 
3137   // Adjust src_region_idx to prepare for decrementing destination counts (the
3138   // destination count is not decremented when a region is copied to itself).
3139   if (src_region_idx == region_idx) {
3140     src_region_idx += 1;
3141   }
3142 
3143   if (bitmap->is_unmarked(closure.source())) {
3144     // The first source word is in the middle of an object; copy the remainder
3145     // of the object or as much as will fit.  The fact that pointer updates were
3146     // deferred will be noted when the object header is processed.
3147     HeapWord* const old_src_addr = closure.source();
3148     closure.copy_partial_obj();
3149     if (closure.is_full()) {
3150       decrement_destination_counts(cm, src_space_id, src_region_idx,
3151                                    closure.source());
3152       region_ptr->set_deferred_obj_addr(NULL);
3153       region_ptr->set_completed();
3154       return;
3155     }
3156 
3157     HeapWord* const end_addr = sd.region_align_down(closure.source());
3158     if (sd.region_align_down(old_src_addr) != end_addr) {
3159       // The partial object was copied from more than one source region.
3160       decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr);
3161 
3162       // Move to the next source region, possibly switching spaces as well.  All
3163       // args except end_addr may be modified.
3164       src_region_idx = next_src_region(closure, src_space_id, src_space_top,
3165                                        end_addr);
3166     }
3167   }
3168 
3169   do {
3170     HeapWord* const cur_addr = closure.source();
3171     HeapWord* const end_addr = MIN2(sd.region_align_up(cur_addr + 1),
3172                                     src_space_top);
3173     IterationStatus status = bitmap->iterate(&closure, cur_addr, end_addr);
3174 
3175     if (status == ParMarkBitMap::incomplete) {
3176       // The last obj that starts in the source region does not end in the
3177       // region.
3178       assert(closure.source() < end_addr, "sanity");
3179       HeapWord* const obj_beg = closure.source();
3180       HeapWord* const range_end = MIN2(obj_beg + closure.words_remaining(),
3181                                        src_space_top);
3182       HeapWord* const obj_end = bitmap->find_obj_end(obj_beg, range_end);
3183       if (obj_end < range_end) {
3184         // The end was found; the entire object will fit.
3185         status = closure.do_addr(obj_beg, bitmap->obj_size(obj_beg, obj_end));
3186         assert(status != ParMarkBitMap::would_overflow, "sanity");
3187       } else {
3188         // The end was not found; the object will not fit.
3189         assert(range_end < src_space_top, "obj cannot cross space boundary");
3190         status = ParMarkBitMap::would_overflow;
3191       }
3192     }
3193 
3194     if (status == ParMarkBitMap::would_overflow) {
3195       // The last object did not fit.  Note that interior oop updates were
3196       // deferred, then copy enough of the object to fill the region.
3197       region_ptr->set_deferred_obj_addr(closure.destination());
3198       status = closure.copy_until_full(); // copies from closure.source()
3199 
3200       decrement_destination_counts(cm, src_space_id, src_region_idx,
3201                                    closure.source());
3202       region_ptr->set_completed();
3203       return;
3204     }
3205 
3206     if (status == ParMarkBitMap::full) {
3207       decrement_destination_counts(cm, src_space_id, src_region_idx,
3208                                    closure.source());
3209       region_ptr->set_deferred_obj_addr(NULL);
3210       region_ptr->set_completed();
3211       return;
3212     }
3213 
3214     decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr);
3215 
3216     // Move to the next source region, possibly switching spaces as well.  All
3217     // args except end_addr may be modified.
3218     src_region_idx = next_src_region(closure, src_space_id, src_space_top,
3219                                      end_addr);
3220   } while (true);
3221 }
3222 
3223 void PSParallelCompact::fill_blocks(size_t region_idx)
3224 {
3225   // Fill in the block table elements for the specified region.  Each block
3226   // table element holds the number of live words in the region that are to the
3227   // left of the first object that starts in the block.  Thus only blocks in
3228   // which an object starts need to be filled.
3229   //
3230   // The algorithm scans the section of the bitmap that corresponds to the
3231   // region, keeping a running total of the live words.  When an object start is
3232   // found, if it's the first to start in the block that contains it, the
3233   // current total is written to the block table element.
3234   const size_t Log2BlockSize = ParallelCompactData::Log2BlockSize;
3235   const size_t Log2RegionSize = ParallelCompactData::Log2RegionSize;
3236   const size_t RegionSize = ParallelCompactData::RegionSize;
3237 
3238   ParallelCompactData& sd = summary_data();
3239   const size_t partial_obj_size = sd.region(region_idx)->partial_obj_size();
3240   if (partial_obj_size >= RegionSize) {
3241     return; // No objects start in this region.
3242   }
3243 
3244   // Ensure the first loop iteration decides that the block has changed.
3245   size_t cur_block = sd.block_count();
3246 
3247   const ParMarkBitMap* const bitmap = mark_bitmap();
3248 
3249   const size_t Log2BitsPerBlock = Log2BlockSize - LogMinObjAlignment;
3250   assert((size_t)1 << Log2BitsPerBlock ==
3251          bitmap->words_to_bits(ParallelCompactData::BlockSize), "sanity");
3252 
3253   size_t beg_bit = bitmap->words_to_bits(region_idx << Log2RegionSize);
3254   const size_t range_end = beg_bit + bitmap->words_to_bits(RegionSize);
3255   size_t live_bits = bitmap->words_to_bits(partial_obj_size);
3256   beg_bit = bitmap->find_obj_beg(beg_bit + live_bits, range_end);
3257   while (beg_bit < range_end) {
3258     const size_t new_block = beg_bit >> Log2BitsPerBlock;
3259     if (new_block != cur_block) {
3260       cur_block = new_block;
3261       sd.block(cur_block)->set_offset(bitmap->bits_to_words(live_bits));
3262     }
3263 
3264     const size_t end_bit = bitmap->find_obj_end(beg_bit, range_end);
3265     if (end_bit < range_end - 1) {
3266       live_bits += end_bit - beg_bit + 1;
3267       beg_bit = bitmap->find_obj_beg(end_bit + 1, range_end);
3268     } else {
3269       return;
3270     }
3271   }
3272 }
3273 
3274 void
3275 PSParallelCompact::move_and_update(ParCompactionManager* cm, SpaceId space_id) {
3276   const MutableSpace* sp = space(space_id);
3277   if (sp->is_empty()) {
3278     return;
3279   }
3280 
3281   ParallelCompactData& sd = PSParallelCompact::summary_data();
3282   ParMarkBitMap* const bitmap = mark_bitmap();
3283   HeapWord* const dp_addr = dense_prefix(space_id);
3284   HeapWord* beg_addr = sp->bottom();
3285   HeapWord* end_addr = sp->top();
3286 
3287   assert(beg_addr <= dp_addr && dp_addr <= end_addr, "bad dense prefix");
3288 
3289   const size_t beg_region = sd.addr_to_region_idx(beg_addr);
3290   const size_t dp_region = sd.addr_to_region_idx(dp_addr);
3291   if (beg_region < dp_region) {
3292     update_and_deadwood_in_dense_prefix(cm, space_id, beg_region, dp_region);
3293   }
3294 
3295   // The destination of the first live object that starts in the region is one
3296   // past the end of the partial object entering the region (if any).
3297   HeapWord* const dest_addr = sd.partial_obj_end(dp_region);
3298   HeapWord* const new_top = _space_info[space_id].new_top();
3299   assert(new_top >= dest_addr, "bad new_top value");
3300   const size_t words = pointer_delta(new_top, dest_addr);
3301 
3302   if (words > 0) {
3303     ObjectStartArray* start_array = _space_info[space_id].start_array();
3304     MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words);
3305 
3306     ParMarkBitMap::IterationStatus status;
3307     status = bitmap->iterate(&closure, dest_addr, end_addr);
3308     assert(status == ParMarkBitMap::full, "iteration not complete");
3309     assert(bitmap->find_obj_beg(closure.source(), end_addr) == end_addr,
3310            "live objects skipped because closure is full");
3311   }
3312 }
3313 
3314 jlong PSParallelCompact::millis_since_last_gc() {
3315   // We need a monotonically non-deccreasing time in ms but
3316   // os::javaTimeMillis() does not guarantee monotonicity.
3317   jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
3318   jlong ret_val = now - _time_of_last_gc;
3319   // XXX See note in genCollectedHeap::millis_since_last_gc().
3320   if (ret_val < 0) {
3321     NOT_PRODUCT(warning("time warp: "INT64_FORMAT, ret_val);)
3322     return 0;
3323   }
3324   return ret_val;
3325 }
3326 
3327 void PSParallelCompact::reset_millis_since_last_gc() {
3328   // We need a monotonically non-deccreasing time in ms but
3329   // os::javaTimeMillis() does not guarantee monotonicity.
3330   _time_of_last_gc = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
3331 }
3332 
3333 ParMarkBitMap::IterationStatus MoveAndUpdateClosure::copy_until_full()
3334 {
3335   if (source() != destination()) {
3336     DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3337     Copy::aligned_conjoint_words(source(), destination(), words_remaining());
3338   }
3339   update_state(words_remaining());
3340   assert(is_full(), "sanity");
3341   return ParMarkBitMap::full;
3342 }
3343 
3344 void MoveAndUpdateClosure::copy_partial_obj()
3345 {
3346   size_t words = words_remaining();
3347 
3348   HeapWord* const range_end = MIN2(source() + words, bitmap()->region_end());
3349   HeapWord* const end_addr = bitmap()->find_obj_end(source(), range_end);
3350   if (end_addr < range_end) {
3351     words = bitmap()->obj_size(source(), end_addr);
3352   }
3353 
3354   // This test is necessary; if omitted, the pointer updates to a partial object
3355   // that crosses the dense prefix boundary could be overwritten.
3356   if (source() != destination()) {
3357     DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3358     Copy::aligned_conjoint_words(source(), destination(), words);
3359   }
3360   update_state(words);
3361 }
3362 
3363 ParMarkBitMapClosure::IterationStatus
3364 MoveAndUpdateClosure::do_addr(HeapWord* addr, size_t words) {
3365   assert(destination() != NULL, "sanity");
3366   assert(bitmap()->obj_size(addr) == words, "bad size");
3367 
3368   _source = addr;
3369   assert(PSParallelCompact::summary_data().calc_new_pointer(source()) ==
3370          destination(), "wrong destination");
3371 
3372   if (words > words_remaining()) {
3373     return ParMarkBitMap::would_overflow;
3374   }
3375 
3376   // The start_array must be updated even if the object is not moving.
3377   if (_start_array != NULL) {
3378     _start_array->allocate_block(destination());
3379   }
3380 
3381   if (destination() != source()) {
3382     DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3383     Copy::aligned_conjoint_words(source(), destination(), words);
3384   }
3385 
3386   oop moved_oop = (oop) destination();
3387   moved_oop->update_contents(compaction_manager());
3388   assert(moved_oop->is_oop_or_null(), "Object should be whole at this point");
3389 
3390   update_state(words);
3391   assert(destination() == (HeapWord*)moved_oop + moved_oop->size(), "sanity");
3392   return is_full() ? ParMarkBitMap::full : ParMarkBitMap::incomplete;
3393 }
3394 
3395 UpdateOnlyClosure::UpdateOnlyClosure(ParMarkBitMap* mbm,
3396                                      ParCompactionManager* cm,
3397                                      PSParallelCompact::SpaceId space_id) :
3398   ParMarkBitMapClosure(mbm, cm),
3399   _space_id(space_id),
3400   _start_array(PSParallelCompact::start_array(space_id))
3401 {
3402 }
3403 
3404 // Updates the references in the object to their new values.
3405 ParMarkBitMapClosure::IterationStatus
3406 UpdateOnlyClosure::do_addr(HeapWord* addr, size_t words) {
3407   do_addr(addr);
3408   return ParMarkBitMap::incomplete;
3409 }