1 /*
   2  * Copyright (c) 2005, 2015, 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/stringTable.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.inline.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/instanceKlass.inline.hpp"
  52 #include "oops/instanceMirrorKlass.inline.hpp"
  53 #include "oops/methodData.hpp"
  54 #include "oops/objArrayKlass.inline.hpp"
  55 #include "oops/oop.inline.hpp"
  56 #include "runtime/atomic.inline.hpp"
  57 #include "runtime/fprofiler.hpp"
  58 #include "runtime/safepoint.hpp"
  59 #include "runtime/vmThread.hpp"
  60 #include "services/management.hpp"
  61 #include "services/memoryService.hpp"
  62 #include "services/memTracker.hpp"
  63 #include "utilities/events.hpp"
  64 #include "utilities/stack.inline.hpp"
  65 
  66 #include <math.h>
  67 
  68 // All sizes are in HeapWords.
  69 const size_t ParallelCompactData::Log2RegionSize  = 16; // 64K words
  70 const size_t ParallelCompactData::RegionSize      = (size_t)1 << Log2RegionSize;
  71 const size_t ParallelCompactData::RegionSizeBytes =
  72   RegionSize << LogHeapWordSize;
  73 const size_t ParallelCompactData::RegionSizeOffsetMask = RegionSize - 1;
  74 const size_t ParallelCompactData::RegionAddrOffsetMask = RegionSizeBytes - 1;
  75 const size_t ParallelCompactData::RegionAddrMask       = ~RegionAddrOffsetMask;
  76 
  77 const size_t ParallelCompactData::Log2BlockSize   = 7; // 128 words
  78 const size_t ParallelCompactData::BlockSize       = (size_t)1 << Log2BlockSize;
  79 const size_t ParallelCompactData::BlockSizeBytes  =
  80   BlockSize << LogHeapWordSize;
  81 const size_t ParallelCompactData::BlockSizeOffsetMask = BlockSize - 1;
  82 const size_t ParallelCompactData::BlockAddrOffsetMask = BlockSizeBytes - 1;
  83 const size_t ParallelCompactData::BlockAddrMask       = ~BlockAddrOffsetMask;
  84 
  85 const size_t ParallelCompactData::BlocksPerRegion = RegionSize / BlockSize;
  86 const size_t ParallelCompactData::Log2BlocksPerRegion =
  87   Log2RegionSize - Log2BlockSize;
  88 
  89 const ParallelCompactData::RegionData::region_sz_t
  90 ParallelCompactData::RegionData::dc_shift = 27;
  91 
  92 const ParallelCompactData::RegionData::region_sz_t
  93 ParallelCompactData::RegionData::dc_mask = ~0U << dc_shift;
  94 
  95 const ParallelCompactData::RegionData::region_sz_t
  96 ParallelCompactData::RegionData::dc_one = 0x1U << dc_shift;
  97 
  98 const ParallelCompactData::RegionData::region_sz_t
  99 ParallelCompactData::RegionData::los_mask = ~dc_mask;
 100 
 101 const ParallelCompactData::RegionData::region_sz_t
 102 ParallelCompactData::RegionData::dc_claimed = 0x8U << dc_shift;
 103 
 104 const ParallelCompactData::RegionData::region_sz_t
 105 ParallelCompactData::RegionData::dc_completed = 0xcU << dc_shift;
 106 
 107 SpaceInfo PSParallelCompact::_space_info[PSParallelCompact::last_space_id];
 108 bool      PSParallelCompact::_print_phases = false;
 109 
 110 ReferenceProcessor* PSParallelCompact::_ref_processor = NULL;
 111 Klass*              PSParallelCompact::_updated_int_array_klass_obj = NULL;
 112 
 113 double PSParallelCompact::_dwl_mean;
 114 double PSParallelCompact::_dwl_std_dev;
 115 double PSParallelCompact::_dwl_first_term;
 116 double PSParallelCompact::_dwl_adjustment;
 117 #ifdef  ASSERT
 118 bool   PSParallelCompact::_dwl_initialized = false;
 119 #endif  // #ifdef ASSERT
 120 
 121 void SplitInfo::record(size_t src_region_idx, size_t partial_obj_size,
 122                        HeapWord* destination)
 123 {
 124   assert(src_region_idx != 0, "invalid src_region_idx");
 125   assert(partial_obj_size != 0, "invalid partial_obj_size argument");
 126   assert(destination != NULL, "invalid destination argument");
 127 
 128   _src_region_idx = src_region_idx;
 129   _partial_obj_size = partial_obj_size;
 130   _destination = destination;
 131 
 132   // These fields may not be updated below, so make sure they're clear.
 133   assert(_dest_region_addr == NULL, "should have been cleared");
 134   assert(_first_src_addr == NULL, "should have been cleared");
 135 
 136   // Determine the number of destination regions for the partial object.
 137   HeapWord* const last_word = destination + partial_obj_size - 1;
 138   const ParallelCompactData& sd = PSParallelCompact::summary_data();
 139   HeapWord* const beg_region_addr = sd.region_align_down(destination);
 140   HeapWord* const end_region_addr = sd.region_align_down(last_word);
 141 
 142   if (beg_region_addr == end_region_addr) {
 143     // One destination region.
 144     _destination_count = 1;
 145     if (end_region_addr == destination) {
 146       // The destination falls on a region boundary, thus the first word of the
 147       // partial object will be the first word copied to the destination region.
 148       _dest_region_addr = end_region_addr;
 149       _first_src_addr = sd.region_to_addr(src_region_idx);
 150     }
 151   } else {
 152     // Two destination regions.  When copied, the partial object will cross a
 153     // destination region boundary, so a word somewhere within the partial
 154     // object will be the first word copied to the second destination region.
 155     _destination_count = 2;
 156     _dest_region_addr = end_region_addr;
 157     const size_t ofs = pointer_delta(end_region_addr, destination);
 158     assert(ofs < _partial_obj_size, "sanity");
 159     _first_src_addr = sd.region_to_addr(src_region_idx) + ofs;
 160   }
 161 }
 162 
 163 void SplitInfo::clear()
 164 {
 165   _src_region_idx = 0;
 166   _partial_obj_size = 0;
 167   _destination = NULL;
 168   _destination_count = 0;
 169   _dest_region_addr = NULL;
 170   _first_src_addr = NULL;
 171   assert(!is_valid(), "sanity");
 172 }
 173 
 174 #ifdef  ASSERT
 175 void SplitInfo::verify_clear()
 176 {
 177   assert(_src_region_idx == 0, "not clear");
 178   assert(_partial_obj_size == 0, "not clear");
 179   assert(_destination == NULL, "not clear");
 180   assert(_destination_count == 0, "not clear");
 181   assert(_dest_region_addr == NULL, "not clear");
 182   assert(_first_src_addr == NULL, "not clear");
 183 }
 184 #endif  // #ifdef ASSERT
 185 
 186 
 187 void PSParallelCompact::print_on_error(outputStream* st) {
 188   _mark_bitmap.print_on_error(st);
 189 }
 190 
 191 #ifndef PRODUCT
 192 const char* PSParallelCompact::space_names[] = {
 193   "old ", "eden", "from", "to  "
 194 };
 195 
 196 void PSParallelCompact::print_region_ranges()
 197 {
 198   tty->print_cr("space  bottom     top        end        new_top");
 199   tty->print_cr("------ ---------- ---------- ---------- ----------");
 200 
 201   for (unsigned int id = 0; id < last_space_id; ++id) {
 202     const MutableSpace* space = _space_info[id].space();
 203     tty->print_cr("%u %s "
 204                   SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " "
 205                   SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " ",
 206                   id, space_names[id],
 207                   summary_data().addr_to_region_idx(space->bottom()),
 208                   summary_data().addr_to_region_idx(space->top()),
 209                   summary_data().addr_to_region_idx(space->end()),
 210                   summary_data().addr_to_region_idx(_space_info[id].new_top()));
 211   }
 212 }
 213 
 214 void
 215 print_generic_summary_region(size_t i, const ParallelCompactData::RegionData* c)
 216 {
 217 #define REGION_IDX_FORMAT        SIZE_FORMAT_W(7)
 218 #define REGION_DATA_FORMAT       SIZE_FORMAT_W(5)
 219 
 220   ParallelCompactData& sd = PSParallelCompact::summary_data();
 221   size_t dci = c->destination() ? sd.addr_to_region_idx(c->destination()) : 0;
 222   tty->print_cr(REGION_IDX_FORMAT " " PTR_FORMAT " "
 223                 REGION_IDX_FORMAT " " PTR_FORMAT " "
 224                 REGION_DATA_FORMAT " " REGION_DATA_FORMAT " "
 225                 REGION_DATA_FORMAT " " REGION_IDX_FORMAT " %d",
 226                 i, p2i(c->data_location()), dci, p2i(c->destination()),
 227                 c->partial_obj_size(), c->live_obj_size(),
 228                 c->data_size(), c->source_region(), c->destination_count());
 229 
 230 #undef  REGION_IDX_FORMAT
 231 #undef  REGION_DATA_FORMAT
 232 }
 233 
 234 void
 235 print_generic_summary_data(ParallelCompactData& summary_data,
 236                            HeapWord* const beg_addr,
 237                            HeapWord* const end_addr)
 238 {
 239   size_t total_words = 0;
 240   size_t i = summary_data.addr_to_region_idx(beg_addr);
 241   const size_t last = summary_data.addr_to_region_idx(end_addr);
 242   HeapWord* pdest = 0;
 243 
 244   while (i <= last) {
 245     ParallelCompactData::RegionData* c = summary_data.region(i);
 246     if (c->data_size() != 0 || c->destination() != pdest) {
 247       print_generic_summary_region(i, c);
 248       total_words += c->data_size();
 249       pdest = c->destination();
 250     }
 251     ++i;
 252   }
 253 
 254   tty->print_cr("summary_data_bytes=" SIZE_FORMAT, total_words * HeapWordSize);
 255 }
 256 
 257 void
 258 print_generic_summary_data(ParallelCompactData& summary_data,
 259                            SpaceInfo* space_info)
 260 {
 261   for (unsigned int id = 0; id < PSParallelCompact::last_space_id; ++id) {
 262     const MutableSpace* space = space_info[id].space();
 263     print_generic_summary_data(summary_data, space->bottom(),
 264                                MAX2(space->top(), space_info[id].new_top()));
 265   }
 266 }
 267 
 268 void
 269 print_initial_summary_region(size_t i,
 270                              const ParallelCompactData::RegionData* c,
 271                              bool newline = true)
 272 {
 273   tty->print(SIZE_FORMAT_W(5) " " PTR_FORMAT " "
 274              SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " "
 275              SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d",
 276              i, p2i(c->destination()),
 277              c->partial_obj_size(), c->live_obj_size(),
 278              c->data_size(), c->source_region(), c->destination_count());
 279   if (newline) tty->cr();
 280 }
 281 
 282 void
 283 print_initial_summary_data(ParallelCompactData& summary_data,
 284                            const MutableSpace* space) {
 285   if (space->top() == space->bottom()) {
 286     return;
 287   }
 288 
 289   const size_t region_size = ParallelCompactData::RegionSize;
 290   typedef ParallelCompactData::RegionData RegionData;
 291   HeapWord* const top_aligned_up = summary_data.region_align_up(space->top());
 292   const size_t end_region = summary_data.addr_to_region_idx(top_aligned_up);
 293   const RegionData* c = summary_data.region(end_region - 1);
 294   HeapWord* end_addr = c->destination() + c->data_size();
 295   const size_t live_in_space = pointer_delta(end_addr, space->bottom());
 296 
 297   // Print (and count) the full regions at the beginning of the space.
 298   size_t full_region_count = 0;
 299   size_t i = summary_data.addr_to_region_idx(space->bottom());
 300   while (i < end_region && summary_data.region(i)->data_size() == region_size) {
 301     print_initial_summary_region(i, summary_data.region(i));
 302     ++full_region_count;
 303     ++i;
 304   }
 305 
 306   size_t live_to_right = live_in_space - full_region_count * region_size;
 307 
 308   double max_reclaimed_ratio = 0.0;
 309   size_t max_reclaimed_ratio_region = 0;
 310   size_t max_dead_to_right = 0;
 311   size_t max_live_to_right = 0;
 312 
 313   // Print the 'reclaimed ratio' for regions while there is something live in
 314   // the region or to the right of it.  The remaining regions are empty (and
 315   // uninteresting), and computing the ratio will result in division by 0.
 316   while (i < end_region && live_to_right > 0) {
 317     c = summary_data.region(i);
 318     HeapWord* const region_addr = summary_data.region_to_addr(i);
 319     const size_t used_to_right = pointer_delta(space->top(), region_addr);
 320     const size_t dead_to_right = used_to_right - live_to_right;
 321     const double reclaimed_ratio = double(dead_to_right) / live_to_right;
 322 
 323     if (reclaimed_ratio > max_reclaimed_ratio) {
 324             max_reclaimed_ratio = reclaimed_ratio;
 325             max_reclaimed_ratio_region = i;
 326             max_dead_to_right = dead_to_right;
 327             max_live_to_right = live_to_right;
 328     }
 329 
 330     print_initial_summary_region(i, c, false);
 331     tty->print_cr(" %12.10f " SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10),
 332                   reclaimed_ratio, dead_to_right, live_to_right);
 333 
 334     live_to_right -= c->data_size();
 335     ++i;
 336   }
 337 
 338   // Any remaining regions are empty.  Print one more if there is one.
 339   if (i < end_region) {
 340     print_initial_summary_region(i, summary_data.region(i));
 341   }
 342 
 343   tty->print_cr("max:  " SIZE_FORMAT_W(4) " d2r=" SIZE_FORMAT_W(10) " "
 344                 "l2r=" SIZE_FORMAT_W(10) " max_ratio=%14.12f",
 345                 max_reclaimed_ratio_region, max_dead_to_right,
 346                 max_live_to_right, max_reclaimed_ratio);
 347 }
 348 
 349 void
 350 print_initial_summary_data(ParallelCompactData& summary_data,
 351                            SpaceInfo* space_info) {
 352   unsigned int id = PSParallelCompact::old_space_id;
 353   const MutableSpace* space;
 354   do {
 355     space = space_info[id].space();
 356     print_initial_summary_data(summary_data, space);
 357   } while (++id < PSParallelCompact::eden_space_id);
 358 
 359   do {
 360     space = space_info[id].space();
 361     print_generic_summary_data(summary_data, space->bottom(), space->top());
 362   } while (++id < PSParallelCompact::last_space_id);
 363 }
 364 #endif  // #ifndef PRODUCT
 365 
 366 #ifdef  ASSERT
 367 size_t add_obj_count;
 368 size_t add_obj_size;
 369 size_t mark_bitmap_count;
 370 size_t mark_bitmap_size;
 371 #endif  // #ifdef ASSERT
 372 
 373 ParallelCompactData::ParallelCompactData()
 374 {
 375   _region_start = 0;
 376 
 377   _region_vspace = 0;
 378   _reserved_byte_size = 0;
 379   _region_data = 0;
 380   _region_count = 0;
 381 
 382   _block_vspace = 0;
 383   _block_data = 0;
 384   _block_count = 0;
 385 }
 386 
 387 bool ParallelCompactData::initialize(MemRegion covered_region)
 388 {
 389   _region_start = covered_region.start();
 390   const size_t region_size = covered_region.word_size();
 391   DEBUG_ONLY(_region_end = _region_start + region_size;)
 392 
 393   assert(region_align_down(_region_start) == _region_start,
 394          "region start not aligned");
 395   assert((region_size & RegionSizeOffsetMask) == 0,
 396          "region size not a multiple of RegionSize");
 397 
 398   bool result = initialize_region_data(region_size) && initialize_block_data();
 399   return result;
 400 }
 401 
 402 PSVirtualSpace*
 403 ParallelCompactData::create_vspace(size_t count, size_t element_size)
 404 {
 405   const size_t raw_bytes = count * element_size;
 406   const size_t page_sz = os::page_size_for_region_aligned(raw_bytes, 10);
 407   const size_t granularity = os::vm_allocation_granularity();
 408   _reserved_byte_size = align_size_up(raw_bytes, MAX2(page_sz, granularity));
 409 
 410   const size_t rs_align = page_sz == (size_t) os::vm_page_size() ? 0 :
 411     MAX2(page_sz, granularity);
 412   ReservedSpace rs(_reserved_byte_size, rs_align, rs_align > 0);
 413   os::trace_page_sizes("par compact", raw_bytes, raw_bytes, page_sz, rs.base(),
 414                        rs.size());
 415 
 416   MemTracker::record_virtual_memory_type((address)rs.base(), mtGC);
 417 
 418   PSVirtualSpace* vspace = new PSVirtualSpace(rs, page_sz);
 419   if (vspace != 0) {
 420     if (vspace->expand_by(_reserved_byte_size)) {
 421       return vspace;
 422     }
 423     delete vspace;
 424     // Release memory reserved in the space.
 425     rs.release();
 426   }
 427 
 428   return 0;
 429 }
 430 
 431 bool ParallelCompactData::initialize_region_data(size_t region_size)
 432 {
 433   const size_t count = (region_size + RegionSizeOffsetMask) >> Log2RegionSize;
 434   _region_vspace = create_vspace(count, sizeof(RegionData));
 435   if (_region_vspace != 0) {
 436     _region_data = (RegionData*)_region_vspace->reserved_low_addr();
 437     _region_count = count;
 438     return true;
 439   }
 440   return false;
 441 }
 442 
 443 bool ParallelCompactData::initialize_block_data()
 444 {
 445   assert(_region_count != 0, "region data must be initialized first");
 446   const size_t count = _region_count << Log2BlocksPerRegion;
 447   _block_vspace = create_vspace(count, sizeof(BlockData));
 448   if (_block_vspace != 0) {
 449     _block_data = (BlockData*)_block_vspace->reserved_low_addr();
 450     _block_count = count;
 451     return true;
 452   }
 453   return false;
 454 }
 455 
 456 void ParallelCompactData::clear()
 457 {
 458   memset(_region_data, 0, _region_vspace->committed_size());
 459   memset(_block_data, 0, _block_vspace->committed_size());
 460 }
 461 
 462 void ParallelCompactData::clear_range(size_t beg_region, size_t end_region) {
 463   assert(beg_region <= _region_count, "beg_region out of range");
 464   assert(end_region <= _region_count, "end_region out of range");
 465   assert(RegionSize % BlockSize == 0, "RegionSize not a multiple of BlockSize");
 466 
 467   const size_t region_cnt = end_region - beg_region;
 468   memset(_region_data + beg_region, 0, region_cnt * sizeof(RegionData));
 469 
 470   const size_t beg_block = beg_region * BlocksPerRegion;
 471   const size_t block_cnt = region_cnt * BlocksPerRegion;
 472   memset(_block_data + beg_block, 0, block_cnt * sizeof(BlockData));
 473 }
 474 
 475 HeapWord* ParallelCompactData::partial_obj_end(size_t region_idx) const
 476 {
 477   const RegionData* cur_cp = region(region_idx);
 478   const RegionData* const end_cp = region(region_count() - 1);
 479 
 480   HeapWord* result = region_to_addr(region_idx);
 481   if (cur_cp < end_cp) {
 482     do {
 483       result += cur_cp->partial_obj_size();
 484     } while (cur_cp->partial_obj_size() == RegionSize && ++cur_cp < end_cp);
 485   }
 486   return result;
 487 }
 488 
 489 void ParallelCompactData::add_obj(HeapWord* addr, size_t len)
 490 {
 491   const size_t obj_ofs = pointer_delta(addr, _region_start);
 492   const size_t beg_region = obj_ofs >> Log2RegionSize;
 493   const size_t end_region = (obj_ofs + len - 1) >> Log2RegionSize;
 494 
 495   DEBUG_ONLY(Atomic::inc_ptr(&add_obj_count);)
 496   DEBUG_ONLY(Atomic::add_ptr(len, &add_obj_size);)
 497 
 498   if (beg_region == end_region) {
 499     // All in one region.
 500     _region_data[beg_region].add_live_obj(len);
 501     return;
 502   }
 503 
 504   // First region.
 505   const size_t beg_ofs = region_offset(addr);
 506   _region_data[beg_region].add_live_obj(RegionSize - beg_ofs);
 507 
 508   Klass* klass = ((oop)addr)->klass();
 509   // Middle regions--completely spanned by this object.
 510   for (size_t region = beg_region + 1; region < end_region; ++region) {
 511     _region_data[region].set_partial_obj_size(RegionSize);
 512     _region_data[region].set_partial_obj_addr(addr);
 513   }
 514 
 515   // Last region.
 516   const size_t end_ofs = region_offset(addr + len - 1);
 517   _region_data[end_region].set_partial_obj_size(end_ofs + 1);
 518   _region_data[end_region].set_partial_obj_addr(addr);
 519 }
 520 
 521 void
 522 ParallelCompactData::summarize_dense_prefix(HeapWord* beg, HeapWord* end)
 523 {
 524   assert(region_offset(beg) == 0, "not RegionSize aligned");
 525   assert(region_offset(end) == 0, "not RegionSize aligned");
 526 
 527   size_t cur_region = addr_to_region_idx(beg);
 528   const size_t end_region = addr_to_region_idx(end);
 529   HeapWord* addr = beg;
 530   while (cur_region < end_region) {
 531     _region_data[cur_region].set_destination(addr);
 532     _region_data[cur_region].set_destination_count(0);
 533     _region_data[cur_region].set_source_region(cur_region);
 534     _region_data[cur_region].set_data_location(addr);
 535 
 536     // Update live_obj_size so the region appears completely full.
 537     size_t live_size = RegionSize - _region_data[cur_region].partial_obj_size();
 538     _region_data[cur_region].set_live_obj_size(live_size);
 539 
 540     ++cur_region;
 541     addr += RegionSize;
 542   }
 543 }
 544 
 545 // Find the point at which a space can be split and, if necessary, record the
 546 // split point.
 547 //
 548 // If the current src region (which overflowed the destination space) doesn't
 549 // have a partial object, the split point is at the beginning of the current src
 550 // region (an "easy" split, no extra bookkeeping required).
 551 //
 552 // If the current src region has a partial object, the split point is in the
 553 // region where that partial object starts (call it the split_region).  If
 554 // split_region has a partial object, then the split point is just after that
 555 // partial object (a "hard" split where we have to record the split data and
 556 // zero the partial_obj_size field).  With a "hard" split, we know that the
 557 // partial_obj ends within split_region because the partial object that caused
 558 // the overflow starts in split_region.  If split_region doesn't have a partial
 559 // obj, then the split is at the beginning of split_region (another "easy"
 560 // split).
 561 HeapWord*
 562 ParallelCompactData::summarize_split_space(size_t src_region,
 563                                            SplitInfo& split_info,
 564                                            HeapWord* destination,
 565                                            HeapWord* target_end,
 566                                            HeapWord** target_next)
 567 {
 568   assert(destination <= target_end, "sanity");
 569   assert(destination + _region_data[src_region].data_size() > target_end,
 570     "region should not fit into target space");
 571   assert(is_region_aligned(target_end), "sanity");
 572 
 573   size_t split_region = src_region;
 574   HeapWord* split_destination = destination;
 575   size_t partial_obj_size = _region_data[src_region].partial_obj_size();
 576 
 577   if (destination + partial_obj_size > target_end) {
 578     // The split point is just after the partial object (if any) in the
 579     // src_region that contains the start of the object that overflowed the
 580     // destination space.
 581     //
 582     // Find the start of the "overflow" object and set split_region to the
 583     // region containing it.
 584     HeapWord* const overflow_obj = _region_data[src_region].partial_obj_addr();
 585     split_region = addr_to_region_idx(overflow_obj);
 586 
 587     // Clear the source_region field of all destination regions whose first word
 588     // came from data after the split point (a non-null source_region field
 589     // implies a region must be filled).
 590     //
 591     // An alternative to the simple loop below:  clear during post_compact(),
 592     // which uses memcpy instead of individual stores, and is easy to
 593     // parallelize.  (The downside is that it clears the entire RegionData
 594     // object as opposed to just one field.)
 595     //
 596     // post_compact() would have to clear the summary data up to the highest
 597     // address that was written during the summary phase, which would be
 598     //
 599     //         max(top, max(new_top, clear_top))
 600     //
 601     // where clear_top is a new field in SpaceInfo.  Would have to set clear_top
 602     // to target_end.
 603     const RegionData* const sr = region(split_region);
 604     const size_t beg_idx =
 605       addr_to_region_idx(region_align_up(sr->destination() +
 606                                          sr->partial_obj_size()));
 607     const size_t end_idx = addr_to_region_idx(target_end);
 608 
 609     if (TraceParallelOldGCSummaryPhase) {
 610         gclog_or_tty->print_cr("split:  clearing source_region field in ["
 611                                SIZE_FORMAT ", " SIZE_FORMAT ")",
 612                                beg_idx, end_idx);
 613     }
 614     for (size_t idx = beg_idx; idx < end_idx; ++idx) {
 615       _region_data[idx].set_source_region(0);
 616     }
 617 
 618     // Set split_destination and partial_obj_size to reflect the split region.
 619     split_destination = sr->destination();
 620     partial_obj_size = sr->partial_obj_size();
 621   }
 622 
 623   // The split is recorded only if a partial object extends onto the region.
 624   if (partial_obj_size != 0) {
 625     _region_data[split_region].set_partial_obj_size(0);
 626     split_info.record(split_region, partial_obj_size, split_destination);
 627   }
 628 
 629   // Setup the continuation addresses.
 630   *target_next = split_destination + partial_obj_size;
 631   HeapWord* const source_next = region_to_addr(split_region) + partial_obj_size;
 632 
 633   if (TraceParallelOldGCSummaryPhase) {
 634     const char * split_type = partial_obj_size == 0 ? "easy" : "hard";
 635     gclog_or_tty->print_cr("%s split:  src=" PTR_FORMAT " src_c=" SIZE_FORMAT
 636                            " pos=" SIZE_FORMAT,
 637                            split_type, p2i(source_next), split_region,
 638                            partial_obj_size);
 639     gclog_or_tty->print_cr("%s split:  dst=" PTR_FORMAT " dst_c=" SIZE_FORMAT
 640                            " tn=" PTR_FORMAT,
 641                            split_type, p2i(split_destination),
 642                            addr_to_region_idx(split_destination),
 643                            p2i(*target_next));
 644 
 645     if (partial_obj_size != 0) {
 646       HeapWord* const po_beg = split_info.destination();
 647       HeapWord* const po_end = po_beg + split_info.partial_obj_size();
 648       gclog_or_tty->print_cr("%s split:  "
 649                              "po_beg=" PTR_FORMAT " " SIZE_FORMAT " "
 650                              "po_end=" PTR_FORMAT " " SIZE_FORMAT,
 651                              split_type,
 652                              p2i(po_beg), addr_to_region_idx(po_beg),
 653                              p2i(po_end), addr_to_region_idx(po_end));
 654     }
 655   }
 656 
 657   return source_next;
 658 }
 659 
 660 bool ParallelCompactData::summarize(SplitInfo& split_info,
 661                                     HeapWord* source_beg, HeapWord* source_end,
 662                                     HeapWord** source_next,
 663                                     HeapWord* target_beg, HeapWord* target_end,
 664                                     HeapWord** target_next)
 665 {
 666   if (TraceParallelOldGCSummaryPhase) {
 667     HeapWord* const source_next_val = source_next == NULL ? NULL : *source_next;
 668     tty->print_cr("sb=" PTR_FORMAT " se=" PTR_FORMAT " sn=" PTR_FORMAT
 669                   "tb=" PTR_FORMAT " te=" PTR_FORMAT " tn=" PTR_FORMAT,
 670                   p2i(source_beg), p2i(source_end), p2i(source_next_val),
 671                   p2i(target_beg), p2i(target_end), p2i(*target_next));
 672   }
 673 
 674   size_t cur_region = addr_to_region_idx(source_beg);
 675   const size_t end_region = addr_to_region_idx(region_align_up(source_end));
 676 
 677   HeapWord *dest_addr = target_beg;
 678   while (cur_region < end_region) {
 679     // The destination must be set even if the region has no data.
 680     _region_data[cur_region].set_destination(dest_addr);
 681 
 682     size_t words = _region_data[cur_region].data_size();
 683     if (words > 0) {
 684       // If cur_region does not fit entirely into the target space, find a point
 685       // at which the source space can be 'split' so that part is copied to the
 686       // target space and the rest is copied elsewhere.
 687       if (dest_addr + words > target_end) {
 688         assert(source_next != NULL, "source_next is NULL when splitting");
 689         *source_next = summarize_split_space(cur_region, split_info, dest_addr,
 690                                              target_end, target_next);
 691         return false;
 692       }
 693 
 694       // Compute the destination_count for cur_region, and if necessary, update
 695       // source_region for a destination region.  The source_region field is
 696       // updated if cur_region is the first (left-most) region to be copied to a
 697       // destination region.
 698       //
 699       // The destination_count calculation is a bit subtle.  A region that has
 700       // data that compacts into itself does not count itself as a destination.
 701       // This maintains the invariant that a zero count means the region is
 702       // available and can be claimed and then filled.
 703       uint destination_count = 0;
 704       if (split_info.is_split(cur_region)) {
 705         // The current region has been split:  the partial object will be copied
 706         // to one destination space and the remaining data will be copied to
 707         // another destination space.  Adjust the initial destination_count and,
 708         // if necessary, set the source_region field if the partial object will
 709         // cross a destination region boundary.
 710         destination_count = split_info.destination_count();
 711         if (destination_count == 2) {
 712           size_t dest_idx = addr_to_region_idx(split_info.dest_region_addr());
 713           _region_data[dest_idx].set_source_region(cur_region);
 714         }
 715       }
 716 
 717       HeapWord* const last_addr = dest_addr + words - 1;
 718       const size_t dest_region_1 = addr_to_region_idx(dest_addr);
 719       const size_t dest_region_2 = addr_to_region_idx(last_addr);
 720 
 721       // Initially assume that the destination regions will be the same and
 722       // adjust the value below if necessary.  Under this assumption, if
 723       // cur_region == dest_region_2, then cur_region will be compacted
 724       // completely into itself.
 725       destination_count += cur_region == dest_region_2 ? 0 : 1;
 726       if (dest_region_1 != dest_region_2) {
 727         // Destination regions differ; adjust destination_count.
 728         destination_count += 1;
 729         // Data from cur_region will be copied to the start of dest_region_2.
 730         _region_data[dest_region_2].set_source_region(cur_region);
 731       } else if (region_offset(dest_addr) == 0) {
 732         // Data from cur_region will be copied to the start of the destination
 733         // region.
 734         _region_data[dest_region_1].set_source_region(cur_region);
 735       }
 736 
 737       _region_data[cur_region].set_destination_count(destination_count);
 738       _region_data[cur_region].set_data_location(region_to_addr(cur_region));
 739       dest_addr += words;
 740     }
 741 
 742     ++cur_region;
 743   }
 744 
 745   *target_next = dest_addr;
 746   return true;
 747 }
 748 
 749 HeapWord* ParallelCompactData::calc_new_pointer(HeapWord* addr) {
 750   assert(addr != NULL, "Should detect NULL oop earlier");
 751   assert(ParallelScavengeHeap::heap()->is_in(addr), "not in heap");
 752   assert(PSParallelCompact::mark_bitmap()->is_marked(addr), "not marked");
 753 
 754   // Region covering the object.
 755   RegionData* const region_ptr = addr_to_region_ptr(addr);
 756   HeapWord* result = region_ptr->destination();
 757 
 758   // If the entire Region is live, the new location is region->destination + the
 759   // offset of the object within in the Region.
 760 
 761   // Run some performance tests to determine if this special case pays off.  It
 762   // is worth it for pointers into the dense prefix.  If the optimization to
 763   // avoid pointer updates in regions that only point to the dense prefix is
 764   // ever implemented, this should be revisited.
 765   if (region_ptr->data_size() == RegionSize) {
 766     result += region_offset(addr);
 767     return result;
 768   }
 769 
 770   // Otherwise, the new location is region->destination + block offset + the
 771   // number of live words in the Block that are (a) to the left of addr and (b)
 772   // due to objects that start in the Block.
 773 
 774   // Fill in the block table if necessary.  This is unsynchronized, so multiple
 775   // threads may fill the block table for a region (harmless, since it is
 776   // idempotent).
 777   if (!region_ptr->blocks_filled()) {
 778     PSParallelCompact::fill_blocks(addr_to_region_idx(addr));
 779     region_ptr->set_blocks_filled();
 780   }
 781 
 782   HeapWord* const search_start = block_align_down(addr);
 783   const size_t block_offset = addr_to_block_ptr(addr)->offset();
 784 
 785   const ParMarkBitMap* bitmap = PSParallelCompact::mark_bitmap();
 786   const size_t live = bitmap->live_words_in_range(search_start, oop(addr));
 787   result += block_offset + live;
 788   DEBUG_ONLY(PSParallelCompact::check_new_location(addr, result));
 789   return result;
 790 }
 791 
 792 #ifdef ASSERT
 793 void ParallelCompactData::verify_clear(const PSVirtualSpace* vspace)
 794 {
 795   const size_t* const beg = (const size_t*)vspace->committed_low_addr();
 796   const size_t* const end = (const size_t*)vspace->committed_high_addr();
 797   for (const size_t* p = beg; p < end; ++p) {
 798     assert(*p == 0, "not zero");
 799   }
 800 }
 801 
 802 void ParallelCompactData::verify_clear()
 803 {
 804   verify_clear(_region_vspace);
 805   verify_clear(_block_vspace);
 806 }
 807 #endif  // #ifdef ASSERT
 808 
 809 STWGCTimer          PSParallelCompact::_gc_timer;
 810 ParallelOldTracer   PSParallelCompact::_gc_tracer;
 811 elapsedTimer        PSParallelCompact::_accumulated_time;
 812 unsigned int        PSParallelCompact::_total_invocations = 0;
 813 unsigned int        PSParallelCompact::_maximum_compaction_gc_num = 0;
 814 jlong               PSParallelCompact::_time_of_last_gc = 0;
 815 CollectorCounters*  PSParallelCompact::_counters = NULL;
 816 ParMarkBitMap       PSParallelCompact::_mark_bitmap;
 817 ParallelCompactData PSParallelCompact::_summary_data;
 818 
 819 PSParallelCompact::IsAliveClosure PSParallelCompact::_is_alive_closure;
 820 
 821 bool PSParallelCompact::IsAliveClosure::do_object_b(oop p) { return mark_bitmap()->is_marked(p); }
 822 
 823 PSParallelCompact::AdjustPointerClosure PSParallelCompact::_adjust_pointer_closure;
 824 PSParallelCompact::AdjustKlassClosure PSParallelCompact::_adjust_klass_closure;
 825 
 826 void PSParallelCompact::FollowStackClosure::do_void() { _compaction_manager->follow_marking_stacks(); }
 827 
 828 void PSParallelCompact::FollowKlassClosure::do_klass(Klass* klass) {
 829   klass->oops_do(_mark_and_push_closure);
 830 }
 831 void PSParallelCompact::AdjustKlassClosure::do_klass(Klass* klass) {
 832   klass->oops_do(&PSParallelCompact::_adjust_pointer_closure);
 833 }
 834 
 835 void PSParallelCompact::post_initialize() {
 836   ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
 837   MemRegion mr = heap->reserved_region();
 838   _ref_processor =
 839     new ReferenceProcessor(mr,            // span
 840                            ParallelRefProcEnabled && (ParallelGCThreads > 1), // mt processing
 841                            (int) ParallelGCThreads, // mt processing degree
 842                            true,          // mt discovery
 843                            (int) ParallelGCThreads, // mt discovery degree
 844                            true,          // atomic_discovery
 845                            &_is_alive_closure); // non-header is alive closure
 846   _counters = new CollectorCounters("PSParallelCompact", 1);
 847 
 848   // Initialize static fields in ParCompactionManager.
 849   ParCompactionManager::initialize(mark_bitmap());
 850 }
 851 
 852 bool PSParallelCompact::initialize() {
 853   ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
 854   MemRegion mr = heap->reserved_region();
 855 
 856   // Was the old gen get allocated successfully?
 857   if (!heap->old_gen()->is_allocated()) {
 858     return false;
 859   }
 860 
 861   initialize_space_info();
 862   initialize_dead_wood_limiter();
 863 
 864   if (!_mark_bitmap.initialize(mr)) {
 865     vm_shutdown_during_initialization(
 866       err_msg("Unable to allocate " SIZE_FORMAT "KB bitmaps for parallel "
 867       "garbage collection for the requested " SIZE_FORMAT "KB heap.",
 868       _mark_bitmap.reserved_byte_size()/K, mr.byte_size()/K));
 869     return false;
 870   }
 871 
 872   if (!_summary_data.initialize(mr)) {
 873     vm_shutdown_during_initialization(
 874       err_msg("Unable to allocate " SIZE_FORMAT "KB card tables for parallel "
 875       "garbage collection for the requested " SIZE_FORMAT "KB heap.",
 876       _summary_data.reserved_byte_size()/K, mr.byte_size()/K));
 877     return false;
 878   }
 879 
 880   return true;
 881 }
 882 
 883 void PSParallelCompact::initialize_space_info()
 884 {
 885   memset(&_space_info, 0, sizeof(_space_info));
 886 
 887   ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
 888   PSYoungGen* young_gen = heap->young_gen();
 889 
 890   _space_info[old_space_id].set_space(heap->old_gen()->object_space());
 891   _space_info[eden_space_id].set_space(young_gen->eden_space());
 892   _space_info[from_space_id].set_space(young_gen->from_space());
 893   _space_info[to_space_id].set_space(young_gen->to_space());
 894 
 895   _space_info[old_space_id].set_start_array(heap->old_gen()->start_array());
 896 }
 897 
 898 void PSParallelCompact::initialize_dead_wood_limiter()
 899 {
 900   const size_t max = 100;
 901   _dwl_mean = double(MIN2(ParallelOldDeadWoodLimiterMean, max)) / 100.0;
 902   _dwl_std_dev = double(MIN2(ParallelOldDeadWoodLimiterStdDev, max)) / 100.0;
 903   _dwl_first_term = 1.0 / (sqrt(2.0 * M_PI) * _dwl_std_dev);
 904   DEBUG_ONLY(_dwl_initialized = true;)
 905   _dwl_adjustment = normal_distribution(1.0);
 906 }
 907 
 908 // Simple class for storing info about the heap at the start of GC, to be used
 909 // after GC for comparison/printing.
 910 class PreGCValues {
 911 public:
 912   PreGCValues() { }
 913   PreGCValues(ParallelScavengeHeap* heap) { fill(heap); }
 914 
 915   void fill(ParallelScavengeHeap* heap) {
 916     _heap_used      = heap->used();
 917     _young_gen_used = heap->young_gen()->used_in_bytes();
 918     _old_gen_used   = heap->old_gen()->used_in_bytes();
 919     _metadata_used  = MetaspaceAux::used_bytes();
 920   };
 921 
 922   size_t heap_used() const      { return _heap_used; }
 923   size_t young_gen_used() const { return _young_gen_used; }
 924   size_t old_gen_used() const   { return _old_gen_used; }
 925   size_t metadata_used() const  { return _metadata_used; }
 926 
 927 private:
 928   size_t _heap_used;
 929   size_t _young_gen_used;
 930   size_t _old_gen_used;
 931   size_t _metadata_used;
 932 };
 933 
 934 void
 935 PSParallelCompact::clear_data_covering_space(SpaceId id)
 936 {
 937   // At this point, top is the value before GC, new_top() is the value that will
 938   // be set at the end of GC.  The marking bitmap is cleared to top; nothing
 939   // should be marked above top.  The summary data is cleared to the larger of
 940   // top & new_top.
 941   MutableSpace* const space = _space_info[id].space();
 942   HeapWord* const bot = space->bottom();
 943   HeapWord* const top = space->top();
 944   HeapWord* const max_top = MAX2(top, _space_info[id].new_top());
 945 
 946   const idx_t beg_bit = _mark_bitmap.addr_to_bit(bot);
 947   const idx_t end_bit = BitMap::word_align_up(_mark_bitmap.addr_to_bit(top));
 948   _mark_bitmap.clear_range(beg_bit, end_bit);
 949 
 950   const size_t beg_region = _summary_data.addr_to_region_idx(bot);
 951   const size_t end_region =
 952     _summary_data.addr_to_region_idx(_summary_data.region_align_up(max_top));
 953   _summary_data.clear_range(beg_region, end_region);
 954 
 955   // Clear the data used to 'split' regions.
 956   SplitInfo& split_info = _space_info[id].split_info();
 957   if (split_info.is_valid()) {
 958     split_info.clear();
 959   }
 960   DEBUG_ONLY(split_info.verify_clear();)
 961 }
 962 
 963 void PSParallelCompact::pre_compact(PreGCValues* pre_gc_values)
 964 {
 965   // Update the from & to space pointers in space_info, since they are swapped
 966   // at each young gen gc.  Do the update unconditionally (even though a
 967   // promotion failure does not swap spaces) because an unknown number of minor
 968   // collections will have swapped the spaces an unknown number of times.
 969   GCTraceTime tm("pre compact", print_phases(), true, &_gc_timer, _gc_tracer.gc_id());
 970   ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
 971   _space_info[from_space_id].set_space(heap->young_gen()->from_space());
 972   _space_info[to_space_id].set_space(heap->young_gen()->to_space());
 973 
 974   pre_gc_values->fill(heap);
 975 
 976   DEBUG_ONLY(add_obj_count = add_obj_size = 0;)
 977   DEBUG_ONLY(mark_bitmap_count = mark_bitmap_size = 0;)
 978 
 979   // Increment the invocation count
 980   heap->increment_total_collections(true);
 981 
 982   // We need to track unique mark sweep invocations as well.
 983   _total_invocations++;
 984 
 985   heap->print_heap_before_gc();
 986   heap->trace_heap_before_gc(&_gc_tracer);
 987 
 988   // Fill in TLABs
 989   heap->accumulate_statistics_all_tlabs();
 990   heap->ensure_parsability(true);  // retire TLABs
 991 
 992   if (VerifyBeforeGC && heap->total_collections() >= VerifyGCStartAt) {
 993     HandleMark hm;  // Discard invalid handles created during verification
 994     Universe::verify(" VerifyBeforeGC:");
 995   }
 996 
 997   // Verify object start arrays
 998   if (VerifyObjectStartArray &&
 999       VerifyBeforeGC) {
1000     heap->old_gen()->verify_object_start_array();
1001   }
1002 
1003   DEBUG_ONLY(mark_bitmap()->verify_clear();)
1004   DEBUG_ONLY(summary_data().verify_clear();)
1005 
1006   // Have worker threads release resources the next time they run a task.
1007   gc_task_manager()->release_all_resources();
1008 }
1009 
1010 void PSParallelCompact::post_compact()
1011 {
1012   GCTraceTime tm("post compact", print_phases(), true, &_gc_timer, _gc_tracer.gc_id());
1013 
1014   for (unsigned int id = old_space_id; id < last_space_id; ++id) {
1015     // Clear the marking bitmap, summary data and split info.
1016     clear_data_covering_space(SpaceId(id));
1017     // Update top().  Must be done after clearing the bitmap and summary data.
1018     _space_info[id].publish_new_top();
1019   }
1020 
1021   MutableSpace* const eden_space = _space_info[eden_space_id].space();
1022   MutableSpace* const from_space = _space_info[from_space_id].space();
1023   MutableSpace* const to_space   = _space_info[to_space_id].space();
1024 
1025   ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
1026   bool eden_empty = eden_space->is_empty();
1027   if (!eden_empty) {
1028     eden_empty = absorb_live_data_from_eden(heap->size_policy(),
1029                                             heap->young_gen(), heap->old_gen());
1030   }
1031 
1032   // Update heap occupancy information which is used as input to the soft ref
1033   // clearing policy at the next gc.
1034   Universe::update_heap_info_at_gc();
1035 
1036   bool young_gen_empty = eden_empty && from_space->is_empty() &&
1037     to_space->is_empty();
1038 
1039   ModRefBarrierSet* modBS = barrier_set_cast<ModRefBarrierSet>(heap->barrier_set());
1040   MemRegion old_mr = heap->old_gen()->reserved();
1041   if (young_gen_empty) {
1042     modBS->clear(MemRegion(old_mr.start(), old_mr.end()));
1043   } else {
1044     modBS->invalidate(MemRegion(old_mr.start(), old_mr.end()));
1045   }
1046 
1047   // Delete metaspaces for unloaded class loaders and clean up loader_data graph
1048   ClassLoaderDataGraph::purge();
1049   MetaspaceAux::verify_metrics();
1050 
1051   CodeCache::gc_epilogue();
1052   JvmtiExport::gc_epilogue();
1053 
1054   COMPILER2_PRESENT(DerivedPointerTable::update_pointers());
1055 
1056   ref_processor()->enqueue_discovered_references(NULL);
1057 
1058   if (ZapUnusedHeapArea) {
1059     heap->gen_mangle_unused_area();
1060   }
1061 
1062   // Update time of last GC
1063   reset_millis_since_last_gc();
1064 }
1065 
1066 HeapWord*
1067 PSParallelCompact::compute_dense_prefix_via_density(const SpaceId id,
1068                                                     bool maximum_compaction)
1069 {
1070   const size_t region_size = ParallelCompactData::RegionSize;
1071   const ParallelCompactData& sd = summary_data();
1072 
1073   const MutableSpace* const space = _space_info[id].space();
1074   HeapWord* const top_aligned_up = sd.region_align_up(space->top());
1075   const RegionData* const beg_cp = sd.addr_to_region_ptr(space->bottom());
1076   const RegionData* const end_cp = sd.addr_to_region_ptr(top_aligned_up);
1077 
1078   // Skip full regions at the beginning of the space--they are necessarily part
1079   // of the dense prefix.
1080   size_t full_count = 0;
1081   const RegionData* cp;
1082   for (cp = beg_cp; cp < end_cp && cp->data_size() == region_size; ++cp) {
1083     ++full_count;
1084   }
1085 
1086   assert(total_invocations() >= _maximum_compaction_gc_num, "sanity");
1087   const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num;
1088   const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval;
1089   if (maximum_compaction || cp == end_cp || interval_ended) {
1090     _maximum_compaction_gc_num = total_invocations();
1091     return sd.region_to_addr(cp);
1092   }
1093 
1094   HeapWord* const new_top = _space_info[id].new_top();
1095   const size_t space_live = pointer_delta(new_top, space->bottom());
1096   const size_t space_used = space->used_in_words();
1097   const size_t space_capacity = space->capacity_in_words();
1098 
1099   const double cur_density = double(space_live) / space_capacity;
1100   const double deadwood_density =
1101     (1.0 - cur_density) * (1.0 - cur_density) * cur_density * cur_density;
1102   const size_t deadwood_goal = size_t(space_capacity * deadwood_density);
1103 
1104   if (TraceParallelOldGCDensePrefix) {
1105     tty->print_cr("cur_dens=%5.3f dw_dens=%5.3f dw_goal=" SIZE_FORMAT,
1106                   cur_density, deadwood_density, deadwood_goal);
1107     tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " "
1108                   "space_cap=" SIZE_FORMAT,
1109                   space_live, space_used,
1110                   space_capacity);
1111   }
1112 
1113   // XXX - Use binary search?
1114   HeapWord* dense_prefix = sd.region_to_addr(cp);
1115   const RegionData* full_cp = cp;
1116   const RegionData* const top_cp = sd.addr_to_region_ptr(space->top() - 1);
1117   while (cp < end_cp) {
1118     HeapWord* region_destination = cp->destination();
1119     const size_t cur_deadwood = pointer_delta(dense_prefix, region_destination);
1120     if (TraceParallelOldGCDensePrefix && Verbose) {
1121       tty->print_cr("c#=" SIZE_FORMAT_W(4) " dst=" PTR_FORMAT " "
1122                     "dp=" PTR_FORMAT " " "cdw=" SIZE_FORMAT_W(8),
1123                     sd.region(cp), p2i(region_destination),
1124                     p2i(dense_prefix), cur_deadwood);
1125     }
1126 
1127     if (cur_deadwood >= deadwood_goal) {
1128       // Found the region that has the correct amount of deadwood to the left.
1129       // This typically occurs after crossing a fairly sparse set of regions, so
1130       // iterate backwards over those sparse regions, looking for the region
1131       // that has the lowest density of live objects 'to the right.'
1132       size_t space_to_left = sd.region(cp) * region_size;
1133       size_t live_to_left = space_to_left - cur_deadwood;
1134       size_t space_to_right = space_capacity - space_to_left;
1135       size_t live_to_right = space_live - live_to_left;
1136       double density_to_right = double(live_to_right) / space_to_right;
1137       while (cp > full_cp) {
1138         --cp;
1139         const size_t prev_region_live_to_right = live_to_right -
1140           cp->data_size();
1141         const size_t prev_region_space_to_right = space_to_right + region_size;
1142         double prev_region_density_to_right =
1143           double(prev_region_live_to_right) / prev_region_space_to_right;
1144         if (density_to_right <= prev_region_density_to_right) {
1145           return dense_prefix;
1146         }
1147         if (TraceParallelOldGCDensePrefix && Verbose) {
1148           tty->print_cr("backing up from c=" SIZE_FORMAT_W(4) " d2r=%10.8f "
1149                         "pc_d2r=%10.8f", sd.region(cp), density_to_right,
1150                         prev_region_density_to_right);
1151         }
1152         dense_prefix -= region_size;
1153         live_to_right = prev_region_live_to_right;
1154         space_to_right = prev_region_space_to_right;
1155         density_to_right = prev_region_density_to_right;
1156       }
1157       return dense_prefix;
1158     }
1159 
1160     dense_prefix += region_size;
1161     ++cp;
1162   }
1163 
1164   return dense_prefix;
1165 }
1166 
1167 #ifndef PRODUCT
1168 void PSParallelCompact::print_dense_prefix_stats(const char* const algorithm,
1169                                                  const SpaceId id,
1170                                                  const bool maximum_compaction,
1171                                                  HeapWord* const addr)
1172 {
1173   const size_t region_idx = summary_data().addr_to_region_idx(addr);
1174   RegionData* const cp = summary_data().region(region_idx);
1175   const MutableSpace* const space = _space_info[id].space();
1176   HeapWord* const new_top = _space_info[id].new_top();
1177 
1178   const size_t space_live = pointer_delta(new_top, space->bottom());
1179   const size_t dead_to_left = pointer_delta(addr, cp->destination());
1180   const size_t space_cap = space->capacity_in_words();
1181   const double dead_to_left_pct = double(dead_to_left) / space_cap;
1182   const size_t live_to_right = new_top - cp->destination();
1183   const size_t dead_to_right = space->top() - addr - live_to_right;
1184 
1185   tty->print_cr("%s=" PTR_FORMAT " dpc=" SIZE_FORMAT_W(5) " "
1186                 "spl=" SIZE_FORMAT " "
1187                 "d2l=" SIZE_FORMAT " d2l%%=%6.4f "
1188                 "d2r=" SIZE_FORMAT " l2r=" SIZE_FORMAT
1189                 " ratio=%10.8f",
1190                 algorithm, p2i(addr), region_idx,
1191                 space_live,
1192                 dead_to_left, dead_to_left_pct,
1193                 dead_to_right, live_to_right,
1194                 double(dead_to_right) / live_to_right);
1195 }
1196 #endif  // #ifndef PRODUCT
1197 
1198 // Return a fraction indicating how much of the generation can be treated as
1199 // "dead wood" (i.e., not reclaimed).  The function uses a normal distribution
1200 // based on the density of live objects in the generation to determine a limit,
1201 // which is then adjusted so the return value is min_percent when the density is
1202 // 1.
1203 //
1204 // The following table shows some return values for a different values of the
1205 // standard deviation (ParallelOldDeadWoodLimiterStdDev); the mean is 0.5 and
1206 // min_percent is 1.
1207 //
1208 //                          fraction allowed as dead wood
1209 //         -----------------------------------------------------------------
1210 // density std_dev=70 std_dev=75 std_dev=80 std_dev=85 std_dev=90 std_dev=95
1211 // ------- ---------- ---------- ---------- ---------- ---------- ----------
1212 // 0.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000
1213 // 0.05000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941
1214 // 0.10000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272
1215 // 0.15000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066
1216 // 0.20000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975
1217 // 0.25000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313
1218 // 0.30000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132
1219 // 0.35000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289
1220 // 0.40000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500
1221 // 0.45000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386
1222 // 0.50000 0.13832410 0.11599237 0.09847664 0.08456518 0.07338887 0.06431510
1223 // 0.55000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386
1224 // 0.60000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500
1225 // 0.65000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289
1226 // 0.70000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132
1227 // 0.75000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313
1228 // 0.80000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975
1229 // 0.85000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066
1230 // 0.90000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272
1231 // 0.95000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941
1232 // 1.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000
1233 
1234 double PSParallelCompact::dead_wood_limiter(double density, size_t min_percent)
1235 {
1236   assert(_dwl_initialized, "uninitialized");
1237 
1238   // The raw limit is the value of the normal distribution at x = density.
1239   const double raw_limit = normal_distribution(density);
1240 
1241   // Adjust the raw limit so it becomes the minimum when the density is 1.
1242   //
1243   // First subtract the adjustment value (which is simply the precomputed value
1244   // normal_distribution(1.0)); this yields a value of 0 when the density is 1.
1245   // Then add the minimum value, so the minimum is returned when the density is
1246   // 1.  Finally, prevent negative values, which occur when the mean is not 0.5.
1247   const double min = double(min_percent) / 100.0;
1248   const double limit = raw_limit - _dwl_adjustment + min;
1249   return MAX2(limit, 0.0);
1250 }
1251 
1252 ParallelCompactData::RegionData*
1253 PSParallelCompact::first_dead_space_region(const RegionData* beg,
1254                                            const RegionData* end)
1255 {
1256   const size_t region_size = ParallelCompactData::RegionSize;
1257   ParallelCompactData& sd = summary_data();
1258   size_t left = sd.region(beg);
1259   size_t right = end > beg ? sd.region(end) - 1 : left;
1260 
1261   // Binary search.
1262   while (left < right) {
1263     // Equivalent to (left + right) / 2, but does not overflow.
1264     const size_t middle = left + (right - left) / 2;
1265     RegionData* const middle_ptr = sd.region(middle);
1266     HeapWord* const dest = middle_ptr->destination();
1267     HeapWord* const addr = sd.region_to_addr(middle);
1268     assert(dest != NULL, "sanity");
1269     assert(dest <= addr, "must move left");
1270 
1271     if (middle > left && dest < addr) {
1272       right = middle - 1;
1273     } else if (middle < right && middle_ptr->data_size() == region_size) {
1274       left = middle + 1;
1275     } else {
1276       return middle_ptr;
1277     }
1278   }
1279   return sd.region(left);
1280 }
1281 
1282 ParallelCompactData::RegionData*
1283 PSParallelCompact::dead_wood_limit_region(const RegionData* beg,
1284                                           const RegionData* end,
1285                                           size_t dead_words)
1286 {
1287   ParallelCompactData& sd = summary_data();
1288   size_t left = sd.region(beg);
1289   size_t right = end > beg ? sd.region(end) - 1 : left;
1290 
1291   // Binary search.
1292   while (left < right) {
1293     // Equivalent to (left + right) / 2, but does not overflow.
1294     const size_t middle = left + (right - left) / 2;
1295     RegionData* const middle_ptr = sd.region(middle);
1296     HeapWord* const dest = middle_ptr->destination();
1297     HeapWord* const addr = sd.region_to_addr(middle);
1298     assert(dest != NULL, "sanity");
1299     assert(dest <= addr, "must move left");
1300 
1301     const size_t dead_to_left = pointer_delta(addr, dest);
1302     if (middle > left && dead_to_left > dead_words) {
1303       right = middle - 1;
1304     } else if (middle < right && dead_to_left < dead_words) {
1305       left = middle + 1;
1306     } else {
1307       return middle_ptr;
1308     }
1309   }
1310   return sd.region(left);
1311 }
1312 
1313 // The result is valid during the summary phase, after the initial summarization
1314 // of each space into itself, and before final summarization.
1315 inline double
1316 PSParallelCompact::reclaimed_ratio(const RegionData* const cp,
1317                                    HeapWord* const bottom,
1318                                    HeapWord* const top,
1319                                    HeapWord* const new_top)
1320 {
1321   ParallelCompactData& sd = summary_data();
1322 
1323   assert(cp != NULL, "sanity");
1324   assert(bottom != NULL, "sanity");
1325   assert(top != NULL, "sanity");
1326   assert(new_top != NULL, "sanity");
1327   assert(top >= new_top, "summary data problem?");
1328   assert(new_top > bottom, "space is empty; should not be here");
1329   assert(new_top >= cp->destination(), "sanity");
1330   assert(top >= sd.region_to_addr(cp), "sanity");
1331 
1332   HeapWord* const destination = cp->destination();
1333   const size_t dense_prefix_live  = pointer_delta(destination, bottom);
1334   const size_t compacted_region_live = pointer_delta(new_top, destination);
1335   const size_t compacted_region_used = pointer_delta(top,
1336                                                      sd.region_to_addr(cp));
1337   const size_t reclaimable = compacted_region_used - compacted_region_live;
1338 
1339   const double divisor = dense_prefix_live + 1.25 * compacted_region_live;
1340   return double(reclaimable) / divisor;
1341 }
1342 
1343 // Return the address of the end of the dense prefix, a.k.a. the start of the
1344 // compacted region.  The address is always on a region boundary.
1345 //
1346 // Completely full regions at the left are skipped, since no compaction can
1347 // occur in those regions.  Then the maximum amount of dead wood to allow is
1348 // computed, based on the density (amount live / capacity) of the generation;
1349 // the region with approximately that amount of dead space to the left is
1350 // identified as the limit region.  Regions between the last completely full
1351 // region and the limit region are scanned and the one that has the best
1352 // (maximum) reclaimed_ratio() is selected.
1353 HeapWord*
1354 PSParallelCompact::compute_dense_prefix(const SpaceId id,
1355                                         bool maximum_compaction)
1356 {
1357   if (ParallelOldGCSplitALot) {
1358     if (_space_info[id].dense_prefix() != _space_info[id].space()->bottom()) {
1359       // The value was chosen to provoke splitting a young gen space; use it.
1360       return _space_info[id].dense_prefix();
1361     }
1362   }
1363 
1364   const size_t region_size = ParallelCompactData::RegionSize;
1365   const ParallelCompactData& sd = summary_data();
1366 
1367   const MutableSpace* const space = _space_info[id].space();
1368   HeapWord* const top = space->top();
1369   HeapWord* const top_aligned_up = sd.region_align_up(top);
1370   HeapWord* const new_top = _space_info[id].new_top();
1371   HeapWord* const new_top_aligned_up = sd.region_align_up(new_top);
1372   HeapWord* const bottom = space->bottom();
1373   const RegionData* const beg_cp = sd.addr_to_region_ptr(bottom);
1374   const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up);
1375   const RegionData* const new_top_cp =
1376     sd.addr_to_region_ptr(new_top_aligned_up);
1377 
1378   // Skip full regions at the beginning of the space--they are necessarily part
1379   // of the dense prefix.
1380   const RegionData* const full_cp = first_dead_space_region(beg_cp, new_top_cp);
1381   assert(full_cp->destination() == sd.region_to_addr(full_cp) ||
1382          space->is_empty(), "no dead space allowed to the left");
1383   assert(full_cp->data_size() < region_size || full_cp == new_top_cp - 1,
1384          "region must have dead space");
1385 
1386   // The gc number is saved whenever a maximum compaction is done, and used to
1387   // determine when the maximum compaction interval has expired.  This avoids
1388   // successive max compactions for different reasons.
1389   assert(total_invocations() >= _maximum_compaction_gc_num, "sanity");
1390   const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num;
1391   const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval ||
1392     total_invocations() == HeapFirstMaximumCompactionCount;
1393   if (maximum_compaction || full_cp == top_cp || interval_ended) {
1394     _maximum_compaction_gc_num = total_invocations();
1395     return sd.region_to_addr(full_cp);
1396   }
1397 
1398   const size_t space_live = pointer_delta(new_top, bottom);
1399   const size_t space_used = space->used_in_words();
1400   const size_t space_capacity = space->capacity_in_words();
1401 
1402   const double density = double(space_live) / double(space_capacity);
1403   const size_t min_percent_free = MarkSweepDeadRatio;
1404   const double limiter = dead_wood_limiter(density, min_percent_free);
1405   const size_t dead_wood_max = space_used - space_live;
1406   const size_t dead_wood_limit = MIN2(size_t(space_capacity * limiter),
1407                                       dead_wood_max);
1408 
1409   if (TraceParallelOldGCDensePrefix) {
1410     tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " "
1411                   "space_cap=" SIZE_FORMAT,
1412                   space_live, space_used,
1413                   space_capacity);
1414     tty->print_cr("dead_wood_limiter(%6.4f, " SIZE_FORMAT ")=%6.4f "
1415                   "dead_wood_max=" SIZE_FORMAT " dead_wood_limit=" SIZE_FORMAT,
1416                   density, min_percent_free, limiter,
1417                   dead_wood_max, dead_wood_limit);
1418   }
1419 
1420   // Locate the region with the desired amount of dead space to the left.
1421   const RegionData* const limit_cp =
1422     dead_wood_limit_region(full_cp, top_cp, dead_wood_limit);
1423 
1424   // Scan from the first region with dead space to the limit region and find the
1425   // one with the best (largest) reclaimed ratio.
1426   double best_ratio = 0.0;
1427   const RegionData* best_cp = full_cp;
1428   for (const RegionData* cp = full_cp; cp < limit_cp; ++cp) {
1429     double tmp_ratio = reclaimed_ratio(cp, bottom, top, new_top);
1430     if (tmp_ratio > best_ratio) {
1431       best_cp = cp;
1432       best_ratio = tmp_ratio;
1433     }
1434   }
1435 
1436 #if     0
1437   // Something to consider:  if the region with the best ratio is 'close to' the
1438   // first region w/free space, choose the first region with free space
1439   // ("first-free").  The first-free region is usually near the start of the
1440   // heap, which means we are copying most of the heap already, so copy a bit
1441   // more to get complete compaction.
1442   if (pointer_delta(best_cp, full_cp, sizeof(RegionData)) < 4) {
1443     _maximum_compaction_gc_num = total_invocations();
1444     best_cp = full_cp;
1445   }
1446 #endif  // #if 0
1447 
1448   return sd.region_to_addr(best_cp);
1449 }
1450 
1451 #ifndef PRODUCT
1452 void
1453 PSParallelCompact::fill_with_live_objects(SpaceId id, HeapWord* const start,
1454                                           size_t words)
1455 {
1456   if (TraceParallelOldGCSummaryPhase) {
1457     tty->print_cr("fill_with_live_objects [" PTR_FORMAT " " PTR_FORMAT ") "
1458                   SIZE_FORMAT, p2i(start), p2i(start + words), words);
1459   }
1460 
1461   ObjectStartArray* const start_array = _space_info[id].start_array();
1462   CollectedHeap::fill_with_objects(start, words);
1463   for (HeapWord* p = start; p < start + words; p += oop(p)->size()) {
1464     _mark_bitmap.mark_obj(p, words);
1465     _summary_data.add_obj(p, words);
1466     start_array->allocate_block(p);
1467   }
1468 }
1469 
1470 void
1471 PSParallelCompact::summarize_new_objects(SpaceId id, HeapWord* start)
1472 {
1473   ParallelCompactData& sd = summary_data();
1474   MutableSpace* space = _space_info[id].space();
1475 
1476   // Find the source and destination start addresses.
1477   HeapWord* const src_addr = sd.region_align_down(start);
1478   HeapWord* dst_addr;
1479   if (src_addr < start) {
1480     dst_addr = sd.addr_to_region_ptr(src_addr)->destination();
1481   } else if (src_addr > space->bottom()) {
1482     // The start (the original top() value) is aligned to a region boundary so
1483     // the associated region does not have a destination.  Compute the
1484     // destination from the previous region.
1485     RegionData* const cp = sd.addr_to_region_ptr(src_addr) - 1;
1486     dst_addr = cp->destination() + cp->data_size();
1487   } else {
1488     // Filling the entire space.
1489     dst_addr = space->bottom();
1490   }
1491   assert(dst_addr != NULL, "sanity");
1492 
1493   // Update the summary data.
1494   bool result = _summary_data.summarize(_space_info[id].split_info(),
1495                                         src_addr, space->top(), NULL,
1496                                         dst_addr, space->end(),
1497                                         _space_info[id].new_top_addr());
1498   assert(result, "should not fail:  bad filler object size");
1499 }
1500 
1501 void
1502 PSParallelCompact::provoke_split_fill_survivor(SpaceId id)
1503 {
1504   if (total_invocations() % (ParallelOldGCSplitInterval * 3) != 0) {
1505     return;
1506   }
1507 
1508   MutableSpace* const space = _space_info[id].space();
1509   if (space->is_empty()) {
1510     HeapWord* b = space->bottom();
1511     HeapWord* t = b + space->capacity_in_words() / 2;
1512     space->set_top(t);
1513     if (ZapUnusedHeapArea) {
1514       space->set_top_for_allocations();
1515     }
1516 
1517     size_t min_size = CollectedHeap::min_fill_size();
1518     size_t obj_len = min_size;
1519     while (b + obj_len <= t) {
1520       CollectedHeap::fill_with_object(b, obj_len);
1521       mark_bitmap()->mark_obj(b, obj_len);
1522       summary_data().add_obj(b, obj_len);
1523       b += obj_len;
1524       obj_len = (obj_len & (min_size*3)) + min_size; // 8 16 24 32 8 16 24 32 ...
1525     }
1526     if (b < t) {
1527       // The loop didn't completely fill to t (top); adjust top downward.
1528       space->set_top(b);
1529       if (ZapUnusedHeapArea) {
1530         space->set_top_for_allocations();
1531       }
1532     }
1533 
1534     HeapWord** nta = _space_info[id].new_top_addr();
1535     bool result = summary_data().summarize(_space_info[id].split_info(),
1536                                            space->bottom(), space->top(), NULL,
1537                                            space->bottom(), space->end(), nta);
1538     assert(result, "space must fit into itself");
1539   }
1540 }
1541 
1542 void
1543 PSParallelCompact::provoke_split(bool & max_compaction)
1544 {
1545   if (total_invocations() % ParallelOldGCSplitInterval != 0) {
1546     return;
1547   }
1548 
1549   const size_t region_size = ParallelCompactData::RegionSize;
1550   ParallelCompactData& sd = summary_data();
1551 
1552   MutableSpace* const eden_space = _space_info[eden_space_id].space();
1553   MutableSpace* const from_space = _space_info[from_space_id].space();
1554   const size_t eden_live = pointer_delta(eden_space->top(),
1555                                          _space_info[eden_space_id].new_top());
1556   const size_t from_live = pointer_delta(from_space->top(),
1557                                          _space_info[from_space_id].new_top());
1558 
1559   const size_t min_fill_size = CollectedHeap::min_fill_size();
1560   const size_t eden_free = pointer_delta(eden_space->end(), eden_space->top());
1561   const size_t eden_fillable = eden_free >= min_fill_size ? eden_free : 0;
1562   const size_t from_free = pointer_delta(from_space->end(), from_space->top());
1563   const size_t from_fillable = from_free >= min_fill_size ? from_free : 0;
1564 
1565   // Choose the space to split; need at least 2 regions live (or fillable).
1566   SpaceId id;
1567   MutableSpace* space;
1568   size_t live_words;
1569   size_t fill_words;
1570   if (eden_live + eden_fillable >= region_size * 2) {
1571     id = eden_space_id;
1572     space = eden_space;
1573     live_words = eden_live;
1574     fill_words = eden_fillable;
1575   } else if (from_live + from_fillable >= region_size * 2) {
1576     id = from_space_id;
1577     space = from_space;
1578     live_words = from_live;
1579     fill_words = from_fillable;
1580   } else {
1581     return; // Give up.
1582   }
1583   assert(fill_words == 0 || fill_words >= min_fill_size, "sanity");
1584 
1585   if (live_words < region_size * 2) {
1586     // Fill from top() to end() w/live objects of mixed sizes.
1587     HeapWord* const fill_start = space->top();
1588     live_words += fill_words;
1589 
1590     space->set_top(fill_start + fill_words);
1591     if (ZapUnusedHeapArea) {
1592       space->set_top_for_allocations();
1593     }
1594 
1595     HeapWord* cur_addr = fill_start;
1596     while (fill_words > 0) {
1597       const size_t r = (size_t)os::random() % (region_size / 2) + min_fill_size;
1598       size_t cur_size = MIN2(align_object_size_(r), fill_words);
1599       if (fill_words - cur_size < min_fill_size) {
1600         cur_size = fill_words; // Avoid leaving a fragment too small to fill.
1601       }
1602 
1603       CollectedHeap::fill_with_object(cur_addr, cur_size);
1604       mark_bitmap()->mark_obj(cur_addr, cur_size);
1605       sd.add_obj(cur_addr, cur_size);
1606 
1607       cur_addr += cur_size;
1608       fill_words -= cur_size;
1609     }
1610 
1611     summarize_new_objects(id, fill_start);
1612   }
1613 
1614   max_compaction = false;
1615 
1616   // Manipulate the old gen so that it has room for about half of the live data
1617   // in the target young gen space (live_words / 2).
1618   id = old_space_id;
1619   space = _space_info[id].space();
1620   const size_t free_at_end = space->free_in_words();
1621   const size_t free_target = align_object_size(live_words / 2);
1622   const size_t dead = pointer_delta(space->top(), _space_info[id].new_top());
1623 
1624   if (free_at_end >= free_target + min_fill_size) {
1625     // Fill space above top() and set the dense prefix so everything survives.
1626     HeapWord* const fill_start = space->top();
1627     const size_t fill_size = free_at_end - free_target;
1628     space->set_top(space->top() + fill_size);
1629     if (ZapUnusedHeapArea) {
1630       space->set_top_for_allocations();
1631     }
1632     fill_with_live_objects(id, fill_start, fill_size);
1633     summarize_new_objects(id, fill_start);
1634     _space_info[id].set_dense_prefix(sd.region_align_down(space->top()));
1635   } else if (dead + free_at_end > free_target) {
1636     // Find a dense prefix that makes the right amount of space available.
1637     HeapWord* cur = sd.region_align_down(space->top());
1638     HeapWord* cur_destination = sd.addr_to_region_ptr(cur)->destination();
1639     size_t dead_to_right = pointer_delta(space->end(), cur_destination);
1640     while (dead_to_right < free_target) {
1641       cur -= region_size;
1642       cur_destination = sd.addr_to_region_ptr(cur)->destination();
1643       dead_to_right = pointer_delta(space->end(), cur_destination);
1644     }
1645     _space_info[id].set_dense_prefix(cur);
1646   }
1647 }
1648 #endif // #ifndef PRODUCT
1649 
1650 void PSParallelCompact::summarize_spaces_quick()
1651 {
1652   for (unsigned int i = 0; i < last_space_id; ++i) {
1653     const MutableSpace* space = _space_info[i].space();
1654     HeapWord** nta = _space_info[i].new_top_addr();
1655     bool result = _summary_data.summarize(_space_info[i].split_info(),
1656                                           space->bottom(), space->top(), NULL,
1657                                           space->bottom(), space->end(), nta);
1658     assert(result, "space must fit into itself");
1659     _space_info[i].set_dense_prefix(space->bottom());
1660   }
1661 
1662 #ifndef PRODUCT
1663   if (ParallelOldGCSplitALot) {
1664     provoke_split_fill_survivor(to_space_id);
1665   }
1666 #endif // #ifndef PRODUCT
1667 }
1668 
1669 void PSParallelCompact::fill_dense_prefix_end(SpaceId id)
1670 {
1671   HeapWord* const dense_prefix_end = dense_prefix(id);
1672   const RegionData* region = _summary_data.addr_to_region_ptr(dense_prefix_end);
1673   const idx_t dense_prefix_bit = _mark_bitmap.addr_to_bit(dense_prefix_end);
1674   if (dead_space_crosses_boundary(region, dense_prefix_bit)) {
1675     // Only enough dead space is filled so that any remaining dead space to the
1676     // left is larger than the minimum filler object.  (The remainder is filled
1677     // during the copy/update phase.)
1678     //
1679     // The size of the dead space to the right of the boundary is not a
1680     // concern, since compaction will be able to use whatever space is
1681     // available.
1682     //
1683     // Here '||' is the boundary, 'x' represents a don't care bit and a box
1684     // surrounds the space to be filled with an object.
1685     //
1686     // In the 32-bit VM, each bit represents two 32-bit words:
1687     //                              +---+
1688     // a) beg_bits:  ...  x   x   x | 0 | ||   0   x  x  ...
1689     //    end_bits:  ...  x   x   x | 0 | ||   0   x  x  ...
1690     //                              +---+
1691     //
1692     // In the 64-bit VM, each bit represents one 64-bit word:
1693     //                              +------------+
1694     // b) beg_bits:  ...  x   x   x | 0   ||   0 | x  x  ...
1695     //    end_bits:  ...  x   x   1 | 0   ||   0 | x  x  ...
1696     //                              +------------+
1697     //                          +-------+
1698     // c) beg_bits:  ...  x   x | 0   0 | ||   0   x  x  ...
1699     //    end_bits:  ...  x   1 | 0   0 | ||   0   x  x  ...
1700     //                          +-------+
1701     //                      +-----------+
1702     // d) beg_bits:  ...  x | 0   0   0 | ||   0   x  x  ...
1703     //    end_bits:  ...  1 | 0   0   0 | ||   0   x  x  ...
1704     //                      +-----------+
1705     //                          +-------+
1706     // e) beg_bits:  ...  0   0 | 0   0 | ||   0   x  x  ...
1707     //    end_bits:  ...  0   0 | 0   0 | ||   0   x  x  ...
1708     //                          +-------+
1709 
1710     // Initially assume case a, c or e will apply.
1711     size_t obj_len = CollectedHeap::min_fill_size();
1712     HeapWord* obj_beg = dense_prefix_end - obj_len;
1713 
1714 #ifdef  _LP64
1715     if (MinObjAlignment > 1) { // object alignment > heap word size
1716       // Cases a, c or e.
1717     } else if (_mark_bitmap.is_obj_end(dense_prefix_bit - 2)) {
1718       // Case b above.
1719       obj_beg = dense_prefix_end - 1;
1720     } else if (!_mark_bitmap.is_obj_end(dense_prefix_bit - 3) &&
1721                _mark_bitmap.is_obj_end(dense_prefix_bit - 4)) {
1722       // Case d above.
1723       obj_beg = dense_prefix_end - 3;
1724       obj_len = 3;
1725     }
1726 #endif  // #ifdef _LP64
1727 
1728     CollectedHeap::fill_with_object(obj_beg, obj_len);
1729     _mark_bitmap.mark_obj(obj_beg, obj_len);
1730     _summary_data.add_obj(obj_beg, obj_len);
1731     assert(start_array(id) != NULL, "sanity");
1732     start_array(id)->allocate_block(obj_beg);
1733   }
1734 }
1735 
1736 void
1737 PSParallelCompact::clear_source_region(HeapWord* beg_addr, HeapWord* end_addr)
1738 {
1739   RegionData* const beg_ptr = _summary_data.addr_to_region_ptr(beg_addr);
1740   HeapWord* const end_aligned_up = _summary_data.region_align_up(end_addr);
1741   RegionData* const end_ptr = _summary_data.addr_to_region_ptr(end_aligned_up);
1742   for (RegionData* cur = beg_ptr; cur < end_ptr; ++cur) {
1743     cur->set_source_region(0);
1744   }
1745 }
1746 
1747 void
1748 PSParallelCompact::summarize_space(SpaceId id, bool maximum_compaction)
1749 {
1750   assert(id < last_space_id, "id out of range");
1751   assert(_space_info[id].dense_prefix() == _space_info[id].space()->bottom() ||
1752          ParallelOldGCSplitALot && id == old_space_id,
1753          "should have been reset in summarize_spaces_quick()");
1754 
1755   const MutableSpace* space = _space_info[id].space();
1756   if (_space_info[id].new_top() != space->bottom()) {
1757     HeapWord* dense_prefix_end = compute_dense_prefix(id, maximum_compaction);
1758     _space_info[id].set_dense_prefix(dense_prefix_end);
1759 
1760 #ifndef PRODUCT
1761     if (TraceParallelOldGCDensePrefix) {
1762       print_dense_prefix_stats("ratio", id, maximum_compaction,
1763                                dense_prefix_end);
1764       HeapWord* addr = compute_dense_prefix_via_density(id, maximum_compaction);
1765       print_dense_prefix_stats("density", id, maximum_compaction, addr);
1766     }
1767 #endif  // #ifndef PRODUCT
1768 
1769     // Recompute the summary data, taking into account the dense prefix.  If
1770     // every last byte will be reclaimed, then the existing summary data which
1771     // compacts everything can be left in place.
1772     if (!maximum_compaction && dense_prefix_end != space->bottom()) {
1773       // If dead space crosses the dense prefix boundary, it is (at least
1774       // partially) filled with a dummy object, marked live and added to the
1775       // summary data.  This simplifies the copy/update phase and must be done
1776       // before the final locations of objects are determined, to prevent
1777       // leaving a fragment of dead space that is too small to fill.
1778       fill_dense_prefix_end(id);
1779 
1780       // Compute the destination of each Region, and thus each object.
1781       _summary_data.summarize_dense_prefix(space->bottom(), dense_prefix_end);
1782       _summary_data.summarize(_space_info[id].split_info(),
1783                               dense_prefix_end, space->top(), NULL,
1784                               dense_prefix_end, space->end(),
1785                               _space_info[id].new_top_addr());
1786     }
1787   }
1788 
1789   if (TraceParallelOldGCSummaryPhase) {
1790     const size_t region_size = ParallelCompactData::RegionSize;
1791     HeapWord* const dense_prefix_end = _space_info[id].dense_prefix();
1792     const size_t dp_region = _summary_data.addr_to_region_idx(dense_prefix_end);
1793     const size_t dp_words = pointer_delta(dense_prefix_end, space->bottom());
1794     HeapWord* const new_top = _space_info[id].new_top();
1795     const HeapWord* nt_aligned_up = _summary_data.region_align_up(new_top);
1796     const size_t cr_words = pointer_delta(nt_aligned_up, dense_prefix_end);
1797     tty->print_cr("id=%d cap=" SIZE_FORMAT " dp=" PTR_FORMAT " "
1798                   "dp_region=" SIZE_FORMAT " " "dp_count=" SIZE_FORMAT " "
1799                   "cr_count=" SIZE_FORMAT " " "nt=" PTR_FORMAT,
1800                   id, space->capacity_in_words(), p2i(dense_prefix_end),
1801                   dp_region, dp_words / region_size,
1802                   cr_words / region_size, p2i(new_top));
1803   }
1804 }
1805 
1806 #ifndef PRODUCT
1807 void PSParallelCompact::summary_phase_msg(SpaceId dst_space_id,
1808                                           HeapWord* dst_beg, HeapWord* dst_end,
1809                                           SpaceId src_space_id,
1810                                           HeapWord* src_beg, HeapWord* src_end)
1811 {
1812   if (TraceParallelOldGCSummaryPhase) {
1813     tty->print_cr("summarizing %d [%s] into %d [%s]:  "
1814                   "src=" PTR_FORMAT "-" PTR_FORMAT " "
1815                   SIZE_FORMAT "-" SIZE_FORMAT " "
1816                   "dst=" PTR_FORMAT "-" PTR_FORMAT " "
1817                   SIZE_FORMAT "-" SIZE_FORMAT,
1818                   src_space_id, space_names[src_space_id],
1819                   dst_space_id, space_names[dst_space_id],
1820                   p2i(src_beg), p2i(src_end),
1821                   _summary_data.addr_to_region_idx(src_beg),
1822                   _summary_data.addr_to_region_idx(src_end),
1823                   p2i(dst_beg), p2i(dst_end),
1824                   _summary_data.addr_to_region_idx(dst_beg),
1825                   _summary_data.addr_to_region_idx(dst_end));
1826   }
1827 }
1828 #endif  // #ifndef PRODUCT
1829 
1830 void PSParallelCompact::summary_phase(ParCompactionManager* cm,
1831                                       bool maximum_compaction)
1832 {
1833   GCTraceTime tm("summary phase", print_phases(), true, &_gc_timer, _gc_tracer.gc_id());
1834   // trace("2");
1835 
1836 #ifdef  ASSERT
1837   if (TraceParallelOldGCMarkingPhase) {
1838     tty->print_cr("add_obj_count=" SIZE_FORMAT " "
1839                   "add_obj_bytes=" SIZE_FORMAT,
1840                   add_obj_count, add_obj_size * HeapWordSize);
1841     tty->print_cr("mark_bitmap_count=" SIZE_FORMAT " "
1842                   "mark_bitmap_bytes=" SIZE_FORMAT,
1843                   mark_bitmap_count, mark_bitmap_size * HeapWordSize);
1844   }
1845 #endif  // #ifdef ASSERT
1846 
1847   // Quick summarization of each space into itself, to see how much is live.
1848   summarize_spaces_quick();
1849 
1850   if (TraceParallelOldGCSummaryPhase) {
1851     tty->print_cr("summary_phase:  after summarizing each space to self");
1852     Universe::print();
1853     NOT_PRODUCT(print_region_ranges());
1854     if (Verbose) {
1855       NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info));
1856     }
1857   }
1858 
1859   // The amount of live data that will end up in old space (assuming it fits).
1860   size_t old_space_total_live = 0;
1861   for (unsigned int id = old_space_id; id < last_space_id; ++id) {
1862     old_space_total_live += pointer_delta(_space_info[id].new_top(),
1863                                           _space_info[id].space()->bottom());
1864   }
1865 
1866   MutableSpace* const old_space = _space_info[old_space_id].space();
1867   const size_t old_capacity = old_space->capacity_in_words();
1868   if (old_space_total_live > old_capacity) {
1869     // XXX - should also try to expand
1870     maximum_compaction = true;
1871   }
1872 #ifndef PRODUCT
1873   if (ParallelOldGCSplitALot && old_space_total_live < old_capacity) {
1874     provoke_split(maximum_compaction);
1875   }
1876 #endif // #ifndef PRODUCT
1877 
1878   // Old generations.
1879   summarize_space(old_space_id, maximum_compaction);
1880 
1881   // Summarize the remaining spaces in the young gen.  The initial target space
1882   // is the old gen.  If a space does not fit entirely into the target, then the
1883   // remainder is compacted into the space itself and that space becomes the new
1884   // target.
1885   SpaceId dst_space_id = old_space_id;
1886   HeapWord* dst_space_end = old_space->end();
1887   HeapWord** new_top_addr = _space_info[dst_space_id].new_top_addr();
1888   for (unsigned int id = eden_space_id; id < last_space_id; ++id) {
1889     const MutableSpace* space = _space_info[id].space();
1890     const size_t live = pointer_delta(_space_info[id].new_top(),
1891                                       space->bottom());
1892     const size_t available = pointer_delta(dst_space_end, *new_top_addr);
1893 
1894     NOT_PRODUCT(summary_phase_msg(dst_space_id, *new_top_addr, dst_space_end,
1895                                   SpaceId(id), space->bottom(), space->top());)
1896     if (live > 0 && live <= available) {
1897       // All the live data will fit.
1898       bool done = _summary_data.summarize(_space_info[id].split_info(),
1899                                           space->bottom(), space->top(),
1900                                           NULL,
1901                                           *new_top_addr, dst_space_end,
1902                                           new_top_addr);
1903       assert(done, "space must fit into old gen");
1904 
1905       // Reset the new_top value for the space.
1906       _space_info[id].set_new_top(space->bottom());
1907     } else if (live > 0) {
1908       // Attempt to fit part of the source space into the target space.
1909       HeapWord* next_src_addr = NULL;
1910       bool done = _summary_data.summarize(_space_info[id].split_info(),
1911                                           space->bottom(), space->top(),
1912                                           &next_src_addr,
1913                                           *new_top_addr, dst_space_end,
1914                                           new_top_addr);
1915       assert(!done, "space should not fit into old gen");
1916       assert(next_src_addr != NULL, "sanity");
1917 
1918       // The source space becomes the new target, so the remainder is compacted
1919       // within the space itself.
1920       dst_space_id = SpaceId(id);
1921       dst_space_end = space->end();
1922       new_top_addr = _space_info[id].new_top_addr();
1923       NOT_PRODUCT(summary_phase_msg(dst_space_id,
1924                                     space->bottom(), dst_space_end,
1925                                     SpaceId(id), next_src_addr, space->top());)
1926       done = _summary_data.summarize(_space_info[id].split_info(),
1927                                      next_src_addr, space->top(),
1928                                      NULL,
1929                                      space->bottom(), dst_space_end,
1930                                      new_top_addr);
1931       assert(done, "space must fit when compacted into itself");
1932       assert(*new_top_addr <= space->top(), "usage should not grow");
1933     }
1934   }
1935 
1936   if (TraceParallelOldGCSummaryPhase) {
1937     tty->print_cr("summary_phase:  after final summarization");
1938     Universe::print();
1939     NOT_PRODUCT(print_region_ranges());
1940     if (Verbose) {
1941       NOT_PRODUCT(print_generic_summary_data(_summary_data, _space_info));
1942     }
1943   }
1944 }
1945 
1946 // This method should contain all heap-specific policy for invoking a full
1947 // collection.  invoke_no_policy() will only attempt to compact the heap; it
1948 // will do nothing further.  If we need to bail out for policy reasons, scavenge
1949 // before full gc, or any other specialized behavior, it needs to be added here.
1950 //
1951 // Note that this method should only be called from the vm_thread while at a
1952 // safepoint.
1953 //
1954 // Note that the all_soft_refs_clear flag in the collector policy
1955 // may be true because this method can be called without intervening
1956 // activity.  For example when the heap space is tight and full measure
1957 // are being taken to free space.
1958 void PSParallelCompact::invoke(bool maximum_heap_compaction) {
1959   assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
1960   assert(Thread::current() == (Thread*)VMThread::vm_thread(),
1961          "should be in vm thread");
1962 
1963   ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
1964   GCCause::Cause gc_cause = heap->gc_cause();
1965   assert(!heap->is_gc_active(), "not reentrant");
1966 
1967   PSAdaptiveSizePolicy* policy = heap->size_policy();
1968   IsGCActiveMark mark;
1969 
1970   if (ScavengeBeforeFullGC) {
1971     PSScavenge::invoke_no_policy();
1972   }
1973 
1974   const bool clear_all_soft_refs =
1975     heap->collector_policy()->should_clear_all_soft_refs();
1976 
1977   PSParallelCompact::invoke_no_policy(clear_all_soft_refs ||
1978                                       maximum_heap_compaction);
1979 }
1980 
1981 // This method contains no policy. You should probably
1982 // be calling invoke() instead.
1983 bool PSParallelCompact::invoke_no_policy(bool maximum_heap_compaction) {
1984   assert(SafepointSynchronize::is_at_safepoint(), "must be at a safepoint");
1985   assert(ref_processor() != NULL, "Sanity");
1986 
1987   if (GC_locker::check_active_before_gc()) {
1988     return false;
1989   }
1990 
1991   ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
1992 
1993   _gc_timer.register_gc_start();
1994   _gc_tracer.report_gc_start(heap->gc_cause(), _gc_timer.gc_start());
1995 
1996   TimeStamp marking_start;
1997   TimeStamp compaction_start;
1998   TimeStamp collection_exit;
1999 
2000   GCCause::Cause gc_cause = heap->gc_cause();
2001   PSYoungGen* young_gen = heap->young_gen();
2002   PSOldGen* old_gen = heap->old_gen();
2003   PSAdaptiveSizePolicy* size_policy = heap->size_policy();
2004 
2005   // The scope of casr should end after code that can change
2006   // CollectorPolicy::_should_clear_all_soft_refs.
2007   ClearedAllSoftRefs casr(maximum_heap_compaction,
2008                           heap->collector_policy());
2009 
2010   if (ZapUnusedHeapArea) {
2011     // Save information needed to minimize mangling
2012     heap->record_gen_tops_before_GC();
2013   }
2014 
2015   heap->pre_full_gc_dump(&_gc_timer);
2016 
2017   _print_phases = PrintGCDetails && PrintParallelOldGCPhaseTimes;
2018 
2019   // Make sure data structures are sane, make the heap parsable, and do other
2020   // miscellaneous bookkeeping.
2021   PreGCValues pre_gc_values;
2022   pre_compact(&pre_gc_values);
2023 
2024   // Get the compaction manager reserved for the VM thread.
2025   ParCompactionManager* const vmthread_cm =
2026     ParCompactionManager::manager_array(gc_task_manager()->workers());
2027 
2028   // Place after pre_compact() where the number of invocations is incremented.
2029   AdaptiveSizePolicyOutput(size_policy, heap->total_collections());
2030 
2031   {
2032     ResourceMark rm;
2033     HandleMark hm;
2034 
2035     // Set the number of GC threads to be used in this collection
2036     gc_task_manager()->set_active_gang();
2037     gc_task_manager()->task_idle_workers();
2038     heap->set_par_threads(gc_task_manager()->active_workers());
2039 
2040     TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
2041     GCTraceTime t1(GCCauseString("Full GC", gc_cause), PrintGC, !PrintGCDetails, NULL, _gc_tracer.gc_id());
2042     TraceCollectorStats tcs(counters());
2043     TraceMemoryManagerStats tms(true /* Full GC */,gc_cause);
2044 
2045     if (TraceOldGenTime) accumulated_time()->start();
2046 
2047     // Let the size policy know we're starting
2048     size_policy->major_collection_begin();
2049 
2050     CodeCache::gc_prologue();
2051 
2052     COMPILER2_PRESENT(DerivedPointerTable::clear());
2053 
2054     ref_processor()->enable_discovery();
2055     ref_processor()->setup_policy(maximum_heap_compaction);
2056 
2057     bool marked_for_unloading = false;
2058 
2059     marking_start.update();
2060     marking_phase(vmthread_cm, maximum_heap_compaction, &_gc_tracer);
2061 
2062     bool max_on_system_gc = UseMaximumCompactionOnSystemGC
2063       && gc_cause == GCCause::_java_lang_system_gc;
2064     summary_phase(vmthread_cm, maximum_heap_compaction || max_on_system_gc);
2065 
2066     COMPILER2_PRESENT(assert(DerivedPointerTable::is_active(), "Sanity"));
2067     COMPILER2_PRESENT(DerivedPointerTable::set_active(false));
2068 
2069     // adjust_roots() updates Universe::_intArrayKlassObj which is
2070     // needed by the compaction for filling holes in the dense prefix.
2071     adjust_roots();
2072 
2073     compaction_start.update();
2074     compact();
2075 
2076     // Reset the mark bitmap, summary data, and do other bookkeeping.  Must be
2077     // done before resizing.
2078     post_compact();
2079 
2080     // Let the size policy know we're done
2081     size_policy->major_collection_end(old_gen->used_in_bytes(), gc_cause);
2082 
2083     if (UseAdaptiveSizePolicy) {
2084       if (PrintAdaptiveSizePolicy) {
2085         gclog_or_tty->print("AdaptiveSizeStart: ");
2086         gclog_or_tty->stamp();
2087         gclog_or_tty->print_cr(" collection: %d ",
2088                        heap->total_collections());
2089         if (Verbose) {
2090           gclog_or_tty->print("old_gen_capacity: " SIZE_FORMAT
2091             " young_gen_capacity: " SIZE_FORMAT,
2092             old_gen->capacity_in_bytes(), young_gen->capacity_in_bytes());
2093         }
2094       }
2095 
2096       // Don't check if the size_policy is ready here.  Let
2097       // the size_policy check that internally.
2098       if (UseAdaptiveGenerationSizePolicyAtMajorCollection &&
2099           ((gc_cause != GCCause::_java_lang_system_gc) ||
2100             UseAdaptiveSizePolicyWithSystemGC)) {
2101         // Swap the survivor spaces if from_space is empty. The
2102         // resize_young_gen() called below is normally used after
2103         // a successful young GC and swapping of survivor spaces;
2104         // otherwise, it will fail to resize the young gen with
2105         // the current implementation.
2106         if (young_gen->from_space()->is_empty()) {
2107           young_gen->from_space()->clear(SpaceDecorator::Mangle);
2108           young_gen->swap_spaces();
2109         }
2110 
2111         // Calculate optimal free space amounts
2112         assert(young_gen->max_size() >
2113           young_gen->from_space()->capacity_in_bytes() +
2114           young_gen->to_space()->capacity_in_bytes(),
2115           "Sizes of space in young gen are out-of-bounds");
2116 
2117         size_t young_live = young_gen->used_in_bytes();
2118         size_t eden_live = young_gen->eden_space()->used_in_bytes();
2119         size_t old_live = old_gen->used_in_bytes();
2120         size_t cur_eden = young_gen->eden_space()->capacity_in_bytes();
2121         size_t max_old_gen_size = old_gen->max_gen_size();
2122         size_t max_eden_size = young_gen->max_size() -
2123           young_gen->from_space()->capacity_in_bytes() -
2124           young_gen->to_space()->capacity_in_bytes();
2125 
2126         // Used for diagnostics
2127         size_policy->clear_generation_free_space_flags();
2128 
2129         size_policy->compute_generations_free_space(young_live,
2130                                                     eden_live,
2131                                                     old_live,
2132                                                     cur_eden,
2133                                                     max_old_gen_size,
2134                                                     max_eden_size,
2135                                                     true /* full gc*/);
2136 
2137         size_policy->check_gc_overhead_limit(young_live,
2138                                              eden_live,
2139                                              max_old_gen_size,
2140                                              max_eden_size,
2141                                              true /* full gc*/,
2142                                              gc_cause,
2143                                              heap->collector_policy());
2144 
2145         size_policy->decay_supplemental_growth(true /* full gc*/);
2146 
2147         heap->resize_old_gen(
2148           size_policy->calculated_old_free_size_in_bytes());
2149 
2150         heap->resize_young_gen(size_policy->calculated_eden_size_in_bytes(),
2151                                size_policy->calculated_survivor_size_in_bytes());
2152       }
2153       if (PrintAdaptiveSizePolicy) {
2154         gclog_or_tty->print_cr("AdaptiveSizeStop: collection: %d ",
2155                        heap->total_collections());
2156       }
2157     }
2158 
2159     if (UsePerfData) {
2160       PSGCAdaptivePolicyCounters* const counters = heap->gc_policy_counters();
2161       counters->update_counters();
2162       counters->update_old_capacity(old_gen->capacity_in_bytes());
2163       counters->update_young_capacity(young_gen->capacity_in_bytes());
2164     }
2165 
2166     heap->resize_all_tlabs();
2167 
2168     // Resize the metaspace capacity after a collection
2169     MetaspaceGC::compute_new_size();
2170 
2171     if (TraceOldGenTime) accumulated_time()->stop();
2172 
2173     if (PrintGC) {
2174       if (PrintGCDetails) {
2175         // No GC timestamp here.  This is after GC so it would be confusing.
2176         young_gen->print_used_change(pre_gc_values.young_gen_used());
2177         old_gen->print_used_change(pre_gc_values.old_gen_used());
2178         heap->print_heap_change(pre_gc_values.heap_used());
2179         MetaspaceAux::print_metaspace_change(pre_gc_values.metadata_used());
2180       } else {
2181         heap->print_heap_change(pre_gc_values.heap_used());
2182       }
2183     }
2184 
2185     // Track memory usage and detect low memory
2186     MemoryService::track_memory_usage();
2187     heap->update_counters();
2188     gc_task_manager()->release_idle_workers();
2189   }
2190 
2191 #ifdef ASSERT
2192   for (size_t i = 0; i < ParallelGCThreads + 1; ++i) {
2193     ParCompactionManager* const cm =
2194       ParCompactionManager::manager_array(int(i));
2195     assert(cm->marking_stack()->is_empty(),       "should be empty");
2196     assert(ParCompactionManager::region_list(int(i))->is_empty(), "should be empty");
2197   }
2198 #endif // ASSERT
2199 
2200   if (VerifyAfterGC && heap->total_collections() >= VerifyGCStartAt) {
2201     HandleMark hm;  // Discard invalid handles created during verification
2202     Universe::verify(" VerifyAfterGC:");
2203   }
2204 
2205   // Re-verify object start arrays
2206   if (VerifyObjectStartArray &&
2207       VerifyAfterGC) {
2208     old_gen->verify_object_start_array();
2209   }
2210 
2211   if (ZapUnusedHeapArea) {
2212     old_gen->object_space()->check_mangled_unused_area_complete();
2213   }
2214 
2215   NOT_PRODUCT(ref_processor()->verify_no_references_recorded());
2216 
2217   collection_exit.update();
2218 
2219   heap->print_heap_after_gc();
2220   heap->trace_heap_after_gc(&_gc_tracer);
2221 
2222   if (PrintGCTaskTimeStamps) {
2223     gclog_or_tty->print_cr("VM-Thread " JLONG_FORMAT " " JLONG_FORMAT " "
2224                            JLONG_FORMAT,
2225                            marking_start.ticks(), compaction_start.ticks(),
2226                            collection_exit.ticks());
2227     gc_task_manager()->print_task_time_stamps();
2228   }
2229 
2230   heap->post_full_gc_dump(&_gc_timer);
2231 
2232 #ifdef TRACESPINNING
2233   ParallelTaskTerminator::print_termination_counts();
2234 #endif
2235 
2236   _gc_timer.register_gc_end();
2237 
2238   _gc_tracer.report_dense_prefix(dense_prefix(old_space_id));
2239   _gc_tracer.report_gc_end(_gc_timer.gc_end(), _gc_timer.time_partitions());
2240 
2241   return true;
2242 }
2243 
2244 bool PSParallelCompact::absorb_live_data_from_eden(PSAdaptiveSizePolicy* size_policy,
2245                                              PSYoungGen* young_gen,
2246                                              PSOldGen* old_gen) {
2247   MutableSpace* const eden_space = young_gen->eden_space();
2248   assert(!eden_space->is_empty(), "eden must be non-empty");
2249   assert(young_gen->virtual_space()->alignment() ==
2250          old_gen->virtual_space()->alignment(), "alignments do not match");
2251 
2252   if (!(UseAdaptiveSizePolicy && UseAdaptiveGCBoundary)) {
2253     return false;
2254   }
2255 
2256   // Both generations must be completely committed.
2257   if (young_gen->virtual_space()->uncommitted_size() != 0) {
2258     return false;
2259   }
2260   if (old_gen->virtual_space()->uncommitted_size() != 0) {
2261     return false;
2262   }
2263 
2264   // Figure out how much to take from eden.  Include the average amount promoted
2265   // in the total; otherwise the next young gen GC will simply bail out to a
2266   // full GC.
2267   const size_t alignment = old_gen->virtual_space()->alignment();
2268   const size_t eden_used = eden_space->used_in_bytes();
2269   const size_t promoted = (size_t)size_policy->avg_promoted()->padded_average();
2270   const size_t absorb_size = align_size_up(eden_used + promoted, alignment);
2271   const size_t eden_capacity = eden_space->capacity_in_bytes();
2272 
2273   if (absorb_size >= eden_capacity) {
2274     return false; // Must leave some space in eden.
2275   }
2276 
2277   const size_t new_young_size = young_gen->capacity_in_bytes() - absorb_size;
2278   if (new_young_size < young_gen->min_gen_size()) {
2279     return false; // Respect young gen minimum size.
2280   }
2281 
2282   if (TraceAdaptiveGCBoundary && Verbose) {
2283     gclog_or_tty->print(" absorbing " SIZE_FORMAT "K:  "
2284                         "eden " SIZE_FORMAT "K->" SIZE_FORMAT "K "
2285                         "from " SIZE_FORMAT "K, to " SIZE_FORMAT "K "
2286                         "young_gen " SIZE_FORMAT "K->" SIZE_FORMAT "K ",
2287                         absorb_size / K,
2288                         eden_capacity / K, (eden_capacity - absorb_size) / K,
2289                         young_gen->from_space()->used_in_bytes() / K,
2290                         young_gen->to_space()->used_in_bytes() / K,
2291                         young_gen->capacity_in_bytes() / K, new_young_size / K);
2292   }
2293 
2294   // Fill the unused part of the old gen.
2295   MutableSpace* const old_space = old_gen->object_space();
2296   HeapWord* const unused_start = old_space->top();
2297   size_t const unused_words = pointer_delta(old_space->end(), unused_start);
2298 
2299   if (unused_words > 0) {
2300     if (unused_words < CollectedHeap::min_fill_size()) {
2301       return false;  // If the old gen cannot be filled, must give up.
2302     }
2303     CollectedHeap::fill_with_objects(unused_start, unused_words);
2304   }
2305 
2306   // Take the live data from eden and set both top and end in the old gen to
2307   // eden top.  (Need to set end because reset_after_change() mangles the region
2308   // from end to virtual_space->high() in debug builds).
2309   HeapWord* const new_top = eden_space->top();
2310   old_gen->virtual_space()->expand_into(young_gen->virtual_space(),
2311                                         absorb_size);
2312   young_gen->reset_after_change();
2313   old_space->set_top(new_top);
2314   old_space->set_end(new_top);
2315   old_gen->reset_after_change();
2316 
2317   // Update the object start array for the filler object and the data from eden.
2318   ObjectStartArray* const start_array = old_gen->start_array();
2319   for (HeapWord* p = unused_start; p < new_top; p += oop(p)->size()) {
2320     start_array->allocate_block(p);
2321   }
2322 
2323   // Could update the promoted average here, but it is not typically updated at
2324   // full GCs and the value to use is unclear.  Something like
2325   //
2326   // cur_promoted_avg + absorb_size / number_of_scavenges_since_last_full_gc.
2327 
2328   size_policy->set_bytes_absorbed_from_eden(absorb_size);
2329   return true;
2330 }
2331 
2332 GCTaskManager* const PSParallelCompact::gc_task_manager() {
2333   assert(ParallelScavengeHeap::gc_task_manager() != NULL,
2334     "shouldn't return NULL");
2335   return ParallelScavengeHeap::gc_task_manager();
2336 }
2337 
2338 void PSParallelCompact::marking_phase(ParCompactionManager* cm,
2339                                       bool maximum_heap_compaction,
2340                                       ParallelOldTracer *gc_tracer) {
2341   // Recursively traverse all live objects and mark them
2342   GCTraceTime tm("marking phase", print_phases(), true, &_gc_timer, _gc_tracer.gc_id());
2343 
2344   ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
2345   uint parallel_gc_threads = heap->gc_task_manager()->workers();
2346   uint active_gc_threads = heap->gc_task_manager()->active_workers();
2347   TaskQueueSetSuper* qset = ParCompactionManager::region_array();
2348   ParallelTaskTerminator terminator(active_gc_threads, qset);
2349 
2350   PSParallelCompact::MarkAndPushClosure mark_and_push_closure(cm);
2351   PSParallelCompact::FollowStackClosure follow_stack_closure(cm);
2352 
2353   // Need new claim bits before marking starts.
2354   ClassLoaderDataGraph::clear_claimed_marks();
2355 
2356   {
2357     GCTraceTime tm_m("par mark", print_phases(), true, &_gc_timer, _gc_tracer.gc_id());
2358 
2359     ParallelScavengeHeap::ParStrongRootsScope psrs;
2360 
2361     GCTaskQueue* q = GCTaskQueue::create();
2362 
2363     q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::universe));
2364     q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jni_handles));
2365     // We scan the thread roots in parallel
2366     Threads::create_thread_roots_marking_tasks(q);
2367     q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::object_synchronizer));
2368     q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::flat_profiler));
2369     q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::management));
2370     q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::system_dictionary));
2371     q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::class_loader_data));
2372     q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jvmti));
2373     q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::code_cache));
2374 
2375     if (active_gc_threads > 1) {
2376       for (uint j = 0; j < active_gc_threads; j++) {
2377         q->enqueue(new StealMarkingTask(&terminator));
2378       }
2379     }
2380 
2381     gc_task_manager()->execute_and_wait(q);
2382   }
2383 
2384   // Process reference objects found during marking
2385   {
2386     GCTraceTime tm_r("reference processing", print_phases(), true, &_gc_timer, _gc_tracer.gc_id());
2387 
2388     ReferenceProcessorStats stats;
2389     if (ref_processor()->processing_is_mt()) {
2390       RefProcTaskExecutor task_executor;
2391       stats = ref_processor()->process_discovered_references(
2392         is_alive_closure(), &mark_and_push_closure, &follow_stack_closure,
2393         &task_executor, &_gc_timer, _gc_tracer.gc_id());
2394     } else {
2395       stats = ref_processor()->process_discovered_references(
2396         is_alive_closure(), &mark_and_push_closure, &follow_stack_closure, NULL,
2397         &_gc_timer, _gc_tracer.gc_id());
2398     }
2399 
2400     gc_tracer->report_gc_reference_stats(stats);
2401   }
2402 
2403   GCTraceTime tm_c("class unloading", print_phases(), true, &_gc_timer, _gc_tracer.gc_id());
2404 
2405   // This is the point where the entire marking should have completed.
2406   assert(cm->marking_stacks_empty(), "Marking should have completed");
2407 
2408   // Follow system dictionary roots and unload classes.
2409   bool purged_class = SystemDictionary::do_unloading(is_alive_closure());
2410 
2411   // Unload nmethods.
2412   CodeCache::do_unloading(is_alive_closure(), purged_class);
2413 
2414   // Prune dead klasses from subklass/sibling/implementor lists.
2415   Klass::clean_weak_klass_links(is_alive_closure());
2416 
2417   // Delete entries for dead interned strings.
2418   StringTable::unlink(is_alive_closure());
2419 
2420   // Clean up unreferenced symbols in symbol table.
2421   SymbolTable::unlink();
2422   _gc_tracer.report_object_count_after_gc(is_alive_closure());
2423 }
2424 
2425 void PSParallelCompact::follow_class_loader(ParCompactionManager* cm,
2426                                             ClassLoaderData* cld) {
2427   PSParallelCompact::MarkAndPushClosure mark_and_push_closure(cm);
2428   PSParallelCompact::FollowKlassClosure follow_klass_closure(&mark_and_push_closure);
2429 
2430   cld->oops_do(&mark_and_push_closure, &follow_klass_closure, true);
2431 }
2432 
2433 // This should be moved to the shared markSweep code!
2434 class PSAlwaysTrueClosure: public BoolObjectClosure {
2435 public:
2436   bool do_object_b(oop p) { return true; }
2437 };
2438 static PSAlwaysTrueClosure always_true;
2439 
2440 void PSParallelCompact::adjust_roots() {
2441   // Adjust the pointers to reflect the new locations
2442   GCTraceTime tm("adjust roots", print_phases(), true, &_gc_timer, _gc_tracer.gc_id());
2443 
2444   // Need new claim bits when tracing through and adjusting pointers.
2445   ClassLoaderDataGraph::clear_claimed_marks();
2446 
2447   // General strong roots.
2448   Universe::oops_do(adjust_pointer_closure());
2449   JNIHandles::oops_do(adjust_pointer_closure());   // Global (strong) JNI handles
2450   CLDToOopClosure adjust_from_cld(adjust_pointer_closure());
2451   Threads::oops_do(adjust_pointer_closure(), &adjust_from_cld, NULL);
2452   ObjectSynchronizer::oops_do(adjust_pointer_closure());
2453   FlatProfiler::oops_do(adjust_pointer_closure());
2454   Management::oops_do(adjust_pointer_closure());
2455   JvmtiExport::oops_do(adjust_pointer_closure());
2456   SystemDictionary::oops_do(adjust_pointer_closure());
2457   ClassLoaderDataGraph::oops_do(adjust_pointer_closure(), adjust_klass_closure(), true);
2458 
2459   // Now adjust pointers in remaining weak roots.  (All of which should
2460   // have been cleared if they pointed to non-surviving objects.)
2461   // Global (weak) JNI handles
2462   JNIHandles::weak_oops_do(&always_true, adjust_pointer_closure());
2463 
2464   CodeBlobToOopClosure adjust_from_blobs(adjust_pointer_closure(), CodeBlobToOopClosure::FixRelocations);
2465   CodeCache::blobs_do(&adjust_from_blobs);
2466   StringTable::oops_do(adjust_pointer_closure());
2467   ref_processor()->weak_oops_do(adjust_pointer_closure());
2468   // Roots were visited so references into the young gen in roots
2469   // may have been scanned.  Process them also.
2470   // Should the reference processor have a span that excludes
2471   // young gen objects?
2472   PSScavenge::reference_processor()->weak_oops_do(adjust_pointer_closure());
2473 }
2474 
2475 void PSParallelCompact::enqueue_region_draining_tasks(GCTaskQueue* q,
2476                                                       uint parallel_gc_threads)
2477 {
2478   GCTraceTime tm("drain task setup", print_phases(), true, &_gc_timer, _gc_tracer.gc_id());
2479 
2480   // Find the threads that are active
2481   unsigned int which = 0;
2482 
2483   const uint task_count = MAX2(parallel_gc_threads, 1U);
2484   for (uint j = 0; j < task_count; j++) {
2485     q->enqueue(new DrainStacksCompactionTask(j));
2486     ParCompactionManager::verify_region_list_empty(j);
2487     // Set the region stacks variables to "no" region stack values
2488     // so that they will be recognized and needing a region stack
2489     // in the stealing tasks if they do not get one by executing
2490     // a draining stack.
2491     ParCompactionManager* cm = ParCompactionManager::manager_array(j);
2492     cm->set_region_stack(NULL);
2493     cm->set_region_stack_index((uint)max_uintx);
2494   }
2495   ParCompactionManager::reset_recycled_stack_index();
2496 
2497   // Find all regions that are available (can be filled immediately) and
2498   // distribute them to the thread stacks.  The iteration is done in reverse
2499   // order (high to low) so the regions will be removed in ascending order.
2500 
2501   const ParallelCompactData& sd = PSParallelCompact::summary_data();
2502 
2503   size_t fillable_regions = 0;   // A count for diagnostic purposes.
2504   // A region index which corresponds to the tasks created above.
2505   // "which" must be 0 <= which < task_count
2506 
2507   which = 0;
2508   // id + 1 is used to test termination so unsigned  can
2509   // be used with an old_space_id == 0.
2510   for (unsigned int id = to_space_id; id + 1 > old_space_id; --id) {
2511     SpaceInfo* const space_info = _space_info + id;
2512     MutableSpace* const space = space_info->space();
2513     HeapWord* const new_top = space_info->new_top();
2514 
2515     const size_t beg_region = sd.addr_to_region_idx(space_info->dense_prefix());
2516     const size_t end_region =
2517       sd.addr_to_region_idx(sd.region_align_up(new_top));
2518 
2519     for (size_t cur = end_region - 1; cur + 1 > beg_region; --cur) {
2520       if (sd.region(cur)->claim_unsafe()) {
2521         ParCompactionManager::region_list_push(which, cur);
2522 
2523         if (TraceParallelOldGCCompactionPhase && Verbose) {
2524           const size_t count_mod_8 = fillable_regions & 7;
2525           if (count_mod_8 == 0) gclog_or_tty->print("fillable: ");
2526           gclog_or_tty->print(" " SIZE_FORMAT_W(7), cur);
2527           if (count_mod_8 == 7) gclog_or_tty->cr();
2528         }
2529 
2530         NOT_PRODUCT(++fillable_regions;)
2531 
2532         // Assign regions to tasks in round-robin fashion.
2533         if (++which == task_count) {
2534           assert(which <= parallel_gc_threads,
2535             "Inconsistent number of workers");
2536           which = 0;
2537         }
2538       }
2539     }
2540   }
2541 
2542   if (TraceParallelOldGCCompactionPhase) {
2543     if (Verbose && (fillable_regions & 7) != 0) gclog_or_tty->cr();
2544     gclog_or_tty->print_cr(SIZE_FORMAT " initially fillable regions", fillable_regions);
2545   }
2546 }
2547 
2548 #define PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING 4
2549 
2550 void PSParallelCompact::enqueue_dense_prefix_tasks(GCTaskQueue* q,
2551                                                     uint parallel_gc_threads) {
2552   GCTraceTime tm("dense prefix task setup", print_phases(), true, &_gc_timer, _gc_tracer.gc_id());
2553 
2554   ParallelCompactData& sd = PSParallelCompact::summary_data();
2555 
2556   // Iterate over all the spaces adding tasks for updating
2557   // regions in the dense prefix.  Assume that 1 gc thread
2558   // will work on opening the gaps and the remaining gc threads
2559   // will work on the dense prefix.
2560   unsigned int space_id;
2561   for (space_id = old_space_id; space_id < last_space_id; ++ space_id) {
2562     HeapWord* const dense_prefix_end = _space_info[space_id].dense_prefix();
2563     const MutableSpace* const space = _space_info[space_id].space();
2564 
2565     if (dense_prefix_end == space->bottom()) {
2566       // There is no dense prefix for this space.
2567       continue;
2568     }
2569 
2570     // The dense prefix is before this region.
2571     size_t region_index_end_dense_prefix =
2572         sd.addr_to_region_idx(dense_prefix_end);
2573     RegionData* const dense_prefix_cp =
2574       sd.region(region_index_end_dense_prefix);
2575     assert(dense_prefix_end == space->end() ||
2576            dense_prefix_cp->available() ||
2577            dense_prefix_cp->claimed(),
2578            "The region after the dense prefix should always be ready to fill");
2579 
2580     size_t region_index_start = sd.addr_to_region_idx(space->bottom());
2581 
2582     // Is there dense prefix work?
2583     size_t total_dense_prefix_regions =
2584       region_index_end_dense_prefix - region_index_start;
2585     // How many regions of the dense prefix should be given to
2586     // each thread?
2587     if (total_dense_prefix_regions > 0) {
2588       uint tasks_for_dense_prefix = 1;
2589       if (total_dense_prefix_regions <=
2590           (parallel_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING)) {
2591         // Don't over partition.  This assumes that
2592         // PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING is a small integer value
2593         // so there are not many regions to process.
2594         tasks_for_dense_prefix = parallel_gc_threads;
2595       } else {
2596         // Over partition
2597         tasks_for_dense_prefix = parallel_gc_threads *
2598           PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING;
2599       }
2600       size_t regions_per_thread = total_dense_prefix_regions /
2601         tasks_for_dense_prefix;
2602       // Give each thread at least 1 region.
2603       if (regions_per_thread == 0) {
2604         regions_per_thread = 1;
2605       }
2606 
2607       for (uint k = 0; k < tasks_for_dense_prefix; k++) {
2608         if (region_index_start >= region_index_end_dense_prefix) {
2609           break;
2610         }
2611         // region_index_end is not processed
2612         size_t region_index_end = MIN2(region_index_start + regions_per_thread,
2613                                        region_index_end_dense_prefix);
2614         q->enqueue(new UpdateDensePrefixTask(SpaceId(space_id),
2615                                              region_index_start,
2616                                              region_index_end));
2617         region_index_start = region_index_end;
2618       }
2619     }
2620     // This gets any part of the dense prefix that did not
2621     // fit evenly.
2622     if (region_index_start < region_index_end_dense_prefix) {
2623       q->enqueue(new UpdateDensePrefixTask(SpaceId(space_id),
2624                                            region_index_start,
2625                                            region_index_end_dense_prefix));
2626     }
2627   }
2628 }
2629 
2630 void PSParallelCompact::enqueue_region_stealing_tasks(
2631                                      GCTaskQueue* q,
2632                                      ParallelTaskTerminator* terminator_ptr,
2633                                      uint parallel_gc_threads) {
2634   GCTraceTime tm("steal task setup", print_phases(), true, &_gc_timer, _gc_tracer.gc_id());
2635 
2636   // Once a thread has drained it's stack, it should try to steal regions from
2637   // other threads.
2638   if (parallel_gc_threads > 1) {
2639     for (uint j = 0; j < parallel_gc_threads; j++) {
2640       q->enqueue(new StealRegionCompactionTask(terminator_ptr));
2641     }
2642   }
2643 }
2644 
2645 #ifdef ASSERT
2646 // Write a histogram of the number of times the block table was filled for a
2647 // region.
2648 void PSParallelCompact::write_block_fill_histogram(outputStream* const out)
2649 {
2650   if (!TraceParallelOldGCCompactionPhase) return;
2651 
2652   typedef ParallelCompactData::RegionData rd_t;
2653   ParallelCompactData& sd = summary_data();
2654 
2655   for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2656     MutableSpace* const spc = _space_info[id].space();
2657     if (spc->bottom() != spc->top()) {
2658       const rd_t* const beg = sd.addr_to_region_ptr(spc->bottom());
2659       HeapWord* const top_aligned_up = sd.region_align_up(spc->top());
2660       const rd_t* const end = sd.addr_to_region_ptr(top_aligned_up);
2661 
2662       size_t histo[5] = { 0, 0, 0, 0, 0 };
2663       const size_t histo_len = sizeof(histo) / sizeof(size_t);
2664       const size_t region_cnt = pointer_delta(end, beg, sizeof(rd_t));
2665 
2666       for (const rd_t* cur = beg; cur < end; ++cur) {
2667         ++histo[MIN2(cur->blocks_filled_count(), histo_len - 1)];
2668       }
2669       out->print("%u %-4s" SIZE_FORMAT_W(5), id, space_names[id], region_cnt);
2670       for (size_t i = 0; i < histo_len; ++i) {
2671         out->print(" " SIZE_FORMAT_W(5) " %5.1f%%",
2672                    histo[i], 100.0 * histo[i] / region_cnt);
2673       }
2674       out->cr();
2675     }
2676   }
2677 }
2678 #endif // #ifdef ASSERT
2679 
2680 void PSParallelCompact::compact() {
2681   // trace("5");
2682   GCTraceTime tm("compaction phase", print_phases(), true, &_gc_timer, _gc_tracer.gc_id());
2683 
2684   ParallelScavengeHeap* heap = ParallelScavengeHeap::heap();
2685   PSOldGen* old_gen = heap->old_gen();
2686   old_gen->start_array()->reset();
2687   uint parallel_gc_threads = heap->gc_task_manager()->workers();
2688   uint active_gc_threads = heap->gc_task_manager()->active_workers();
2689   TaskQueueSetSuper* qset = ParCompactionManager::region_array();
2690   ParallelTaskTerminator terminator(active_gc_threads, qset);
2691 
2692   GCTaskQueue* q = GCTaskQueue::create();
2693   enqueue_region_draining_tasks(q, active_gc_threads);
2694   enqueue_dense_prefix_tasks(q, active_gc_threads);
2695   enqueue_region_stealing_tasks(q, &terminator, active_gc_threads);
2696 
2697   {
2698     GCTraceTime tm_pc("par compact", print_phases(), true, &_gc_timer, _gc_tracer.gc_id());
2699 
2700     gc_task_manager()->execute_and_wait(q);
2701 
2702 #ifdef  ASSERT
2703     // Verify that all regions have been processed before the deferred updates.
2704     for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2705       verify_complete(SpaceId(id));
2706     }
2707 #endif
2708   }
2709 
2710   {
2711     // Update the deferred objects, if any.  Any compaction manager can be used.
2712     GCTraceTime tm_du("deferred updates", print_phases(), true, &_gc_timer, _gc_tracer.gc_id());
2713     ParCompactionManager* cm = ParCompactionManager::manager_array(0);
2714     for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2715       update_deferred_objects(cm, SpaceId(id));
2716     }
2717   }
2718 
2719   DEBUG_ONLY(write_block_fill_histogram(gclog_or_tty));
2720 }
2721 
2722 #ifdef  ASSERT
2723 void PSParallelCompact::verify_complete(SpaceId space_id) {
2724   // All Regions between space bottom() to new_top() should be marked as filled
2725   // and all Regions between new_top() and top() should be available (i.e.,
2726   // should have been emptied).
2727   ParallelCompactData& sd = summary_data();
2728   SpaceInfo si = _space_info[space_id];
2729   HeapWord* new_top_addr = sd.region_align_up(si.new_top());
2730   HeapWord* old_top_addr = sd.region_align_up(si.space()->top());
2731   const size_t beg_region = sd.addr_to_region_idx(si.space()->bottom());
2732   const size_t new_top_region = sd.addr_to_region_idx(new_top_addr);
2733   const size_t old_top_region = sd.addr_to_region_idx(old_top_addr);
2734 
2735   bool issued_a_warning = false;
2736 
2737   size_t cur_region;
2738   for (cur_region = beg_region; cur_region < new_top_region; ++cur_region) {
2739     const RegionData* const c = sd.region(cur_region);
2740     if (!c->completed()) {
2741       warning("region " SIZE_FORMAT " not filled:  "
2742               "destination_count=%u",
2743               cur_region, c->destination_count());
2744       issued_a_warning = true;
2745     }
2746   }
2747 
2748   for (cur_region = new_top_region; cur_region < old_top_region; ++cur_region) {
2749     const RegionData* const c = sd.region(cur_region);
2750     if (!c->available()) {
2751       warning("region " SIZE_FORMAT " not empty:   "
2752               "destination_count=%u",
2753               cur_region, c->destination_count());
2754       issued_a_warning = true;
2755     }
2756   }
2757 
2758   if (issued_a_warning) {
2759     print_region_ranges();
2760   }
2761 }
2762 #endif  // #ifdef ASSERT
2763 
2764 inline void UpdateOnlyClosure::do_addr(HeapWord* addr) {
2765   _start_array->allocate_block(addr);
2766   compaction_manager()->update_contents(oop(addr));
2767 }
2768 
2769 // Update interior oops in the ranges of regions [beg_region, end_region).
2770 void
2771 PSParallelCompact::update_and_deadwood_in_dense_prefix(ParCompactionManager* cm,
2772                                                        SpaceId space_id,
2773                                                        size_t beg_region,
2774                                                        size_t end_region) {
2775   ParallelCompactData& sd = summary_data();
2776   ParMarkBitMap* const mbm = mark_bitmap();
2777 
2778   HeapWord* beg_addr = sd.region_to_addr(beg_region);
2779   HeapWord* const end_addr = sd.region_to_addr(end_region);
2780   assert(beg_region <= end_region, "bad region range");
2781   assert(end_addr <= dense_prefix(space_id), "not in the dense prefix");
2782 
2783 #ifdef  ASSERT
2784   // Claim the regions to avoid triggering an assert when they are marked as
2785   // filled.
2786   for (size_t claim_region = beg_region; claim_region < end_region; ++claim_region) {
2787     assert(sd.region(claim_region)->claim_unsafe(), "claim() failed");
2788   }
2789 #endif  // #ifdef ASSERT
2790 
2791   if (beg_addr != space(space_id)->bottom()) {
2792     // Find the first live object or block of dead space that *starts* in this
2793     // range of regions.  If a partial object crosses onto the region, skip it;
2794     // it will be marked for 'deferred update' when the object head is
2795     // processed.  If dead space crosses onto the region, it is also skipped; it
2796     // will be filled when the prior region is processed.  If neither of those
2797     // apply, the first word in the region is the start of a live object or dead
2798     // space.
2799     assert(beg_addr > space(space_id)->bottom(), "sanity");
2800     const RegionData* const cp = sd.region(beg_region);
2801     if (cp->partial_obj_size() != 0) {
2802       beg_addr = sd.partial_obj_end(beg_region);
2803     } else if (dead_space_crosses_boundary(cp, mbm->addr_to_bit(beg_addr))) {
2804       beg_addr = mbm->find_obj_beg(beg_addr, end_addr);
2805     }
2806   }
2807 
2808   if (beg_addr < end_addr) {
2809     // A live object or block of dead space starts in this range of Regions.
2810      HeapWord* const dense_prefix_end = dense_prefix(space_id);
2811 
2812     // Create closures and iterate.
2813     UpdateOnlyClosure update_closure(mbm, cm, space_id);
2814     FillClosure fill_closure(cm, space_id);
2815     ParMarkBitMap::IterationStatus status;
2816     status = mbm->iterate(&update_closure, &fill_closure, beg_addr, end_addr,
2817                           dense_prefix_end);
2818     if (status == ParMarkBitMap::incomplete) {
2819       update_closure.do_addr(update_closure.source());
2820     }
2821   }
2822 
2823   // Mark the regions as filled.
2824   RegionData* const beg_cp = sd.region(beg_region);
2825   RegionData* const end_cp = sd.region(end_region);
2826   for (RegionData* cp = beg_cp; cp < end_cp; ++cp) {
2827     cp->set_completed();
2828   }
2829 }
2830 
2831 // Return the SpaceId for the space containing addr.  If addr is not in the
2832 // heap, last_space_id is returned.  In debug mode it expects the address to be
2833 // in the heap and asserts such.
2834 PSParallelCompact::SpaceId PSParallelCompact::space_id(HeapWord* addr) {
2835   assert(ParallelScavengeHeap::heap()->is_in_reserved(addr), "addr not in the heap");
2836 
2837   for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2838     if (_space_info[id].space()->contains(addr)) {
2839       return SpaceId(id);
2840     }
2841   }
2842 
2843   assert(false, "no space contains the addr");
2844   return last_space_id;
2845 }
2846 
2847 void PSParallelCompact::update_deferred_objects(ParCompactionManager* cm,
2848                                                 SpaceId id) {
2849   assert(id < last_space_id, "bad space id");
2850 
2851   ParallelCompactData& sd = summary_data();
2852   const SpaceInfo* const space_info = _space_info + id;
2853   ObjectStartArray* const start_array = space_info->start_array();
2854 
2855   const MutableSpace* const space = space_info->space();
2856   assert(space_info->dense_prefix() >= space->bottom(), "dense_prefix not set");
2857   HeapWord* const beg_addr = space_info->dense_prefix();
2858   HeapWord* const end_addr = sd.region_align_up(space_info->new_top());
2859 
2860   const RegionData* const beg_region = sd.addr_to_region_ptr(beg_addr);
2861   const RegionData* const end_region = sd.addr_to_region_ptr(end_addr);
2862   const RegionData* cur_region;
2863   for (cur_region = beg_region; cur_region < end_region; ++cur_region) {
2864     HeapWord* const addr = cur_region->deferred_obj_addr();
2865     if (addr != NULL) {
2866       if (start_array != NULL) {
2867         start_array->allocate_block(addr);
2868       }
2869       cm->update_contents(oop(addr));
2870       assert(oop(addr)->is_oop_or_null(), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(oop(addr))));
2871     }
2872   }
2873 }
2874 
2875 // Skip over count live words starting from beg, and return the address of the
2876 // next live word.  Unless marked, the word corresponding to beg is assumed to
2877 // be dead.  Callers must either ensure beg does not correspond to the middle of
2878 // an object, or account for those live words in some other way.  Callers must
2879 // also ensure that there are enough live words in the range [beg, end) to skip.
2880 HeapWord*
2881 PSParallelCompact::skip_live_words(HeapWord* beg, HeapWord* end, size_t count)
2882 {
2883   assert(count > 0, "sanity");
2884 
2885   ParMarkBitMap* m = mark_bitmap();
2886   idx_t bits_to_skip = m->words_to_bits(count);
2887   idx_t cur_beg = m->addr_to_bit(beg);
2888   const idx_t search_end = BitMap::word_align_up(m->addr_to_bit(end));
2889 
2890   do {
2891     cur_beg = m->find_obj_beg(cur_beg, search_end);
2892     idx_t cur_end = m->find_obj_end(cur_beg, search_end);
2893     const size_t obj_bits = cur_end - cur_beg + 1;
2894     if (obj_bits > bits_to_skip) {
2895       return m->bit_to_addr(cur_beg + bits_to_skip);
2896     }
2897     bits_to_skip -= obj_bits;
2898     cur_beg = cur_end + 1;
2899   } while (bits_to_skip > 0);
2900 
2901   // Skipping the desired number of words landed just past the end of an object.
2902   // Find the start of the next object.
2903   cur_beg = m->find_obj_beg(cur_beg, search_end);
2904   assert(cur_beg < m->addr_to_bit(end), "not enough live words to skip");
2905   return m->bit_to_addr(cur_beg);
2906 }
2907 
2908 HeapWord* PSParallelCompact::first_src_addr(HeapWord* const dest_addr,
2909                                             SpaceId src_space_id,
2910                                             size_t src_region_idx)
2911 {
2912   assert(summary_data().is_region_aligned(dest_addr), "not aligned");
2913 
2914   const SplitInfo& split_info = _space_info[src_space_id].split_info();
2915   if (split_info.dest_region_addr() == dest_addr) {
2916     // The partial object ending at the split point contains the first word to
2917     // be copied to dest_addr.
2918     return split_info.first_src_addr();
2919   }
2920 
2921   const ParallelCompactData& sd = summary_data();
2922   ParMarkBitMap* const bitmap = mark_bitmap();
2923   const size_t RegionSize = ParallelCompactData::RegionSize;
2924 
2925   assert(sd.is_region_aligned(dest_addr), "not aligned");
2926   const RegionData* const src_region_ptr = sd.region(src_region_idx);
2927   const size_t partial_obj_size = src_region_ptr->partial_obj_size();
2928   HeapWord* const src_region_destination = src_region_ptr->destination();
2929 
2930   assert(dest_addr >= src_region_destination, "wrong src region");
2931   assert(src_region_ptr->data_size() > 0, "src region cannot be empty");
2932 
2933   HeapWord* const src_region_beg = sd.region_to_addr(src_region_idx);
2934   HeapWord* const src_region_end = src_region_beg + RegionSize;
2935 
2936   HeapWord* addr = src_region_beg;
2937   if (dest_addr == src_region_destination) {
2938     // Return the first live word in the source region.
2939     if (partial_obj_size == 0) {
2940       addr = bitmap->find_obj_beg(addr, src_region_end);
2941       assert(addr < src_region_end, "no objects start in src region");
2942     }
2943     return addr;
2944   }
2945 
2946   // Must skip some live data.
2947   size_t words_to_skip = dest_addr - src_region_destination;
2948   assert(src_region_ptr->data_size() > words_to_skip, "wrong src region");
2949 
2950   if (partial_obj_size >= words_to_skip) {
2951     // All the live words to skip are part of the partial object.
2952     addr += words_to_skip;
2953     if (partial_obj_size == words_to_skip) {
2954       // Find the first live word past the partial object.
2955       addr = bitmap->find_obj_beg(addr, src_region_end);
2956       assert(addr < src_region_end, "wrong src region");
2957     }
2958     return addr;
2959   }
2960 
2961   // Skip over the partial object (if any).
2962   if (partial_obj_size != 0) {
2963     words_to_skip -= partial_obj_size;
2964     addr += partial_obj_size;
2965   }
2966 
2967   // Skip over live words due to objects that start in the region.
2968   addr = skip_live_words(addr, src_region_end, words_to_skip);
2969   assert(addr < src_region_end, "wrong src region");
2970   return addr;
2971 }
2972 
2973 void PSParallelCompact::decrement_destination_counts(ParCompactionManager* cm,
2974                                                      SpaceId src_space_id,
2975                                                      size_t beg_region,
2976                                                      HeapWord* end_addr)
2977 {
2978   ParallelCompactData& sd = summary_data();
2979 
2980 #ifdef ASSERT
2981   MutableSpace* const src_space = _space_info[src_space_id].space();
2982   HeapWord* const beg_addr = sd.region_to_addr(beg_region);
2983   assert(src_space->contains(beg_addr) || beg_addr == src_space->end(),
2984          "src_space_id does not match beg_addr");
2985   assert(src_space->contains(end_addr) || end_addr == src_space->end(),
2986          "src_space_id does not match end_addr");
2987 #endif // #ifdef ASSERT
2988 
2989   RegionData* const beg = sd.region(beg_region);
2990   RegionData* const end = sd.addr_to_region_ptr(sd.region_align_up(end_addr));
2991 
2992   // Regions up to new_top() are enqueued if they become available.
2993   HeapWord* const new_top = _space_info[src_space_id].new_top();
2994   RegionData* const enqueue_end =
2995     sd.addr_to_region_ptr(sd.region_align_up(new_top));
2996 
2997   for (RegionData* cur = beg; cur < end; ++cur) {
2998     assert(cur->data_size() > 0, "region must have live data");
2999     cur->decrement_destination_count();
3000     if (cur < enqueue_end && cur->available() && cur->claim()) {
3001       cm->push_region(sd.region(cur));
3002     }
3003   }
3004 }
3005 
3006 size_t PSParallelCompact::next_src_region(MoveAndUpdateClosure& closure,
3007                                           SpaceId& src_space_id,
3008                                           HeapWord*& src_space_top,
3009                                           HeapWord* end_addr)
3010 {
3011   typedef ParallelCompactData::RegionData RegionData;
3012 
3013   ParallelCompactData& sd = PSParallelCompact::summary_data();
3014   const size_t region_size = ParallelCompactData::RegionSize;
3015 
3016   size_t src_region_idx = 0;
3017 
3018   // Skip empty regions (if any) up to the top of the space.
3019   HeapWord* const src_aligned_up = sd.region_align_up(end_addr);
3020   RegionData* src_region_ptr = sd.addr_to_region_ptr(src_aligned_up);
3021   HeapWord* const top_aligned_up = sd.region_align_up(src_space_top);
3022   const RegionData* const top_region_ptr =
3023     sd.addr_to_region_ptr(top_aligned_up);
3024   while (src_region_ptr < top_region_ptr && src_region_ptr->data_size() == 0) {
3025     ++src_region_ptr;
3026   }
3027 
3028   if (src_region_ptr < top_region_ptr) {
3029     // The next source region is in the current space.  Update src_region_idx
3030     // and the source address to match src_region_ptr.
3031     src_region_idx = sd.region(src_region_ptr);
3032     HeapWord* const src_region_addr = sd.region_to_addr(src_region_idx);
3033     if (src_region_addr > closure.source()) {
3034       closure.set_source(src_region_addr);
3035     }
3036     return src_region_idx;
3037   }
3038 
3039   // Switch to a new source space and find the first non-empty region.
3040   unsigned int space_id = src_space_id + 1;
3041   assert(space_id < last_space_id, "not enough spaces");
3042 
3043   HeapWord* const destination = closure.destination();
3044 
3045   do {
3046     MutableSpace* space = _space_info[space_id].space();
3047     HeapWord* const bottom = space->bottom();
3048     const RegionData* const bottom_cp = sd.addr_to_region_ptr(bottom);
3049 
3050     // Iterate over the spaces that do not compact into themselves.
3051     if (bottom_cp->destination() != bottom) {
3052       HeapWord* const top_aligned_up = sd.region_align_up(space->top());
3053       const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up);
3054 
3055       for (const RegionData* src_cp = bottom_cp; src_cp < top_cp; ++src_cp) {
3056         if (src_cp->live_obj_size() > 0) {
3057           // Found it.
3058           assert(src_cp->destination() == destination,
3059                  "first live obj in the space must match the destination");
3060           assert(src_cp->partial_obj_size() == 0,
3061                  "a space cannot begin with a partial obj");
3062 
3063           src_space_id = SpaceId(space_id);
3064           src_space_top = space->top();
3065           const size_t src_region_idx = sd.region(src_cp);
3066           closure.set_source(sd.region_to_addr(src_region_idx));
3067           return src_region_idx;
3068         } else {
3069           assert(src_cp->data_size() == 0, "sanity");
3070         }
3071       }
3072     }
3073   } while (++space_id < last_space_id);
3074 
3075   assert(false, "no source region was found");
3076   return 0;
3077 }
3078 
3079 void PSParallelCompact::fill_region(ParCompactionManager* cm, size_t region_idx)
3080 {
3081   typedef ParMarkBitMap::IterationStatus IterationStatus;
3082   const size_t RegionSize = ParallelCompactData::RegionSize;
3083   ParMarkBitMap* const bitmap = mark_bitmap();
3084   ParallelCompactData& sd = summary_data();
3085   RegionData* const region_ptr = sd.region(region_idx);
3086 
3087   // Get the items needed to construct the closure.
3088   HeapWord* dest_addr = sd.region_to_addr(region_idx);
3089   SpaceId dest_space_id = space_id(dest_addr);
3090   ObjectStartArray* start_array = _space_info[dest_space_id].start_array();
3091   HeapWord* new_top = _space_info[dest_space_id].new_top();
3092   assert(dest_addr < new_top, "sanity");
3093   const size_t words = MIN2(pointer_delta(new_top, dest_addr), RegionSize);
3094 
3095   // Get the source region and related info.
3096   size_t src_region_idx = region_ptr->source_region();
3097   SpaceId src_space_id = space_id(sd.region_to_addr(src_region_idx));
3098   HeapWord* src_space_top = _space_info[src_space_id].space()->top();
3099 
3100   MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words);
3101   closure.set_source(first_src_addr(dest_addr, src_space_id, src_region_idx));
3102 
3103   // Adjust src_region_idx to prepare for decrementing destination counts (the
3104   // destination count is not decremented when a region is copied to itself).
3105   if (src_region_idx == region_idx) {
3106     src_region_idx += 1;
3107   }
3108 
3109   if (bitmap->is_unmarked(closure.source())) {
3110     // The first source word is in the middle of an object; copy the remainder
3111     // of the object or as much as will fit.  The fact that pointer updates were
3112     // deferred will be noted when the object header is processed.
3113     HeapWord* const old_src_addr = closure.source();
3114     closure.copy_partial_obj();
3115     if (closure.is_full()) {
3116       decrement_destination_counts(cm, src_space_id, src_region_idx,
3117                                    closure.source());
3118       region_ptr->set_deferred_obj_addr(NULL);
3119       region_ptr->set_completed();
3120       return;
3121     }
3122 
3123     HeapWord* const end_addr = sd.region_align_down(closure.source());
3124     if (sd.region_align_down(old_src_addr) != end_addr) {
3125       // The partial object was copied from more than one source region.
3126       decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr);
3127 
3128       // Move to the next source region, possibly switching spaces as well.  All
3129       // args except end_addr may be modified.
3130       src_region_idx = next_src_region(closure, src_space_id, src_space_top,
3131                                        end_addr);
3132     }
3133   }
3134 
3135   do {
3136     HeapWord* const cur_addr = closure.source();
3137     HeapWord* const end_addr = MIN2(sd.region_align_up(cur_addr + 1),
3138                                     src_space_top);
3139     IterationStatus status = bitmap->iterate(&closure, cur_addr, end_addr);
3140 
3141     if (status == ParMarkBitMap::incomplete) {
3142       // The last obj that starts in the source region does not end in the
3143       // region.
3144       assert(closure.source() < end_addr, "sanity");
3145       HeapWord* const obj_beg = closure.source();
3146       HeapWord* const range_end = MIN2(obj_beg + closure.words_remaining(),
3147                                        src_space_top);
3148       HeapWord* const obj_end = bitmap->find_obj_end(obj_beg, range_end);
3149       if (obj_end < range_end) {
3150         // The end was found; the entire object will fit.
3151         status = closure.do_addr(obj_beg, bitmap->obj_size(obj_beg, obj_end));
3152         assert(status != ParMarkBitMap::would_overflow, "sanity");
3153       } else {
3154         // The end was not found; the object will not fit.
3155         assert(range_end < src_space_top, "obj cannot cross space boundary");
3156         status = ParMarkBitMap::would_overflow;
3157       }
3158     }
3159 
3160     if (status == ParMarkBitMap::would_overflow) {
3161       // The last object did not fit.  Note that interior oop updates were
3162       // deferred, then copy enough of the object to fill the region.
3163       region_ptr->set_deferred_obj_addr(closure.destination());
3164       status = closure.copy_until_full(); // copies from closure.source()
3165 
3166       decrement_destination_counts(cm, src_space_id, src_region_idx,
3167                                    closure.source());
3168       region_ptr->set_completed();
3169       return;
3170     }
3171 
3172     if (status == ParMarkBitMap::full) {
3173       decrement_destination_counts(cm, src_space_id, src_region_idx,
3174                                    closure.source());
3175       region_ptr->set_deferred_obj_addr(NULL);
3176       region_ptr->set_completed();
3177       return;
3178     }
3179 
3180     decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr);
3181 
3182     // Move to the next source region, possibly switching spaces as well.  All
3183     // args except end_addr may be modified.
3184     src_region_idx = next_src_region(closure, src_space_id, src_space_top,
3185                                      end_addr);
3186   } while (true);
3187 }
3188 
3189 void PSParallelCompact::fill_blocks(size_t region_idx)
3190 {
3191   // Fill in the block table elements for the specified region.  Each block
3192   // table element holds the number of live words in the region that are to the
3193   // left of the first object that starts in the block.  Thus only blocks in
3194   // which an object starts need to be filled.
3195   //
3196   // The algorithm scans the section of the bitmap that corresponds to the
3197   // region, keeping a running total of the live words.  When an object start is
3198   // found, if it's the first to start in the block that contains it, the
3199   // current total is written to the block table element.
3200   const size_t Log2BlockSize = ParallelCompactData::Log2BlockSize;
3201   const size_t Log2RegionSize = ParallelCompactData::Log2RegionSize;
3202   const size_t RegionSize = ParallelCompactData::RegionSize;
3203 
3204   ParallelCompactData& sd = summary_data();
3205   const size_t partial_obj_size = sd.region(region_idx)->partial_obj_size();
3206   if (partial_obj_size >= RegionSize) {
3207     return; // No objects start in this region.
3208   }
3209 
3210   // Ensure the first loop iteration decides that the block has changed.
3211   size_t cur_block = sd.block_count();
3212 
3213   const ParMarkBitMap* const bitmap = mark_bitmap();
3214 
3215   const size_t Log2BitsPerBlock = Log2BlockSize - LogMinObjAlignment;
3216   assert((size_t)1 << Log2BitsPerBlock ==
3217          bitmap->words_to_bits(ParallelCompactData::BlockSize), "sanity");
3218 
3219   size_t beg_bit = bitmap->words_to_bits(region_idx << Log2RegionSize);
3220   const size_t range_end = beg_bit + bitmap->words_to_bits(RegionSize);
3221   size_t live_bits = bitmap->words_to_bits(partial_obj_size);
3222   beg_bit = bitmap->find_obj_beg(beg_bit + live_bits, range_end);
3223   while (beg_bit < range_end) {
3224     const size_t new_block = beg_bit >> Log2BitsPerBlock;
3225     if (new_block != cur_block) {
3226       cur_block = new_block;
3227       sd.block(cur_block)->set_offset(bitmap->bits_to_words(live_bits));
3228     }
3229 
3230     const size_t end_bit = bitmap->find_obj_end(beg_bit, range_end);
3231     if (end_bit < range_end - 1) {
3232       live_bits += end_bit - beg_bit + 1;
3233       beg_bit = bitmap->find_obj_beg(end_bit + 1, range_end);
3234     } else {
3235       return;
3236     }
3237   }
3238 }
3239 
3240 void
3241 PSParallelCompact::move_and_update(ParCompactionManager* cm, SpaceId space_id) {
3242   const MutableSpace* sp = space(space_id);
3243   if (sp->is_empty()) {
3244     return;
3245   }
3246 
3247   ParallelCompactData& sd = PSParallelCompact::summary_data();
3248   ParMarkBitMap* const bitmap = mark_bitmap();
3249   HeapWord* const dp_addr = dense_prefix(space_id);
3250   HeapWord* beg_addr = sp->bottom();
3251   HeapWord* end_addr = sp->top();
3252 
3253   assert(beg_addr <= dp_addr && dp_addr <= end_addr, "bad dense prefix");
3254 
3255   const size_t beg_region = sd.addr_to_region_idx(beg_addr);
3256   const size_t dp_region = sd.addr_to_region_idx(dp_addr);
3257   if (beg_region < dp_region) {
3258     update_and_deadwood_in_dense_prefix(cm, space_id, beg_region, dp_region);
3259   }
3260 
3261   // The destination of the first live object that starts in the region is one
3262   // past the end of the partial object entering the region (if any).
3263   HeapWord* const dest_addr = sd.partial_obj_end(dp_region);
3264   HeapWord* const new_top = _space_info[space_id].new_top();
3265   assert(new_top >= dest_addr, "bad new_top value");
3266   const size_t words = pointer_delta(new_top, dest_addr);
3267 
3268   if (words > 0) {
3269     ObjectStartArray* start_array = _space_info[space_id].start_array();
3270     MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words);
3271 
3272     ParMarkBitMap::IterationStatus status;
3273     status = bitmap->iterate(&closure, dest_addr, end_addr);
3274     assert(status == ParMarkBitMap::full, "iteration not complete");
3275     assert(bitmap->find_obj_beg(closure.source(), end_addr) == end_addr,
3276            "live objects skipped because closure is full");
3277   }
3278 }
3279 
3280 jlong PSParallelCompact::millis_since_last_gc() {
3281   // We need a monotonically non-decreasing time in ms but
3282   // os::javaTimeMillis() does not guarantee monotonicity.
3283   jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
3284   jlong ret_val = now - _time_of_last_gc;
3285   // XXX See note in genCollectedHeap::millis_since_last_gc().
3286   if (ret_val < 0) {
3287     NOT_PRODUCT(warning("time warp: " JLONG_FORMAT, ret_val);)
3288     return 0;
3289   }
3290   return ret_val;
3291 }
3292 
3293 void PSParallelCompact::reset_millis_since_last_gc() {
3294   // We need a monotonically non-decreasing time in ms but
3295   // os::javaTimeMillis() does not guarantee monotonicity.
3296   _time_of_last_gc = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
3297 }
3298 
3299 ParMarkBitMap::IterationStatus MoveAndUpdateClosure::copy_until_full()
3300 {
3301   if (source() != destination()) {
3302     DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3303     Copy::aligned_conjoint_words(source(), destination(), words_remaining());
3304   }
3305   update_state(words_remaining());
3306   assert(is_full(), "sanity");
3307   return ParMarkBitMap::full;
3308 }
3309 
3310 void MoveAndUpdateClosure::copy_partial_obj()
3311 {
3312   size_t words = words_remaining();
3313 
3314   HeapWord* const range_end = MIN2(source() + words, bitmap()->region_end());
3315   HeapWord* const end_addr = bitmap()->find_obj_end(source(), range_end);
3316   if (end_addr < range_end) {
3317     words = bitmap()->obj_size(source(), end_addr);
3318   }
3319 
3320   // This test is necessary; if omitted, the pointer updates to a partial object
3321   // that crosses the dense prefix boundary could be overwritten.
3322   if (source() != destination()) {
3323     DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3324     Copy::aligned_conjoint_words(source(), destination(), words);
3325   }
3326   update_state(words);
3327 }
3328 
3329 void InstanceKlass::oop_pc_update_pointers(oop obj) {
3330   oop_oop_iterate_oop_maps<true>(obj, PSParallelCompact::adjust_pointer_closure());
3331 }
3332 
3333 void InstanceMirrorKlass::oop_pc_update_pointers(oop obj) {
3334   InstanceKlass::oop_pc_update_pointers(obj);
3335 
3336   oop_oop_iterate_statics<true>(obj, PSParallelCompact::adjust_pointer_closure());
3337 }
3338 
3339 void InstanceClassLoaderKlass::oop_pc_update_pointers(oop obj) {
3340   InstanceKlass::oop_pc_update_pointers(obj);
3341 }
3342 
3343 #ifdef ASSERT
3344 template <class T> static void trace_reference_gc(const char *s, oop obj,
3345                                                   T* referent_addr,
3346                                                   T* next_addr,
3347                                                   T* discovered_addr) {
3348   if(TraceReferenceGC && PrintGCDetails) {
3349     gclog_or_tty->print_cr("%s obj " PTR_FORMAT, s, p2i(obj));
3350     gclog_or_tty->print_cr("     referent_addr/* " PTR_FORMAT " / "
3351                            PTR_FORMAT, p2i(referent_addr),
3352                            referent_addr ? p2i(oopDesc::load_decode_heap_oop(referent_addr)) : NULL);
3353     gclog_or_tty->print_cr("     next_addr/* " PTR_FORMAT " / "
3354                            PTR_FORMAT, p2i(next_addr),
3355                            next_addr ? p2i(oopDesc::load_decode_heap_oop(next_addr)) : NULL);
3356     gclog_or_tty->print_cr("     discovered_addr/* " PTR_FORMAT " / "
3357                            PTR_FORMAT, p2i(discovered_addr),
3358                            discovered_addr ? p2i(oopDesc::load_decode_heap_oop(discovered_addr)) : NULL);
3359   }
3360 }
3361 #endif
3362 
3363 template <class T>
3364 static void oop_pc_update_pointers_specialized(oop obj) {
3365   T* referent_addr = (T*)java_lang_ref_Reference::referent_addr(obj);
3366   PSParallelCompact::adjust_pointer(referent_addr);
3367   T* next_addr = (T*)java_lang_ref_Reference::next_addr(obj);
3368   PSParallelCompact::adjust_pointer(next_addr);
3369   T* discovered_addr = (T*)java_lang_ref_Reference::discovered_addr(obj);
3370   PSParallelCompact::adjust_pointer(discovered_addr);
3371   debug_only(trace_reference_gc("InstanceRefKlass::oop_update_ptrs", obj,
3372                                 referent_addr, next_addr, discovered_addr);)
3373 }
3374 
3375 void InstanceRefKlass::oop_pc_update_pointers(oop obj) {
3376   InstanceKlass::oop_pc_update_pointers(obj);
3377 
3378   if (UseCompressedOops) {
3379     oop_pc_update_pointers_specialized<narrowOop>(obj);
3380   } else {
3381     oop_pc_update_pointers_specialized<oop>(obj);
3382   }
3383 }
3384 
3385 void ObjArrayKlass::oop_pc_update_pointers(oop obj) {
3386   assert(obj->is_objArray(), "obj must be obj array");
3387   oop_oop_iterate_elements<true>(objArrayOop(obj), PSParallelCompact::adjust_pointer_closure());
3388 }
3389 
3390 void TypeArrayKlass::oop_pc_update_pointers(oop obj) {
3391   assert(obj->is_typeArray(),"must be a type array");
3392 }
3393 
3394 ParMarkBitMapClosure::IterationStatus
3395 MoveAndUpdateClosure::do_addr(HeapWord* addr, size_t words) {
3396   assert(destination() != NULL, "sanity");
3397   assert(bitmap()->obj_size(addr) == words, "bad size");
3398 
3399   _source = addr;
3400   assert(PSParallelCompact::summary_data().calc_new_pointer(source()) ==
3401          destination(), "wrong destination");
3402 
3403   if (words > words_remaining()) {
3404     return ParMarkBitMap::would_overflow;
3405   }
3406 
3407   // The start_array must be updated even if the object is not moving.
3408   if (_start_array != NULL) {
3409     _start_array->allocate_block(destination());
3410   }
3411 
3412   if (destination() != source()) {
3413     DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3414     Copy::aligned_conjoint_words(source(), destination(), words);
3415   }
3416 
3417   oop moved_oop = (oop) destination();
3418   compaction_manager()->update_contents(moved_oop);
3419   assert(moved_oop->is_oop_or_null(), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(moved_oop)));
3420 
3421   update_state(words);
3422   assert(destination() == (HeapWord*)moved_oop + moved_oop->size(), "sanity");
3423   return is_full() ? ParMarkBitMap::full : ParMarkBitMap::incomplete;
3424 }
3425 
3426 UpdateOnlyClosure::UpdateOnlyClosure(ParMarkBitMap* mbm,
3427                                      ParCompactionManager* cm,
3428                                      PSParallelCompact::SpaceId space_id) :
3429   ParMarkBitMapClosure(mbm, cm),
3430   _space_id(space_id),
3431   _start_array(PSParallelCompact::start_array(space_id))
3432 {
3433 }
3434 
3435 // Updates the references in the object to their new values.
3436 ParMarkBitMapClosure::IterationStatus
3437 UpdateOnlyClosure::do_addr(HeapWord* addr, size_t words) {
3438   do_addr(addr);
3439   return ParMarkBitMap::incomplete;
3440 }