1 /*
   2  * Copyright (c) 2005, 2014, 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.hpp"
  38 #include "gc_implementation/parallelScavenge/psPromotionManager.inline.hpp"
  39 #include "gc_implementation/parallelScavenge/psScavenge.hpp"
  40 #include "gc_implementation/parallelScavenge/psYoungGen.hpp"
  41 #include "gc_implementation/shared/gcHeapSummary.hpp"
  42 #include "gc_implementation/shared/gcTimer.hpp"
  43 #include "gc_implementation/shared/gcTrace.hpp"
  44 #include "gc_implementation/shared/gcTraceTime.hpp"
  45 #include "gc_implementation/shared/isGCActiveMark.hpp"
  46 #include "gc_implementation/shared/spaceDecorator.hpp"
  47 #include "gc_interface/gcCause.hpp"
  48 #include "memory/gcLocker.inline.hpp"
  49 #include "memory/referencePolicy.hpp"
  50 #include "memory/referenceProcessor.hpp"
  51 #include "oops/methodData.hpp"
  52 #include "oops/oop.inline.hpp"
  53 #include "oops/oop.pcgc.inline.hpp"
  54 #include "runtime/atomic.inline.hpp"
  55 #include "runtime/fprofiler.hpp"
  56 #include "runtime/safepoint.hpp"
  57 #include "runtime/vmThread.hpp"
  58 #include "services/management.hpp"
  59 #include "services/memoryService.hpp"
  60 #include "services/memTracker.hpp"
  61 #include "utilities/events.hpp"
  62 #include "utilities/stack.inline.hpp"
  63 
  64 #include <math.h>
  65 
  66 PRAGMA_FORMAT_MUTE_WARNINGS_FOR_GCC
  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, c->data_location(), dci, 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, 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, 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, split_destination,
 642                            addr_to_region_idx(split_destination),
 643                            *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                              po_beg, addr_to_region_idx(po_beg),
 653                              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                   source_beg, source_end, source_next_val,
 671                   target_beg, target_end, *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(PSParallelCompact::gc_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 void PSParallelCompact::KeepAliveClosure::do_oop(oop* p)       { PSParallelCompact::KeepAliveClosure::do_oop_work(p); }
 824 void PSParallelCompact::KeepAliveClosure::do_oop(narrowOop* p) { PSParallelCompact::KeepAliveClosure::do_oop_work(p); }
 825 
 826 PSParallelCompact::AdjustPointerClosure PSParallelCompact::_adjust_pointer_closure;
 827 PSParallelCompact::AdjustKlassClosure PSParallelCompact::_adjust_klass_closure;
 828 
 829 void PSParallelCompact::AdjustPointerClosure::do_oop(oop* p)       { adjust_pointer(p); }
 830 void PSParallelCompact::AdjustPointerClosure::do_oop(narrowOop* p) { adjust_pointer(p); }
 831 
 832 void PSParallelCompact::FollowStackClosure::do_void() { _compaction_manager->follow_marking_stacks(); }
 833 
 834 void PSParallelCompact::MarkAndPushClosure::do_oop(oop* p)       {
 835   mark_and_push(_compaction_manager, p);
 836 }
 837 void PSParallelCompact::MarkAndPushClosure::do_oop(narrowOop* p) { mark_and_push(_compaction_manager, p); }
 838 
 839 void PSParallelCompact::FollowKlassClosure::do_klass(Klass* klass) {
 840   klass->oops_do(_mark_and_push_closure);
 841 }
 842 void PSParallelCompact::AdjustKlassClosure::do_klass(Klass* klass) {
 843   klass->oops_do(&PSParallelCompact::_adjust_pointer_closure);
 844 }
 845 
 846 void PSParallelCompact::post_initialize() {
 847   ParallelScavengeHeap* heap = gc_heap();
 848   assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity");
 849 
 850   MemRegion mr = heap->reserved_region();
 851   _ref_processor =
 852     new ReferenceProcessor(mr,            // span
 853                            ParallelRefProcEnabled && (ParallelGCThreads > 1), // mt processing
 854                            (int) ParallelGCThreads, // mt processing degree
 855                            true,          // mt discovery
 856                            (int) ParallelGCThreads, // mt discovery degree
 857                            true,          // atomic_discovery
 858                            &_is_alive_closure); // non-header is alive closure
 859   _counters = new CollectorCounters("PSParallelCompact", 1);
 860 
 861   // Initialize static fields in ParCompactionManager.
 862   ParCompactionManager::initialize(mark_bitmap());
 863 }
 864 
 865 bool PSParallelCompact::initialize() {
 866   ParallelScavengeHeap* heap = gc_heap();
 867   assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity");
 868   MemRegion mr = heap->reserved_region();
 869 
 870   // Was the old gen get allocated successfully?
 871   if (!heap->old_gen()->is_allocated()) {
 872     return false;
 873   }
 874 
 875   initialize_space_info();
 876   initialize_dead_wood_limiter();
 877 
 878   if (!_mark_bitmap.initialize(mr)) {
 879     vm_shutdown_during_initialization(
 880       err_msg("Unable to allocate " SIZE_FORMAT "KB bitmaps for parallel "
 881       "garbage collection for the requested " SIZE_FORMAT "KB heap.",
 882       _mark_bitmap.reserved_byte_size()/K, mr.byte_size()/K));
 883     return false;
 884   }
 885 
 886   if (!_summary_data.initialize(mr)) {
 887     vm_shutdown_during_initialization(
 888       err_msg("Unable to allocate " SIZE_FORMAT "KB card tables for parallel "
 889       "garbage collection for the requested " SIZE_FORMAT "KB heap.",
 890       _summary_data.reserved_byte_size()/K, mr.byte_size()/K));
 891     return false;
 892   }
 893 
 894   return true;
 895 }
 896 
 897 void PSParallelCompact::initialize_space_info()
 898 {
 899   memset(&_space_info, 0, sizeof(_space_info));
 900 
 901   ParallelScavengeHeap* heap = gc_heap();
 902   PSYoungGen* young_gen = heap->young_gen();
 903 
 904   _space_info[old_space_id].set_space(heap->old_gen()->object_space());
 905   _space_info[eden_space_id].set_space(young_gen->eden_space());
 906   _space_info[from_space_id].set_space(young_gen->from_space());
 907   _space_info[to_space_id].set_space(young_gen->to_space());
 908 
 909   _space_info[old_space_id].set_start_array(heap->old_gen()->start_array());
 910 }
 911 
 912 void PSParallelCompact::initialize_dead_wood_limiter()
 913 {
 914   const size_t max = 100;
 915   _dwl_mean = double(MIN2(ParallelOldDeadWoodLimiterMean, max)) / 100.0;
 916   _dwl_std_dev = double(MIN2(ParallelOldDeadWoodLimiterStdDev, max)) / 100.0;
 917   _dwl_first_term = 1.0 / (sqrt(2.0 * M_PI) * _dwl_std_dev);
 918   DEBUG_ONLY(_dwl_initialized = true;)
 919   _dwl_adjustment = normal_distribution(1.0);
 920 }
 921 
 922 // Simple class for storing info about the heap at the start of GC, to be used
 923 // after GC for comparison/printing.
 924 class PreGCValues {
 925 public:
 926   PreGCValues() { }
 927   PreGCValues(ParallelScavengeHeap* heap) { fill(heap); }
 928 
 929   void fill(ParallelScavengeHeap* heap) {
 930     _heap_used      = heap->used();
 931     _young_gen_used = heap->young_gen()->used_in_bytes();
 932     _old_gen_used   = heap->old_gen()->used_in_bytes();
 933     _metadata_used  = MetaspaceAux::used_bytes();
 934   };
 935 
 936   size_t heap_used() const      { return _heap_used; }
 937   size_t young_gen_used() const { return _young_gen_used; }
 938   size_t old_gen_used() const   { return _old_gen_used; }
 939   size_t metadata_used() const  { return _metadata_used; }
 940 
 941 private:
 942   size_t _heap_used;
 943   size_t _young_gen_used;
 944   size_t _old_gen_used;
 945   size_t _metadata_used;
 946 };
 947 
 948 void
 949 PSParallelCompact::clear_data_covering_space(SpaceId id)
 950 {
 951   // At this point, top is the value before GC, new_top() is the value that will
 952   // be set at the end of GC.  The marking bitmap is cleared to top; nothing
 953   // should be marked above top.  The summary data is cleared to the larger of
 954   // top & new_top.
 955   MutableSpace* const space = _space_info[id].space();
 956   HeapWord* const bot = space->bottom();
 957   HeapWord* const top = space->top();
 958   HeapWord* const max_top = MAX2(top, _space_info[id].new_top());
 959 
 960   const idx_t beg_bit = _mark_bitmap.addr_to_bit(bot);
 961   const idx_t end_bit = BitMap::word_align_up(_mark_bitmap.addr_to_bit(top));
 962   _mark_bitmap.clear_range(beg_bit, end_bit);
 963 
 964   const size_t beg_region = _summary_data.addr_to_region_idx(bot);
 965   const size_t end_region =
 966     _summary_data.addr_to_region_idx(_summary_data.region_align_up(max_top));
 967   _summary_data.clear_range(beg_region, end_region);
 968 
 969   // Clear the data used to 'split' regions.
 970   SplitInfo& split_info = _space_info[id].split_info();
 971   if (split_info.is_valid()) {
 972     split_info.clear();
 973   }
 974   DEBUG_ONLY(split_info.verify_clear();)
 975 }
 976 
 977 void PSParallelCompact::pre_compact(PreGCValues* pre_gc_values)
 978 {
 979   // Update the from & to space pointers in space_info, since they are swapped
 980   // at each young gen gc.  Do the update unconditionally (even though a
 981   // promotion failure does not swap spaces) because an unknown number of minor
 982   // collections will have swapped the spaces an unknown number of times.
 983   GCTraceTime tm("pre compact", print_phases(), true, &_gc_timer, _gc_tracer.gc_id());
 984   ParallelScavengeHeap* heap = gc_heap();
 985   _space_info[from_space_id].set_space(heap->young_gen()->from_space());
 986   _space_info[to_space_id].set_space(heap->young_gen()->to_space());
 987 
 988   pre_gc_values->fill(heap);
 989 
 990   DEBUG_ONLY(add_obj_count = add_obj_size = 0;)
 991   DEBUG_ONLY(mark_bitmap_count = mark_bitmap_size = 0;)
 992 
 993   // Increment the invocation count
 994   heap->increment_total_collections(true);
 995 
 996   // We need to track unique mark sweep invocations as well.
 997   _total_invocations++;
 998 
 999   heap->print_heap_before_gc();
1000   heap->trace_heap_before_gc(&_gc_tracer);
1001 
1002   // Fill in TLABs
1003   heap->accumulate_statistics_all_tlabs();
1004   heap->ensure_parsability(true);  // retire TLABs
1005 
1006   if (VerifyBeforeGC && heap->total_collections() >= VerifyGCStartAt) {
1007     HandleMark hm;  // Discard invalid handles created during verification
1008     Universe::verify(" VerifyBeforeGC:");
1009   }
1010 
1011   // Verify object start arrays
1012   if (VerifyObjectStartArray &&
1013       VerifyBeforeGC) {
1014     heap->old_gen()->verify_object_start_array();
1015   }
1016 
1017   DEBUG_ONLY(mark_bitmap()->verify_clear();)
1018   DEBUG_ONLY(summary_data().verify_clear();)
1019 
1020   // Have worker threads release resources the next time they run a task.
1021   gc_task_manager()->release_all_resources();
1022 }
1023 
1024 void PSParallelCompact::post_compact()
1025 {
1026   GCTraceTime tm("post compact", print_phases(), true, &_gc_timer, _gc_tracer.gc_id());
1027 
1028   for (unsigned int id = old_space_id; id < last_space_id; ++id) {
1029     // Clear the marking bitmap, summary data and split info.
1030     clear_data_covering_space(SpaceId(id));
1031     // Update top().  Must be done after clearing the bitmap and summary data.
1032     _space_info[id].publish_new_top();
1033   }
1034 
1035   MutableSpace* const eden_space = _space_info[eden_space_id].space();
1036   MutableSpace* const from_space = _space_info[from_space_id].space();
1037   MutableSpace* const to_space   = _space_info[to_space_id].space();
1038 
1039   ParallelScavengeHeap* heap = gc_heap();
1040   bool eden_empty = eden_space->is_empty();
1041   if (!eden_empty) {
1042     eden_empty = absorb_live_data_from_eden(heap->size_policy(),
1043                                             heap->young_gen(), heap->old_gen());
1044   }
1045 
1046   // Update heap occupancy information which is used as input to the soft ref
1047   // clearing policy at the next gc.
1048   Universe::update_heap_info_at_gc();
1049 
1050   bool young_gen_empty = eden_empty && from_space->is_empty() &&
1051     to_space->is_empty();
1052 
1053   BarrierSet* bs = heap->barrier_set();
1054   if (bs->is_a(BarrierSet::ModRef)) {
1055     ModRefBarrierSet* modBS = (ModRefBarrierSet*)bs;
1056     MemRegion old_mr = heap->old_gen()->reserved();
1057 
1058     if (young_gen_empty) {
1059       modBS->clear(MemRegion(old_mr.start(), old_mr.end()));
1060     } else {
1061       modBS->invalidate(MemRegion(old_mr.start(), old_mr.end()));
1062     }
1063   }
1064 
1065   // Delete metaspaces for unloaded class loaders and clean up loader_data graph
1066   ClassLoaderDataGraph::purge();
1067   MetaspaceAux::verify_metrics();
1068 
1069   CodeCache::gc_epilogue();
1070   JvmtiExport::gc_epilogue();
1071 
1072   COMPILER2_PRESENT(DerivedPointerTable::update_pointers());
1073 
1074   ref_processor()->enqueue_discovered_references(NULL);
1075 
1076   if (ZapUnusedHeapArea) {
1077     heap->gen_mangle_unused_area();
1078   }
1079 
1080   // Update time of last GC
1081   reset_millis_since_last_gc();
1082 }
1083 
1084 HeapWord*
1085 PSParallelCompact::compute_dense_prefix_via_density(const SpaceId id,
1086                                                     bool maximum_compaction)
1087 {
1088   const size_t region_size = ParallelCompactData::RegionSize;
1089   const ParallelCompactData& sd = summary_data();
1090 
1091   const MutableSpace* const space = _space_info[id].space();
1092   HeapWord* const top_aligned_up = sd.region_align_up(space->top());
1093   const RegionData* const beg_cp = sd.addr_to_region_ptr(space->bottom());
1094   const RegionData* const end_cp = sd.addr_to_region_ptr(top_aligned_up);
1095 
1096   // Skip full regions at the beginning of the space--they are necessarily part
1097   // of the dense prefix.
1098   size_t full_count = 0;
1099   const RegionData* cp;
1100   for (cp = beg_cp; cp < end_cp && cp->data_size() == region_size; ++cp) {
1101     ++full_count;
1102   }
1103 
1104   assert(total_invocations() >= _maximum_compaction_gc_num, "sanity");
1105   const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num;
1106   const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval;
1107   if (maximum_compaction || cp == end_cp || interval_ended) {
1108     _maximum_compaction_gc_num = total_invocations();
1109     return sd.region_to_addr(cp);
1110   }
1111 
1112   HeapWord* const new_top = _space_info[id].new_top();
1113   const size_t space_live = pointer_delta(new_top, space->bottom());
1114   const size_t space_used = space->used_in_words();
1115   const size_t space_capacity = space->capacity_in_words();
1116 
1117   const double cur_density = double(space_live) / space_capacity;
1118   const double deadwood_density =
1119     (1.0 - cur_density) * (1.0 - cur_density) * cur_density * cur_density;
1120   const size_t deadwood_goal = size_t(space_capacity * deadwood_density);
1121 
1122   if (TraceParallelOldGCDensePrefix) {
1123     tty->print_cr("cur_dens=%5.3f dw_dens=%5.3f dw_goal=" SIZE_FORMAT,
1124                   cur_density, deadwood_density, deadwood_goal);
1125     tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " "
1126                   "space_cap=" SIZE_FORMAT,
1127                   space_live, space_used,
1128                   space_capacity);
1129   }
1130 
1131   // XXX - Use binary search?
1132   HeapWord* dense_prefix = sd.region_to_addr(cp);
1133   const RegionData* full_cp = cp;
1134   const RegionData* const top_cp = sd.addr_to_region_ptr(space->top() - 1);
1135   while (cp < end_cp) {
1136     HeapWord* region_destination = cp->destination();
1137     const size_t cur_deadwood = pointer_delta(dense_prefix, region_destination);
1138     if (TraceParallelOldGCDensePrefix && Verbose) {
1139       tty->print_cr("c#=" SIZE_FORMAT_W(4) " dst=" PTR_FORMAT " "
1140                     "dp=" SIZE_FORMAT_W(8) " " "cdw=" SIZE_FORMAT_W(8),
1141                     sd.region(cp), region_destination,
1142                     dense_prefix, cur_deadwood);
1143     }
1144 
1145     if (cur_deadwood >= deadwood_goal) {
1146       // Found the region that has the correct amount of deadwood to the left.
1147       // This typically occurs after crossing a fairly sparse set of regions, so
1148       // iterate backwards over those sparse regions, looking for the region
1149       // that has the lowest density of live objects 'to the right.'
1150       size_t space_to_left = sd.region(cp) * region_size;
1151       size_t live_to_left = space_to_left - cur_deadwood;
1152       size_t space_to_right = space_capacity - space_to_left;
1153       size_t live_to_right = space_live - live_to_left;
1154       double density_to_right = double(live_to_right) / space_to_right;
1155       while (cp > full_cp) {
1156         --cp;
1157         const size_t prev_region_live_to_right = live_to_right -
1158           cp->data_size();
1159         const size_t prev_region_space_to_right = space_to_right + region_size;
1160         double prev_region_density_to_right =
1161           double(prev_region_live_to_right) / prev_region_space_to_right;
1162         if (density_to_right <= prev_region_density_to_right) {
1163           return dense_prefix;
1164         }
1165         if (TraceParallelOldGCDensePrefix && Verbose) {
1166           tty->print_cr("backing up from c=" SIZE_FORMAT_W(4) " d2r=%10.8f "
1167                         "pc_d2r=%10.8f", sd.region(cp), density_to_right,
1168                         prev_region_density_to_right);
1169         }
1170         dense_prefix -= region_size;
1171         live_to_right = prev_region_live_to_right;
1172         space_to_right = prev_region_space_to_right;
1173         density_to_right = prev_region_density_to_right;
1174       }
1175       return dense_prefix;
1176     }
1177 
1178     dense_prefix += region_size;
1179     ++cp;
1180   }
1181 
1182   return dense_prefix;
1183 }
1184 
1185 #ifndef PRODUCT
1186 void PSParallelCompact::print_dense_prefix_stats(const char* const algorithm,
1187                                                  const SpaceId id,
1188                                                  const bool maximum_compaction,
1189                                                  HeapWord* const addr)
1190 {
1191   const size_t region_idx = summary_data().addr_to_region_idx(addr);
1192   RegionData* const cp = summary_data().region(region_idx);
1193   const MutableSpace* const space = _space_info[id].space();
1194   HeapWord* const new_top = _space_info[id].new_top();
1195 
1196   const size_t space_live = pointer_delta(new_top, space->bottom());
1197   const size_t dead_to_left = pointer_delta(addr, cp->destination());
1198   const size_t space_cap = space->capacity_in_words();
1199   const double dead_to_left_pct = double(dead_to_left) / space_cap;
1200   const size_t live_to_right = new_top - cp->destination();
1201   const size_t dead_to_right = space->top() - addr - live_to_right;
1202 
1203   tty->print_cr("%s=" PTR_FORMAT " dpc=" SIZE_FORMAT_W(5) " "
1204                 "spl=" SIZE_FORMAT " "
1205                 "d2l=" SIZE_FORMAT " d2l%%=%6.4f "
1206                 "d2r=" SIZE_FORMAT " l2r=" SIZE_FORMAT
1207                 " ratio=%10.8f",
1208                 algorithm, addr, region_idx,
1209                 space_live,
1210                 dead_to_left, dead_to_left_pct,
1211                 dead_to_right, live_to_right,
1212                 double(dead_to_right) / live_to_right);
1213 }
1214 #endif  // #ifndef PRODUCT
1215 
1216 // Return a fraction indicating how much of the generation can be treated as
1217 // "dead wood" (i.e., not reclaimed).  The function uses a normal distribution
1218 // based on the density of live objects in the generation to determine a limit,
1219 // which is then adjusted so the return value is min_percent when the density is
1220 // 1.
1221 //
1222 // The following table shows some return values for a different values of the
1223 // standard deviation (ParallelOldDeadWoodLimiterStdDev); the mean is 0.5 and
1224 // min_percent is 1.
1225 //
1226 //                          fraction allowed as dead wood
1227 //         -----------------------------------------------------------------
1228 // density std_dev=70 std_dev=75 std_dev=80 std_dev=85 std_dev=90 std_dev=95
1229 // ------- ---------- ---------- ---------- ---------- ---------- ----------
1230 // 0.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000
1231 // 0.05000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941
1232 // 0.10000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272
1233 // 0.15000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066
1234 // 0.20000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975
1235 // 0.25000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313
1236 // 0.30000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132
1237 // 0.35000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289
1238 // 0.40000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500
1239 // 0.45000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386
1240 // 0.50000 0.13832410 0.11599237 0.09847664 0.08456518 0.07338887 0.06431510
1241 // 0.55000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386
1242 // 0.60000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500
1243 // 0.65000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289
1244 // 0.70000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132
1245 // 0.75000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313
1246 // 0.80000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975
1247 // 0.85000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066
1248 // 0.90000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272
1249 // 0.95000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941
1250 // 1.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000
1251 
1252 double PSParallelCompact::dead_wood_limiter(double density, size_t min_percent)
1253 {
1254   assert(_dwl_initialized, "uninitialized");
1255 
1256   // The raw limit is the value of the normal distribution at x = density.
1257   const double raw_limit = normal_distribution(density);
1258 
1259   // Adjust the raw limit so it becomes the minimum when the density is 1.
1260   //
1261   // First subtract the adjustment value (which is simply the precomputed value
1262   // normal_distribution(1.0)); this yields a value of 0 when the density is 1.
1263   // Then add the minimum value, so the minimum is returned when the density is
1264   // 1.  Finally, prevent negative values, which occur when the mean is not 0.5.
1265   const double min = double(min_percent) / 100.0;
1266   const double limit = raw_limit - _dwl_adjustment + min;
1267   return MAX2(limit, 0.0);
1268 }
1269 
1270 ParallelCompactData::RegionData*
1271 PSParallelCompact::first_dead_space_region(const RegionData* beg,
1272                                            const RegionData* end)
1273 {
1274   const size_t region_size = ParallelCompactData::RegionSize;
1275   ParallelCompactData& sd = summary_data();
1276   size_t left = sd.region(beg);
1277   size_t right = end > beg ? sd.region(end) - 1 : left;
1278 
1279   // Binary search.
1280   while (left < right) {
1281     // Equivalent to (left + right) / 2, but does not overflow.
1282     const size_t middle = left + (right - left) / 2;
1283     RegionData* const middle_ptr = sd.region(middle);
1284     HeapWord* const dest = middle_ptr->destination();
1285     HeapWord* const addr = sd.region_to_addr(middle);
1286     assert(dest != NULL, "sanity");
1287     assert(dest <= addr, "must move left");
1288 
1289     if (middle > left && dest < addr) {
1290       right = middle - 1;
1291     } else if (middle < right && middle_ptr->data_size() == region_size) {
1292       left = middle + 1;
1293     } else {
1294       return middle_ptr;
1295     }
1296   }
1297   return sd.region(left);
1298 }
1299 
1300 ParallelCompactData::RegionData*
1301 PSParallelCompact::dead_wood_limit_region(const RegionData* beg,
1302                                           const RegionData* end,
1303                                           size_t dead_words)
1304 {
1305   ParallelCompactData& sd = summary_data();
1306   size_t left = sd.region(beg);
1307   size_t right = end > beg ? sd.region(end) - 1 : left;
1308 
1309   // Binary search.
1310   while (left < right) {
1311     // Equivalent to (left + right) / 2, but does not overflow.
1312     const size_t middle = left + (right - left) / 2;
1313     RegionData* const middle_ptr = sd.region(middle);
1314     HeapWord* const dest = middle_ptr->destination();
1315     HeapWord* const addr = sd.region_to_addr(middle);
1316     assert(dest != NULL, "sanity");
1317     assert(dest <= addr, "must move left");
1318 
1319     const size_t dead_to_left = pointer_delta(addr, dest);
1320     if (middle > left && dead_to_left > dead_words) {
1321       right = middle - 1;
1322     } else if (middle < right && dead_to_left < dead_words) {
1323       left = middle + 1;
1324     } else {
1325       return middle_ptr;
1326     }
1327   }
1328   return sd.region(left);
1329 }
1330 
1331 // The result is valid during the summary phase, after the initial summarization
1332 // of each space into itself, and before final summarization.
1333 inline double
1334 PSParallelCompact::reclaimed_ratio(const RegionData* const cp,
1335                                    HeapWord* const bottom,
1336                                    HeapWord* const top,
1337                                    HeapWord* const new_top)
1338 {
1339   ParallelCompactData& sd = summary_data();
1340 
1341   assert(cp != NULL, "sanity");
1342   assert(bottom != NULL, "sanity");
1343   assert(top != NULL, "sanity");
1344   assert(new_top != NULL, "sanity");
1345   assert(top >= new_top, "summary data problem?");
1346   assert(new_top > bottom, "space is empty; should not be here");
1347   assert(new_top >= cp->destination(), "sanity");
1348   assert(top >= sd.region_to_addr(cp), "sanity");
1349 
1350   HeapWord* const destination = cp->destination();
1351   const size_t dense_prefix_live  = pointer_delta(destination, bottom);
1352   const size_t compacted_region_live = pointer_delta(new_top, destination);
1353   const size_t compacted_region_used = pointer_delta(top,
1354                                                      sd.region_to_addr(cp));
1355   const size_t reclaimable = compacted_region_used - compacted_region_live;
1356 
1357   const double divisor = dense_prefix_live + 1.25 * compacted_region_live;
1358   return double(reclaimable) / divisor;
1359 }
1360 
1361 // Return the address of the end of the dense prefix, a.k.a. the start of the
1362 // compacted region.  The address is always on a region boundary.
1363 //
1364 // Completely full regions at the left are skipped, since no compaction can
1365 // occur in those regions.  Then the maximum amount of dead wood to allow is
1366 // computed, based on the density (amount live / capacity) of the generation;
1367 // the region with approximately that amount of dead space to the left is
1368 // identified as the limit region.  Regions between the last completely full
1369 // region and the limit region are scanned and the one that has the best
1370 // (maximum) reclaimed_ratio() is selected.
1371 HeapWord*
1372 PSParallelCompact::compute_dense_prefix(const SpaceId id,
1373                                         bool maximum_compaction)
1374 {
1375   if (ParallelOldGCSplitALot) {
1376     if (_space_info[id].dense_prefix() != _space_info[id].space()->bottom()) {
1377       // The value was chosen to provoke splitting a young gen space; use it.
1378       return _space_info[id].dense_prefix();
1379     }
1380   }
1381 
1382   const size_t region_size = ParallelCompactData::RegionSize;
1383   const ParallelCompactData& sd = summary_data();
1384 
1385   const MutableSpace* const space = _space_info[id].space();
1386   HeapWord* const top = space->top();
1387   HeapWord* const top_aligned_up = sd.region_align_up(top);
1388   HeapWord* const new_top = _space_info[id].new_top();
1389   HeapWord* const new_top_aligned_up = sd.region_align_up(new_top);
1390   HeapWord* const bottom = space->bottom();
1391   const RegionData* const beg_cp = sd.addr_to_region_ptr(bottom);
1392   const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up);
1393   const RegionData* const new_top_cp =
1394     sd.addr_to_region_ptr(new_top_aligned_up);
1395 
1396   // Skip full regions at the beginning of the space--they are necessarily part
1397   // of the dense prefix.
1398   const RegionData* const full_cp = first_dead_space_region(beg_cp, new_top_cp);
1399   assert(full_cp->destination() == sd.region_to_addr(full_cp) ||
1400          space->is_empty(), "no dead space allowed to the left");
1401   assert(full_cp->data_size() < region_size || full_cp == new_top_cp - 1,
1402          "region must have dead space");
1403 
1404   // The gc number is saved whenever a maximum compaction is done, and used to
1405   // determine when the maximum compaction interval has expired.  This avoids
1406   // successive max compactions for different reasons.
1407   assert(total_invocations() >= _maximum_compaction_gc_num, "sanity");
1408   const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num;
1409   const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval ||
1410     total_invocations() == HeapFirstMaximumCompactionCount;
1411   if (maximum_compaction || full_cp == top_cp || interval_ended) {
1412     _maximum_compaction_gc_num = total_invocations();
1413     return sd.region_to_addr(full_cp);
1414   }
1415 
1416   const size_t space_live = pointer_delta(new_top, bottom);
1417   const size_t space_used = space->used_in_words();
1418   const size_t space_capacity = space->capacity_in_words();
1419 
1420   const double density = double(space_live) / double(space_capacity);
1421   const size_t min_percent_free = MarkSweepDeadRatio;
1422   const double limiter = dead_wood_limiter(density, min_percent_free);
1423   const size_t dead_wood_max = space_used - space_live;
1424   const size_t dead_wood_limit = MIN2(size_t(space_capacity * limiter),
1425                                       dead_wood_max);
1426 
1427   if (TraceParallelOldGCDensePrefix) {
1428     tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " "
1429                   "space_cap=" SIZE_FORMAT,
1430                   space_live, space_used,
1431                   space_capacity);
1432     tty->print_cr("dead_wood_limiter(%6.4f, " SIZE_FORMAT ")=%6.4f "
1433                   "dead_wood_max=" SIZE_FORMAT " dead_wood_limit=" SIZE_FORMAT,
1434                   density, min_percent_free, limiter,
1435                   dead_wood_max, dead_wood_limit);
1436   }
1437 
1438   // Locate the region with the desired amount of dead space to the left.
1439   const RegionData* const limit_cp =
1440     dead_wood_limit_region(full_cp, top_cp, dead_wood_limit);
1441 
1442   // Scan from the first region with dead space to the limit region and find the
1443   // one with the best (largest) reclaimed ratio.
1444   double best_ratio = 0.0;
1445   const RegionData* best_cp = full_cp;
1446   for (const RegionData* cp = full_cp; cp < limit_cp; ++cp) {
1447     double tmp_ratio = reclaimed_ratio(cp, bottom, top, new_top);
1448     if (tmp_ratio > best_ratio) {
1449       best_cp = cp;
1450       best_ratio = tmp_ratio;
1451     }
1452   }
1453 
1454 #if     0
1455   // Something to consider:  if the region with the best ratio is 'close to' the
1456   // first region w/free space, choose the first region with free space
1457   // ("first-free").  The first-free region is usually near the start of the
1458   // heap, which means we are copying most of the heap already, so copy a bit
1459   // more to get complete compaction.
1460   if (pointer_delta(best_cp, full_cp, sizeof(RegionData)) < 4) {
1461     _maximum_compaction_gc_num = total_invocations();
1462     best_cp = full_cp;
1463   }
1464 #endif  // #if 0
1465 
1466   return sd.region_to_addr(best_cp);
1467 }
1468 
1469 #ifndef PRODUCT
1470 void
1471 PSParallelCompact::fill_with_live_objects(SpaceId id, HeapWord* const start,
1472                                           size_t words)
1473 {
1474   if (TraceParallelOldGCSummaryPhase) {
1475     tty->print_cr("fill_with_live_objects [" PTR_FORMAT " " PTR_FORMAT ") "
1476                   SIZE_FORMAT, start, start + words, words);
1477   }
1478 
1479   ObjectStartArray* const start_array = _space_info[id].start_array();
1480   CollectedHeap::fill_with_objects(start, words);
1481   for (HeapWord* p = start; p < start + words; p += oop(p)->size()) {
1482     _mark_bitmap.mark_obj(p, words);
1483     _summary_data.add_obj(p, words);
1484     start_array->allocate_block(p);
1485   }
1486 }
1487 
1488 void
1489 PSParallelCompact::summarize_new_objects(SpaceId id, HeapWord* start)
1490 {
1491   ParallelCompactData& sd = summary_data();
1492   MutableSpace* space = _space_info[id].space();
1493 
1494   // Find the source and destination start addresses.
1495   HeapWord* const src_addr = sd.region_align_down(start);
1496   HeapWord* dst_addr;
1497   if (src_addr < start) {
1498     dst_addr = sd.addr_to_region_ptr(src_addr)->destination();
1499   } else if (src_addr > space->bottom()) {
1500     // The start (the original top() value) is aligned to a region boundary so
1501     // the associated region does not have a destination.  Compute the
1502     // destination from the previous region.
1503     RegionData* const cp = sd.addr_to_region_ptr(src_addr) - 1;
1504     dst_addr = cp->destination() + cp->data_size();
1505   } else {
1506     // Filling the entire space.
1507     dst_addr = space->bottom();
1508   }
1509   assert(dst_addr != NULL, "sanity");
1510 
1511   // Update the summary data.
1512   bool result = _summary_data.summarize(_space_info[id].split_info(),
1513                                         src_addr, space->top(), NULL,
1514                                         dst_addr, space->end(),
1515                                         _space_info[id].new_top_addr());
1516   assert(result, "should not fail:  bad filler object size");
1517 }
1518 
1519 void
1520 PSParallelCompact::provoke_split_fill_survivor(SpaceId id)
1521 {
1522   if (total_invocations() % (ParallelOldGCSplitInterval * 3) != 0) {
1523     return;
1524   }
1525 
1526   MutableSpace* const space = _space_info[id].space();
1527   if (space->is_empty()) {
1528     HeapWord* b = space->bottom();
1529     HeapWord* t = b + space->capacity_in_words() / 2;
1530     space->set_top(t);
1531     if (ZapUnusedHeapArea) {
1532       space->set_top_for_allocations();
1533     }
1534 
1535     size_t min_size = CollectedHeap::min_fill_size();
1536     size_t obj_len = min_size;
1537     while (b + obj_len <= t) {
1538       CollectedHeap::fill_with_object(b, obj_len);
1539       mark_bitmap()->mark_obj(b, obj_len);
1540       summary_data().add_obj(b, obj_len);
1541       b += obj_len;
1542       obj_len = (obj_len & (min_size*3)) + min_size; // 8 16 24 32 8 16 24 32 ...
1543     }
1544     if (b < t) {
1545       // The loop didn't completely fill to t (top); adjust top downward.
1546       space->set_top(b);
1547       if (ZapUnusedHeapArea) {
1548         space->set_top_for_allocations();
1549       }
1550     }
1551 
1552     HeapWord** nta = _space_info[id].new_top_addr();
1553     bool result = summary_data().summarize(_space_info[id].split_info(),
1554                                            space->bottom(), space->top(), NULL,
1555                                            space->bottom(), space->end(), nta);
1556     assert(result, "space must fit into itself");
1557   }
1558 }
1559 
1560 void
1561 PSParallelCompact::provoke_split(bool & max_compaction)
1562 {
1563   if (total_invocations() % ParallelOldGCSplitInterval != 0) {
1564     return;
1565   }
1566 
1567   const size_t region_size = ParallelCompactData::RegionSize;
1568   ParallelCompactData& sd = summary_data();
1569 
1570   MutableSpace* const eden_space = _space_info[eden_space_id].space();
1571   MutableSpace* const from_space = _space_info[from_space_id].space();
1572   const size_t eden_live = pointer_delta(eden_space->top(),
1573                                          _space_info[eden_space_id].new_top());
1574   const size_t from_live = pointer_delta(from_space->top(),
1575                                          _space_info[from_space_id].new_top());
1576 
1577   const size_t min_fill_size = CollectedHeap::min_fill_size();
1578   const size_t eden_free = pointer_delta(eden_space->end(), eden_space->top());
1579   const size_t eden_fillable = eden_free >= min_fill_size ? eden_free : 0;
1580   const size_t from_free = pointer_delta(from_space->end(), from_space->top());
1581   const size_t from_fillable = from_free >= min_fill_size ? from_free : 0;
1582 
1583   // Choose the space to split; need at least 2 regions live (or fillable).
1584   SpaceId id;
1585   MutableSpace* space;
1586   size_t live_words;
1587   size_t fill_words;
1588   if (eden_live + eden_fillable >= region_size * 2) {
1589     id = eden_space_id;
1590     space = eden_space;
1591     live_words = eden_live;
1592     fill_words = eden_fillable;
1593   } else if (from_live + from_fillable >= region_size * 2) {
1594     id = from_space_id;
1595     space = from_space;
1596     live_words = from_live;
1597     fill_words = from_fillable;
1598   } else {
1599     return; // Give up.
1600   }
1601   assert(fill_words == 0 || fill_words >= min_fill_size, "sanity");
1602 
1603   if (live_words < region_size * 2) {
1604     // Fill from top() to end() w/live objects of mixed sizes.
1605     HeapWord* const fill_start = space->top();
1606     live_words += fill_words;
1607 
1608     space->set_top(fill_start + fill_words);
1609     if (ZapUnusedHeapArea) {
1610       space->set_top_for_allocations();
1611     }
1612 
1613     HeapWord* cur_addr = fill_start;
1614     while (fill_words > 0) {
1615       const size_t r = (size_t)os::random() % (region_size / 2) + min_fill_size;
1616       size_t cur_size = MIN2(align_object_size_(r), fill_words);
1617       if (fill_words - cur_size < min_fill_size) {
1618         cur_size = fill_words; // Avoid leaving a fragment too small to fill.
1619       }
1620 
1621       CollectedHeap::fill_with_object(cur_addr, cur_size);
1622       mark_bitmap()->mark_obj(cur_addr, cur_size);
1623       sd.add_obj(cur_addr, cur_size);
1624 
1625       cur_addr += cur_size;
1626       fill_words -= cur_size;
1627     }
1628 
1629     summarize_new_objects(id, fill_start);
1630   }
1631 
1632   max_compaction = false;
1633 
1634   // Manipulate the old gen so that it has room for about half of the live data
1635   // in the target young gen space (live_words / 2).
1636   id = old_space_id;
1637   space = _space_info[id].space();
1638   const size_t free_at_end = space->free_in_words();
1639   const size_t free_target = align_object_size(live_words / 2);
1640   const size_t dead = pointer_delta(space->top(), _space_info[id].new_top());
1641 
1642   if (free_at_end >= free_target + min_fill_size) {
1643     // Fill space above top() and set the dense prefix so everything survives.
1644     HeapWord* const fill_start = space->top();
1645     const size_t fill_size = free_at_end - free_target;
1646     space->set_top(space->top() + fill_size);
1647     if (ZapUnusedHeapArea) {
1648       space->set_top_for_allocations();
1649     }
1650     fill_with_live_objects(id, fill_start, fill_size);
1651     summarize_new_objects(id, fill_start);
1652     _space_info[id].set_dense_prefix(sd.region_align_down(space->top()));
1653   } else if (dead + free_at_end > free_target) {
1654     // Find a dense prefix that makes the right amount of space available.
1655     HeapWord* cur = sd.region_align_down(space->top());
1656     HeapWord* cur_destination = sd.addr_to_region_ptr(cur)->destination();
1657     size_t dead_to_right = pointer_delta(space->end(), cur_destination);
1658     while (dead_to_right < free_target) {
1659       cur -= region_size;
1660       cur_destination = sd.addr_to_region_ptr(cur)->destination();
1661       dead_to_right = pointer_delta(space->end(), cur_destination);
1662     }
1663     _space_info[id].set_dense_prefix(cur);
1664   }
1665 }
1666 #endif // #ifndef PRODUCT
1667 
1668 void PSParallelCompact::summarize_spaces_quick()
1669 {
1670   for (unsigned int i = 0; i < last_space_id; ++i) {
1671     const MutableSpace* space = _space_info[i].space();
1672     HeapWord** nta = _space_info[i].new_top_addr();
1673     bool result = _summary_data.summarize(_space_info[i].split_info(),
1674                                           space->bottom(), space->top(), NULL,
1675                                           space->bottom(), space->end(), nta);
1676     assert(result, "space must fit into itself");
1677     _space_info[i].set_dense_prefix(space->bottom());
1678   }
1679 
1680 #ifndef PRODUCT
1681   if (ParallelOldGCSplitALot) {
1682     provoke_split_fill_survivor(to_space_id);
1683   }
1684 #endif // #ifndef PRODUCT
1685 }
1686 
1687 void PSParallelCompact::fill_dense_prefix_end(SpaceId id)
1688 {
1689   HeapWord* const dense_prefix_end = dense_prefix(id);
1690   const RegionData* region = _summary_data.addr_to_region_ptr(dense_prefix_end);
1691   const idx_t dense_prefix_bit = _mark_bitmap.addr_to_bit(dense_prefix_end);
1692   if (dead_space_crosses_boundary(region, dense_prefix_bit)) {
1693     // Only enough dead space is filled so that any remaining dead space to the
1694     // left is larger than the minimum filler object.  (The remainder is filled
1695     // during the copy/update phase.)
1696     //
1697     // The size of the dead space to the right of the boundary is not a
1698     // concern, since compaction will be able to use whatever space is
1699     // available.
1700     //
1701     // Here '||' is the boundary, 'x' represents a don't care bit and a box
1702     // surrounds the space to be filled with an object.
1703     //
1704     // In the 32-bit VM, each bit represents two 32-bit words:
1705     //                              +---+
1706     // a) beg_bits:  ...  x   x   x | 0 | ||   0   x  x  ...
1707     //    end_bits:  ...  x   x   x | 0 | ||   0   x  x  ...
1708     //                              +---+
1709     //
1710     // In the 64-bit VM, each bit represents one 64-bit word:
1711     //                              +------------+
1712     // b) beg_bits:  ...  x   x   x | 0   ||   0 | x  x  ...
1713     //    end_bits:  ...  x   x   1 | 0   ||   0 | x  x  ...
1714     //                              +------------+
1715     //                          +-------+
1716     // c) beg_bits:  ...  x   x | 0   0 | ||   0   x  x  ...
1717     //    end_bits:  ...  x   1 | 0   0 | ||   0   x  x  ...
1718     //                          +-------+
1719     //                      +-----------+
1720     // d) beg_bits:  ...  x | 0   0   0 | ||   0   x  x  ...
1721     //    end_bits:  ...  1 | 0   0   0 | ||   0   x  x  ...
1722     //                      +-----------+
1723     //                          +-------+
1724     // e) beg_bits:  ...  0   0 | 0   0 | ||   0   x  x  ...
1725     //    end_bits:  ...  0   0 | 0   0 | ||   0   x  x  ...
1726     //                          +-------+
1727 
1728     // Initially assume case a, c or e will apply.
1729     size_t obj_len = CollectedHeap::min_fill_size();
1730     HeapWord* obj_beg = dense_prefix_end - obj_len;
1731 
1732 #ifdef  _LP64
1733     if (MinObjAlignment > 1) { // object alignment > heap word size
1734       // Cases a, c or e.
1735     } else if (_mark_bitmap.is_obj_end(dense_prefix_bit - 2)) {
1736       // Case b above.
1737       obj_beg = dense_prefix_end - 1;
1738     } else if (!_mark_bitmap.is_obj_end(dense_prefix_bit - 3) &&
1739                _mark_bitmap.is_obj_end(dense_prefix_bit - 4)) {
1740       // Case d above.
1741       obj_beg = dense_prefix_end - 3;
1742       obj_len = 3;
1743     }
1744 #endif  // #ifdef _LP64
1745 
1746     CollectedHeap::fill_with_object(obj_beg, obj_len);
1747     _mark_bitmap.mark_obj(obj_beg, obj_len);
1748     _summary_data.add_obj(obj_beg, obj_len);
1749     assert(start_array(id) != NULL, "sanity");
1750     start_array(id)->allocate_block(obj_beg);
1751   }
1752 }
1753 
1754 void
1755 PSParallelCompact::clear_source_region(HeapWord* beg_addr, HeapWord* end_addr)
1756 {
1757   RegionData* const beg_ptr = _summary_data.addr_to_region_ptr(beg_addr);
1758   HeapWord* const end_aligned_up = _summary_data.region_align_up(end_addr);
1759   RegionData* const end_ptr = _summary_data.addr_to_region_ptr(end_aligned_up);
1760   for (RegionData* cur = beg_ptr; cur < end_ptr; ++cur) {
1761     cur->set_source_region(0);
1762   }
1763 }
1764 
1765 void
1766 PSParallelCompact::summarize_space(SpaceId id, bool maximum_compaction)
1767 {
1768   assert(id < last_space_id, "id out of range");
1769   assert(_space_info[id].dense_prefix() == _space_info[id].space()->bottom() ||
1770          ParallelOldGCSplitALot && id == old_space_id,
1771          "should have been reset in summarize_spaces_quick()");
1772 
1773   const MutableSpace* space = _space_info[id].space();
1774   if (_space_info[id].new_top() != space->bottom()) {
1775     HeapWord* dense_prefix_end = compute_dense_prefix(id, maximum_compaction);
1776     _space_info[id].set_dense_prefix(dense_prefix_end);
1777 
1778 #ifndef PRODUCT
1779     if (TraceParallelOldGCDensePrefix) {
1780       print_dense_prefix_stats("ratio", id, maximum_compaction,
1781                                dense_prefix_end);
1782       HeapWord* addr = compute_dense_prefix_via_density(id, maximum_compaction);
1783       print_dense_prefix_stats("density", id, maximum_compaction, addr);
1784     }
1785 #endif  // #ifndef PRODUCT
1786 
1787     // Recompute the summary data, taking into account the dense prefix.  If
1788     // every last byte will be reclaimed, then the existing summary data which
1789     // compacts everything can be left in place.
1790     if (!maximum_compaction && dense_prefix_end != space->bottom()) {
1791       // If dead space crosses the dense prefix boundary, it is (at least
1792       // partially) filled with a dummy object, marked live and added to the
1793       // summary data.  This simplifies the copy/update phase and must be done
1794       // before the final locations of objects are determined, to prevent
1795       // leaving a fragment of dead space that is too small to fill.
1796       fill_dense_prefix_end(id);
1797 
1798       // Compute the destination of each Region, and thus each object.
1799       _summary_data.summarize_dense_prefix(space->bottom(), dense_prefix_end);
1800       _summary_data.summarize(_space_info[id].split_info(),
1801                               dense_prefix_end, space->top(), NULL,
1802                               dense_prefix_end, space->end(),
1803                               _space_info[id].new_top_addr());
1804     }
1805   }
1806 
1807   if (TraceParallelOldGCSummaryPhase) {
1808     const size_t region_size = ParallelCompactData::RegionSize;
1809     HeapWord* const dense_prefix_end = _space_info[id].dense_prefix();
1810     const size_t dp_region = _summary_data.addr_to_region_idx(dense_prefix_end);
1811     const size_t dp_words = pointer_delta(dense_prefix_end, space->bottom());
1812     HeapWord* const new_top = _space_info[id].new_top();
1813     const HeapWord* nt_aligned_up = _summary_data.region_align_up(new_top);
1814     const size_t cr_words = pointer_delta(nt_aligned_up, dense_prefix_end);
1815     tty->print_cr("id=%d cap=" SIZE_FORMAT " dp=" PTR_FORMAT " "
1816                   "dp_region=" SIZE_FORMAT " " "dp_count=" SIZE_FORMAT " "
1817                   "cr_count=" SIZE_FORMAT " " "nt=" PTR_FORMAT,
1818                   id, space->capacity_in_words(), dense_prefix_end,
1819                   dp_region, dp_words / region_size,
1820                   cr_words / region_size, new_top);
1821   }
1822 }
1823 
1824 #ifndef PRODUCT
1825 void PSParallelCompact::summary_phase_msg(SpaceId dst_space_id,
1826                                           HeapWord* dst_beg, HeapWord* dst_end,
1827                                           SpaceId src_space_id,
1828                                           HeapWord* src_beg, HeapWord* src_end)
1829 {
1830   if (TraceParallelOldGCSummaryPhase) {
1831     tty->print_cr("summarizing %d [%s] into %d [%s]:  "
1832                   "src=" PTR_FORMAT "-" PTR_FORMAT " "
1833                   SIZE_FORMAT "-" SIZE_FORMAT " "
1834                   "dst=" PTR_FORMAT "-" PTR_FORMAT " "
1835                   SIZE_FORMAT "-" SIZE_FORMAT,
1836                   src_space_id, space_names[src_space_id],
1837                   dst_space_id, space_names[dst_space_id],
1838                   src_beg, src_end,
1839                   _summary_data.addr_to_region_idx(src_beg),
1840                   _summary_data.addr_to_region_idx(src_end),
1841                   dst_beg, dst_end,
1842                   _summary_data.addr_to_region_idx(dst_beg),
1843                   _summary_data.addr_to_region_idx(dst_end));
1844   }
1845 }
1846 #endif  // #ifndef PRODUCT
1847 
1848 void PSParallelCompact::summary_phase(ParCompactionManager* cm,
1849                                       bool maximum_compaction)
1850 {
1851   GCTraceTime tm("summary phase", print_phases(), true, &_gc_timer, _gc_tracer.gc_id());
1852   // trace("2");
1853 
1854 #ifdef  ASSERT
1855   if (TraceParallelOldGCMarkingPhase) {
1856     tty->print_cr("add_obj_count=" SIZE_FORMAT " "
1857                   "add_obj_bytes=" SIZE_FORMAT,
1858                   add_obj_count, add_obj_size * HeapWordSize);
1859     tty->print_cr("mark_bitmap_count=" SIZE_FORMAT " "
1860                   "mark_bitmap_bytes=" SIZE_FORMAT,
1861                   mark_bitmap_count, mark_bitmap_size * HeapWordSize);
1862   }
1863 #endif  // #ifdef ASSERT
1864 
1865   // Quick summarization of each space into itself, to see how much is live.
1866   summarize_spaces_quick();
1867 
1868   if (TraceParallelOldGCSummaryPhase) {
1869     tty->print_cr("summary_phase:  after summarizing each space to self");
1870     Universe::print();
1871     NOT_PRODUCT(print_region_ranges());
1872     if (Verbose) {
1873       NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info));
1874     }
1875   }
1876 
1877   // The amount of live data that will end up in old space (assuming it fits).
1878   size_t old_space_total_live = 0;
1879   for (unsigned int id = old_space_id; id < last_space_id; ++id) {
1880     old_space_total_live += pointer_delta(_space_info[id].new_top(),
1881                                           _space_info[id].space()->bottom());
1882   }
1883 
1884   MutableSpace* const old_space = _space_info[old_space_id].space();
1885   const size_t old_capacity = old_space->capacity_in_words();
1886   if (old_space_total_live > old_capacity) {
1887     // XXX - should also try to expand
1888     maximum_compaction = true;
1889   }
1890 #ifndef PRODUCT
1891   if (ParallelOldGCSplitALot && old_space_total_live < old_capacity) {
1892     provoke_split(maximum_compaction);
1893   }
1894 #endif // #ifndef PRODUCT
1895 
1896   // Old generations.
1897   summarize_space(old_space_id, maximum_compaction);
1898 
1899   // Summarize the remaining spaces in the young gen.  The initial target space
1900   // is the old gen.  If a space does not fit entirely into the target, then the
1901   // remainder is compacted into the space itself and that space becomes the new
1902   // target.
1903   SpaceId dst_space_id = old_space_id;
1904   HeapWord* dst_space_end = old_space->end();
1905   HeapWord** new_top_addr = _space_info[dst_space_id].new_top_addr();
1906   for (unsigned int id = eden_space_id; id < last_space_id; ++id) {
1907     const MutableSpace* space = _space_info[id].space();
1908     const size_t live = pointer_delta(_space_info[id].new_top(),
1909                                       space->bottom());
1910     const size_t available = pointer_delta(dst_space_end, *new_top_addr);
1911 
1912     NOT_PRODUCT(summary_phase_msg(dst_space_id, *new_top_addr, dst_space_end,
1913                                   SpaceId(id), space->bottom(), space->top());)
1914     if (live > 0 && live <= available) {
1915       // All the live data will fit.
1916       bool done = _summary_data.summarize(_space_info[id].split_info(),
1917                                           space->bottom(), space->top(),
1918                                           NULL,
1919                                           *new_top_addr, dst_space_end,
1920                                           new_top_addr);
1921       assert(done, "space must fit into old gen");
1922 
1923       // Reset the new_top value for the space.
1924       _space_info[id].set_new_top(space->bottom());
1925     } else if (live > 0) {
1926       // Attempt to fit part of the source space into the target space.
1927       HeapWord* next_src_addr = NULL;
1928       bool done = _summary_data.summarize(_space_info[id].split_info(),
1929                                           space->bottom(), space->top(),
1930                                           &next_src_addr,
1931                                           *new_top_addr, dst_space_end,
1932                                           new_top_addr);
1933       assert(!done, "space should not fit into old gen");
1934       assert(next_src_addr != NULL, "sanity");
1935 
1936       // The source space becomes the new target, so the remainder is compacted
1937       // within the space itself.
1938       dst_space_id = SpaceId(id);
1939       dst_space_end = space->end();
1940       new_top_addr = _space_info[id].new_top_addr();
1941       NOT_PRODUCT(summary_phase_msg(dst_space_id,
1942                                     space->bottom(), dst_space_end,
1943                                     SpaceId(id), next_src_addr, space->top());)
1944       done = _summary_data.summarize(_space_info[id].split_info(),
1945                                      next_src_addr, space->top(),
1946                                      NULL,
1947                                      space->bottom(), dst_space_end,
1948                                      new_top_addr);
1949       assert(done, "space must fit when compacted into itself");
1950       assert(*new_top_addr <= space->top(), "usage should not grow");
1951     }
1952   }
1953 
1954   if (TraceParallelOldGCSummaryPhase) {
1955     tty->print_cr("summary_phase:  after final summarization");
1956     Universe::print();
1957     NOT_PRODUCT(print_region_ranges());
1958     if (Verbose) {
1959       NOT_PRODUCT(print_generic_summary_data(_summary_data, _space_info));
1960     }
1961   }
1962 }
1963 
1964 // This method should contain all heap-specific policy for invoking a full
1965 // collection.  invoke_no_policy() will only attempt to compact the heap; it
1966 // will do nothing further.  If we need to bail out for policy reasons, scavenge
1967 // before full gc, or any other specialized behavior, it needs to be added here.
1968 //
1969 // Note that this method should only be called from the vm_thread while at a
1970 // safepoint.
1971 //
1972 // Note that the all_soft_refs_clear flag in the collector policy
1973 // may be true because this method can be called without intervening
1974 // activity.  For example when the heap space is tight and full measure
1975 // are being taken to free space.
1976 void PSParallelCompact::invoke(bool maximum_heap_compaction) {
1977   assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
1978   assert(Thread::current() == (Thread*)VMThread::vm_thread(),
1979          "should be in vm thread");
1980 
1981   ParallelScavengeHeap* heap = gc_heap();
1982   GCCause::Cause gc_cause = heap->gc_cause();
1983   assert(!heap->is_gc_active(), "not reentrant");
1984 
1985   PSAdaptiveSizePolicy* policy = heap->size_policy();
1986   IsGCActiveMark mark;
1987 
1988   if (ScavengeBeforeFullGC) {
1989     PSScavenge::invoke_no_policy();
1990   }
1991 
1992   const bool clear_all_soft_refs =
1993     heap->collector_policy()->should_clear_all_soft_refs();
1994 
1995   PSParallelCompact::invoke_no_policy(clear_all_soft_refs ||
1996                                       maximum_heap_compaction);
1997 }
1998 
1999 // This method contains no policy. You should probably
2000 // be calling invoke() instead.
2001 bool PSParallelCompact::invoke_no_policy(bool maximum_heap_compaction) {
2002   assert(SafepointSynchronize::is_at_safepoint(), "must be at a safepoint");
2003   assert(ref_processor() != NULL, "Sanity");
2004 
2005   if (GC_locker::check_active_before_gc()) {
2006     return false;
2007   }
2008 
2009   ParallelScavengeHeap* heap = gc_heap();
2010 
2011   _gc_timer.register_gc_start();
2012   _gc_tracer.report_gc_start(heap->gc_cause(), _gc_timer.gc_start());
2013 
2014   TimeStamp marking_start;
2015   TimeStamp compaction_start;
2016   TimeStamp collection_exit;
2017 
2018   GCCause::Cause gc_cause = heap->gc_cause();
2019   PSYoungGen* young_gen = heap->young_gen();
2020   PSOldGen* old_gen = heap->old_gen();
2021   PSAdaptiveSizePolicy* size_policy = heap->size_policy();
2022 
2023   // The scope of casr should end after code that can change
2024   // CollectorPolicy::_should_clear_all_soft_refs.
2025   ClearedAllSoftRefs casr(maximum_heap_compaction,
2026                           heap->collector_policy());
2027 
2028   if (ZapUnusedHeapArea) {
2029     // Save information needed to minimize mangling
2030     heap->record_gen_tops_before_GC();
2031   }
2032 
2033   heap->pre_full_gc_dump(&_gc_timer);
2034 
2035   _print_phases = PrintGCDetails && PrintParallelOldGCPhaseTimes;
2036 
2037   // Make sure data structures are sane, make the heap parsable, and do other
2038   // miscellaneous bookkeeping.
2039   PreGCValues pre_gc_values;
2040   pre_compact(&pre_gc_values);
2041 
2042   // Get the compaction manager reserved for the VM thread.
2043   ParCompactionManager* const vmthread_cm =
2044     ParCompactionManager::manager_array(gc_task_manager()->workers());
2045 
2046   // Place after pre_compact() where the number of invocations is incremented.
2047   AdaptiveSizePolicyOutput(size_policy, heap->total_collections());
2048 
2049   {
2050     ResourceMark rm;
2051     HandleMark hm;
2052 
2053     // Set the number of GC threads to be used in this collection
2054     gc_task_manager()->set_active_gang();
2055     gc_task_manager()->task_idle_workers();
2056     heap->set_par_threads(gc_task_manager()->active_workers());
2057 
2058     TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
2059     GCTraceTime t1(GCCauseString("Full GC", gc_cause), PrintGC, !PrintGCDetails, NULL, _gc_tracer.gc_id());
2060     TraceCollectorStats tcs(counters());
2061     TraceMemoryManagerStats tms(true /* Full GC */,gc_cause);
2062 
2063     if (TraceOldGenTime) accumulated_time()->start();
2064 
2065     // Let the size policy know we're starting
2066     size_policy->major_collection_begin();
2067 
2068     CodeCache::gc_prologue();
2069 
2070     COMPILER2_PRESENT(DerivedPointerTable::clear());
2071 
2072     ref_processor()->enable_discovery();
2073     ref_processor()->setup_policy(maximum_heap_compaction);
2074 
2075     bool marked_for_unloading = false;
2076 
2077     marking_start.update();
2078     marking_phase(vmthread_cm, maximum_heap_compaction, &_gc_tracer);
2079 
2080     bool max_on_system_gc = UseMaximumCompactionOnSystemGC
2081       && gc_cause == GCCause::_java_lang_system_gc;
2082     summary_phase(vmthread_cm, maximum_heap_compaction || max_on_system_gc);
2083 
2084     COMPILER2_PRESENT(assert(DerivedPointerTable::is_active(), "Sanity"));
2085     COMPILER2_PRESENT(DerivedPointerTable::set_active(false));
2086 
2087     // adjust_roots() updates Universe::_intArrayKlassObj which is
2088     // needed by the compaction for filling holes in the dense prefix.
2089     adjust_roots();
2090 
2091     compaction_start.update();
2092     compact();
2093 
2094     // Reset the mark bitmap, summary data, and do other bookkeeping.  Must be
2095     // done before resizing.
2096     post_compact();
2097 
2098     // Let the size policy know we're done
2099     size_policy->major_collection_end(old_gen->used_in_bytes(), gc_cause);
2100 
2101     if (UseAdaptiveSizePolicy) {
2102       if (PrintAdaptiveSizePolicy) {
2103         gclog_or_tty->print("AdaptiveSizeStart: ");
2104         gclog_or_tty->stamp();
2105         gclog_or_tty->print_cr(" collection: %d ",
2106                        heap->total_collections());
2107         if (Verbose) {
2108           gclog_or_tty->print("old_gen_capacity: " SIZE_FORMAT
2109             " young_gen_capacity: " SIZE_FORMAT,
2110             old_gen->capacity_in_bytes(), young_gen->capacity_in_bytes());
2111         }
2112       }
2113 
2114       // Don't check if the size_policy is ready here.  Let
2115       // the size_policy check that internally.
2116       if (UseAdaptiveGenerationSizePolicyAtMajorCollection &&
2117           ((gc_cause != GCCause::_java_lang_system_gc) ||
2118             UseAdaptiveSizePolicyWithSystemGC)) {
2119         // Swap the survivor spaces if from_space is empty. The
2120         // resize_young_gen() called below is normally used after
2121         // a successful young GC and swapping of survivor spaces;
2122         // otherwise, it will fail to resize the young gen with
2123         // the current implementation.
2124         if (young_gen->from_space()->is_empty()) {
2125           young_gen->from_space()->clear(SpaceDecorator::Mangle);
2126           young_gen->swap_spaces();
2127         }
2128 
2129         // Calculate optimal free space amounts
2130         assert(young_gen->max_size() >
2131           young_gen->from_space()->capacity_in_bytes() +
2132           young_gen->to_space()->capacity_in_bytes(),
2133           "Sizes of space in young gen are out-of-bounds");
2134 
2135         size_t young_live = young_gen->used_in_bytes();
2136         size_t eden_live = young_gen->eden_space()->used_in_bytes();
2137         size_t old_live = old_gen->used_in_bytes();
2138         size_t cur_eden = young_gen->eden_space()->capacity_in_bytes();
2139         size_t max_old_gen_size = old_gen->max_gen_size();
2140         size_t max_eden_size = young_gen->max_size() -
2141           young_gen->from_space()->capacity_in_bytes() -
2142           young_gen->to_space()->capacity_in_bytes();
2143 
2144         // Used for diagnostics
2145         size_policy->clear_generation_free_space_flags();
2146 
2147         size_policy->compute_generations_free_space(young_live,
2148                                                     eden_live,
2149                                                     old_live,
2150                                                     cur_eden,
2151                                                     max_old_gen_size,
2152                                                     max_eden_size,
2153                                                     true /* full gc*/);
2154 
2155         size_policy->check_gc_overhead_limit(young_live,
2156                                              eden_live,
2157                                              max_old_gen_size,
2158                                              max_eden_size,
2159                                              true /* full gc*/,
2160                                              gc_cause,
2161                                              heap->collector_policy());
2162 
2163         size_policy->decay_supplemental_growth(true /* full gc*/);
2164 
2165         heap->resize_old_gen(
2166           size_policy->calculated_old_free_size_in_bytes());
2167 
2168         heap->resize_young_gen(size_policy->calculated_eden_size_in_bytes(),
2169                                size_policy->calculated_survivor_size_in_bytes());
2170       }
2171       if (PrintAdaptiveSizePolicy) {
2172         gclog_or_tty->print_cr("AdaptiveSizeStop: collection: %d ",
2173                        heap->total_collections());
2174       }
2175     }
2176 
2177     if (UsePerfData) {
2178       PSGCAdaptivePolicyCounters* const counters = heap->gc_policy_counters();
2179       counters->update_counters();
2180       counters->update_old_capacity(old_gen->capacity_in_bytes());
2181       counters->update_young_capacity(young_gen->capacity_in_bytes());
2182     }
2183 
2184     heap->resize_all_tlabs();
2185 
2186     // Resize the metaspace capacity after a collection
2187     MetaspaceGC::compute_new_size();
2188 
2189     if (TraceOldGenTime) accumulated_time()->stop();
2190 
2191     if (PrintGC) {
2192       if (PrintGCDetails) {
2193         // No GC timestamp here.  This is after GC so it would be confusing.
2194         young_gen->print_used_change(pre_gc_values.young_gen_used());
2195         old_gen->print_used_change(pre_gc_values.old_gen_used());
2196         heap->print_heap_change(pre_gc_values.heap_used());
2197         MetaspaceAux::print_metaspace_change(pre_gc_values.metadata_used());
2198       } else {
2199         heap->print_heap_change(pre_gc_values.heap_used());
2200       }
2201     }
2202 
2203     // Track memory usage and detect low memory
2204     MemoryService::track_memory_usage();
2205     heap->update_counters();
2206     gc_task_manager()->release_idle_workers();
2207   }
2208 
2209 #ifdef ASSERT
2210   for (size_t i = 0; i < ParallelGCThreads + 1; ++i) {
2211     ParCompactionManager* const cm =
2212       ParCompactionManager::manager_array(int(i));
2213     assert(cm->marking_stack()->is_empty(),       "should be empty");
2214     assert(ParCompactionManager::region_list(int(i))->is_empty(), "should be empty");
2215   }
2216 #endif // ASSERT
2217 
2218   if (VerifyAfterGC && heap->total_collections() >= VerifyGCStartAt) {
2219     HandleMark hm;  // Discard invalid handles created during verification
2220     Universe::verify(" VerifyAfterGC:");
2221   }
2222 
2223   // Re-verify object start arrays
2224   if (VerifyObjectStartArray &&
2225       VerifyAfterGC) {
2226     old_gen->verify_object_start_array();
2227   }
2228 
2229   if (ZapUnusedHeapArea) {
2230     old_gen->object_space()->check_mangled_unused_area_complete();
2231   }
2232 
2233   NOT_PRODUCT(ref_processor()->verify_no_references_recorded());
2234 
2235   collection_exit.update();
2236 
2237   heap->print_heap_after_gc();
2238   heap->trace_heap_after_gc(&_gc_tracer);
2239 
2240   if (PrintGCTaskTimeStamps) {
2241     gclog_or_tty->print_cr("VM-Thread " INT64_FORMAT " " INT64_FORMAT " "
2242                            INT64_FORMAT,
2243                            marking_start.ticks(), compaction_start.ticks(),
2244                            collection_exit.ticks());
2245     gc_task_manager()->print_task_time_stamps();
2246   }
2247 
2248   heap->post_full_gc_dump(&_gc_timer);
2249 
2250 #ifdef TRACESPINNING
2251   ParallelTaskTerminator::print_termination_counts();
2252 #endif
2253 
2254   _gc_timer.register_gc_end();
2255 
2256   _gc_tracer.report_dense_prefix(dense_prefix(old_space_id));
2257   _gc_tracer.report_gc_end(_gc_timer.gc_end(), _gc_timer.time_partitions());
2258 
2259   return true;
2260 }
2261 
2262 bool PSParallelCompact::absorb_live_data_from_eden(PSAdaptiveSizePolicy* size_policy,
2263                                              PSYoungGen* young_gen,
2264                                              PSOldGen* old_gen) {
2265   MutableSpace* const eden_space = young_gen->eden_space();
2266   assert(!eden_space->is_empty(), "eden must be non-empty");
2267   assert(young_gen->virtual_space()->alignment() ==
2268          old_gen->virtual_space()->alignment(), "alignments do not match");
2269 
2270   if (!(UseAdaptiveSizePolicy && UseAdaptiveGCBoundary)) {
2271     return false;
2272   }
2273 
2274   // Both generations must be completely committed.
2275   if (young_gen->virtual_space()->uncommitted_size() != 0) {
2276     return false;
2277   }
2278   if (old_gen->virtual_space()->uncommitted_size() != 0) {
2279     return false;
2280   }
2281 
2282   // Figure out how much to take from eden.  Include the average amount promoted
2283   // in the total; otherwise the next young gen GC will simply bail out to a
2284   // full GC.
2285   const size_t alignment = old_gen->virtual_space()->alignment();
2286   const size_t eden_used = eden_space->used_in_bytes();
2287   const size_t promoted = (size_t)size_policy->avg_promoted()->padded_average();
2288   const size_t absorb_size = align_size_up(eden_used + promoted, alignment);
2289   const size_t eden_capacity = eden_space->capacity_in_bytes();
2290 
2291   if (absorb_size >= eden_capacity) {
2292     return false; // Must leave some space in eden.
2293   }
2294 
2295   const size_t new_young_size = young_gen->capacity_in_bytes() - absorb_size;
2296   if (new_young_size < young_gen->min_gen_size()) {
2297     return false; // Respect young gen minimum size.
2298   }
2299 
2300   if (TraceAdaptiveGCBoundary && Verbose) {
2301     gclog_or_tty->print(" absorbing " SIZE_FORMAT "K:  "
2302                         "eden " SIZE_FORMAT "K->" SIZE_FORMAT "K "
2303                         "from " SIZE_FORMAT "K, to " SIZE_FORMAT "K "
2304                         "young_gen " SIZE_FORMAT "K->" SIZE_FORMAT "K ",
2305                         absorb_size / K,
2306                         eden_capacity / K, (eden_capacity - absorb_size) / K,
2307                         young_gen->from_space()->used_in_bytes() / K,
2308                         young_gen->to_space()->used_in_bytes() / K,
2309                         young_gen->capacity_in_bytes() / K, new_young_size / K);
2310   }
2311 
2312   // Fill the unused part of the old gen.
2313   MutableSpace* const old_space = old_gen->object_space();
2314   HeapWord* const unused_start = old_space->top();
2315   size_t const unused_words = pointer_delta(old_space->end(), unused_start);
2316 
2317   if (unused_words > 0) {
2318     if (unused_words < CollectedHeap::min_fill_size()) {
2319       return false;  // If the old gen cannot be filled, must give up.
2320     }
2321     CollectedHeap::fill_with_objects(unused_start, unused_words);
2322   }
2323 
2324   // Take the live data from eden and set both top and end in the old gen to
2325   // eden top.  (Need to set end because reset_after_change() mangles the region
2326   // from end to virtual_space->high() in debug builds).
2327   HeapWord* const new_top = eden_space->top();
2328   old_gen->virtual_space()->expand_into(young_gen->virtual_space(),
2329                                         absorb_size);
2330   young_gen->reset_after_change();
2331   old_space->set_top(new_top);
2332   old_space->set_end(new_top);
2333   old_gen->reset_after_change();
2334 
2335   // Update the object start array for the filler object and the data from eden.
2336   ObjectStartArray* const start_array = old_gen->start_array();
2337   for (HeapWord* p = unused_start; p < new_top; p += oop(p)->size()) {
2338     start_array->allocate_block(p);
2339   }
2340 
2341   // Could update the promoted average here, but it is not typically updated at
2342   // full GCs and the value to use is unclear.  Something like
2343   //
2344   // cur_promoted_avg + absorb_size / number_of_scavenges_since_last_full_gc.
2345 
2346   size_policy->set_bytes_absorbed_from_eden(absorb_size);
2347   return true;
2348 }
2349 
2350 GCTaskManager* const PSParallelCompact::gc_task_manager() {
2351   assert(ParallelScavengeHeap::gc_task_manager() != NULL,
2352     "shouldn't return NULL");
2353   return ParallelScavengeHeap::gc_task_manager();
2354 }
2355 
2356 void PSParallelCompact::marking_phase(ParCompactionManager* cm,
2357                                       bool maximum_heap_compaction,
2358                                       ParallelOldTracer *gc_tracer) {
2359   // Recursively traverse all live objects and mark them
2360   GCTraceTime tm("marking phase", print_phases(), true, &_gc_timer, _gc_tracer.gc_id());
2361 
2362   ParallelScavengeHeap* heap = gc_heap();
2363   uint parallel_gc_threads = heap->gc_task_manager()->workers();
2364   uint active_gc_threads = heap->gc_task_manager()->active_workers();
2365   TaskQueueSetSuper* qset = ParCompactionManager::region_array();
2366   ParallelTaskTerminator terminator(active_gc_threads, qset);
2367 
2368   PSParallelCompact::MarkAndPushClosure mark_and_push_closure(cm);
2369   PSParallelCompact::FollowStackClosure follow_stack_closure(cm);
2370 
2371   // Need new claim bits before marking starts.
2372   ClassLoaderDataGraph::clear_claimed_marks();
2373 
2374   {
2375     GCTraceTime tm_m("par mark", print_phases(), true, &_gc_timer, _gc_tracer.gc_id());
2376 
2377     ParallelScavengeHeap::ParStrongRootsScope psrs;
2378 
2379     GCTaskQueue* q = GCTaskQueue::create();
2380 
2381     q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::universe));
2382     q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jni_handles));
2383     // We scan the thread roots in parallel
2384     Threads::create_thread_roots_marking_tasks(q);
2385     q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::object_synchronizer));
2386     q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::flat_profiler));
2387     q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::management));
2388     q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::system_dictionary));
2389     q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::class_loader_data));
2390     q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jvmti));
2391     q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::code_cache));
2392 
2393     if (active_gc_threads > 1) {
2394       for (uint j = 0; j < active_gc_threads; j++) {
2395         q->enqueue(new StealMarkingTask(&terminator));
2396       }
2397     }
2398 
2399     gc_task_manager()->execute_and_wait(q);
2400   }
2401 
2402   // Process reference objects found during marking
2403   {
2404     GCTraceTime tm_r("reference processing", print_phases(), true, &_gc_timer, _gc_tracer.gc_id());
2405 
2406     ReferenceProcessorStats stats;
2407     if (ref_processor()->processing_is_mt()) {
2408       RefProcTaskExecutor task_executor;
2409       stats = ref_processor()->process_discovered_references(
2410         is_alive_closure(), &mark_and_push_closure, &follow_stack_closure,
2411         &task_executor, &_gc_timer, _gc_tracer.gc_id());
2412     } else {
2413       stats = ref_processor()->process_discovered_references(
2414         is_alive_closure(), &mark_and_push_closure, &follow_stack_closure, NULL,
2415         &_gc_timer, _gc_tracer.gc_id());
2416     }
2417 
2418     gc_tracer->report_gc_reference_stats(stats);
2419   }
2420 
2421   GCTraceTime tm_c("class unloading", print_phases(), true, &_gc_timer, _gc_tracer.gc_id());
2422 
2423   // This is the point where the entire marking should have completed.
2424   assert(cm->marking_stacks_empty(), "Marking should have completed");
2425 
2426   // Follow system dictionary roots and unload classes.
2427   bool purged_class = SystemDictionary::do_unloading(is_alive_closure());
2428 
2429   // Unload nmethods.
2430   CodeCache::do_unloading(is_alive_closure(), purged_class);
2431 
2432   // Prune dead klasses from subklass/sibling/implementor lists.
2433   Klass::clean_weak_klass_links(is_alive_closure());
2434 
2435   // Delete entries for dead interned strings.
2436   StringTable::unlink(is_alive_closure());
2437 
2438   // Clean up unreferenced symbols in symbol table.
2439   SymbolTable::unlink();
2440   _gc_tracer.report_object_count_after_gc(is_alive_closure());
2441 }
2442 
2443 void PSParallelCompact::follow_class_loader(ParCompactionManager* cm,
2444                                             ClassLoaderData* cld) {
2445   PSParallelCompact::MarkAndPushClosure mark_and_push_closure(cm);
2446   PSParallelCompact::FollowKlassClosure follow_klass_closure(&mark_and_push_closure);
2447 
2448   cld->oops_do(&mark_and_push_closure, &follow_klass_closure, true);
2449 }
2450 
2451 // This should be moved to the shared markSweep code!
2452 class PSAlwaysTrueClosure: public BoolObjectClosure {
2453 public:
2454   bool do_object_b(oop p) { return true; }
2455 };
2456 static PSAlwaysTrueClosure always_true;
2457 
2458 void PSParallelCompact::adjust_roots() {
2459   // Adjust the pointers to reflect the new locations
2460   GCTraceTime tm("adjust roots", print_phases(), true, &_gc_timer, _gc_tracer.gc_id());
2461 
2462   // Need new claim bits when tracing through and adjusting pointers.
2463   ClassLoaderDataGraph::clear_claimed_marks();
2464 
2465   // General strong roots.
2466   Universe::oops_do(adjust_pointer_closure());
2467   JNIHandles::oops_do(adjust_pointer_closure());   // Global (strong) JNI handles
2468   CLDToOopClosure adjust_from_cld(adjust_pointer_closure());
2469   Threads::oops_do(adjust_pointer_closure(), &adjust_from_cld, NULL);
2470   ObjectSynchronizer::oops_do(adjust_pointer_closure());
2471   FlatProfiler::oops_do(adjust_pointer_closure());
2472   Management::oops_do(adjust_pointer_closure());
2473   JvmtiExport::oops_do(adjust_pointer_closure());
2474   SystemDictionary::oops_do(adjust_pointer_closure());
2475   ClassLoaderDataGraph::oops_do(adjust_pointer_closure(), adjust_klass_closure(), true);
2476 
2477   // Now adjust pointers in remaining weak roots.  (All of which should
2478   // have been cleared if they pointed to non-surviving objects.)
2479   // Global (weak) JNI handles
2480   JNIHandles::weak_oops_do(&always_true, adjust_pointer_closure());
2481 
2482   CodeBlobToOopClosure adjust_from_blobs(adjust_pointer_closure(), CodeBlobToOopClosure::FixRelocations);
2483   CodeCache::blobs_do(&adjust_from_blobs);
2484   StringTable::oops_do(adjust_pointer_closure());
2485   ref_processor()->weak_oops_do(adjust_pointer_closure());
2486   // Roots were visited so references into the young gen in roots
2487   // may have been scanned.  Process them also.
2488   // Should the reference processor have a span that excludes
2489   // young gen objects?
2490   PSScavenge::reference_processor()->weak_oops_do(adjust_pointer_closure());
2491 }
2492 
2493 void PSParallelCompact::enqueue_region_draining_tasks(GCTaskQueue* q,
2494                                                       uint parallel_gc_threads)
2495 {
2496   GCTraceTime tm("drain task setup", print_phases(), true, &_gc_timer, _gc_tracer.gc_id());
2497 
2498   // Find the threads that are active
2499   unsigned int which = 0;
2500 
2501   const uint task_count = MAX2(parallel_gc_threads, 1U);
2502   for (uint j = 0; j < task_count; j++) {
2503     q->enqueue(new DrainStacksCompactionTask(j));
2504     ParCompactionManager::verify_region_list_empty(j);
2505     // Set the region stacks variables to "no" region stack values
2506     // so that they will be recognized and needing a region stack
2507     // in the stealing tasks if they do not get one by executing
2508     // a draining stack.
2509     ParCompactionManager* cm = ParCompactionManager::manager_array(j);
2510     cm->set_region_stack(NULL);
2511     cm->set_region_stack_index((uint)max_uintx);
2512   }
2513   ParCompactionManager::reset_recycled_stack_index();
2514 
2515   // Find all regions that are available (can be filled immediately) and
2516   // distribute them to the thread stacks.  The iteration is done in reverse
2517   // order (high to low) so the regions will be removed in ascending order.
2518 
2519   const ParallelCompactData& sd = PSParallelCompact::summary_data();
2520 
2521   size_t fillable_regions = 0;   // A count for diagnostic purposes.
2522   // A region index which corresponds to the tasks created above.
2523   // "which" must be 0 <= which < task_count
2524 
2525   which = 0;
2526   // id + 1 is used to test termination so unsigned  can
2527   // be used with an old_space_id == 0.
2528   for (unsigned int id = to_space_id; id + 1 > old_space_id; --id) {
2529     SpaceInfo* const space_info = _space_info + id;
2530     MutableSpace* const space = space_info->space();
2531     HeapWord* const new_top = space_info->new_top();
2532 
2533     const size_t beg_region = sd.addr_to_region_idx(space_info->dense_prefix());
2534     const size_t end_region =
2535       sd.addr_to_region_idx(sd.region_align_up(new_top));
2536 
2537     for (size_t cur = end_region - 1; cur + 1 > beg_region; --cur) {
2538       if (sd.region(cur)->claim_unsafe()) {
2539         ParCompactionManager::region_list_push(which, cur);
2540 
2541         if (TraceParallelOldGCCompactionPhase && Verbose) {
2542           const size_t count_mod_8 = fillable_regions & 7;
2543           if (count_mod_8 == 0) gclog_or_tty->print("fillable: ");
2544           gclog_or_tty->print(" " SIZE_FORMAT_W(7), cur);
2545           if (count_mod_8 == 7) gclog_or_tty->cr();
2546         }
2547 
2548         NOT_PRODUCT(++fillable_regions;)
2549 
2550         // Assign regions to tasks in round-robin fashion.
2551         if (++which == task_count) {
2552           assert(which <= parallel_gc_threads,
2553             "Inconsistent number of workers");
2554           which = 0;
2555         }
2556       }
2557     }
2558   }
2559 
2560   if (TraceParallelOldGCCompactionPhase) {
2561     if (Verbose && (fillable_regions & 7) != 0) gclog_or_tty->cr();
2562     gclog_or_tty->print_cr(SIZE_FORMAT " initially fillable regions", fillable_regions);
2563   }
2564 }
2565 
2566 #define PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING 4
2567 
2568 void PSParallelCompact::enqueue_dense_prefix_tasks(GCTaskQueue* q,
2569                                                     uint parallel_gc_threads) {
2570   GCTraceTime tm("dense prefix task setup", print_phases(), true, &_gc_timer, _gc_tracer.gc_id());
2571 
2572   ParallelCompactData& sd = PSParallelCompact::summary_data();
2573 
2574   // Iterate over all the spaces adding tasks for updating
2575   // regions in the dense prefix.  Assume that 1 gc thread
2576   // will work on opening the gaps and the remaining gc threads
2577   // will work on the dense prefix.
2578   unsigned int space_id;
2579   for (space_id = old_space_id; space_id < last_space_id; ++ space_id) {
2580     HeapWord* const dense_prefix_end = _space_info[space_id].dense_prefix();
2581     const MutableSpace* const space = _space_info[space_id].space();
2582 
2583     if (dense_prefix_end == space->bottom()) {
2584       // There is no dense prefix for this space.
2585       continue;
2586     }
2587 
2588     // The dense prefix is before this region.
2589     size_t region_index_end_dense_prefix =
2590         sd.addr_to_region_idx(dense_prefix_end);
2591     RegionData* const dense_prefix_cp =
2592       sd.region(region_index_end_dense_prefix);
2593     assert(dense_prefix_end == space->end() ||
2594            dense_prefix_cp->available() ||
2595            dense_prefix_cp->claimed(),
2596            "The region after the dense prefix should always be ready to fill");
2597 
2598     size_t region_index_start = sd.addr_to_region_idx(space->bottom());
2599 
2600     // Is there dense prefix work?
2601     size_t total_dense_prefix_regions =
2602       region_index_end_dense_prefix - region_index_start;
2603     // How many regions of the dense prefix should be given to
2604     // each thread?
2605     if (total_dense_prefix_regions > 0) {
2606       uint tasks_for_dense_prefix = 1;
2607       if (total_dense_prefix_regions <=
2608           (parallel_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING)) {
2609         // Don't over partition.  This assumes that
2610         // PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING is a small integer value
2611         // so there are not many regions to process.
2612         tasks_for_dense_prefix = parallel_gc_threads;
2613       } else {
2614         // Over partition
2615         tasks_for_dense_prefix = parallel_gc_threads *
2616           PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING;
2617       }
2618       size_t regions_per_thread = total_dense_prefix_regions /
2619         tasks_for_dense_prefix;
2620       // Give each thread at least 1 region.
2621       if (regions_per_thread == 0) {
2622         regions_per_thread = 1;
2623       }
2624 
2625       for (uint k = 0; k < tasks_for_dense_prefix; k++) {
2626         if (region_index_start >= region_index_end_dense_prefix) {
2627           break;
2628         }
2629         // region_index_end is not processed
2630         size_t region_index_end = MIN2(region_index_start + regions_per_thread,
2631                                        region_index_end_dense_prefix);
2632         q->enqueue(new UpdateDensePrefixTask(SpaceId(space_id),
2633                                              region_index_start,
2634                                              region_index_end));
2635         region_index_start = region_index_end;
2636       }
2637     }
2638     // This gets any part of the dense prefix that did not
2639     // fit evenly.
2640     if (region_index_start < region_index_end_dense_prefix) {
2641       q->enqueue(new UpdateDensePrefixTask(SpaceId(space_id),
2642                                            region_index_start,
2643                                            region_index_end_dense_prefix));
2644     }
2645   }
2646 }
2647 
2648 void PSParallelCompact::enqueue_region_stealing_tasks(
2649                                      GCTaskQueue* q,
2650                                      ParallelTaskTerminator* terminator_ptr,
2651                                      uint parallel_gc_threads) {
2652   GCTraceTime tm("steal task setup", print_phases(), true, &_gc_timer, _gc_tracer.gc_id());
2653 
2654   // Once a thread has drained it's stack, it should try to steal regions from
2655   // other threads.
2656   if (parallel_gc_threads > 1) {
2657     for (uint j = 0; j < parallel_gc_threads; j++) {
2658       q->enqueue(new StealRegionCompactionTask(terminator_ptr));
2659     }
2660   }
2661 }
2662 
2663 #ifdef ASSERT
2664 // Write a histogram of the number of times the block table was filled for a
2665 // region.
2666 void PSParallelCompact::write_block_fill_histogram(outputStream* const out)
2667 {
2668   if (!TraceParallelOldGCCompactionPhase) return;
2669 
2670   typedef ParallelCompactData::RegionData rd_t;
2671   ParallelCompactData& sd = summary_data();
2672 
2673   for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2674     MutableSpace* const spc = _space_info[id].space();
2675     if (spc->bottom() != spc->top()) {
2676       const rd_t* const beg = sd.addr_to_region_ptr(spc->bottom());
2677       HeapWord* const top_aligned_up = sd.region_align_up(spc->top());
2678       const rd_t* const end = sd.addr_to_region_ptr(top_aligned_up);
2679 
2680       size_t histo[5] = { 0, 0, 0, 0, 0 };
2681       const size_t histo_len = sizeof(histo) / sizeof(size_t);
2682       const size_t region_cnt = pointer_delta(end, beg, sizeof(rd_t));
2683 
2684       for (const rd_t* cur = beg; cur < end; ++cur) {
2685         ++histo[MIN2(cur->blocks_filled_count(), histo_len - 1)];
2686       }
2687       out->print("%u %-4s" SIZE_FORMAT_W(5), id, space_names[id], region_cnt);
2688       for (size_t i = 0; i < histo_len; ++i) {
2689         out->print(" " SIZE_FORMAT_W(5) " %5.1f%%",
2690                    histo[i], 100.0 * histo[i] / region_cnt);
2691       }
2692       out->cr();
2693     }
2694   }
2695 }
2696 #endif // #ifdef ASSERT
2697 
2698 void PSParallelCompact::compact() {
2699   // trace("5");
2700   GCTraceTime tm("compaction phase", print_phases(), true, &_gc_timer, _gc_tracer.gc_id());
2701 
2702   ParallelScavengeHeap* heap = (ParallelScavengeHeap*)Universe::heap();
2703   assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity");
2704   PSOldGen* old_gen = heap->old_gen();
2705   old_gen->start_array()->reset();
2706   uint parallel_gc_threads = heap->gc_task_manager()->workers();
2707   uint active_gc_threads = heap->gc_task_manager()->active_workers();
2708   TaskQueueSetSuper* qset = ParCompactionManager::region_array();
2709   ParallelTaskTerminator terminator(active_gc_threads, qset);
2710 
2711   GCTaskQueue* q = GCTaskQueue::create();
2712   enqueue_region_draining_tasks(q, active_gc_threads);
2713   enqueue_dense_prefix_tasks(q, active_gc_threads);
2714   enqueue_region_stealing_tasks(q, &terminator, active_gc_threads);
2715 
2716   {
2717     GCTraceTime tm_pc("par compact", print_phases(), true, &_gc_timer, _gc_tracer.gc_id());
2718 
2719     gc_task_manager()->execute_and_wait(q);
2720 
2721 #ifdef  ASSERT
2722     // Verify that all regions have been processed before the deferred updates.
2723     for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2724       verify_complete(SpaceId(id));
2725     }
2726 #endif
2727   }
2728 
2729   {
2730     // Update the deferred objects, if any.  Any compaction manager can be used.
2731     GCTraceTime tm_du("deferred updates", print_phases(), true, &_gc_timer, _gc_tracer.gc_id());
2732     ParCompactionManager* cm = ParCompactionManager::manager_array(0);
2733     for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2734       update_deferred_objects(cm, SpaceId(id));
2735     }
2736   }
2737 
2738   DEBUG_ONLY(write_block_fill_histogram(gclog_or_tty));
2739 }
2740 
2741 #ifdef  ASSERT
2742 void PSParallelCompact::verify_complete(SpaceId space_id) {
2743   // All Regions between space bottom() to new_top() should be marked as filled
2744   // and all Regions between new_top() and top() should be available (i.e.,
2745   // should have been emptied).
2746   ParallelCompactData& sd = summary_data();
2747   SpaceInfo si = _space_info[space_id];
2748   HeapWord* new_top_addr = sd.region_align_up(si.new_top());
2749   HeapWord* old_top_addr = sd.region_align_up(si.space()->top());
2750   const size_t beg_region = sd.addr_to_region_idx(si.space()->bottom());
2751   const size_t new_top_region = sd.addr_to_region_idx(new_top_addr);
2752   const size_t old_top_region = sd.addr_to_region_idx(old_top_addr);
2753 
2754   bool issued_a_warning = false;
2755 
2756   size_t cur_region;
2757   for (cur_region = beg_region; cur_region < new_top_region; ++cur_region) {
2758     const RegionData* const c = sd.region(cur_region);
2759     if (!c->completed()) {
2760       warning("region " SIZE_FORMAT " not filled:  "
2761               "destination_count=" SIZE_FORMAT,
2762               cur_region, c->destination_count());
2763       issued_a_warning = true;
2764     }
2765   }
2766 
2767   for (cur_region = new_top_region; cur_region < old_top_region; ++cur_region) {
2768     const RegionData* const c = sd.region(cur_region);
2769     if (!c->available()) {
2770       warning("region " SIZE_FORMAT " not empty:   "
2771               "destination_count=" SIZE_FORMAT,
2772               cur_region, c->destination_count());
2773       issued_a_warning = true;
2774     }
2775   }
2776 
2777   if (issued_a_warning) {
2778     print_region_ranges();
2779   }
2780 }
2781 #endif  // #ifdef ASSERT
2782 
2783 // Update interior oops in the ranges of regions [beg_region, end_region).
2784 void
2785 PSParallelCompact::update_and_deadwood_in_dense_prefix(ParCompactionManager* cm,
2786                                                        SpaceId space_id,
2787                                                        size_t beg_region,
2788                                                        size_t end_region) {
2789   ParallelCompactData& sd = summary_data();
2790   ParMarkBitMap* const mbm = mark_bitmap();
2791 
2792   HeapWord* beg_addr = sd.region_to_addr(beg_region);
2793   HeapWord* const end_addr = sd.region_to_addr(end_region);
2794   assert(beg_region <= end_region, "bad region range");
2795   assert(end_addr <= dense_prefix(space_id), "not in the dense prefix");
2796 
2797 #ifdef  ASSERT
2798   // Claim the regions to avoid triggering an assert when they are marked as
2799   // filled.
2800   for (size_t claim_region = beg_region; claim_region < end_region; ++claim_region) {
2801     assert(sd.region(claim_region)->claim_unsafe(), "claim() failed");
2802   }
2803 #endif  // #ifdef ASSERT
2804 
2805   if (beg_addr != space(space_id)->bottom()) {
2806     // Find the first live object or block of dead space that *starts* in this
2807     // range of regions.  If a partial object crosses onto the region, skip it;
2808     // it will be marked for 'deferred update' when the object head is
2809     // processed.  If dead space crosses onto the region, it is also skipped; it
2810     // will be filled when the prior region is processed.  If neither of those
2811     // apply, the first word in the region is the start of a live object or dead
2812     // space.
2813     assert(beg_addr > space(space_id)->bottom(), "sanity");
2814     const RegionData* const cp = sd.region(beg_region);
2815     if (cp->partial_obj_size() != 0) {
2816       beg_addr = sd.partial_obj_end(beg_region);
2817     } else if (dead_space_crosses_boundary(cp, mbm->addr_to_bit(beg_addr))) {
2818       beg_addr = mbm->find_obj_beg(beg_addr, end_addr);
2819     }
2820   }
2821 
2822   if (beg_addr < end_addr) {
2823     // A live object or block of dead space starts in this range of Regions.
2824      HeapWord* const dense_prefix_end = dense_prefix(space_id);
2825 
2826     // Create closures and iterate.
2827     UpdateOnlyClosure update_closure(mbm, cm, space_id);
2828     FillClosure fill_closure(cm, space_id);
2829     ParMarkBitMap::IterationStatus status;
2830     status = mbm->iterate(&update_closure, &fill_closure, beg_addr, end_addr,
2831                           dense_prefix_end);
2832     if (status == ParMarkBitMap::incomplete) {
2833       update_closure.do_addr(update_closure.source());
2834     }
2835   }
2836 
2837   // Mark the regions as filled.
2838   RegionData* const beg_cp = sd.region(beg_region);
2839   RegionData* const end_cp = sd.region(end_region);
2840   for (RegionData* cp = beg_cp; cp < end_cp; ++cp) {
2841     cp->set_completed();
2842   }
2843 }
2844 
2845 // Return the SpaceId for the space containing addr.  If addr is not in the
2846 // heap, last_space_id is returned.  In debug mode it expects the address to be
2847 // in the heap and asserts such.
2848 PSParallelCompact::SpaceId PSParallelCompact::space_id(HeapWord* addr) {
2849   assert(Universe::heap()->is_in_reserved(addr), "addr not in the heap");
2850 
2851   for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2852     if (_space_info[id].space()->contains(addr)) {
2853       return SpaceId(id);
2854     }
2855   }
2856 
2857   assert(false, "no space contains the addr");
2858   return last_space_id;
2859 }
2860 
2861 void PSParallelCompact::update_deferred_objects(ParCompactionManager* cm,
2862                                                 SpaceId id) {
2863   assert(id < last_space_id, "bad space id");
2864 
2865   ParallelCompactData& sd = summary_data();
2866   const SpaceInfo* const space_info = _space_info + id;
2867   ObjectStartArray* const start_array = space_info->start_array();
2868 
2869   const MutableSpace* const space = space_info->space();
2870   assert(space_info->dense_prefix() >= space->bottom(), "dense_prefix not set");
2871   HeapWord* const beg_addr = space_info->dense_prefix();
2872   HeapWord* const end_addr = sd.region_align_up(space_info->new_top());
2873 
2874   const RegionData* const beg_region = sd.addr_to_region_ptr(beg_addr);
2875   const RegionData* const end_region = sd.addr_to_region_ptr(end_addr);
2876   const RegionData* cur_region;
2877   for (cur_region = beg_region; cur_region < end_region; ++cur_region) {
2878     HeapWord* const addr = cur_region->deferred_obj_addr();
2879     if (addr != NULL) {
2880       if (start_array != NULL) {
2881         start_array->allocate_block(addr);
2882       }
2883       oop(addr)->update_contents(cm);
2884       assert(oop(addr)->is_oop_or_null(), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(oop(addr))));
2885     }
2886   }
2887 }
2888 
2889 // Skip over count live words starting from beg, and return the address of the
2890 // next live word.  Unless marked, the word corresponding to beg is assumed to
2891 // be dead.  Callers must either ensure beg does not correspond to the middle of
2892 // an object, or account for those live words in some other way.  Callers must
2893 // also ensure that there are enough live words in the range [beg, end) to skip.
2894 HeapWord*
2895 PSParallelCompact::skip_live_words(HeapWord* beg, HeapWord* end, size_t count)
2896 {
2897   assert(count > 0, "sanity");
2898 
2899   ParMarkBitMap* m = mark_bitmap();
2900   idx_t bits_to_skip = m->words_to_bits(count);
2901   idx_t cur_beg = m->addr_to_bit(beg);
2902   const idx_t search_end = BitMap::word_align_up(m->addr_to_bit(end));
2903 
2904   do {
2905     cur_beg = m->find_obj_beg(cur_beg, search_end);
2906     idx_t cur_end = m->find_obj_end(cur_beg, search_end);
2907     const size_t obj_bits = cur_end - cur_beg + 1;
2908     if (obj_bits > bits_to_skip) {
2909       return m->bit_to_addr(cur_beg + bits_to_skip);
2910     }
2911     bits_to_skip -= obj_bits;
2912     cur_beg = cur_end + 1;
2913   } while (bits_to_skip > 0);
2914 
2915   // Skipping the desired number of words landed just past the end of an object.
2916   // Find the start of the next object.
2917   cur_beg = m->find_obj_beg(cur_beg, search_end);
2918   assert(cur_beg < m->addr_to_bit(end), "not enough live words to skip");
2919   return m->bit_to_addr(cur_beg);
2920 }
2921 
2922 HeapWord* PSParallelCompact::first_src_addr(HeapWord* const dest_addr,
2923                                             SpaceId src_space_id,
2924                                             size_t src_region_idx)
2925 {
2926   assert(summary_data().is_region_aligned(dest_addr), "not aligned");
2927 
2928   const SplitInfo& split_info = _space_info[src_space_id].split_info();
2929   if (split_info.dest_region_addr() == dest_addr) {
2930     // The partial object ending at the split point contains the first word to
2931     // be copied to dest_addr.
2932     return split_info.first_src_addr();
2933   }
2934 
2935   const ParallelCompactData& sd = summary_data();
2936   ParMarkBitMap* const bitmap = mark_bitmap();
2937   const size_t RegionSize = ParallelCompactData::RegionSize;
2938 
2939   assert(sd.is_region_aligned(dest_addr), "not aligned");
2940   const RegionData* const src_region_ptr = sd.region(src_region_idx);
2941   const size_t partial_obj_size = src_region_ptr->partial_obj_size();
2942   HeapWord* const src_region_destination = src_region_ptr->destination();
2943 
2944   assert(dest_addr >= src_region_destination, "wrong src region");
2945   assert(src_region_ptr->data_size() > 0, "src region cannot be empty");
2946 
2947   HeapWord* const src_region_beg = sd.region_to_addr(src_region_idx);
2948   HeapWord* const src_region_end = src_region_beg + RegionSize;
2949 
2950   HeapWord* addr = src_region_beg;
2951   if (dest_addr == src_region_destination) {
2952     // Return the first live word in the source region.
2953     if (partial_obj_size == 0) {
2954       addr = bitmap->find_obj_beg(addr, src_region_end);
2955       assert(addr < src_region_end, "no objects start in src region");
2956     }
2957     return addr;
2958   }
2959 
2960   // Must skip some live data.
2961   size_t words_to_skip = dest_addr - src_region_destination;
2962   assert(src_region_ptr->data_size() > words_to_skip, "wrong src region");
2963 
2964   if (partial_obj_size >= words_to_skip) {
2965     // All the live words to skip are part of the partial object.
2966     addr += words_to_skip;
2967     if (partial_obj_size == words_to_skip) {
2968       // Find the first live word past the partial object.
2969       addr = bitmap->find_obj_beg(addr, src_region_end);
2970       assert(addr < src_region_end, "wrong src region");
2971     }
2972     return addr;
2973   }
2974 
2975   // Skip over the partial object (if any).
2976   if (partial_obj_size != 0) {
2977     words_to_skip -= partial_obj_size;
2978     addr += partial_obj_size;
2979   }
2980 
2981   // Skip over live words due to objects that start in the region.
2982   addr = skip_live_words(addr, src_region_end, words_to_skip);
2983   assert(addr < src_region_end, "wrong src region");
2984   return addr;
2985 }
2986 
2987 void PSParallelCompact::decrement_destination_counts(ParCompactionManager* cm,
2988                                                      SpaceId src_space_id,
2989                                                      size_t beg_region,
2990                                                      HeapWord* end_addr)
2991 {
2992   ParallelCompactData& sd = summary_data();
2993 
2994 #ifdef ASSERT
2995   MutableSpace* const src_space = _space_info[src_space_id].space();
2996   HeapWord* const beg_addr = sd.region_to_addr(beg_region);
2997   assert(src_space->contains(beg_addr) || beg_addr == src_space->end(),
2998          "src_space_id does not match beg_addr");
2999   assert(src_space->contains(end_addr) || end_addr == src_space->end(),
3000          "src_space_id does not match end_addr");
3001 #endif // #ifdef ASSERT
3002 
3003   RegionData* const beg = sd.region(beg_region);
3004   RegionData* const end = sd.addr_to_region_ptr(sd.region_align_up(end_addr));
3005 
3006   // Regions up to new_top() are enqueued if they become available.
3007   HeapWord* const new_top = _space_info[src_space_id].new_top();
3008   RegionData* const enqueue_end =
3009     sd.addr_to_region_ptr(sd.region_align_up(new_top));
3010 
3011   for (RegionData* cur = beg; cur < end; ++cur) {
3012     assert(cur->data_size() > 0, "region must have live data");
3013     cur->decrement_destination_count();
3014     if (cur < enqueue_end && cur->available() && cur->claim()) {
3015       cm->push_region(sd.region(cur));
3016     }
3017   }
3018 }
3019 
3020 size_t PSParallelCompact::next_src_region(MoveAndUpdateClosure& closure,
3021                                           SpaceId& src_space_id,
3022                                           HeapWord*& src_space_top,
3023                                           HeapWord* end_addr)
3024 {
3025   typedef ParallelCompactData::RegionData RegionData;
3026 
3027   ParallelCompactData& sd = PSParallelCompact::summary_data();
3028   const size_t region_size = ParallelCompactData::RegionSize;
3029 
3030   size_t src_region_idx = 0;
3031 
3032   // Skip empty regions (if any) up to the top of the space.
3033   HeapWord* const src_aligned_up = sd.region_align_up(end_addr);
3034   RegionData* src_region_ptr = sd.addr_to_region_ptr(src_aligned_up);
3035   HeapWord* const top_aligned_up = sd.region_align_up(src_space_top);
3036   const RegionData* const top_region_ptr =
3037     sd.addr_to_region_ptr(top_aligned_up);
3038   while (src_region_ptr < top_region_ptr && src_region_ptr->data_size() == 0) {
3039     ++src_region_ptr;
3040   }
3041 
3042   if (src_region_ptr < top_region_ptr) {
3043     // The next source region is in the current space.  Update src_region_idx
3044     // and the source address to match src_region_ptr.
3045     src_region_idx = sd.region(src_region_ptr);
3046     HeapWord* const src_region_addr = sd.region_to_addr(src_region_idx);
3047     if (src_region_addr > closure.source()) {
3048       closure.set_source(src_region_addr);
3049     }
3050     return src_region_idx;
3051   }
3052 
3053   // Switch to a new source space and find the first non-empty region.
3054   unsigned int space_id = src_space_id + 1;
3055   assert(space_id < last_space_id, "not enough spaces");
3056 
3057   HeapWord* const destination = closure.destination();
3058 
3059   do {
3060     MutableSpace* space = _space_info[space_id].space();
3061     HeapWord* const bottom = space->bottom();
3062     const RegionData* const bottom_cp = sd.addr_to_region_ptr(bottom);
3063 
3064     // Iterate over the spaces that do not compact into themselves.
3065     if (bottom_cp->destination() != bottom) {
3066       HeapWord* const top_aligned_up = sd.region_align_up(space->top());
3067       const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up);
3068 
3069       for (const RegionData* src_cp = bottom_cp; src_cp < top_cp; ++src_cp) {
3070         if (src_cp->live_obj_size() > 0) {
3071           // Found it.
3072           assert(src_cp->destination() == destination,
3073                  "first live obj in the space must match the destination");
3074           assert(src_cp->partial_obj_size() == 0,
3075                  "a space cannot begin with a partial obj");
3076 
3077           src_space_id = SpaceId(space_id);
3078           src_space_top = space->top();
3079           const size_t src_region_idx = sd.region(src_cp);
3080           closure.set_source(sd.region_to_addr(src_region_idx));
3081           return src_region_idx;
3082         } else {
3083           assert(src_cp->data_size() == 0, "sanity");
3084         }
3085       }
3086     }
3087   } while (++space_id < last_space_id);
3088 
3089   assert(false, "no source region was found");
3090   return 0;
3091 }
3092 
3093 void PSParallelCompact::fill_region(ParCompactionManager* cm, size_t region_idx)
3094 {
3095   typedef ParMarkBitMap::IterationStatus IterationStatus;
3096   const size_t RegionSize = ParallelCompactData::RegionSize;
3097   ParMarkBitMap* const bitmap = mark_bitmap();
3098   ParallelCompactData& sd = summary_data();
3099   RegionData* const region_ptr = sd.region(region_idx);
3100 
3101   // Get the items needed to construct the closure.
3102   HeapWord* dest_addr = sd.region_to_addr(region_idx);
3103   SpaceId dest_space_id = space_id(dest_addr);
3104   ObjectStartArray* start_array = _space_info[dest_space_id].start_array();
3105   HeapWord* new_top = _space_info[dest_space_id].new_top();
3106   assert(dest_addr < new_top, "sanity");
3107   const size_t words = MIN2(pointer_delta(new_top, dest_addr), RegionSize);
3108 
3109   // Get the source region and related info.
3110   size_t src_region_idx = region_ptr->source_region();
3111   SpaceId src_space_id = space_id(sd.region_to_addr(src_region_idx));
3112   HeapWord* src_space_top = _space_info[src_space_id].space()->top();
3113 
3114   MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words);
3115   closure.set_source(first_src_addr(dest_addr, src_space_id, src_region_idx));
3116 
3117   // Adjust src_region_idx to prepare for decrementing destination counts (the
3118   // destination count is not decremented when a region is copied to itself).
3119   if (src_region_idx == region_idx) {
3120     src_region_idx += 1;
3121   }
3122 
3123   if (bitmap->is_unmarked(closure.source())) {
3124     // The first source word is in the middle of an object; copy the remainder
3125     // of the object or as much as will fit.  The fact that pointer updates were
3126     // deferred will be noted when the object header is processed.
3127     HeapWord* const old_src_addr = closure.source();
3128     closure.copy_partial_obj();
3129     if (closure.is_full()) {
3130       decrement_destination_counts(cm, src_space_id, src_region_idx,
3131                                    closure.source());
3132       region_ptr->set_deferred_obj_addr(NULL);
3133       region_ptr->set_completed();
3134       return;
3135     }
3136 
3137     HeapWord* const end_addr = sd.region_align_down(closure.source());
3138     if (sd.region_align_down(old_src_addr) != end_addr) {
3139       // The partial object was copied from more than one source region.
3140       decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr);
3141 
3142       // Move to the next source region, possibly switching spaces as well.  All
3143       // args except end_addr may be modified.
3144       src_region_idx = next_src_region(closure, src_space_id, src_space_top,
3145                                        end_addr);
3146     }
3147   }
3148 
3149   do {
3150     HeapWord* const cur_addr = closure.source();
3151     HeapWord* const end_addr = MIN2(sd.region_align_up(cur_addr + 1),
3152                                     src_space_top);
3153     IterationStatus status = bitmap->iterate(&closure, cur_addr, end_addr);
3154 
3155     if (status == ParMarkBitMap::incomplete) {
3156       // The last obj that starts in the source region does not end in the
3157       // region.
3158       assert(closure.source() < end_addr, "sanity");
3159       HeapWord* const obj_beg = closure.source();
3160       HeapWord* const range_end = MIN2(obj_beg + closure.words_remaining(),
3161                                        src_space_top);
3162       HeapWord* const obj_end = bitmap->find_obj_end(obj_beg, range_end);
3163       if (obj_end < range_end) {
3164         // The end was found; the entire object will fit.
3165         status = closure.do_addr(obj_beg, bitmap->obj_size(obj_beg, obj_end));
3166         assert(status != ParMarkBitMap::would_overflow, "sanity");
3167       } else {
3168         // The end was not found; the object will not fit.
3169         assert(range_end < src_space_top, "obj cannot cross space boundary");
3170         status = ParMarkBitMap::would_overflow;
3171       }
3172     }
3173 
3174     if (status == ParMarkBitMap::would_overflow) {
3175       // The last object did not fit.  Note that interior oop updates were
3176       // deferred, then copy enough of the object to fill the region.
3177       region_ptr->set_deferred_obj_addr(closure.destination());
3178       status = closure.copy_until_full(); // copies from closure.source()
3179 
3180       decrement_destination_counts(cm, src_space_id, src_region_idx,
3181                                    closure.source());
3182       region_ptr->set_completed();
3183       return;
3184     }
3185 
3186     if (status == ParMarkBitMap::full) {
3187       decrement_destination_counts(cm, src_space_id, src_region_idx,
3188                                    closure.source());
3189       region_ptr->set_deferred_obj_addr(NULL);
3190       region_ptr->set_completed();
3191       return;
3192     }
3193 
3194     decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr);
3195 
3196     // Move to the next source region, possibly switching spaces as well.  All
3197     // args except end_addr may be modified.
3198     src_region_idx = next_src_region(closure, src_space_id, src_space_top,
3199                                      end_addr);
3200   } while (true);
3201 }
3202 
3203 void PSParallelCompact::fill_blocks(size_t region_idx)
3204 {
3205   // Fill in the block table elements for the specified region.  Each block
3206   // table element holds the number of live words in the region that are to the
3207   // left of the first object that starts in the block.  Thus only blocks in
3208   // which an object starts need to be filled.
3209   //
3210   // The algorithm scans the section of the bitmap that corresponds to the
3211   // region, keeping a running total of the live words.  When an object start is
3212   // found, if it's the first to start in the block that contains it, the
3213   // current total is written to the block table element.
3214   const size_t Log2BlockSize = ParallelCompactData::Log2BlockSize;
3215   const size_t Log2RegionSize = ParallelCompactData::Log2RegionSize;
3216   const size_t RegionSize = ParallelCompactData::RegionSize;
3217 
3218   ParallelCompactData& sd = summary_data();
3219   const size_t partial_obj_size = sd.region(region_idx)->partial_obj_size();
3220   if (partial_obj_size >= RegionSize) {
3221     return; // No objects start in this region.
3222   }
3223 
3224   // Ensure the first loop iteration decides that the block has changed.
3225   size_t cur_block = sd.block_count();
3226 
3227   const ParMarkBitMap* const bitmap = mark_bitmap();
3228 
3229   const size_t Log2BitsPerBlock = Log2BlockSize - LogMinObjAlignment;
3230   assert((size_t)1 << Log2BitsPerBlock ==
3231          bitmap->words_to_bits(ParallelCompactData::BlockSize), "sanity");
3232 
3233   size_t beg_bit = bitmap->words_to_bits(region_idx << Log2RegionSize);
3234   const size_t range_end = beg_bit + bitmap->words_to_bits(RegionSize);
3235   size_t live_bits = bitmap->words_to_bits(partial_obj_size);
3236   beg_bit = bitmap->find_obj_beg(beg_bit + live_bits, range_end);
3237   while (beg_bit < range_end) {
3238     const size_t new_block = beg_bit >> Log2BitsPerBlock;
3239     if (new_block != cur_block) {
3240       cur_block = new_block;
3241       sd.block(cur_block)->set_offset(bitmap->bits_to_words(live_bits));
3242     }
3243 
3244     const size_t end_bit = bitmap->find_obj_end(beg_bit, range_end);
3245     if (end_bit < range_end - 1) {
3246       live_bits += end_bit - beg_bit + 1;
3247       beg_bit = bitmap->find_obj_beg(end_bit + 1, range_end);
3248     } else {
3249       return;
3250     }
3251   }
3252 }
3253 
3254 void
3255 PSParallelCompact::move_and_update(ParCompactionManager* cm, SpaceId space_id) {
3256   const MutableSpace* sp = space(space_id);
3257   if (sp->is_empty()) {
3258     return;
3259   }
3260 
3261   ParallelCompactData& sd = PSParallelCompact::summary_data();
3262   ParMarkBitMap* const bitmap = mark_bitmap();
3263   HeapWord* const dp_addr = dense_prefix(space_id);
3264   HeapWord* beg_addr = sp->bottom();
3265   HeapWord* end_addr = sp->top();
3266 
3267   assert(beg_addr <= dp_addr && dp_addr <= end_addr, "bad dense prefix");
3268 
3269   const size_t beg_region = sd.addr_to_region_idx(beg_addr);
3270   const size_t dp_region = sd.addr_to_region_idx(dp_addr);
3271   if (beg_region < dp_region) {
3272     update_and_deadwood_in_dense_prefix(cm, space_id, beg_region, dp_region);
3273   }
3274 
3275   // The destination of the first live object that starts in the region is one
3276   // past the end of the partial object entering the region (if any).
3277   HeapWord* const dest_addr = sd.partial_obj_end(dp_region);
3278   HeapWord* const new_top = _space_info[space_id].new_top();
3279   assert(new_top >= dest_addr, "bad new_top value");
3280   const size_t words = pointer_delta(new_top, dest_addr);
3281 
3282   if (words > 0) {
3283     ObjectStartArray* start_array = _space_info[space_id].start_array();
3284     MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words);
3285 
3286     ParMarkBitMap::IterationStatus status;
3287     status = bitmap->iterate(&closure, dest_addr, end_addr);
3288     assert(status == ParMarkBitMap::full, "iteration not complete");
3289     assert(bitmap->find_obj_beg(closure.source(), end_addr) == end_addr,
3290            "live objects skipped because closure is full");
3291   }
3292 }
3293 
3294 jlong PSParallelCompact::millis_since_last_gc() {
3295   // We need a monotonically non-decreasing time in ms but
3296   // os::javaTimeMillis() does not guarantee monotonicity.
3297   jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
3298   jlong ret_val = now - _time_of_last_gc;
3299   // XXX See note in genCollectedHeap::millis_since_last_gc().
3300   if (ret_val < 0) {
3301     NOT_PRODUCT(warning("time warp: "INT64_FORMAT, ret_val);)
3302     return 0;
3303   }
3304   return ret_val;
3305 }
3306 
3307 void PSParallelCompact::reset_millis_since_last_gc() {
3308   // We need a monotonically non-decreasing time in ms but
3309   // os::javaTimeMillis() does not guarantee monotonicity.
3310   _time_of_last_gc = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
3311 }
3312 
3313 ParMarkBitMap::IterationStatus MoveAndUpdateClosure::copy_until_full()
3314 {
3315   if (source() != destination()) {
3316     DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3317     Copy::aligned_conjoint_words(source(), destination(), words_remaining());
3318   }
3319   update_state(words_remaining());
3320   assert(is_full(), "sanity");
3321   return ParMarkBitMap::full;
3322 }
3323 
3324 void MoveAndUpdateClosure::copy_partial_obj()
3325 {
3326   size_t words = words_remaining();
3327 
3328   HeapWord* const range_end = MIN2(source() + words, bitmap()->region_end());
3329   HeapWord* const end_addr = bitmap()->find_obj_end(source(), range_end);
3330   if (end_addr < range_end) {
3331     words = bitmap()->obj_size(source(), end_addr);
3332   }
3333 
3334   // This test is necessary; if omitted, the pointer updates to a partial object
3335   // that crosses the dense prefix boundary could be overwritten.
3336   if (source() != destination()) {
3337     DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3338     Copy::aligned_conjoint_words(source(), destination(), words);
3339   }
3340   update_state(words);
3341 }
3342 
3343 ParMarkBitMapClosure::IterationStatus
3344 MoveAndUpdateClosure::do_addr(HeapWord* addr, size_t words) {
3345   assert(destination() != NULL, "sanity");
3346   assert(bitmap()->obj_size(addr) == words, "bad size");
3347 
3348   _source = addr;
3349   assert(PSParallelCompact::summary_data().calc_new_pointer(source()) ==
3350          destination(), "wrong destination");
3351 
3352   if (words > words_remaining()) {
3353     return ParMarkBitMap::would_overflow;
3354   }
3355 
3356   // The start_array must be updated even if the object is not moving.
3357   if (_start_array != NULL) {
3358     _start_array->allocate_block(destination());
3359   }
3360 
3361   if (destination() != source()) {
3362     DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3363     Copy::aligned_conjoint_words(source(), destination(), words);
3364   }
3365 
3366   oop moved_oop = (oop) destination();
3367   moved_oop->update_contents(compaction_manager());
3368   assert(moved_oop->is_oop_or_null(), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(moved_oop)));
3369 
3370   update_state(words);
3371   assert(destination() == (HeapWord*)moved_oop + moved_oop->size(), "sanity");
3372   return is_full() ? ParMarkBitMap::full : ParMarkBitMap::incomplete;
3373 }
3374 
3375 UpdateOnlyClosure::UpdateOnlyClosure(ParMarkBitMap* mbm,
3376                                      ParCompactionManager* cm,
3377                                      PSParallelCompact::SpaceId space_id) :
3378   ParMarkBitMapClosure(mbm, cm),
3379   _space_id(space_id),
3380   _start_array(PSParallelCompact::start_array(space_id))
3381 {
3382 }
3383 
3384 // Updates the references in the object to their new values.
3385 ParMarkBitMapClosure::IterationStatus
3386 UpdateOnlyClosure::do_addr(HeapWord* addr, size_t words) {
3387   do_addr(addr);
3388   return ParMarkBitMap::incomplete;
3389 }