1 /*
   2  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #include "precompiled.hpp"
  26 #include "classfile/symbolTable.hpp"
  27 #include "classfile/systemDictionary.hpp"
  28 #include "code/codeCache.hpp"
  29 #include "gc_implementation/parallelScavenge/gcTaskManager.hpp"
  30 #include "gc_implementation/parallelScavenge/generationSizer.hpp"
  31 #include "gc_implementation/parallelScavenge/parallelScavengeHeap.inline.hpp"
  32 #include "gc_implementation/parallelScavenge/pcTasks.hpp"
  33 #include "gc_implementation/parallelScavenge/psAdaptiveSizePolicy.hpp"
  34 #include "gc_implementation/parallelScavenge/psCompactionManager.inline.hpp"
  35 #include "gc_implementation/parallelScavenge/psMarkSweep.hpp"
  36 #include "gc_implementation/parallelScavenge/psMarkSweepDecorator.hpp"
  37 #include "gc_implementation/parallelScavenge/psOldGen.hpp"
  38 #include "gc_implementation/parallelScavenge/psParallelCompact.hpp"
  39 #include "gc_implementation/parallelScavenge/psPermGen.hpp"
  40 #include "gc_implementation/parallelScavenge/psPromotionManager.inline.hpp"
  41 #include "gc_implementation/parallelScavenge/psScavenge.hpp"
  42 #include "gc_implementation/parallelScavenge/psYoungGen.hpp"
  43 #include "gc_implementation/shared/isGCActiveMark.hpp"
  44 #include "gc_interface/gcCause.hpp"
  45 #include "memory/gcLocker.inline.hpp"
  46 #include "memory/referencePolicy.hpp"
  47 #include "memory/referenceProcessor.hpp"
  48 #include "oops/methodDataOop.hpp"
  49 #include "oops/oop.inline.hpp"
  50 #include "oops/oop.pcgc.inline.hpp"
  51 #include "runtime/fprofiler.hpp"
  52 #include "runtime/safepoint.hpp"
  53 #include "runtime/vmThread.hpp"
  54 #include "services/management.hpp"
  55 #include "services/memoryService.hpp"
  56 #include "utilities/events.hpp"
  57 #include "utilities/stack.inline.hpp"
  58 
  59 #include <math.h>
  60 
  61 // All sizes are in HeapWords.
  62 const size_t ParallelCompactData::Log2RegionSize  = 9; // 512 words
  63 const size_t ParallelCompactData::RegionSize      = (size_t)1 << Log2RegionSize;
  64 const size_t ParallelCompactData::RegionSizeBytes =
  65   RegionSize << LogHeapWordSize;
  66 const size_t ParallelCompactData::RegionSizeOffsetMask = RegionSize - 1;
  67 const size_t ParallelCompactData::RegionAddrOffsetMask = RegionSizeBytes - 1;
  68 const size_t ParallelCompactData::RegionAddrMask  = ~RegionAddrOffsetMask;
  69 
  70 const ParallelCompactData::RegionData::region_sz_t
  71 ParallelCompactData::RegionData::dc_shift = 27;
  72 
  73 const ParallelCompactData::RegionData::region_sz_t
  74 ParallelCompactData::RegionData::dc_mask = ~0U << dc_shift;
  75 
  76 const ParallelCompactData::RegionData::region_sz_t
  77 ParallelCompactData::RegionData::dc_one = 0x1U << dc_shift;
  78 
  79 const ParallelCompactData::RegionData::region_sz_t
  80 ParallelCompactData::RegionData::los_mask = ~dc_mask;
  81 
  82 const ParallelCompactData::RegionData::region_sz_t
  83 ParallelCompactData::RegionData::dc_claimed = 0x8U << dc_shift;
  84 
  85 const ParallelCompactData::RegionData::region_sz_t
  86 ParallelCompactData::RegionData::dc_completed = 0xcU << dc_shift;
  87 
  88 SpaceInfo PSParallelCompact::_space_info[PSParallelCompact::last_space_id];
  89 bool      PSParallelCompact::_print_phases = false;
  90 
  91 ReferenceProcessor* PSParallelCompact::_ref_processor = NULL;
  92 klassOop            PSParallelCompact::_updated_int_array_klass_obj = NULL;
  93 
  94 double PSParallelCompact::_dwl_mean;
  95 double PSParallelCompact::_dwl_std_dev;
  96 double PSParallelCompact::_dwl_first_term;
  97 double PSParallelCompact::_dwl_adjustment;
  98 #ifdef  ASSERT
  99 bool   PSParallelCompact::_dwl_initialized = false;
 100 #endif  // #ifdef ASSERT
 101 
 102 #ifdef VALIDATE_MARK_SWEEP
 103 GrowableArray<void*>*   PSParallelCompact::_root_refs_stack = NULL;
 104 GrowableArray<oop> *    PSParallelCompact::_live_oops = NULL;
 105 GrowableArray<oop> *    PSParallelCompact::_live_oops_moved_to = NULL;
 106 GrowableArray<size_t>*  PSParallelCompact::_live_oops_size = NULL;
 107 size_t                  PSParallelCompact::_live_oops_index = 0;
 108 size_t                  PSParallelCompact::_live_oops_index_at_perm = 0;
 109 GrowableArray<void*>*   PSParallelCompact::_other_refs_stack = NULL;
 110 GrowableArray<void*>*   PSParallelCompact::_adjusted_pointers = NULL;
 111 bool                    PSParallelCompact::_pointer_tracking = false;
 112 bool                    PSParallelCompact::_root_tracking = true;
 113 
 114 GrowableArray<HeapWord*>* PSParallelCompact::_cur_gc_live_oops = NULL;
 115 GrowableArray<HeapWord*>* PSParallelCompact::_cur_gc_live_oops_moved_to = NULL;
 116 GrowableArray<size_t>   * PSParallelCompact::_cur_gc_live_oops_size = NULL;
 117 GrowableArray<HeapWord*>* PSParallelCompact::_last_gc_live_oops = NULL;
 118 GrowableArray<HeapWord*>* PSParallelCompact::_last_gc_live_oops_moved_to = NULL;
 119 GrowableArray<size_t>   * PSParallelCompact::_last_gc_live_oops_size = NULL;
 120 #endif
 121 
 122 void SplitInfo::record(size_t src_region_idx, size_t partial_obj_size,
 123                        HeapWord* destination)
 124 {
 125   assert(src_region_idx != 0, "invalid src_region_idx");
 126   assert(partial_obj_size != 0, "invalid partial_obj_size argument");
 127   assert(destination != NULL, "invalid destination argument");
 128 
 129   _src_region_idx = src_region_idx;
 130   _partial_obj_size = partial_obj_size;
 131   _destination = destination;
 132 
 133   // These fields may not be updated below, so make sure they're clear.
 134   assert(_dest_region_addr == NULL, "should have been cleared");
 135   assert(_first_src_addr == NULL, "should have been cleared");
 136 
 137   // Determine the number of destination regions for the partial object.
 138   HeapWord* const last_word = destination + partial_obj_size - 1;
 139   const ParallelCompactData& sd = PSParallelCompact::summary_data();
 140   HeapWord* const beg_region_addr = sd.region_align_down(destination);
 141   HeapWord* const end_region_addr = sd.region_align_down(last_word);
 142 
 143   if (beg_region_addr == end_region_addr) {
 144     // One destination region.
 145     _destination_count = 1;
 146     if (end_region_addr == destination) {
 147       // The destination falls on a region boundary, thus the first word of the
 148       // partial object will be the first word copied to the destination region.
 149       _dest_region_addr = end_region_addr;
 150       _first_src_addr = sd.region_to_addr(src_region_idx);
 151     }
 152   } else {
 153     // Two destination regions.  When copied, the partial object will cross a
 154     // destination region boundary, so a word somewhere within the partial
 155     // object will be the first word copied to the second destination region.
 156     _destination_count = 2;
 157     _dest_region_addr = end_region_addr;
 158     const size_t ofs = pointer_delta(end_region_addr, destination);
 159     assert(ofs < _partial_obj_size, "sanity");
 160     _first_src_addr = sd.region_to_addr(src_region_idx) + ofs;
 161   }
 162 }
 163 
 164 void SplitInfo::clear()
 165 {
 166   _src_region_idx = 0;
 167   _partial_obj_size = 0;
 168   _destination = NULL;
 169   _destination_count = 0;
 170   _dest_region_addr = NULL;
 171   _first_src_addr = NULL;
 172   assert(!is_valid(), "sanity");
 173 }
 174 
 175 #ifdef  ASSERT
 176 void SplitInfo::verify_clear()
 177 {
 178   assert(_src_region_idx == 0, "not clear");
 179   assert(_partial_obj_size == 0, "not clear");
 180   assert(_destination == NULL, "not clear");
 181   assert(_destination_count == 0, "not clear");
 182   assert(_dest_region_addr == NULL, "not clear");
 183   assert(_first_src_addr == NULL, "not clear");
 184 }
 185 #endif  // #ifdef ASSERT
 186 
 187 
 188 #ifndef PRODUCT
 189 const char* PSParallelCompact::space_names[] = {
 190   "perm", "old ", "eden", "from", "to  "
 191 };
 192 
 193 void PSParallelCompact::print_region_ranges()
 194 {
 195   tty->print_cr("space  bottom     top        end        new_top");
 196   tty->print_cr("------ ---------- ---------- ---------- ----------");
 197 
 198   for (unsigned int id = 0; id < last_space_id; ++id) {
 199     const MutableSpace* space = _space_info[id].space();
 200     tty->print_cr("%u %s "
 201                   SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " "
 202                   SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " ",
 203                   id, space_names[id],
 204                   summary_data().addr_to_region_idx(space->bottom()),
 205                   summary_data().addr_to_region_idx(space->top()),
 206                   summary_data().addr_to_region_idx(space->end()),
 207                   summary_data().addr_to_region_idx(_space_info[id].new_top()));
 208   }
 209 }
 210 
 211 void
 212 print_generic_summary_region(size_t i, const ParallelCompactData::RegionData* c)
 213 {
 214 #define REGION_IDX_FORMAT        SIZE_FORMAT_W(7)
 215 #define REGION_DATA_FORMAT       SIZE_FORMAT_W(5)
 216 
 217   ParallelCompactData& sd = PSParallelCompact::summary_data();
 218   size_t dci = c->destination() ? sd.addr_to_region_idx(c->destination()) : 0;
 219   tty->print_cr(REGION_IDX_FORMAT " " PTR_FORMAT " "
 220                 REGION_IDX_FORMAT " " PTR_FORMAT " "
 221                 REGION_DATA_FORMAT " " REGION_DATA_FORMAT " "
 222                 REGION_DATA_FORMAT " " REGION_IDX_FORMAT " %d",
 223                 i, c->data_location(), dci, c->destination(),
 224                 c->partial_obj_size(), c->live_obj_size(),
 225                 c->data_size(), c->source_region(), c->destination_count());
 226 
 227 #undef  REGION_IDX_FORMAT
 228 #undef  REGION_DATA_FORMAT
 229 }
 230 
 231 void
 232 print_generic_summary_data(ParallelCompactData& summary_data,
 233                            HeapWord* const beg_addr,
 234                            HeapWord* const end_addr)
 235 {
 236   size_t total_words = 0;
 237   size_t i = summary_data.addr_to_region_idx(beg_addr);
 238   const size_t last = summary_data.addr_to_region_idx(end_addr);
 239   HeapWord* pdest = 0;
 240 
 241   while (i <= last) {
 242     ParallelCompactData::RegionData* c = summary_data.region(i);
 243     if (c->data_size() != 0 || c->destination() != pdest) {
 244       print_generic_summary_region(i, c);
 245       total_words += c->data_size();
 246       pdest = c->destination();
 247     }
 248     ++i;
 249   }
 250 
 251   tty->print_cr("summary_data_bytes=" SIZE_FORMAT, total_words * HeapWordSize);
 252 }
 253 
 254 void
 255 print_generic_summary_data(ParallelCompactData& summary_data,
 256                            SpaceInfo* space_info)
 257 {
 258   for (unsigned int id = 0; id < PSParallelCompact::last_space_id; ++id) {
 259     const MutableSpace* space = space_info[id].space();
 260     print_generic_summary_data(summary_data, space->bottom(),
 261                                MAX2(space->top(), space_info[id].new_top()));
 262   }
 263 }
 264 
 265 void
 266 print_initial_summary_region(size_t i,
 267                              const ParallelCompactData::RegionData* c,
 268                              bool newline = true)
 269 {
 270   tty->print(SIZE_FORMAT_W(5) " " PTR_FORMAT " "
 271              SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " "
 272              SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d",
 273              i, c->destination(),
 274              c->partial_obj_size(), c->live_obj_size(),
 275              c->data_size(), c->source_region(), c->destination_count());
 276   if (newline) tty->cr();
 277 }
 278 
 279 void
 280 print_initial_summary_data(ParallelCompactData& summary_data,
 281                            const MutableSpace* space) {
 282   if (space->top() == space->bottom()) {
 283     return;
 284   }
 285 
 286   const size_t region_size = ParallelCompactData::RegionSize;
 287   typedef ParallelCompactData::RegionData RegionData;
 288   HeapWord* const top_aligned_up = summary_data.region_align_up(space->top());
 289   const size_t end_region = summary_data.addr_to_region_idx(top_aligned_up);
 290   const RegionData* c = summary_data.region(end_region - 1);
 291   HeapWord* end_addr = c->destination() + c->data_size();
 292   const size_t live_in_space = pointer_delta(end_addr, space->bottom());
 293 
 294   // Print (and count) the full regions at the beginning of the space.
 295   size_t full_region_count = 0;
 296   size_t i = summary_data.addr_to_region_idx(space->bottom());
 297   while (i < end_region && summary_data.region(i)->data_size() == region_size) {
 298     print_initial_summary_region(i, summary_data.region(i));
 299     ++full_region_count;
 300     ++i;
 301   }
 302 
 303   size_t live_to_right = live_in_space - full_region_count * region_size;
 304 
 305   double max_reclaimed_ratio = 0.0;
 306   size_t max_reclaimed_ratio_region = 0;
 307   size_t max_dead_to_right = 0;
 308   size_t max_live_to_right = 0;
 309 
 310   // Print the 'reclaimed ratio' for regions while there is something live in
 311   // the region or to the right of it.  The remaining regions are empty (and
 312   // uninteresting), and computing the ratio will result in division by 0.
 313   while (i < end_region && live_to_right > 0) {
 314     c = summary_data.region(i);
 315     HeapWord* const region_addr = summary_data.region_to_addr(i);
 316     const size_t used_to_right = pointer_delta(space->top(), region_addr);
 317     const size_t dead_to_right = used_to_right - live_to_right;
 318     const double reclaimed_ratio = double(dead_to_right) / live_to_right;
 319 
 320     if (reclaimed_ratio > max_reclaimed_ratio) {
 321             max_reclaimed_ratio = reclaimed_ratio;
 322             max_reclaimed_ratio_region = i;
 323             max_dead_to_right = dead_to_right;
 324             max_live_to_right = live_to_right;
 325     }
 326 
 327     print_initial_summary_region(i, c, false);
 328     tty->print_cr(" %12.10f " SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10),
 329                   reclaimed_ratio, dead_to_right, live_to_right);
 330 
 331     live_to_right -= c->data_size();
 332     ++i;
 333   }
 334 
 335   // Any remaining regions are empty.  Print one more if there is one.
 336   if (i < end_region) {
 337     print_initial_summary_region(i, summary_data.region(i));
 338   }
 339 
 340   tty->print_cr("max:  " SIZE_FORMAT_W(4) " d2r=" SIZE_FORMAT_W(10) " "
 341                 "l2r=" SIZE_FORMAT_W(10) " max_ratio=%14.12f",
 342                 max_reclaimed_ratio_region, max_dead_to_right,
 343                 max_live_to_right, max_reclaimed_ratio);
 344 }
 345 
 346 void
 347 print_initial_summary_data(ParallelCompactData& summary_data,
 348                            SpaceInfo* space_info) {
 349   unsigned int id = PSParallelCompact::perm_space_id;
 350   const MutableSpace* space;
 351   do {
 352     space = space_info[id].space();
 353     print_initial_summary_data(summary_data, space);
 354   } while (++id < PSParallelCompact::eden_space_id);
 355 
 356   do {
 357     space = space_info[id].space();
 358     print_generic_summary_data(summary_data, space->bottom(), space->top());
 359   } while (++id < PSParallelCompact::last_space_id);
 360 }
 361 #endif  // #ifndef PRODUCT
 362 
 363 #ifdef  ASSERT
 364 size_t add_obj_count;
 365 size_t add_obj_size;
 366 size_t mark_bitmap_count;
 367 size_t mark_bitmap_size;
 368 #endif  // #ifdef ASSERT
 369 
 370 ParallelCompactData::ParallelCompactData()
 371 {
 372   _region_start = 0;
 373 
 374   _region_vspace = 0;
 375   _region_data = 0;
 376   _region_count = 0;
 377 }
 378 
 379 bool ParallelCompactData::initialize(MemRegion covered_region)
 380 {
 381   _region_start = covered_region.start();
 382   const size_t region_size = covered_region.word_size();
 383   DEBUG_ONLY(_region_end = _region_start + region_size;)
 384 
 385   assert(region_align_down(_region_start) == _region_start,
 386          "region start not aligned");
 387   assert((region_size & RegionSizeOffsetMask) == 0,
 388          "region size not a multiple of RegionSize");
 389 
 390   bool result = initialize_region_data(region_size);
 391 
 392   return result;
 393 }
 394 
 395 PSVirtualSpace*
 396 ParallelCompactData::create_vspace(size_t count, size_t element_size)
 397 {
 398   const size_t raw_bytes = count * element_size;
 399   const size_t page_sz = os::page_size_for_region(raw_bytes, raw_bytes, 10);
 400   const size_t granularity = os::vm_allocation_granularity();
 401   const size_t bytes = align_size_up(raw_bytes, MAX2(page_sz, granularity));
 402 
 403   const size_t rs_align = page_sz == (size_t) os::vm_page_size() ? 0 :
 404     MAX2(page_sz, granularity);
 405   ReservedSpace rs(bytes, rs_align, rs_align > 0);
 406   os::trace_page_sizes("par compact", raw_bytes, raw_bytes, page_sz, rs.base(),
 407                        rs.size());
 408   PSVirtualSpace* vspace = new PSVirtualSpace(rs, page_sz);
 409   if (vspace != 0) {
 410     if (vspace->expand_by(bytes)) {
 411       return vspace;
 412     }
 413     delete vspace;
 414     // Release memory reserved in the space.
 415     rs.release();
 416   }
 417 
 418   return 0;
 419 }
 420 
 421 bool ParallelCompactData::initialize_region_data(size_t region_size)
 422 {
 423   const size_t count = (region_size + RegionSizeOffsetMask) >> Log2RegionSize;
 424   _region_vspace = create_vspace(count, sizeof(RegionData));
 425   if (_region_vspace != 0) {
 426     _region_data = (RegionData*)_region_vspace->reserved_low_addr();
 427     _region_count = count;
 428     return true;
 429   }
 430   return false;
 431 }
 432 
 433 void ParallelCompactData::clear()
 434 {
 435   memset(_region_data, 0, _region_vspace->committed_size());
 436 }
 437 
 438 void ParallelCompactData::clear_range(size_t beg_region, size_t end_region) {
 439   assert(beg_region <= _region_count, "beg_region out of range");
 440   assert(end_region <= _region_count, "end_region out of range");
 441 
 442   const size_t region_cnt = end_region - beg_region;
 443   memset(_region_data + beg_region, 0, region_cnt * sizeof(RegionData));
 444 }
 445 
 446 HeapWord* ParallelCompactData::partial_obj_end(size_t region_idx) const
 447 {
 448   const RegionData* cur_cp = region(region_idx);
 449   const RegionData* const end_cp = region(region_count() - 1);
 450 
 451   HeapWord* result = region_to_addr(region_idx);
 452   if (cur_cp < end_cp) {
 453     do {
 454       result += cur_cp->partial_obj_size();
 455     } while (cur_cp->partial_obj_size() == RegionSize && ++cur_cp < end_cp);
 456   }
 457   return result;
 458 }
 459 
 460 void ParallelCompactData::add_obj(HeapWord* addr, size_t len)
 461 {
 462   const size_t obj_ofs = pointer_delta(addr, _region_start);
 463   const size_t beg_region = obj_ofs >> Log2RegionSize;
 464   const size_t end_region = (obj_ofs + len - 1) >> Log2RegionSize;
 465 
 466   DEBUG_ONLY(Atomic::inc_ptr(&add_obj_count);)
 467   DEBUG_ONLY(Atomic::add_ptr(len, &add_obj_size);)
 468 
 469   if (beg_region == end_region) {
 470     // All in one region.
 471     _region_data[beg_region].add_live_obj(len);
 472     return;
 473   }
 474 
 475   // First region.
 476   const size_t beg_ofs = region_offset(addr);
 477   _region_data[beg_region].add_live_obj(RegionSize - beg_ofs);
 478 
 479   klassOop klass = ((oop)addr)->klass();
 480   // Middle regions--completely spanned by this object.
 481   for (size_t region = beg_region + 1; region < end_region; ++region) {
 482     _region_data[region].set_partial_obj_size(RegionSize);
 483     _region_data[region].set_partial_obj_addr(addr);
 484   }
 485 
 486   // Last region.
 487   const size_t end_ofs = region_offset(addr + len - 1);
 488   _region_data[end_region].set_partial_obj_size(end_ofs + 1);
 489   _region_data[end_region].set_partial_obj_addr(addr);
 490 }
 491 
 492 void
 493 ParallelCompactData::summarize_dense_prefix(HeapWord* beg, HeapWord* end)
 494 {
 495   assert(region_offset(beg) == 0, "not RegionSize aligned");
 496   assert(region_offset(end) == 0, "not RegionSize aligned");
 497 
 498   size_t cur_region = addr_to_region_idx(beg);
 499   const size_t end_region = addr_to_region_idx(end);
 500   HeapWord* addr = beg;
 501   while (cur_region < end_region) {
 502     _region_data[cur_region].set_destination(addr);
 503     _region_data[cur_region].set_destination_count(0);
 504     _region_data[cur_region].set_source_region(cur_region);
 505     _region_data[cur_region].set_data_location(addr);
 506 
 507     // Update live_obj_size so the region appears completely full.
 508     size_t live_size = RegionSize - _region_data[cur_region].partial_obj_size();
 509     _region_data[cur_region].set_live_obj_size(live_size);
 510 
 511     ++cur_region;
 512     addr += RegionSize;
 513   }
 514 }
 515 
 516 // Find the point at which a space can be split and, if necessary, record the
 517 // split point.
 518 //
 519 // If the current src region (which overflowed the destination space) doesn't
 520 // have a partial object, the split point is at the beginning of the current src
 521 // region (an "easy" split, no extra bookkeeping required).
 522 //
 523 // If the current src region has a partial object, the split point is in the
 524 // region where that partial object starts (call it the split_region).  If
 525 // split_region has a partial object, then the split point is just after that
 526 // partial object (a "hard" split where we have to record the split data and
 527 // zero the partial_obj_size field).  With a "hard" split, we know that the
 528 // partial_obj ends within split_region because the partial object that caused
 529 // the overflow starts in split_region.  If split_region doesn't have a partial
 530 // obj, then the split is at the beginning of split_region (another "easy"
 531 // split).
 532 HeapWord*
 533 ParallelCompactData::summarize_split_space(size_t src_region,
 534                                            SplitInfo& split_info,
 535                                            HeapWord* destination,
 536                                            HeapWord* target_end,
 537                                            HeapWord** target_next)
 538 {
 539   assert(destination <= target_end, "sanity");
 540   assert(destination + _region_data[src_region].data_size() > target_end,
 541     "region should not fit into target space");
 542   assert(is_region_aligned(target_end), "sanity");
 543 
 544   size_t split_region = src_region;
 545   HeapWord* split_destination = destination;
 546   size_t partial_obj_size = _region_data[src_region].partial_obj_size();
 547 
 548   if (destination + partial_obj_size > target_end) {
 549     // The split point is just after the partial object (if any) in the
 550     // src_region that contains the start of the object that overflowed the
 551     // destination space.
 552     //
 553     // Find the start of the "overflow" object and set split_region to the
 554     // region containing it.
 555     HeapWord* const overflow_obj = _region_data[src_region].partial_obj_addr();
 556     split_region = addr_to_region_idx(overflow_obj);
 557 
 558     // Clear the source_region field of all destination regions whose first word
 559     // came from data after the split point (a non-null source_region field
 560     // implies a region must be filled).
 561     //
 562     // An alternative to the simple loop below:  clear during post_compact(),
 563     // which uses memcpy instead of individual stores, and is easy to
 564     // parallelize.  (The downside is that it clears the entire RegionData
 565     // object as opposed to just one field.)
 566     //
 567     // post_compact() would have to clear the summary data up to the highest
 568     // address that was written during the summary phase, which would be
 569     //
 570     //         max(top, max(new_top, clear_top))
 571     //
 572     // where clear_top is a new field in SpaceInfo.  Would have to set clear_top
 573     // to target_end.
 574     const RegionData* const sr = region(split_region);
 575     const size_t beg_idx =
 576       addr_to_region_idx(region_align_up(sr->destination() +
 577                                          sr->partial_obj_size()));
 578     const size_t end_idx = addr_to_region_idx(target_end);
 579 
 580     if (TraceParallelOldGCSummaryPhase) {
 581         gclog_or_tty->print_cr("split:  clearing source_region field in ["
 582                                SIZE_FORMAT ", " SIZE_FORMAT ")",
 583                                beg_idx, end_idx);
 584     }
 585     for (size_t idx = beg_idx; idx < end_idx; ++idx) {
 586       _region_data[idx].set_source_region(0);
 587     }
 588 
 589     // Set split_destination and partial_obj_size to reflect the split region.
 590     split_destination = sr->destination();
 591     partial_obj_size = sr->partial_obj_size();
 592   }
 593 
 594   // The split is recorded only if a partial object extends onto the region.
 595   if (partial_obj_size != 0) {
 596     _region_data[split_region].set_partial_obj_size(0);
 597     split_info.record(split_region, partial_obj_size, split_destination);
 598   }
 599 
 600   // Setup the continuation addresses.
 601   *target_next = split_destination + partial_obj_size;
 602   HeapWord* const source_next = region_to_addr(split_region) + partial_obj_size;
 603 
 604   if (TraceParallelOldGCSummaryPhase) {
 605     const char * split_type = partial_obj_size == 0 ? "easy" : "hard";
 606     gclog_or_tty->print_cr("%s split:  src=" PTR_FORMAT " src_c=" SIZE_FORMAT
 607                            " pos=" SIZE_FORMAT,
 608                            split_type, source_next, split_region,
 609                            partial_obj_size);
 610     gclog_or_tty->print_cr("%s split:  dst=" PTR_FORMAT " dst_c=" SIZE_FORMAT
 611                            " tn=" PTR_FORMAT,
 612                            split_type, split_destination,
 613                            addr_to_region_idx(split_destination),
 614                            *target_next);
 615 
 616     if (partial_obj_size != 0) {
 617       HeapWord* const po_beg = split_info.destination();
 618       HeapWord* const po_end = po_beg + split_info.partial_obj_size();
 619       gclog_or_tty->print_cr("%s split:  "
 620                              "po_beg=" PTR_FORMAT " " SIZE_FORMAT " "
 621                              "po_end=" PTR_FORMAT " " SIZE_FORMAT,
 622                              split_type,
 623                              po_beg, addr_to_region_idx(po_beg),
 624                              po_end, addr_to_region_idx(po_end));
 625     }
 626   }
 627 
 628   return source_next;
 629 }
 630 
 631 bool ParallelCompactData::summarize(SplitInfo& split_info,
 632                                     HeapWord* source_beg, HeapWord* source_end,
 633                                     HeapWord** source_next,
 634                                     HeapWord* target_beg, HeapWord* target_end,
 635                                     HeapWord** target_next)
 636 {
 637   if (TraceParallelOldGCSummaryPhase) {
 638     HeapWord* const source_next_val = source_next == NULL ? NULL : *source_next;
 639     tty->print_cr("sb=" PTR_FORMAT " se=" PTR_FORMAT " sn=" PTR_FORMAT
 640                   "tb=" PTR_FORMAT " te=" PTR_FORMAT " tn=" PTR_FORMAT,
 641                   source_beg, source_end, source_next_val,
 642                   target_beg, target_end, *target_next);
 643   }
 644 
 645   size_t cur_region = addr_to_region_idx(source_beg);
 646   const size_t end_region = addr_to_region_idx(region_align_up(source_end));
 647 
 648   HeapWord *dest_addr = target_beg;
 649   while (cur_region < end_region) {
 650     // The destination must be set even if the region has no data.
 651     _region_data[cur_region].set_destination(dest_addr);
 652 
 653     size_t words = _region_data[cur_region].data_size();
 654     if (words > 0) {
 655       // If cur_region does not fit entirely into the target space, find a point
 656       // at which the source space can be 'split' so that part is copied to the
 657       // target space and the rest is copied elsewhere.
 658       if (dest_addr + words > target_end) {
 659         assert(source_next != NULL, "source_next is NULL when splitting");
 660         *source_next = summarize_split_space(cur_region, split_info, dest_addr,
 661                                              target_end, target_next);
 662         return false;
 663       }
 664 
 665       // Compute the destination_count for cur_region, and if necessary, update
 666       // source_region for a destination region.  The source_region field is
 667       // updated if cur_region is the first (left-most) region to be copied to a
 668       // destination region.
 669       //
 670       // The destination_count calculation is a bit subtle.  A region that has
 671       // data that compacts into itself does not count itself as a destination.
 672       // This maintains the invariant that a zero count means the region is
 673       // available and can be claimed and then filled.
 674       uint destination_count = 0;
 675       if (split_info.is_split(cur_region)) {
 676         // The current region has been split:  the partial object will be copied
 677         // to one destination space and the remaining data will be copied to
 678         // another destination space.  Adjust the initial destination_count and,
 679         // if necessary, set the source_region field if the partial object will
 680         // cross a destination region boundary.
 681         destination_count = split_info.destination_count();
 682         if (destination_count == 2) {
 683           size_t dest_idx = addr_to_region_idx(split_info.dest_region_addr());
 684           _region_data[dest_idx].set_source_region(cur_region);
 685         }
 686       }
 687 
 688       HeapWord* const last_addr = dest_addr + words - 1;
 689       const size_t dest_region_1 = addr_to_region_idx(dest_addr);
 690       const size_t dest_region_2 = addr_to_region_idx(last_addr);
 691 
 692       // Initially assume that the destination regions will be the same and
 693       // adjust the value below if necessary.  Under this assumption, if
 694       // cur_region == dest_region_2, then cur_region will be compacted
 695       // completely into itself.
 696       destination_count += cur_region == dest_region_2 ? 0 : 1;
 697       if (dest_region_1 != dest_region_2) {
 698         // Destination regions differ; adjust destination_count.
 699         destination_count += 1;
 700         // Data from cur_region will be copied to the start of dest_region_2.
 701         _region_data[dest_region_2].set_source_region(cur_region);
 702       } else if (region_offset(dest_addr) == 0) {
 703         // Data from cur_region will be copied to the start of the destination
 704         // region.
 705         _region_data[dest_region_1].set_source_region(cur_region);
 706       }
 707 
 708       _region_data[cur_region].set_destination_count(destination_count);
 709       _region_data[cur_region].set_data_location(region_to_addr(cur_region));
 710       dest_addr += words;
 711     }
 712 
 713     ++cur_region;
 714   }
 715 
 716   *target_next = dest_addr;
 717   return true;
 718 }
 719 
 720 HeapWord* ParallelCompactData::calc_new_pointer(HeapWord* addr) {
 721   assert(addr != NULL, "Should detect NULL oop earlier");
 722   assert(PSParallelCompact::gc_heap()->is_in(addr), "addr not in heap");
 723 #ifdef ASSERT
 724   if (PSParallelCompact::mark_bitmap()->is_unmarked(addr)) {
 725     gclog_or_tty->print_cr("calc_new_pointer:: addr " PTR_FORMAT, addr);
 726   }
 727 #endif
 728   assert(PSParallelCompact::mark_bitmap()->is_marked(addr), "obj not marked");
 729 
 730   // Region covering the object.
 731   size_t region_index = addr_to_region_idx(addr);
 732   const RegionData* const region_ptr = region(region_index);
 733   HeapWord* const region_addr = region_align_down(addr);
 734 
 735   assert(addr < region_addr + RegionSize, "Region does not cover object");
 736   assert(addr_to_region_ptr(region_addr) == region_ptr, "sanity check");
 737 
 738   HeapWord* result = region_ptr->destination();
 739 
 740   // If all the data in the region is live, then the new location of the object
 741   // can be calculated from the destination of the region plus the offset of the
 742   // object in the region.
 743   if (region_ptr->data_size() == RegionSize) {
 744     result += pointer_delta(addr, region_addr);
 745     DEBUG_ONLY(PSParallelCompact::check_new_location(addr, result);)
 746     return result;
 747   }
 748 
 749   // The new location of the object is
 750   //    region destination +
 751   //    size of the partial object extending onto the region +
 752   //    sizes of the live objects in the Region that are to the left of addr
 753   const size_t partial_obj_size = region_ptr->partial_obj_size();
 754   HeapWord* const search_start = region_addr + partial_obj_size;
 755 
 756   const ParMarkBitMap* bitmap = PSParallelCompact::mark_bitmap();
 757   size_t live_to_left = bitmap->live_words_in_range(search_start, oop(addr));
 758 
 759   result += partial_obj_size + live_to_left;
 760   DEBUG_ONLY(PSParallelCompact::check_new_location(addr, result);)
 761   return result;
 762 }
 763 
 764 klassOop ParallelCompactData::calc_new_klass(klassOop old_klass) {
 765   klassOop updated_klass;
 766   if (PSParallelCompact::should_update_klass(old_klass)) {
 767     updated_klass = (klassOop) calc_new_pointer(old_klass);
 768   } else {
 769     updated_klass = old_klass;
 770   }
 771 
 772   return updated_klass;
 773 }
 774 
 775 #ifdef  ASSERT
 776 void ParallelCompactData::verify_clear(const PSVirtualSpace* vspace)
 777 {
 778   const size_t* const beg = (const size_t*)vspace->committed_low_addr();
 779   const size_t* const end = (const size_t*)vspace->committed_high_addr();
 780   for (const size_t* p = beg; p < end; ++p) {
 781     assert(*p == 0, "not zero");
 782   }
 783 }
 784 
 785 void ParallelCompactData::verify_clear()
 786 {
 787   verify_clear(_region_vspace);
 788 }
 789 #endif  // #ifdef ASSERT
 790 
 791 #ifdef NOT_PRODUCT
 792 ParallelCompactData::RegionData* debug_region(size_t region_index) {
 793   ParallelCompactData& sd = PSParallelCompact::summary_data();
 794   return sd.region(region_index);
 795 }
 796 #endif
 797 
 798 elapsedTimer        PSParallelCompact::_accumulated_time;
 799 unsigned int        PSParallelCompact::_total_invocations = 0;
 800 unsigned int        PSParallelCompact::_maximum_compaction_gc_num = 0;
 801 jlong               PSParallelCompact::_time_of_last_gc = 0;
 802 CollectorCounters*  PSParallelCompact::_counters = NULL;
 803 ParMarkBitMap       PSParallelCompact::_mark_bitmap;
 804 ParallelCompactData PSParallelCompact::_summary_data;
 805 
 806 PSParallelCompact::IsAliveClosure PSParallelCompact::_is_alive_closure;
 807 
 808 void PSParallelCompact::IsAliveClosure::do_object(oop p)   { ShouldNotReachHere(); }
 809 bool PSParallelCompact::IsAliveClosure::do_object_b(oop p) { return mark_bitmap()->is_marked(p); }
 810 
 811 void PSParallelCompact::KeepAliveClosure::do_oop(oop* p)       { PSParallelCompact::KeepAliveClosure::do_oop_work(p); }
 812 void PSParallelCompact::KeepAliveClosure::do_oop(narrowOop* p) { PSParallelCompact::KeepAliveClosure::do_oop_work(p); }
 813 
 814 PSParallelCompact::AdjustPointerClosure PSParallelCompact::_adjust_root_pointer_closure(true);
 815 PSParallelCompact::AdjustPointerClosure PSParallelCompact::_adjust_pointer_closure(false);
 816 
 817 void PSParallelCompact::AdjustPointerClosure::do_oop(oop* p)       { adjust_pointer(p, _is_root); }
 818 void PSParallelCompact::AdjustPointerClosure::do_oop(narrowOop* p) { adjust_pointer(p, _is_root); }
 819 
 820 void PSParallelCompact::FollowStackClosure::do_void() { _compaction_manager->follow_marking_stacks(); }
 821 
 822 void PSParallelCompact::MarkAndPushClosure::do_oop(oop* p)       { mark_and_push(_compaction_manager, p); }
 823 void PSParallelCompact::MarkAndPushClosure::do_oop(narrowOop* p) { mark_and_push(_compaction_manager, p); }
 824 
 825 void PSParallelCompact::post_initialize() {
 826   ParallelScavengeHeap* heap = gc_heap();
 827   assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity");
 828 
 829   MemRegion mr = heap->reserved_region();
 830   _ref_processor = ReferenceProcessor::create_ref_processor(
 831     mr,                         // span
 832     true,                       // atomic_discovery
 833     true,                       // mt_discovery
 834     &_is_alive_closure,
 835     ParallelGCThreads,
 836     ParallelRefProcEnabled);
 837   _counters = new CollectorCounters("PSParallelCompact", 1);
 838 
 839   // Initialize static fields in ParCompactionManager.
 840   ParCompactionManager::initialize(mark_bitmap());
 841 }
 842 
 843 bool PSParallelCompact::initialize() {
 844   ParallelScavengeHeap* heap = gc_heap();
 845   assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity");
 846   MemRegion mr = heap->reserved_region();
 847 
 848   // Was the old gen get allocated successfully?
 849   if (!heap->old_gen()->is_allocated()) {
 850     return false;
 851   }
 852 
 853   initialize_space_info();
 854   initialize_dead_wood_limiter();
 855 
 856   if (!_mark_bitmap.initialize(mr)) {
 857     vm_shutdown_during_initialization("Unable to allocate bit map for "
 858       "parallel garbage collection for the requested heap size.");
 859     return false;
 860   }
 861 
 862   if (!_summary_data.initialize(mr)) {
 863     vm_shutdown_during_initialization("Unable to allocate tables for "
 864       "parallel garbage collection for the requested heap size.");
 865     return false;
 866   }
 867 
 868   return true;
 869 }
 870 
 871 void PSParallelCompact::initialize_space_info()
 872 {
 873   memset(&_space_info, 0, sizeof(_space_info));
 874 
 875   ParallelScavengeHeap* heap = gc_heap();
 876   PSYoungGen* young_gen = heap->young_gen();
 877   MutableSpace* perm_space = heap->perm_gen()->object_space();
 878 
 879   _space_info[perm_space_id].set_space(perm_space);
 880   _space_info[old_space_id].set_space(heap->old_gen()->object_space());
 881   _space_info[eden_space_id].set_space(young_gen->eden_space());
 882   _space_info[from_space_id].set_space(young_gen->from_space());
 883   _space_info[to_space_id].set_space(young_gen->to_space());
 884 
 885   _space_info[perm_space_id].set_start_array(heap->perm_gen()->start_array());
 886   _space_info[old_space_id].set_start_array(heap->old_gen()->start_array());
 887 
 888   _space_info[perm_space_id].set_min_dense_prefix(perm_space->top());
 889   if (TraceParallelOldGCDensePrefix) {
 890     tty->print_cr("perm min_dense_prefix=" PTR_FORMAT,
 891                   _space_info[perm_space_id].min_dense_prefix());
 892   }
 893 }
 894 
 895 void PSParallelCompact::initialize_dead_wood_limiter()
 896 {
 897   const size_t max = 100;
 898   _dwl_mean = double(MIN2(ParallelOldDeadWoodLimiterMean, max)) / 100.0;
 899   _dwl_std_dev = double(MIN2(ParallelOldDeadWoodLimiterStdDev, max)) / 100.0;
 900   _dwl_first_term = 1.0 / (sqrt(2.0 * M_PI) * _dwl_std_dev);
 901   DEBUG_ONLY(_dwl_initialized = true;)
 902   _dwl_adjustment = normal_distribution(1.0);
 903 }
 904 
 905 // Simple class for storing info about the heap at the start of GC, to be used
 906 // after GC for comparison/printing.
 907 class PreGCValues {
 908 public:
 909   PreGCValues() { }
 910   PreGCValues(ParallelScavengeHeap* heap) { fill(heap); }
 911 
 912   void fill(ParallelScavengeHeap* heap) {
 913     _heap_used      = heap->used();
 914     _young_gen_used = heap->young_gen()->used_in_bytes();
 915     _old_gen_used   = heap->old_gen()->used_in_bytes();
 916     _perm_gen_used  = heap->perm_gen()->used_in_bytes();
 917   };
 918 
 919   size_t heap_used() const      { return _heap_used; }
 920   size_t young_gen_used() const { return _young_gen_used; }
 921   size_t old_gen_used() const   { return _old_gen_used; }
 922   size_t perm_gen_used() const  { return _perm_gen_used; }
 923 
 924 private:
 925   size_t _heap_used;
 926   size_t _young_gen_used;
 927   size_t _old_gen_used;
 928   size_t _perm_gen_used;
 929 };
 930 
 931 void
 932 PSParallelCompact::clear_data_covering_space(SpaceId id)
 933 {
 934   // At this point, top is the value before GC, new_top() is the value that will
 935   // be set at the end of GC.  The marking bitmap is cleared to top; nothing
 936   // should be marked above top.  The summary data is cleared to the larger of
 937   // top & new_top.
 938   MutableSpace* const space = _space_info[id].space();
 939   HeapWord* const bot = space->bottom();
 940   HeapWord* const top = space->top();
 941   HeapWord* const max_top = MAX2(top, _space_info[id].new_top());
 942 
 943   const idx_t beg_bit = _mark_bitmap.addr_to_bit(bot);
 944   const idx_t end_bit = BitMap::word_align_up(_mark_bitmap.addr_to_bit(top));
 945   _mark_bitmap.clear_range(beg_bit, end_bit);
 946 
 947   const size_t beg_region = _summary_data.addr_to_region_idx(bot);
 948   const size_t end_region =
 949     _summary_data.addr_to_region_idx(_summary_data.region_align_up(max_top));
 950   _summary_data.clear_range(beg_region, end_region);
 951 
 952   // Clear the data used to 'split' regions.
 953   SplitInfo& split_info = _space_info[id].split_info();
 954   if (split_info.is_valid()) {
 955     split_info.clear();
 956   }
 957   DEBUG_ONLY(split_info.verify_clear();)
 958 }
 959 
 960 void PSParallelCompact::pre_compact(PreGCValues* pre_gc_values)
 961 {
 962   // Update the from & to space pointers in space_info, since they are swapped
 963   // at each young gen gc.  Do the update unconditionally (even though a
 964   // promotion failure does not swap spaces) because an unknown number of minor
 965   // collections will have swapped the spaces an unknown number of times.
 966   TraceTime tm("pre compact", print_phases(), true, gclog_or_tty);
 967   ParallelScavengeHeap* heap = gc_heap();
 968   _space_info[from_space_id].set_space(heap->young_gen()->from_space());
 969   _space_info[to_space_id].set_space(heap->young_gen()->to_space());
 970 
 971   pre_gc_values->fill(heap);
 972 
 973   ParCompactionManager::reset();
 974   NOT_PRODUCT(_mark_bitmap.reset_counters());
 975   DEBUG_ONLY(add_obj_count = add_obj_size = 0;)
 976   DEBUG_ONLY(mark_bitmap_count = mark_bitmap_size = 0;)
 977 
 978   // Increment the invocation count
 979   heap->increment_total_collections(true);
 980 
 981   // We need to track unique mark sweep invocations as well.
 982   _total_invocations++;
 983 
 984   if (PrintHeapAtGC) {
 985     Universe::print_heap_before_gc();
 986   }
 987 
 988   // Fill in TLABs
 989   heap->accumulate_statistics_all_tlabs();
 990   heap->ensure_parsability(true);  // retire TLABs
 991 
 992   if (VerifyBeforeGC && heap->total_collections() >= VerifyGCStartAt) {
 993     HandleMark hm;  // Discard invalid handles created during verification
 994     gclog_or_tty->print(" VerifyBeforeGC:");
 995     Universe::verify(true);
 996   }
 997 
 998   // Verify object start arrays
 999   if (VerifyObjectStartArray &&
1000       VerifyBeforeGC) {
1001     heap->old_gen()->verify_object_start_array();
1002     heap->perm_gen()->verify_object_start_array();
1003   }
1004 
1005   DEBUG_ONLY(mark_bitmap()->verify_clear();)
1006   DEBUG_ONLY(summary_data().verify_clear();)
1007 
1008   // Have worker threads release resources the next time they run a task.
1009   gc_task_manager()->release_all_resources();
1010 }
1011 
1012 void PSParallelCompact::post_compact()
1013 {
1014   TraceTime tm("post compact", print_phases(), true, gclog_or_tty);
1015 
1016   for (unsigned int id = perm_space_id; id < last_space_id; ++id) {
1017     // Clear the marking bitmap, summary data and split info.
1018     clear_data_covering_space(SpaceId(id));
1019     // Update top().  Must be done after clearing the bitmap and summary data.
1020     _space_info[id].publish_new_top();
1021   }
1022 
1023   MutableSpace* const eden_space = _space_info[eden_space_id].space();
1024   MutableSpace* const from_space = _space_info[from_space_id].space();
1025   MutableSpace* const to_space   = _space_info[to_space_id].space();
1026 
1027   ParallelScavengeHeap* heap = gc_heap();
1028   bool eden_empty = eden_space->is_empty();
1029   if (!eden_empty) {
1030     eden_empty = absorb_live_data_from_eden(heap->size_policy(),
1031                                             heap->young_gen(), heap->old_gen());
1032   }
1033 
1034   // Update heap occupancy information which is used as input to the soft ref
1035   // clearing policy at the next gc.
1036   Universe::update_heap_info_at_gc();
1037 
1038   bool young_gen_empty = eden_empty && from_space->is_empty() &&
1039     to_space->is_empty();
1040 
1041   BarrierSet* bs = heap->barrier_set();
1042   if (bs->is_a(BarrierSet::ModRef)) {
1043     ModRefBarrierSet* modBS = (ModRefBarrierSet*)bs;
1044     MemRegion old_mr = heap->old_gen()->reserved();
1045     MemRegion perm_mr = heap->perm_gen()->reserved();
1046     assert(perm_mr.end() <= old_mr.start(), "Generations out of order");
1047 
1048     if (young_gen_empty) {
1049       modBS->clear(MemRegion(perm_mr.start(), old_mr.end()));
1050     } else {
1051       modBS->invalidate(MemRegion(perm_mr.start(), old_mr.end()));
1052     }
1053   }
1054 
1055   Threads::gc_epilogue();
1056   CodeCache::gc_epilogue();
1057   JvmtiExport::gc_epilogue();
1058 
1059   COMPILER2_PRESENT(DerivedPointerTable::update_pointers());
1060 
1061   ref_processor()->enqueue_discovered_references(NULL);
1062 
1063   if (ZapUnusedHeapArea) {
1064     heap->gen_mangle_unused_area();
1065   }
1066 
1067   // Update time of last GC
1068   reset_millis_since_last_gc();
1069 }
1070 
1071 HeapWord*
1072 PSParallelCompact::compute_dense_prefix_via_density(const SpaceId id,
1073                                                     bool maximum_compaction)
1074 {
1075   const size_t region_size = ParallelCompactData::RegionSize;
1076   const ParallelCompactData& sd = summary_data();
1077 
1078   const MutableSpace* const space = _space_info[id].space();
1079   HeapWord* const top_aligned_up = sd.region_align_up(space->top());
1080   const RegionData* const beg_cp = sd.addr_to_region_ptr(space->bottom());
1081   const RegionData* const end_cp = sd.addr_to_region_ptr(top_aligned_up);
1082 
1083   // Skip full regions at the beginning of the space--they are necessarily part
1084   // of the dense prefix.
1085   size_t full_count = 0;
1086   const RegionData* cp;
1087   for (cp = beg_cp; cp < end_cp && cp->data_size() == region_size; ++cp) {
1088     ++full_count;
1089   }
1090 
1091   assert(total_invocations() >= _maximum_compaction_gc_num, "sanity");
1092   const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num;
1093   const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval;
1094   if (maximum_compaction || cp == end_cp || interval_ended) {
1095     _maximum_compaction_gc_num = total_invocations();
1096     return sd.region_to_addr(cp);
1097   }
1098 
1099   HeapWord* const new_top = _space_info[id].new_top();
1100   const size_t space_live = pointer_delta(new_top, space->bottom());
1101   const size_t space_used = space->used_in_words();
1102   const size_t space_capacity = space->capacity_in_words();
1103 
1104   const double cur_density = double(space_live) / space_capacity;
1105   const double deadwood_density =
1106     (1.0 - cur_density) * (1.0 - cur_density) * cur_density * cur_density;
1107   const size_t deadwood_goal = size_t(space_capacity * deadwood_density);
1108 
1109   if (TraceParallelOldGCDensePrefix) {
1110     tty->print_cr("cur_dens=%5.3f dw_dens=%5.3f dw_goal=" SIZE_FORMAT,
1111                   cur_density, deadwood_density, deadwood_goal);
1112     tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " "
1113                   "space_cap=" SIZE_FORMAT,
1114                   space_live, space_used,
1115                   space_capacity);
1116   }
1117 
1118   // XXX - Use binary search?
1119   HeapWord* dense_prefix = sd.region_to_addr(cp);
1120   const RegionData* full_cp = cp;
1121   const RegionData* const top_cp = sd.addr_to_region_ptr(space->top() - 1);
1122   while (cp < end_cp) {
1123     HeapWord* region_destination = cp->destination();
1124     const size_t cur_deadwood = pointer_delta(dense_prefix, region_destination);
1125     if (TraceParallelOldGCDensePrefix && Verbose) {
1126       tty->print_cr("c#=" SIZE_FORMAT_W(4) " dst=" PTR_FORMAT " "
1127                     "dp=" SIZE_FORMAT_W(8) " " "cdw=" SIZE_FORMAT_W(8),
1128                     sd.region(cp), region_destination,
1129                     dense_prefix, cur_deadwood);
1130     }
1131 
1132     if (cur_deadwood >= deadwood_goal) {
1133       // Found the region that has the correct amount of deadwood to the left.
1134       // This typically occurs after crossing a fairly sparse set of regions, so
1135       // iterate backwards over those sparse regions, looking for the region
1136       // that has the lowest density of live objects 'to the right.'
1137       size_t space_to_left = sd.region(cp) * region_size;
1138       size_t live_to_left = space_to_left - cur_deadwood;
1139       size_t space_to_right = space_capacity - space_to_left;
1140       size_t live_to_right = space_live - live_to_left;
1141       double density_to_right = double(live_to_right) / space_to_right;
1142       while (cp > full_cp) {
1143         --cp;
1144         const size_t prev_region_live_to_right = live_to_right -
1145           cp->data_size();
1146         const size_t prev_region_space_to_right = space_to_right + region_size;
1147         double prev_region_density_to_right =
1148           double(prev_region_live_to_right) / prev_region_space_to_right;
1149         if (density_to_right <= prev_region_density_to_right) {
1150           return dense_prefix;
1151         }
1152         if (TraceParallelOldGCDensePrefix && Verbose) {
1153           tty->print_cr("backing up from c=" SIZE_FORMAT_W(4) " d2r=%10.8f "
1154                         "pc_d2r=%10.8f", sd.region(cp), density_to_right,
1155                         prev_region_density_to_right);
1156         }
1157         dense_prefix -= region_size;
1158         live_to_right = prev_region_live_to_right;
1159         space_to_right = prev_region_space_to_right;
1160         density_to_right = prev_region_density_to_right;
1161       }
1162       return dense_prefix;
1163     }
1164 
1165     dense_prefix += region_size;
1166     ++cp;
1167   }
1168 
1169   return dense_prefix;
1170 }
1171 
1172 #ifndef PRODUCT
1173 void PSParallelCompact::print_dense_prefix_stats(const char* const algorithm,
1174                                                  const SpaceId id,
1175                                                  const bool maximum_compaction,
1176                                                  HeapWord* const addr)
1177 {
1178   const size_t region_idx = summary_data().addr_to_region_idx(addr);
1179   RegionData* const cp = summary_data().region(region_idx);
1180   const MutableSpace* const space = _space_info[id].space();
1181   HeapWord* const new_top = _space_info[id].new_top();
1182 
1183   const size_t space_live = pointer_delta(new_top, space->bottom());
1184   const size_t dead_to_left = pointer_delta(addr, cp->destination());
1185   const size_t space_cap = space->capacity_in_words();
1186   const double dead_to_left_pct = double(dead_to_left) / space_cap;
1187   const size_t live_to_right = new_top - cp->destination();
1188   const size_t dead_to_right = space->top() - addr - live_to_right;
1189 
1190   tty->print_cr("%s=" PTR_FORMAT " dpc=" SIZE_FORMAT_W(5) " "
1191                 "spl=" SIZE_FORMAT " "
1192                 "d2l=" SIZE_FORMAT " d2l%%=%6.4f "
1193                 "d2r=" SIZE_FORMAT " l2r=" SIZE_FORMAT
1194                 " ratio=%10.8f",
1195                 algorithm, addr, region_idx,
1196                 space_live,
1197                 dead_to_left, dead_to_left_pct,
1198                 dead_to_right, live_to_right,
1199                 double(dead_to_right) / live_to_right);
1200 }
1201 #endif  // #ifndef PRODUCT
1202 
1203 // Return a fraction indicating how much of the generation can be treated as
1204 // "dead wood" (i.e., not reclaimed).  The function uses a normal distribution
1205 // based on the density of live objects in the generation to determine a limit,
1206 // which is then adjusted so the return value is min_percent when the density is
1207 // 1.
1208 //
1209 // The following table shows some return values for a different values of the
1210 // standard deviation (ParallelOldDeadWoodLimiterStdDev); the mean is 0.5 and
1211 // min_percent is 1.
1212 //
1213 //                          fraction allowed as dead wood
1214 //         -----------------------------------------------------------------
1215 // density std_dev=70 std_dev=75 std_dev=80 std_dev=85 std_dev=90 std_dev=95
1216 // ------- ---------- ---------- ---------- ---------- ---------- ----------
1217 // 0.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000
1218 // 0.05000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941
1219 // 0.10000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272
1220 // 0.15000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066
1221 // 0.20000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975
1222 // 0.25000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313
1223 // 0.30000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132
1224 // 0.35000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289
1225 // 0.40000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500
1226 // 0.45000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386
1227 // 0.50000 0.13832410 0.11599237 0.09847664 0.08456518 0.07338887 0.06431510
1228 // 0.55000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386
1229 // 0.60000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500
1230 // 0.65000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289
1231 // 0.70000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132
1232 // 0.75000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313
1233 // 0.80000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975
1234 // 0.85000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066
1235 // 0.90000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272
1236 // 0.95000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941
1237 // 1.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000
1238 
1239 double PSParallelCompact::dead_wood_limiter(double density, size_t min_percent)
1240 {
1241   assert(_dwl_initialized, "uninitialized");
1242 
1243   // The raw limit is the value of the normal distribution at x = density.
1244   const double raw_limit = normal_distribution(density);
1245 
1246   // Adjust the raw limit so it becomes the minimum when the density is 1.
1247   //
1248   // First subtract the adjustment value (which is simply the precomputed value
1249   // normal_distribution(1.0)); this yields a value of 0 when the density is 1.
1250   // Then add the minimum value, so the minimum is returned when the density is
1251   // 1.  Finally, prevent negative values, which occur when the mean is not 0.5.
1252   const double min = double(min_percent) / 100.0;
1253   const double limit = raw_limit - _dwl_adjustment + min;
1254   return MAX2(limit, 0.0);
1255 }
1256 
1257 ParallelCompactData::RegionData*
1258 PSParallelCompact::first_dead_space_region(const RegionData* beg,
1259                                            const RegionData* end)
1260 {
1261   const size_t region_size = ParallelCompactData::RegionSize;
1262   ParallelCompactData& sd = summary_data();
1263   size_t left = sd.region(beg);
1264   size_t right = end > beg ? sd.region(end) - 1 : left;
1265 
1266   // Binary search.
1267   while (left < right) {
1268     // Equivalent to (left + right) / 2, but does not overflow.
1269     const size_t middle = left + (right - left) / 2;
1270     RegionData* const middle_ptr = sd.region(middle);
1271     HeapWord* const dest = middle_ptr->destination();
1272     HeapWord* const addr = sd.region_to_addr(middle);
1273     assert(dest != NULL, "sanity");
1274     assert(dest <= addr, "must move left");
1275 
1276     if (middle > left && dest < addr) {
1277       right = middle - 1;
1278     } else if (middle < right && middle_ptr->data_size() == region_size) {
1279       left = middle + 1;
1280     } else {
1281       return middle_ptr;
1282     }
1283   }
1284   return sd.region(left);
1285 }
1286 
1287 ParallelCompactData::RegionData*
1288 PSParallelCompact::dead_wood_limit_region(const RegionData* beg,
1289                                           const RegionData* end,
1290                                           size_t dead_words)
1291 {
1292   ParallelCompactData& sd = summary_data();
1293   size_t left = sd.region(beg);
1294   size_t right = end > beg ? sd.region(end) - 1 : left;
1295 
1296   // Binary search.
1297   while (left < right) {
1298     // Equivalent to (left + right) / 2, but does not overflow.
1299     const size_t middle = left + (right - left) / 2;
1300     RegionData* const middle_ptr = sd.region(middle);
1301     HeapWord* const dest = middle_ptr->destination();
1302     HeapWord* const addr = sd.region_to_addr(middle);
1303     assert(dest != NULL, "sanity");
1304     assert(dest <= addr, "must move left");
1305 
1306     const size_t dead_to_left = pointer_delta(addr, dest);
1307     if (middle > left && dead_to_left > dead_words) {
1308       right = middle - 1;
1309     } else if (middle < right && dead_to_left < dead_words) {
1310       left = middle + 1;
1311     } else {
1312       return middle_ptr;
1313     }
1314   }
1315   return sd.region(left);
1316 }
1317 
1318 // The result is valid during the summary phase, after the initial summarization
1319 // of each space into itself, and before final summarization.
1320 inline double
1321 PSParallelCompact::reclaimed_ratio(const RegionData* const cp,
1322                                    HeapWord* const bottom,
1323                                    HeapWord* const top,
1324                                    HeapWord* const new_top)
1325 {
1326   ParallelCompactData& sd = summary_data();
1327 
1328   assert(cp != NULL, "sanity");
1329   assert(bottom != NULL, "sanity");
1330   assert(top != NULL, "sanity");
1331   assert(new_top != NULL, "sanity");
1332   assert(top >= new_top, "summary data problem?");
1333   assert(new_top > bottom, "space is empty; should not be here");
1334   assert(new_top >= cp->destination(), "sanity");
1335   assert(top >= sd.region_to_addr(cp), "sanity");
1336 
1337   HeapWord* const destination = cp->destination();
1338   const size_t dense_prefix_live  = pointer_delta(destination, bottom);
1339   const size_t compacted_region_live = pointer_delta(new_top, destination);
1340   const size_t compacted_region_used = pointer_delta(top,
1341                                                      sd.region_to_addr(cp));
1342   const size_t reclaimable = compacted_region_used - compacted_region_live;
1343 
1344   const double divisor = dense_prefix_live + 1.25 * compacted_region_live;
1345   return double(reclaimable) / divisor;
1346 }
1347 
1348 // Return the address of the end of the dense prefix, a.k.a. the start of the
1349 // compacted region.  The address is always on a region boundary.
1350 //
1351 // Completely full regions at the left are skipped, since no compaction can
1352 // occur in those regions.  Then the maximum amount of dead wood to allow is
1353 // computed, based on the density (amount live / capacity) of the generation;
1354 // the region with approximately that amount of dead space to the left is
1355 // identified as the limit region.  Regions between the last completely full
1356 // region and the limit region are scanned and the one that has the best
1357 // (maximum) reclaimed_ratio() is selected.
1358 HeapWord*
1359 PSParallelCompact::compute_dense_prefix(const SpaceId id,
1360                                         bool maximum_compaction)
1361 {
1362   if (ParallelOldGCSplitALot) {
1363     if (_space_info[id].dense_prefix() != _space_info[id].space()->bottom()) {
1364       // The value was chosen to provoke splitting a young gen space; use it.
1365       return _space_info[id].dense_prefix();
1366     }
1367   }
1368 
1369   const size_t region_size = ParallelCompactData::RegionSize;
1370   const ParallelCompactData& sd = summary_data();
1371 
1372   const MutableSpace* const space = _space_info[id].space();
1373   HeapWord* const top = space->top();
1374   HeapWord* const top_aligned_up = sd.region_align_up(top);
1375   HeapWord* const new_top = _space_info[id].new_top();
1376   HeapWord* const new_top_aligned_up = sd.region_align_up(new_top);
1377   HeapWord* const bottom = space->bottom();
1378   const RegionData* const beg_cp = sd.addr_to_region_ptr(bottom);
1379   const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up);
1380   const RegionData* const new_top_cp =
1381     sd.addr_to_region_ptr(new_top_aligned_up);
1382 
1383   // Skip full regions at the beginning of the space--they are necessarily part
1384   // of the dense prefix.
1385   const RegionData* const full_cp = first_dead_space_region(beg_cp, new_top_cp);
1386   assert(full_cp->destination() == sd.region_to_addr(full_cp) ||
1387          space->is_empty(), "no dead space allowed to the left");
1388   assert(full_cp->data_size() < region_size || full_cp == new_top_cp - 1,
1389          "region must have dead space");
1390 
1391   // The gc number is saved whenever a maximum compaction is done, and used to
1392   // determine when the maximum compaction interval has expired.  This avoids
1393   // successive max compactions for different reasons.
1394   assert(total_invocations() >= _maximum_compaction_gc_num, "sanity");
1395   const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num;
1396   const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval ||
1397     total_invocations() == HeapFirstMaximumCompactionCount;
1398   if (maximum_compaction || full_cp == top_cp || interval_ended) {
1399     _maximum_compaction_gc_num = total_invocations();
1400     return sd.region_to_addr(full_cp);
1401   }
1402 
1403   const size_t space_live = pointer_delta(new_top, bottom);
1404   const size_t space_used = space->used_in_words();
1405   const size_t space_capacity = space->capacity_in_words();
1406 
1407   const double density = double(space_live) / double(space_capacity);
1408   const size_t min_percent_free =
1409           id == perm_space_id ? PermMarkSweepDeadRatio : MarkSweepDeadRatio;
1410   const double limiter = dead_wood_limiter(density, min_percent_free);
1411   const size_t dead_wood_max = space_used - space_live;
1412   const size_t dead_wood_limit = MIN2(size_t(space_capacity * limiter),
1413                                       dead_wood_max);
1414 
1415   if (TraceParallelOldGCDensePrefix) {
1416     tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " "
1417                   "space_cap=" SIZE_FORMAT,
1418                   space_live, space_used,
1419                   space_capacity);
1420     tty->print_cr("dead_wood_limiter(%6.4f, %d)=%6.4f "
1421                   "dead_wood_max=" SIZE_FORMAT " dead_wood_limit=" SIZE_FORMAT,
1422                   density, min_percent_free, limiter,
1423                   dead_wood_max, dead_wood_limit);
1424   }
1425 
1426   // Locate the region with the desired amount of dead space to the left.
1427   const RegionData* const limit_cp =
1428     dead_wood_limit_region(full_cp, top_cp, dead_wood_limit);
1429 
1430   // Scan from the first region with dead space to the limit region and find the
1431   // one with the best (largest) reclaimed ratio.
1432   double best_ratio = 0.0;
1433   const RegionData* best_cp = full_cp;
1434   for (const RegionData* cp = full_cp; cp < limit_cp; ++cp) {
1435     double tmp_ratio = reclaimed_ratio(cp, bottom, top, new_top);
1436     if (tmp_ratio > best_ratio) {
1437       best_cp = cp;
1438       best_ratio = tmp_ratio;
1439     }
1440   }
1441 
1442 #if     0
1443   // Something to consider:  if the region with the best ratio is 'close to' the
1444   // first region w/free space, choose the first region with free space
1445   // ("first-free").  The first-free region is usually near the start of the
1446   // heap, which means we are copying most of the heap already, so copy a bit
1447   // more to get complete compaction.
1448   if (pointer_delta(best_cp, full_cp, sizeof(RegionData)) < 4) {
1449     _maximum_compaction_gc_num = total_invocations();
1450     best_cp = full_cp;
1451   }
1452 #endif  // #if 0
1453 
1454   return sd.region_to_addr(best_cp);
1455 }
1456 
1457 #ifndef PRODUCT
1458 void
1459 PSParallelCompact::fill_with_live_objects(SpaceId id, HeapWord* const start,
1460                                           size_t words)
1461 {
1462   if (TraceParallelOldGCSummaryPhase) {
1463     tty->print_cr("fill_with_live_objects [" PTR_FORMAT " " PTR_FORMAT ") "
1464                   SIZE_FORMAT, start, start + words, words);
1465   }
1466 
1467   ObjectStartArray* const start_array = _space_info[id].start_array();
1468   CollectedHeap::fill_with_objects(start, words);
1469   for (HeapWord* p = start; p < start + words; p += oop(p)->size()) {
1470     _mark_bitmap.mark_obj(p, words);
1471     _summary_data.add_obj(p, words);
1472     start_array->allocate_block(p);
1473   }
1474 }
1475 
1476 void
1477 PSParallelCompact::summarize_new_objects(SpaceId id, HeapWord* start)
1478 {
1479   ParallelCompactData& sd = summary_data();
1480   MutableSpace* space = _space_info[id].space();
1481 
1482   // Find the source and destination start addresses.
1483   HeapWord* const src_addr = sd.region_align_down(start);
1484   HeapWord* dst_addr;
1485   if (src_addr < start) {
1486     dst_addr = sd.addr_to_region_ptr(src_addr)->destination();
1487   } else if (src_addr > space->bottom()) {
1488     // The start (the original top() value) is aligned to a region boundary so
1489     // the associated region does not have a destination.  Compute the
1490     // destination from the previous region.
1491     RegionData* const cp = sd.addr_to_region_ptr(src_addr) - 1;
1492     dst_addr = cp->destination() + cp->data_size();
1493   } else {
1494     // Filling the entire space.
1495     dst_addr = space->bottom();
1496   }
1497   assert(dst_addr != NULL, "sanity");
1498 
1499   // Update the summary data.
1500   bool result = _summary_data.summarize(_space_info[id].split_info(),
1501                                         src_addr, space->top(), NULL,
1502                                         dst_addr, space->end(),
1503                                         _space_info[id].new_top_addr());
1504   assert(result, "should not fail:  bad filler object size");
1505 }
1506 
1507 void
1508 PSParallelCompact::provoke_split_fill_survivor(SpaceId id)
1509 {
1510   if (total_invocations() % (ParallelOldGCSplitInterval * 3) != 0) {
1511     return;
1512   }
1513 
1514   MutableSpace* const space = _space_info[id].space();
1515   if (space->is_empty()) {
1516     HeapWord* b = space->bottom();
1517     HeapWord* t = b + space->capacity_in_words() / 2;
1518     space->set_top(t);
1519     if (ZapUnusedHeapArea) {
1520       space->set_top_for_allocations();
1521     }
1522 
1523     size_t min_size = CollectedHeap::min_fill_size();
1524     size_t obj_len = min_size;
1525     while (b + obj_len <= t) {
1526       CollectedHeap::fill_with_object(b, obj_len);
1527       mark_bitmap()->mark_obj(b, obj_len);
1528       summary_data().add_obj(b, obj_len);
1529       b += obj_len;
1530       obj_len = (obj_len & (min_size*3)) + min_size; // 8 16 24 32 8 16 24 32 ...
1531     }
1532     if (b < t) {
1533       // The loop didn't completely fill to t (top); adjust top downward.
1534       space->set_top(b);
1535       if (ZapUnusedHeapArea) {
1536         space->set_top_for_allocations();
1537       }
1538     }
1539 
1540     HeapWord** nta = _space_info[id].new_top_addr();
1541     bool result = summary_data().summarize(_space_info[id].split_info(),
1542                                            space->bottom(), space->top(), NULL,
1543                                            space->bottom(), space->end(), nta);
1544     assert(result, "space must fit into itself");
1545   }
1546 }
1547 
1548 void
1549 PSParallelCompact::provoke_split(bool & max_compaction)
1550 {
1551   if (total_invocations() % ParallelOldGCSplitInterval != 0) {
1552     return;
1553   }
1554 
1555   const size_t region_size = ParallelCompactData::RegionSize;
1556   ParallelCompactData& sd = summary_data();
1557 
1558   MutableSpace* const eden_space = _space_info[eden_space_id].space();
1559   MutableSpace* const from_space = _space_info[from_space_id].space();
1560   const size_t eden_live = pointer_delta(eden_space->top(),
1561                                          _space_info[eden_space_id].new_top());
1562   const size_t from_live = pointer_delta(from_space->top(),
1563                                          _space_info[from_space_id].new_top());
1564 
1565   const size_t min_fill_size = CollectedHeap::min_fill_size();
1566   const size_t eden_free = pointer_delta(eden_space->end(), eden_space->top());
1567   const size_t eden_fillable = eden_free >= min_fill_size ? eden_free : 0;
1568   const size_t from_free = pointer_delta(from_space->end(), from_space->top());
1569   const size_t from_fillable = from_free >= min_fill_size ? from_free : 0;
1570 
1571   // Choose the space to split; need at least 2 regions live (or fillable).
1572   SpaceId id;
1573   MutableSpace* space;
1574   size_t live_words;
1575   size_t fill_words;
1576   if (eden_live + eden_fillable >= region_size * 2) {
1577     id = eden_space_id;
1578     space = eden_space;
1579     live_words = eden_live;
1580     fill_words = eden_fillable;
1581   } else if (from_live + from_fillable >= region_size * 2) {
1582     id = from_space_id;
1583     space = from_space;
1584     live_words = from_live;
1585     fill_words = from_fillable;
1586   } else {
1587     return; // Give up.
1588   }
1589   assert(fill_words == 0 || fill_words >= min_fill_size, "sanity");
1590 
1591   if (live_words < region_size * 2) {
1592     // Fill from top() to end() w/live objects of mixed sizes.
1593     HeapWord* const fill_start = space->top();
1594     live_words += fill_words;
1595 
1596     space->set_top(fill_start + fill_words);
1597     if (ZapUnusedHeapArea) {
1598       space->set_top_for_allocations();
1599     }
1600 
1601     HeapWord* cur_addr = fill_start;
1602     while (fill_words > 0) {
1603       const size_t r = (size_t)os::random() % (region_size / 2) + min_fill_size;
1604       size_t cur_size = MIN2(align_object_size_(r), fill_words);
1605       if (fill_words - cur_size < min_fill_size) {
1606         cur_size = fill_words; // Avoid leaving a fragment too small to fill.
1607       }
1608 
1609       CollectedHeap::fill_with_object(cur_addr, cur_size);
1610       mark_bitmap()->mark_obj(cur_addr, cur_size);
1611       sd.add_obj(cur_addr, cur_size);
1612 
1613       cur_addr += cur_size;
1614       fill_words -= cur_size;
1615     }
1616 
1617     summarize_new_objects(id, fill_start);
1618   }
1619 
1620   max_compaction = false;
1621 
1622   // Manipulate the old gen so that it has room for about half of the live data
1623   // in the target young gen space (live_words / 2).
1624   id = old_space_id;
1625   space = _space_info[id].space();
1626   const size_t free_at_end = space->free_in_words();
1627   const size_t free_target = align_object_size(live_words / 2);
1628   const size_t dead = pointer_delta(space->top(), _space_info[id].new_top());
1629 
1630   if (free_at_end >= free_target + min_fill_size) {
1631     // Fill space above top() and set the dense prefix so everything survives.
1632     HeapWord* const fill_start = space->top();
1633     const size_t fill_size = free_at_end - free_target;
1634     space->set_top(space->top() + fill_size);
1635     if (ZapUnusedHeapArea) {
1636       space->set_top_for_allocations();
1637     }
1638     fill_with_live_objects(id, fill_start, fill_size);
1639     summarize_new_objects(id, fill_start);
1640     _space_info[id].set_dense_prefix(sd.region_align_down(space->top()));
1641   } else if (dead + free_at_end > free_target) {
1642     // Find a dense prefix that makes the right amount of space available.
1643     HeapWord* cur = sd.region_align_down(space->top());
1644     HeapWord* cur_destination = sd.addr_to_region_ptr(cur)->destination();
1645     size_t dead_to_right = pointer_delta(space->end(), cur_destination);
1646     while (dead_to_right < free_target) {
1647       cur -= region_size;
1648       cur_destination = sd.addr_to_region_ptr(cur)->destination();
1649       dead_to_right = pointer_delta(space->end(), cur_destination);
1650     }
1651     _space_info[id].set_dense_prefix(cur);
1652   }
1653 }
1654 #endif // #ifndef PRODUCT
1655 
1656 void PSParallelCompact::summarize_spaces_quick()
1657 {
1658   for (unsigned int i = 0; i < last_space_id; ++i) {
1659     const MutableSpace* space = _space_info[i].space();
1660     HeapWord** nta = _space_info[i].new_top_addr();
1661     bool result = _summary_data.summarize(_space_info[i].split_info(),
1662                                           space->bottom(), space->top(), NULL,
1663                                           space->bottom(), space->end(), nta);
1664     assert(result, "space must fit into itself");
1665     _space_info[i].set_dense_prefix(space->bottom());
1666   }
1667 
1668 #ifndef PRODUCT
1669   if (ParallelOldGCSplitALot) {
1670     provoke_split_fill_survivor(to_space_id);
1671   }
1672 #endif // #ifndef PRODUCT
1673 }
1674 
1675 void PSParallelCompact::fill_dense_prefix_end(SpaceId id)
1676 {
1677   HeapWord* const dense_prefix_end = dense_prefix(id);
1678   const RegionData* region = _summary_data.addr_to_region_ptr(dense_prefix_end);
1679   const idx_t dense_prefix_bit = _mark_bitmap.addr_to_bit(dense_prefix_end);
1680   if (dead_space_crosses_boundary(region, dense_prefix_bit)) {
1681     // Only enough dead space is filled so that any remaining dead space to the
1682     // left is larger than the minimum filler object.  (The remainder is filled
1683     // during the copy/update phase.)
1684     //
1685     // The size of the dead space to the right of the boundary is not a
1686     // concern, since compaction will be able to use whatever space is
1687     // available.
1688     //
1689     // Here '||' is the boundary, 'x' represents a don't care bit and a box
1690     // surrounds the space to be filled with an object.
1691     //
1692     // In the 32-bit VM, each bit represents two 32-bit words:
1693     //                              +---+
1694     // a) beg_bits:  ...  x   x   x | 0 | ||   0   x  x  ...
1695     //    end_bits:  ...  x   x   x | 0 | ||   0   x  x  ...
1696     //                              +---+
1697     //
1698     // In the 64-bit VM, each bit represents one 64-bit word:
1699     //                              +------------+
1700     // b) beg_bits:  ...  x   x   x | 0   ||   0 | x  x  ...
1701     //    end_bits:  ...  x   x   1 | 0   ||   0 | x  x  ...
1702     //                              +------------+
1703     //                          +-------+
1704     // c) beg_bits:  ...  x   x | 0   0 | ||   0   x  x  ...
1705     //    end_bits:  ...  x   1 | 0   0 | ||   0   x  x  ...
1706     //                          +-------+
1707     //                      +-----------+
1708     // d) beg_bits:  ...  x | 0   0   0 | ||   0   x  x  ...
1709     //    end_bits:  ...  1 | 0   0   0 | ||   0   x  x  ...
1710     //                      +-----------+
1711     //                          +-------+
1712     // e) beg_bits:  ...  0   0 | 0   0 | ||   0   x  x  ...
1713     //    end_bits:  ...  0   0 | 0   0 | ||   0   x  x  ...
1714     //                          +-------+
1715 
1716     // Initially assume case a, c or e will apply.
1717     size_t obj_len = CollectedHeap::min_fill_size();
1718     HeapWord* obj_beg = dense_prefix_end - obj_len;
1719 
1720 #ifdef  _LP64
1721     if (MinObjAlignment > 1) { // object alignment > heap word size
1722       // Cases a, c or e.
1723     } else if (_mark_bitmap.is_obj_end(dense_prefix_bit - 2)) {
1724       // Case b above.
1725       obj_beg = dense_prefix_end - 1;
1726     } else if (!_mark_bitmap.is_obj_end(dense_prefix_bit - 3) &&
1727                _mark_bitmap.is_obj_end(dense_prefix_bit - 4)) {
1728       // Case d above.
1729       obj_beg = dense_prefix_end - 3;
1730       obj_len = 3;
1731     }
1732 #endif  // #ifdef _LP64
1733 
1734     CollectedHeap::fill_with_object(obj_beg, obj_len);
1735     _mark_bitmap.mark_obj(obj_beg, obj_len);
1736     _summary_data.add_obj(obj_beg, obj_len);
1737     assert(start_array(id) != NULL, "sanity");
1738     start_array(id)->allocate_block(obj_beg);
1739   }
1740 }
1741 
1742 void
1743 PSParallelCompact::clear_source_region(HeapWord* beg_addr, HeapWord* end_addr)
1744 {
1745   RegionData* const beg_ptr = _summary_data.addr_to_region_ptr(beg_addr);
1746   HeapWord* const end_aligned_up = _summary_data.region_align_up(end_addr);
1747   RegionData* const end_ptr = _summary_data.addr_to_region_ptr(end_aligned_up);
1748   for (RegionData* cur = beg_ptr; cur < end_ptr; ++cur) {
1749     cur->set_source_region(0);
1750   }
1751 }
1752 
1753 void
1754 PSParallelCompact::summarize_space(SpaceId id, bool maximum_compaction)
1755 {
1756   assert(id < last_space_id, "id out of range");
1757   assert(_space_info[id].dense_prefix() == _space_info[id].space()->bottom() ||
1758          ParallelOldGCSplitALot && id == old_space_id,
1759          "should have been reset in summarize_spaces_quick()");
1760 
1761   const MutableSpace* space = _space_info[id].space();
1762   if (_space_info[id].new_top() != space->bottom()) {
1763     HeapWord* dense_prefix_end = compute_dense_prefix(id, maximum_compaction);
1764     _space_info[id].set_dense_prefix(dense_prefix_end);
1765 
1766 #ifndef PRODUCT
1767     if (TraceParallelOldGCDensePrefix) {
1768       print_dense_prefix_stats("ratio", id, maximum_compaction,
1769                                dense_prefix_end);
1770       HeapWord* addr = compute_dense_prefix_via_density(id, maximum_compaction);
1771       print_dense_prefix_stats("density", id, maximum_compaction, addr);
1772     }
1773 #endif  // #ifndef PRODUCT
1774 
1775     // Recompute the summary data, taking into account the dense prefix.  If
1776     // every last byte will be reclaimed, then the existing summary data which
1777     // compacts everything can be left in place.
1778     if (!maximum_compaction && dense_prefix_end != space->bottom()) {
1779       // If dead space crosses the dense prefix boundary, it is (at least
1780       // partially) filled with a dummy object, marked live and added to the
1781       // summary data.  This simplifies the copy/update phase and must be done
1782       // before the final locations of objects are determined, to prevent
1783       // leaving a fragment of dead space that is too small to fill.
1784       fill_dense_prefix_end(id);
1785 
1786       // Compute the destination of each Region, and thus each object.
1787       _summary_data.summarize_dense_prefix(space->bottom(), dense_prefix_end);
1788       _summary_data.summarize(_space_info[id].split_info(),
1789                               dense_prefix_end, space->top(), NULL,
1790                               dense_prefix_end, space->end(),
1791                               _space_info[id].new_top_addr());
1792     }
1793   }
1794 
1795   if (TraceParallelOldGCSummaryPhase) {
1796     const size_t region_size = ParallelCompactData::RegionSize;
1797     HeapWord* const dense_prefix_end = _space_info[id].dense_prefix();
1798     const size_t dp_region = _summary_data.addr_to_region_idx(dense_prefix_end);
1799     const size_t dp_words = pointer_delta(dense_prefix_end, space->bottom());
1800     HeapWord* const new_top = _space_info[id].new_top();
1801     const HeapWord* nt_aligned_up = _summary_data.region_align_up(new_top);
1802     const size_t cr_words = pointer_delta(nt_aligned_up, dense_prefix_end);
1803     tty->print_cr("id=%d cap=" SIZE_FORMAT " dp=" PTR_FORMAT " "
1804                   "dp_region=" SIZE_FORMAT " " "dp_count=" SIZE_FORMAT " "
1805                   "cr_count=" SIZE_FORMAT " " "nt=" PTR_FORMAT,
1806                   id, space->capacity_in_words(), dense_prefix_end,
1807                   dp_region, dp_words / region_size,
1808                   cr_words / region_size, new_top);
1809   }
1810 }
1811 
1812 #ifndef PRODUCT
1813 void PSParallelCompact::summary_phase_msg(SpaceId dst_space_id,
1814                                           HeapWord* dst_beg, HeapWord* dst_end,
1815                                           SpaceId src_space_id,
1816                                           HeapWord* src_beg, HeapWord* src_end)
1817 {
1818   if (TraceParallelOldGCSummaryPhase) {
1819     tty->print_cr("summarizing %d [%s] into %d [%s]:  "
1820                   "src=" PTR_FORMAT "-" PTR_FORMAT " "
1821                   SIZE_FORMAT "-" SIZE_FORMAT " "
1822                   "dst=" PTR_FORMAT "-" PTR_FORMAT " "
1823                   SIZE_FORMAT "-" SIZE_FORMAT,
1824                   src_space_id, space_names[src_space_id],
1825                   dst_space_id, space_names[dst_space_id],
1826                   src_beg, src_end,
1827                   _summary_data.addr_to_region_idx(src_beg),
1828                   _summary_data.addr_to_region_idx(src_end),
1829                   dst_beg, dst_end,
1830                   _summary_data.addr_to_region_idx(dst_beg),
1831                   _summary_data.addr_to_region_idx(dst_end));
1832   }
1833 }
1834 #endif  // #ifndef PRODUCT
1835 
1836 void PSParallelCompact::summary_phase(ParCompactionManager* cm,
1837                                       bool maximum_compaction)
1838 {
1839   EventMark m("2 summarize");
1840   TraceTime tm("summary phase", print_phases(), true, gclog_or_tty);
1841   // trace("2");
1842 
1843 #ifdef  ASSERT
1844   if (TraceParallelOldGCMarkingPhase) {
1845     tty->print_cr("add_obj_count=" SIZE_FORMAT " "
1846                   "add_obj_bytes=" SIZE_FORMAT,
1847                   add_obj_count, add_obj_size * HeapWordSize);
1848     tty->print_cr("mark_bitmap_count=" SIZE_FORMAT " "
1849                   "mark_bitmap_bytes=" SIZE_FORMAT,
1850                   mark_bitmap_count, mark_bitmap_size * HeapWordSize);
1851   }
1852 #endif  // #ifdef ASSERT
1853 
1854   // Quick summarization of each space into itself, to see how much is live.
1855   summarize_spaces_quick();
1856 
1857   if (TraceParallelOldGCSummaryPhase) {
1858     tty->print_cr("summary_phase:  after summarizing each space to self");
1859     Universe::print();
1860     NOT_PRODUCT(print_region_ranges());
1861     if (Verbose) {
1862       NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info));
1863     }
1864   }
1865 
1866   // The amount of live data that will end up in old space (assuming it fits).
1867   size_t old_space_total_live = 0;
1868   assert(perm_space_id < old_space_id, "should not count perm data here");
1869   for (unsigned int id = old_space_id; id < last_space_id; ++id) {
1870     old_space_total_live += pointer_delta(_space_info[id].new_top(),
1871                                           _space_info[id].space()->bottom());
1872   }
1873 
1874   MutableSpace* const old_space = _space_info[old_space_id].space();
1875   const size_t old_capacity = old_space->capacity_in_words();
1876   if (old_space_total_live > old_capacity) {
1877     // XXX - should also try to expand
1878     maximum_compaction = true;
1879   }
1880 #ifndef PRODUCT
1881   if (ParallelOldGCSplitALot && old_space_total_live < old_capacity) {
1882     provoke_split(maximum_compaction);
1883   }
1884 #endif // #ifndef PRODUCT
1885 
1886   // Permanent and Old generations.
1887   summarize_space(perm_space_id, maximum_compaction);
1888   summarize_space(old_space_id, maximum_compaction);
1889 
1890   // Summarize the remaining spaces in the young gen.  The initial target space
1891   // is the old gen.  If a space does not fit entirely into the target, then the
1892   // remainder is compacted into the space itself and that space becomes the new
1893   // target.
1894   SpaceId dst_space_id = old_space_id;
1895   HeapWord* dst_space_end = old_space->end();
1896   HeapWord** new_top_addr = _space_info[dst_space_id].new_top_addr();
1897   for (unsigned int id = eden_space_id; id < last_space_id; ++id) {
1898     const MutableSpace* space = _space_info[id].space();
1899     const size_t live = pointer_delta(_space_info[id].new_top(),
1900                                       space->bottom());
1901     const size_t available = pointer_delta(dst_space_end, *new_top_addr);
1902 
1903     NOT_PRODUCT(summary_phase_msg(dst_space_id, *new_top_addr, dst_space_end,
1904                                   SpaceId(id), space->bottom(), space->top());)
1905     if (live > 0 && live <= available) {
1906       // All the live data will fit.
1907       bool done = _summary_data.summarize(_space_info[id].split_info(),
1908                                           space->bottom(), space->top(),
1909                                           NULL,
1910                                           *new_top_addr, dst_space_end,
1911                                           new_top_addr);
1912       assert(done, "space must fit into old gen");
1913 
1914       // Reset the new_top value for the space.
1915       _space_info[id].set_new_top(space->bottom());
1916     } else if (live > 0) {
1917       // Attempt to fit part of the source space into the target space.
1918       HeapWord* next_src_addr = NULL;
1919       bool done = _summary_data.summarize(_space_info[id].split_info(),
1920                                           space->bottom(), space->top(),
1921                                           &next_src_addr,
1922                                           *new_top_addr, dst_space_end,
1923                                           new_top_addr);
1924       assert(!done, "space should not fit into old gen");
1925       assert(next_src_addr != NULL, "sanity");
1926 
1927       // The source space becomes the new target, so the remainder is compacted
1928       // within the space itself.
1929       dst_space_id = SpaceId(id);
1930       dst_space_end = space->end();
1931       new_top_addr = _space_info[id].new_top_addr();
1932       NOT_PRODUCT(summary_phase_msg(dst_space_id,
1933                                     space->bottom(), dst_space_end,
1934                                     SpaceId(id), next_src_addr, space->top());)
1935       done = _summary_data.summarize(_space_info[id].split_info(),
1936                                      next_src_addr, space->top(),
1937                                      NULL,
1938                                      space->bottom(), dst_space_end,
1939                                      new_top_addr);
1940       assert(done, "space must fit when compacted into itself");
1941       assert(*new_top_addr <= space->top(), "usage should not grow");
1942     }
1943   }
1944 
1945   if (TraceParallelOldGCSummaryPhase) {
1946     tty->print_cr("summary_phase:  after final summarization");
1947     Universe::print();
1948     NOT_PRODUCT(print_region_ranges());
1949     if (Verbose) {
1950       NOT_PRODUCT(print_generic_summary_data(_summary_data, _space_info));
1951     }
1952   }
1953 }
1954 
1955 // This method should contain all heap-specific policy for invoking a full
1956 // collection.  invoke_no_policy() will only attempt to compact the heap; it
1957 // will do nothing further.  If we need to bail out for policy reasons, scavenge
1958 // before full gc, or any other specialized behavior, it needs to be added here.
1959 //
1960 // Note that this method should only be called from the vm_thread while at a
1961 // safepoint.
1962 //
1963 // Note that the all_soft_refs_clear flag in the collector policy
1964 // may be true because this method can be called without intervening
1965 // activity.  For example when the heap space is tight and full measure
1966 // are being taken to free space.
1967 void PSParallelCompact::invoke(bool maximum_heap_compaction) {
1968   assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
1969   assert(Thread::current() == (Thread*)VMThread::vm_thread(),
1970          "should be in vm thread");
1971 
1972   ParallelScavengeHeap* heap = gc_heap();
1973   GCCause::Cause gc_cause = heap->gc_cause();
1974   assert(!heap->is_gc_active(), "not reentrant");
1975 
1976   PSAdaptiveSizePolicy* policy = heap->size_policy();
1977   IsGCActiveMark mark;
1978 
1979   if (ScavengeBeforeFullGC) {
1980     PSScavenge::invoke_no_policy();
1981   }
1982 
1983   const bool clear_all_soft_refs =
1984     heap->collector_policy()->should_clear_all_soft_refs();
1985 
1986   PSParallelCompact::invoke_no_policy(clear_all_soft_refs ||
1987                                       maximum_heap_compaction);
1988 }
1989 
1990 bool ParallelCompactData::region_contains(size_t region_index, HeapWord* addr) {
1991   size_t addr_region_index = addr_to_region_idx(addr);
1992   return region_index == addr_region_index;
1993 }
1994 
1995 // This method contains no policy. You should probably
1996 // be calling invoke() instead.
1997 void PSParallelCompact::invoke_no_policy(bool maximum_heap_compaction) {
1998   assert(SafepointSynchronize::is_at_safepoint(), "must be at a safepoint");
1999   assert(ref_processor() != NULL, "Sanity");
2000 
2001   if (GC_locker::check_active_before_gc()) {
2002     return;
2003   }
2004 
2005   TimeStamp marking_start;
2006   TimeStamp compaction_start;
2007   TimeStamp collection_exit;
2008 
2009   ParallelScavengeHeap* heap = gc_heap();
2010   GCCause::Cause gc_cause = heap->gc_cause();
2011   PSYoungGen* young_gen = heap->young_gen();
2012   PSOldGen* old_gen = heap->old_gen();
2013   PSPermGen* perm_gen = heap->perm_gen();
2014   PSAdaptiveSizePolicy* size_policy = heap->size_policy();
2015 
2016   // The scope of casr should end after code that can change
2017   // CollectorPolicy::_should_clear_all_soft_refs.
2018   ClearedAllSoftRefs casr(maximum_heap_compaction,
2019                           heap->collector_policy());
2020 
2021   if (ZapUnusedHeapArea) {
2022     // Save information needed to minimize mangling
2023     heap->record_gen_tops_before_GC();
2024   }
2025 
2026   heap->pre_full_gc_dump();
2027 
2028   _print_phases = PrintGCDetails && PrintParallelOldGCPhaseTimes;
2029 
2030   // Make sure data structures are sane, make the heap parsable, and do other
2031   // miscellaneous bookkeeping.
2032   PreGCValues pre_gc_values;
2033   pre_compact(&pre_gc_values);
2034 
2035   // Get the compaction manager reserved for the VM thread.
2036   ParCompactionManager* const vmthread_cm =
2037     ParCompactionManager::manager_array(gc_task_manager()->workers());
2038 
2039   // Place after pre_compact() where the number of invocations is incremented.
2040   AdaptiveSizePolicyOutput(size_policy, heap->total_collections());
2041 
2042   {
2043     ResourceMark rm;
2044     HandleMark hm;
2045 
2046     const bool is_system_gc = gc_cause == GCCause::_java_lang_system_gc;
2047 
2048     // This is useful for debugging but don't change the output the
2049     // the customer sees.
2050     const char* gc_cause_str = "Full GC";
2051     if (is_system_gc && PrintGCDetails) {
2052       gc_cause_str = "Full GC (System)";
2053     }
2054     gclog_or_tty->date_stamp(PrintGC && PrintGCDateStamps);
2055     TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
2056     TraceTime t1(gc_cause_str, PrintGC, !PrintGCDetails, gclog_or_tty);
2057     TraceCollectorStats tcs(counters());
2058     TraceMemoryManagerStats tms(true /* Full GC */);
2059 
2060     if (TraceGen1Time) accumulated_time()->start();
2061 
2062     // Let the size policy know we're starting
2063     size_policy->major_collection_begin();
2064 
2065     // When collecting the permanent generation methodOops may be moving,
2066     // so we either have to flush all bcp data or convert it into bci.
2067     CodeCache::gc_prologue();
2068     Threads::gc_prologue();
2069 
2070     NOT_PRODUCT(ref_processor()->verify_no_references_recorded());
2071     COMPILER2_PRESENT(DerivedPointerTable::clear());
2072 
2073     ref_processor()->enable_discovery();
2074     ref_processor()->setup_policy(maximum_heap_compaction);
2075 
2076     bool marked_for_unloading = false;
2077 
2078     marking_start.update();
2079     marking_phase(vmthread_cm, maximum_heap_compaction);
2080 
2081 #ifndef PRODUCT
2082     if (TraceParallelOldGCMarkingPhase) {
2083       gclog_or_tty->print_cr("marking_phase: cas_tries %d  cas_retries %d "
2084         "cas_by_another %d",
2085         mark_bitmap()->cas_tries(), mark_bitmap()->cas_retries(),
2086         mark_bitmap()->cas_by_another());
2087     }
2088 #endif  // #ifndef PRODUCT
2089 
2090     bool max_on_system_gc = UseMaximumCompactionOnSystemGC && is_system_gc;
2091     summary_phase(vmthread_cm, maximum_heap_compaction || max_on_system_gc);
2092 
2093     COMPILER2_PRESENT(assert(DerivedPointerTable::is_active(), "Sanity"));
2094     COMPILER2_PRESENT(DerivedPointerTable::set_active(false));
2095 
2096     // adjust_roots() updates Universe::_intArrayKlassObj which is
2097     // needed by the compaction for filling holes in the dense prefix.
2098     adjust_roots();
2099 
2100     compaction_start.update();
2101     // Does the perm gen always have to be done serially because
2102     // klasses are used in the update of an object?
2103     compact_perm(vmthread_cm);
2104 
2105     if (UseParallelOldGCCompacting) {
2106       compact();
2107     } else {
2108       compact_serial(vmthread_cm);
2109     }
2110 
2111     // Reset the mark bitmap, summary data, and do other bookkeeping.  Must be
2112     // done before resizing.
2113     post_compact();
2114 
2115     // Let the size policy know we're done
2116     size_policy->major_collection_end(old_gen->used_in_bytes(), gc_cause);
2117 
2118     if (UseAdaptiveSizePolicy) {
2119       if (PrintAdaptiveSizePolicy) {
2120         gclog_or_tty->print("AdaptiveSizeStart: ");
2121         gclog_or_tty->stamp();
2122         gclog_or_tty->print_cr(" collection: %d ",
2123                        heap->total_collections());
2124         if (Verbose) {
2125           gclog_or_tty->print("old_gen_capacity: %d young_gen_capacity: %d"
2126             " perm_gen_capacity: %d ",
2127             old_gen->capacity_in_bytes(), young_gen->capacity_in_bytes(),
2128             perm_gen->capacity_in_bytes());
2129         }
2130       }
2131 
2132       // Don't check if the size_policy is ready here.  Let
2133       // the size_policy check that internally.
2134       if (UseAdaptiveGenerationSizePolicyAtMajorCollection &&
2135           ((gc_cause != GCCause::_java_lang_system_gc) ||
2136             UseAdaptiveSizePolicyWithSystemGC)) {
2137         // Calculate optimal free space amounts
2138         assert(young_gen->max_size() >
2139           young_gen->from_space()->capacity_in_bytes() +
2140           young_gen->to_space()->capacity_in_bytes(),
2141           "Sizes of space in young gen are out-of-bounds");
2142         size_t max_eden_size = young_gen->max_size() -
2143           young_gen->from_space()->capacity_in_bytes() -
2144           young_gen->to_space()->capacity_in_bytes();
2145         size_policy->compute_generation_free_space(
2146                               young_gen->used_in_bytes(),
2147                               young_gen->eden_space()->used_in_bytes(),
2148                               old_gen->used_in_bytes(),
2149                               perm_gen->used_in_bytes(),
2150                               young_gen->eden_space()->capacity_in_bytes(),
2151                               old_gen->max_gen_size(),
2152                               max_eden_size,
2153                               true /* full gc*/,
2154                               gc_cause,
2155                               heap->collector_policy());
2156 
2157         heap->resize_old_gen(
2158           size_policy->calculated_old_free_size_in_bytes());
2159 
2160         // Don't resize the young generation at an major collection.  A
2161         // desired young generation size may have been calculated but
2162         // resizing the young generation complicates the code because the
2163         // resizing of the old generation may have moved the boundary
2164         // between the young generation and the old generation.  Let the
2165         // young generation resizing happen at the minor collections.
2166       }
2167       if (PrintAdaptiveSizePolicy) {
2168         gclog_or_tty->print_cr("AdaptiveSizeStop: collection: %d ",
2169                        heap->total_collections());
2170       }
2171     }
2172 
2173     if (UsePerfData) {
2174       PSGCAdaptivePolicyCounters* const counters = heap->gc_policy_counters();
2175       counters->update_counters();
2176       counters->update_old_capacity(old_gen->capacity_in_bytes());
2177       counters->update_young_capacity(young_gen->capacity_in_bytes());
2178     }
2179 
2180     heap->resize_all_tlabs();
2181 
2182     // We collected the perm gen, so we'll resize it here.
2183     perm_gen->compute_new_size(pre_gc_values.perm_gen_used());
2184 
2185     if (TraceGen1Time) accumulated_time()->stop();
2186 
2187     if (PrintGC) {
2188       if (PrintGCDetails) {
2189         // No GC timestamp here.  This is after GC so it would be confusing.
2190         young_gen->print_used_change(pre_gc_values.young_gen_used());
2191         old_gen->print_used_change(pre_gc_values.old_gen_used());
2192         heap->print_heap_change(pre_gc_values.heap_used());
2193         // Print perm gen last (print_heap_change() excludes the perm gen).
2194         perm_gen->print_used_change(pre_gc_values.perm_gen_used());
2195       } else {
2196         heap->print_heap_change(pre_gc_values.heap_used());
2197       }
2198     }
2199 
2200     // Track memory usage and detect low memory
2201     MemoryService::track_memory_usage();
2202     heap->update_counters();
2203   }
2204 
2205 #ifdef ASSERT
2206   for (size_t i = 0; i < ParallelGCThreads + 1; ++i) {
2207     ParCompactionManager* const cm =
2208       ParCompactionManager::manager_array(int(i));
2209     assert(cm->marking_stack()->is_empty(),       "should be empty");
2210     assert(cm->region_stack()->is_empty(),        "should be empty");
2211     assert(cm->revisit_klass_stack()->is_empty(), "should be empty");
2212   }
2213 #endif // ASSERT
2214 
2215   if (VerifyAfterGC && heap->total_collections() >= VerifyGCStartAt) {
2216     HandleMark hm;  // Discard invalid handles created during verification
2217     gclog_or_tty->print(" VerifyAfterGC:");
2218     Universe::verify(false);
2219   }
2220 
2221   // Re-verify object start arrays
2222   if (VerifyObjectStartArray &&
2223       VerifyAfterGC) {
2224     old_gen->verify_object_start_array();
2225     perm_gen->verify_object_start_array();
2226   }
2227 
2228   if (ZapUnusedHeapArea) {
2229     old_gen->object_space()->check_mangled_unused_area_complete();
2230     perm_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   if (PrintHeapAtGC) {
2238     Universe::print_heap_after_gc();
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();
2249 
2250 #ifdef TRACESPINNING
2251   ParallelTaskTerminator::print_termination_counts();
2252 #endif
2253 }
2254 
2255 bool PSParallelCompact::absorb_live_data_from_eden(PSAdaptiveSizePolicy* size_policy,
2256                                              PSYoungGen* young_gen,
2257                                              PSOldGen* old_gen) {
2258   MutableSpace* const eden_space = young_gen->eden_space();
2259   assert(!eden_space->is_empty(), "eden must be non-empty");
2260   assert(young_gen->virtual_space()->alignment() ==
2261          old_gen->virtual_space()->alignment(), "alignments do not match");
2262 
2263   if (!(UseAdaptiveSizePolicy && UseAdaptiveGCBoundary)) {
2264     return false;
2265   }
2266 
2267   // Both generations must be completely committed.
2268   if (young_gen->virtual_space()->uncommitted_size() != 0) {
2269     return false;
2270   }
2271   if (old_gen->virtual_space()->uncommitted_size() != 0) {
2272     return false;
2273   }
2274 
2275   // Figure out how much to take from eden.  Include the average amount promoted
2276   // in the total; otherwise the next young gen GC will simply bail out to a
2277   // full GC.
2278   const size_t alignment = old_gen->virtual_space()->alignment();
2279   const size_t eden_used = eden_space->used_in_bytes();
2280   const size_t promoted = (size_t)size_policy->avg_promoted()->padded_average();
2281   const size_t absorb_size = align_size_up(eden_used + promoted, alignment);
2282   const size_t eden_capacity = eden_space->capacity_in_bytes();
2283 
2284   if (absorb_size >= eden_capacity) {
2285     return false; // Must leave some space in eden.
2286   }
2287 
2288   const size_t new_young_size = young_gen->capacity_in_bytes() - absorb_size;
2289   if (new_young_size < young_gen->min_gen_size()) {
2290     return false; // Respect young gen minimum size.
2291   }
2292 
2293   if (TraceAdaptiveGCBoundary && Verbose) {
2294     gclog_or_tty->print(" absorbing " SIZE_FORMAT "K:  "
2295                         "eden " SIZE_FORMAT "K->" SIZE_FORMAT "K "
2296                         "from " SIZE_FORMAT "K, to " SIZE_FORMAT "K "
2297                         "young_gen " SIZE_FORMAT "K->" SIZE_FORMAT "K ",
2298                         absorb_size / K,
2299                         eden_capacity / K, (eden_capacity - absorb_size) / K,
2300                         young_gen->from_space()->used_in_bytes() / K,
2301                         young_gen->to_space()->used_in_bytes() / K,
2302                         young_gen->capacity_in_bytes() / K, new_young_size / K);
2303   }
2304 
2305   // Fill the unused part of the old gen.
2306   MutableSpace* const old_space = old_gen->object_space();
2307   HeapWord* const unused_start = old_space->top();
2308   size_t const unused_words = pointer_delta(old_space->end(), unused_start);
2309 
2310   if (unused_words > 0) {
2311     if (unused_words < CollectedHeap::min_fill_size()) {
2312       return false;  // If the old gen cannot be filled, must give up.
2313     }
2314     CollectedHeap::fill_with_objects(unused_start, unused_words);
2315   }
2316 
2317   // Take the live data from eden and set both top and end in the old gen to
2318   // eden top.  (Need to set end because reset_after_change() mangles the region
2319   // from end to virtual_space->high() in debug builds).
2320   HeapWord* const new_top = eden_space->top();
2321   old_gen->virtual_space()->expand_into(young_gen->virtual_space(),
2322                                         absorb_size);
2323   young_gen->reset_after_change();
2324   old_space->set_top(new_top);
2325   old_space->set_end(new_top);
2326   old_gen->reset_after_change();
2327 
2328   // Update the object start array for the filler object and the data from eden.
2329   ObjectStartArray* const start_array = old_gen->start_array();
2330   for (HeapWord* p = unused_start; p < new_top; p += oop(p)->size()) {
2331     start_array->allocate_block(p);
2332   }
2333 
2334   // Could update the promoted average here, but it is not typically updated at
2335   // full GCs and the value to use is unclear.  Something like
2336   //
2337   // cur_promoted_avg + absorb_size / number_of_scavenges_since_last_full_gc.
2338 
2339   size_policy->set_bytes_absorbed_from_eden(absorb_size);
2340   return true;
2341 }
2342 
2343 GCTaskManager* const PSParallelCompact::gc_task_manager() {
2344   assert(ParallelScavengeHeap::gc_task_manager() != NULL,
2345     "shouldn't return NULL");
2346   return ParallelScavengeHeap::gc_task_manager();
2347 }
2348 
2349 void PSParallelCompact::marking_phase(ParCompactionManager* cm,
2350                                       bool maximum_heap_compaction) {
2351   // Recursively traverse all live objects and mark them
2352   EventMark m("1 mark object");
2353   TraceTime tm("marking phase", print_phases(), true, gclog_or_tty);
2354 
2355   ParallelScavengeHeap* heap = gc_heap();
2356   uint parallel_gc_threads = heap->gc_task_manager()->workers();
2357   TaskQueueSetSuper* qset = ParCompactionManager::region_array();
2358   ParallelTaskTerminator terminator(parallel_gc_threads, qset);
2359 
2360   PSParallelCompact::MarkAndPushClosure mark_and_push_closure(cm);
2361   PSParallelCompact::FollowStackClosure follow_stack_closure(cm);
2362 
2363   {
2364     TraceTime tm_m("par mark", print_phases(), true, gclog_or_tty);
2365     ParallelScavengeHeap::ParStrongRootsScope psrs;
2366 
2367     GCTaskQueue* q = GCTaskQueue::create();
2368 
2369     q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::universe));
2370     q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jni_handles));
2371     // We scan the thread roots in parallel
2372     Threads::create_thread_roots_marking_tasks(q);
2373     q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::object_synchronizer));
2374     q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::flat_profiler));
2375     q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::management));
2376     q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::system_dictionary));
2377     q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jvmti));
2378     q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::code_cache));
2379 
2380     if (parallel_gc_threads > 1) {
2381       for (uint j = 0; j < parallel_gc_threads; j++) {
2382         q->enqueue(new StealMarkingTask(&terminator));
2383       }
2384     }
2385 
2386     WaitForBarrierGCTask* fin = WaitForBarrierGCTask::create();
2387     q->enqueue(fin);
2388 
2389     gc_task_manager()->add_list(q);
2390 
2391     fin->wait_for();
2392 
2393     // We have to release the barrier tasks!
2394     WaitForBarrierGCTask::destroy(fin);
2395   }
2396 
2397   // Process reference objects found during marking
2398   {
2399     TraceTime tm_r("reference processing", print_phases(), true, gclog_or_tty);
2400     if (ref_processor()->processing_is_mt()) {
2401       RefProcTaskExecutor task_executor;
2402       ref_processor()->process_discovered_references(
2403         is_alive_closure(), &mark_and_push_closure, &follow_stack_closure,
2404         &task_executor);
2405     } else {
2406       ref_processor()->process_discovered_references(
2407         is_alive_closure(), &mark_and_push_closure, &follow_stack_closure, NULL);
2408     }
2409   }
2410 
2411   TraceTime tm_c("class unloading", print_phases(), true, gclog_or_tty);
2412   // Follow system dictionary roots and unload classes.
2413   bool purged_class = SystemDictionary::do_unloading(is_alive_closure());
2414 
2415   // Follow code cache roots.
2416   CodeCache::do_unloading(is_alive_closure(), &mark_and_push_closure,
2417                           purged_class);
2418   cm->follow_marking_stacks(); // Flush marking stack.
2419 
2420   // Update subklass/sibling/implementor links of live klasses
2421   // revisit_klass_stack is used in follow_weak_klass_links().
2422   follow_weak_klass_links();
2423 
2424   // Revisit memoized MDO's and clear any unmarked weak refs
2425   follow_mdo_weak_refs();
2426 
2427   // Visit interned string tables and delete unmarked oops
2428   StringTable::unlink(is_alive_closure());
2429   // Clean up unreferenced symbols in symbol table.
2430   SymbolTable::unlink();
2431 
2432   assert(cm->marking_stacks_empty(), "marking stacks should be empty");
2433 }
2434 
2435 // This should be moved to the shared markSweep code!
2436 class PSAlwaysTrueClosure: public BoolObjectClosure {
2437 public:
2438   void do_object(oop p) { ShouldNotReachHere(); }
2439   bool do_object_b(oop p) { return true; }
2440 };
2441 static PSAlwaysTrueClosure always_true;
2442 
2443 void PSParallelCompact::adjust_roots() {
2444   // Adjust the pointers to reflect the new locations
2445   EventMark m("3 adjust roots");
2446   TraceTime tm("adjust roots", print_phases(), true, gclog_or_tty);
2447 
2448   // General strong roots.
2449   Universe::oops_do(adjust_root_pointer_closure());
2450   ReferenceProcessor::oops_do(adjust_root_pointer_closure());
2451   JNIHandles::oops_do(adjust_root_pointer_closure());   // Global (strong) JNI handles
2452   Threads::oops_do(adjust_root_pointer_closure(), NULL);
2453   ObjectSynchronizer::oops_do(adjust_root_pointer_closure());
2454   FlatProfiler::oops_do(adjust_root_pointer_closure());
2455   Management::oops_do(adjust_root_pointer_closure());
2456   JvmtiExport::oops_do(adjust_root_pointer_closure());
2457   // SO_AllClasses
2458   SystemDictionary::oops_do(adjust_root_pointer_closure());
2459 
2460   // Now adjust pointers in remaining weak roots.  (All of which should
2461   // have been cleared if they pointed to non-surviving objects.)
2462   // Global (weak) JNI handles
2463   JNIHandles::weak_oops_do(&always_true, adjust_root_pointer_closure());
2464 
2465   CodeCache::oops_do(adjust_pointer_closure());
2466   StringTable::oops_do(adjust_root_pointer_closure());
2467   ref_processor()->weak_oops_do(adjust_root_pointer_closure());
2468   // Roots were visited so references into the young gen in roots
2469   // may have been scanned.  Process them also.
2470   // Should the reference processor have a span that excludes
2471   // young gen objects?
2472   PSScavenge::reference_processor()->weak_oops_do(
2473                                               adjust_root_pointer_closure());
2474 }
2475 
2476 void PSParallelCompact::compact_perm(ParCompactionManager* cm) {
2477   EventMark m("4 compact perm");
2478   TraceTime tm("compact perm gen", print_phases(), true, gclog_or_tty);
2479   // trace("4");
2480 
2481   gc_heap()->perm_gen()->start_array()->reset();
2482   move_and_update(cm, perm_space_id);
2483 }
2484 
2485 void PSParallelCompact::enqueue_region_draining_tasks(GCTaskQueue* q,
2486                                                       uint parallel_gc_threads)
2487 {
2488   TraceTime tm("drain task setup", print_phases(), true, gclog_or_tty);
2489 
2490   const unsigned int task_count = MAX2(parallel_gc_threads, 1U);
2491   for (unsigned int j = 0; j < task_count; j++) {
2492     q->enqueue(new DrainStacksCompactionTask(j));
2493   }
2494 
2495   // Find all regions that are available (can be filled immediately) and
2496   // distribute them to the thread stacks.  The iteration is done in reverse
2497   // order (high to low) so the regions will be removed in ascending order.
2498 
2499   const ParallelCompactData& sd = PSParallelCompact::summary_data();
2500 
2501   size_t fillable_regions = 0;   // A count for diagnostic purposes.
2502   unsigned int which = 0;       // The worker thread number.
2503 
2504   for (unsigned int id = to_space_id; id > perm_space_id; --id) {
2505     SpaceInfo* const space_info = _space_info + id;
2506     MutableSpace* const space = space_info->space();
2507     HeapWord* const new_top = space_info->new_top();
2508 
2509     const size_t beg_region = sd.addr_to_region_idx(space_info->dense_prefix());
2510     const size_t end_region =
2511       sd.addr_to_region_idx(sd.region_align_up(new_top));
2512     assert(end_region > 0, "perm gen cannot be empty");
2513 
2514     for (size_t cur = end_region - 1; cur >= beg_region; --cur) {
2515       if (sd.region(cur)->claim_unsafe()) {
2516         ParCompactionManager* cm = ParCompactionManager::manager_array(which);
2517         cm->push_region(cur);
2518 
2519         if (TraceParallelOldGCCompactionPhase && Verbose) {
2520           const size_t count_mod_8 = fillable_regions & 7;
2521           if (count_mod_8 == 0) gclog_or_tty->print("fillable: ");
2522           gclog_or_tty->print(" " SIZE_FORMAT_W(7), cur);
2523           if (count_mod_8 == 7) gclog_or_tty->cr();
2524         }
2525 
2526         NOT_PRODUCT(++fillable_regions;)
2527 
2528         // Assign regions to threads in round-robin fashion.
2529         if (++which == task_count) {
2530           which = 0;
2531         }
2532       }
2533     }
2534   }
2535 
2536   if (TraceParallelOldGCCompactionPhase) {
2537     if (Verbose && (fillable_regions & 7) != 0) gclog_or_tty->cr();
2538     gclog_or_tty->print_cr("%u initially fillable regions", fillable_regions);
2539   }
2540 }
2541 
2542 #define PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING 4
2543 
2544 void PSParallelCompact::enqueue_dense_prefix_tasks(GCTaskQueue* q,
2545                                                     uint parallel_gc_threads) {
2546   TraceTime tm("dense prefix task setup", print_phases(), true, gclog_or_tty);
2547 
2548   ParallelCompactData& sd = PSParallelCompact::summary_data();
2549 
2550   // Iterate over all the spaces adding tasks for updating
2551   // regions in the dense prefix.  Assume that 1 gc thread
2552   // will work on opening the gaps and the remaining gc threads
2553   // will work on the dense prefix.
2554   unsigned int space_id;
2555   for (space_id = old_space_id; space_id < last_space_id; ++ space_id) {
2556     HeapWord* const dense_prefix_end = _space_info[space_id].dense_prefix();
2557     const MutableSpace* const space = _space_info[space_id].space();
2558 
2559     if (dense_prefix_end == space->bottom()) {
2560       // There is no dense prefix for this space.
2561       continue;
2562     }
2563 
2564     // The dense prefix is before this region.
2565     size_t region_index_end_dense_prefix =
2566         sd.addr_to_region_idx(dense_prefix_end);
2567     RegionData* const dense_prefix_cp =
2568       sd.region(region_index_end_dense_prefix);
2569     assert(dense_prefix_end == space->end() ||
2570            dense_prefix_cp->available() ||
2571            dense_prefix_cp->claimed(),
2572            "The region after the dense prefix should always be ready to fill");
2573 
2574     size_t region_index_start = sd.addr_to_region_idx(space->bottom());
2575 
2576     // Is there dense prefix work?
2577     size_t total_dense_prefix_regions =
2578       region_index_end_dense_prefix - region_index_start;
2579     // How many regions of the dense prefix should be given to
2580     // each thread?
2581     if (total_dense_prefix_regions > 0) {
2582       uint tasks_for_dense_prefix = 1;
2583       if (UseParallelDensePrefixUpdate) {
2584         if (total_dense_prefix_regions <=
2585             (parallel_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING)) {
2586           // Don't over partition.  This assumes that
2587           // PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING is a small integer value
2588           // so there are not many regions to process.
2589           tasks_for_dense_prefix = parallel_gc_threads;
2590         } else {
2591           // Over partition
2592           tasks_for_dense_prefix = parallel_gc_threads *
2593             PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING;
2594         }
2595       }
2596       size_t regions_per_thread = total_dense_prefix_regions /
2597         tasks_for_dense_prefix;
2598       // Give each thread at least 1 region.
2599       if (regions_per_thread == 0) {
2600         regions_per_thread = 1;
2601       }
2602 
2603       for (uint k = 0; k < tasks_for_dense_prefix; k++) {
2604         if (region_index_start >= region_index_end_dense_prefix) {
2605           break;
2606         }
2607         // region_index_end is not processed
2608         size_t region_index_end = MIN2(region_index_start + regions_per_thread,
2609                                        region_index_end_dense_prefix);
2610         q->enqueue(new UpdateDensePrefixTask(SpaceId(space_id),
2611                                              region_index_start,
2612                                              region_index_end));
2613         region_index_start = region_index_end;
2614       }
2615     }
2616     // This gets any part of the dense prefix that did not
2617     // fit evenly.
2618     if (region_index_start < region_index_end_dense_prefix) {
2619       q->enqueue(new UpdateDensePrefixTask(SpaceId(space_id),
2620                                            region_index_start,
2621                                            region_index_end_dense_prefix));
2622     }
2623   }
2624 }
2625 
2626 void PSParallelCompact::enqueue_region_stealing_tasks(
2627                                      GCTaskQueue* q,
2628                                      ParallelTaskTerminator* terminator_ptr,
2629                                      uint parallel_gc_threads) {
2630   TraceTime tm("steal task setup", print_phases(), true, gclog_or_tty);
2631 
2632   // Once a thread has drained it's stack, it should try to steal regions from
2633   // other threads.
2634   if (parallel_gc_threads > 1) {
2635     for (uint j = 0; j < parallel_gc_threads; j++) {
2636       q->enqueue(new StealRegionCompactionTask(terminator_ptr));
2637     }
2638   }
2639 }
2640 
2641 void PSParallelCompact::compact() {
2642   EventMark m("5 compact");
2643   // trace("5");
2644   TraceTime tm("compaction phase", print_phases(), true, gclog_or_tty);
2645 
2646   ParallelScavengeHeap* heap = (ParallelScavengeHeap*)Universe::heap();
2647   assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity");
2648   PSOldGen* old_gen = heap->old_gen();
2649   old_gen->start_array()->reset();
2650   uint parallel_gc_threads = heap->gc_task_manager()->workers();
2651   TaskQueueSetSuper* qset = ParCompactionManager::region_array();
2652   ParallelTaskTerminator terminator(parallel_gc_threads, qset);
2653 
2654   GCTaskQueue* q = GCTaskQueue::create();
2655   enqueue_region_draining_tasks(q, parallel_gc_threads);
2656   enqueue_dense_prefix_tasks(q, parallel_gc_threads);
2657   enqueue_region_stealing_tasks(q, &terminator, parallel_gc_threads);
2658 
2659   {
2660     TraceTime tm_pc("par compact", print_phases(), true, gclog_or_tty);
2661 
2662     WaitForBarrierGCTask* fin = WaitForBarrierGCTask::create();
2663     q->enqueue(fin);
2664 
2665     gc_task_manager()->add_list(q);
2666 
2667     fin->wait_for();
2668 
2669     // We have to release the barrier tasks!
2670     WaitForBarrierGCTask::destroy(fin);
2671 
2672 #ifdef  ASSERT
2673     // Verify that all regions have been processed before the deferred updates.
2674     // Note that perm_space_id is skipped; this type of verification is not
2675     // valid until the perm gen is compacted by regions.
2676     for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2677       verify_complete(SpaceId(id));
2678     }
2679 #endif
2680   }
2681 
2682   {
2683     // Update the deferred objects, if any.  Any compaction manager can be used.
2684     TraceTime tm_du("deferred updates", print_phases(), true, gclog_or_tty);
2685     ParCompactionManager* cm = ParCompactionManager::manager_array(0);
2686     for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2687       update_deferred_objects(cm, SpaceId(id));
2688     }
2689   }
2690 }
2691 
2692 #ifdef  ASSERT
2693 void PSParallelCompact::verify_complete(SpaceId space_id) {
2694   // All Regions between space bottom() to new_top() should be marked as filled
2695   // and all Regions between new_top() and top() should be available (i.e.,
2696   // should have been emptied).
2697   ParallelCompactData& sd = summary_data();
2698   SpaceInfo si = _space_info[space_id];
2699   HeapWord* new_top_addr = sd.region_align_up(si.new_top());
2700   HeapWord* old_top_addr = sd.region_align_up(si.space()->top());
2701   const size_t beg_region = sd.addr_to_region_idx(si.space()->bottom());
2702   const size_t new_top_region = sd.addr_to_region_idx(new_top_addr);
2703   const size_t old_top_region = sd.addr_to_region_idx(old_top_addr);
2704 
2705   bool issued_a_warning = false;
2706 
2707   size_t cur_region;
2708   for (cur_region = beg_region; cur_region < new_top_region; ++cur_region) {
2709     const RegionData* const c = sd.region(cur_region);
2710     if (!c->completed()) {
2711       warning("region " SIZE_FORMAT " not filled:  "
2712               "destination_count=" SIZE_FORMAT,
2713               cur_region, c->destination_count());
2714       issued_a_warning = true;
2715     }
2716   }
2717 
2718   for (cur_region = new_top_region; cur_region < old_top_region; ++cur_region) {
2719     const RegionData* const c = sd.region(cur_region);
2720     if (!c->available()) {
2721       warning("region " SIZE_FORMAT " not empty:   "
2722               "destination_count=" SIZE_FORMAT,
2723               cur_region, c->destination_count());
2724       issued_a_warning = true;
2725     }
2726   }
2727 
2728   if (issued_a_warning) {
2729     print_region_ranges();
2730   }
2731 }
2732 #endif  // #ifdef ASSERT
2733 
2734 void PSParallelCompact::compact_serial(ParCompactionManager* cm) {
2735   EventMark m("5 compact serial");
2736   TraceTime tm("compact serial", print_phases(), true, gclog_or_tty);
2737 
2738   ParallelScavengeHeap* heap = (ParallelScavengeHeap*)Universe::heap();
2739   assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity");
2740 
2741   PSYoungGen* young_gen = heap->young_gen();
2742   PSOldGen* old_gen = heap->old_gen();
2743 
2744   old_gen->start_array()->reset();
2745   old_gen->move_and_update(cm);
2746   young_gen->move_and_update(cm);
2747 }
2748 
2749 void
2750 PSParallelCompact::follow_weak_klass_links() {
2751   // All klasses on the revisit stack are marked at this point.
2752   // Update and follow all subklass, sibling and implementor links.
2753   if (PrintRevisitStats) {
2754     gclog_or_tty->print_cr("#classes in system dictionary = %d",
2755                            SystemDictionary::number_of_classes());
2756   }
2757   for (uint i = 0; i < ParallelGCThreads + 1; i++) {
2758     ParCompactionManager* cm = ParCompactionManager::manager_array(i);
2759     KeepAliveClosure keep_alive_closure(cm);
2760     Stack<Klass*>* const rks = cm->revisit_klass_stack();
2761     if (PrintRevisitStats) {
2762       gclog_or_tty->print_cr("Revisit klass stack[%u] length = " SIZE_FORMAT,
2763                              i, rks->size());
2764     }
2765     while (!rks->is_empty()) {
2766       Klass* const k = rks->pop();
2767       k->follow_weak_klass_links(is_alive_closure(), &keep_alive_closure);
2768     }
2769 
2770     cm->follow_marking_stacks();
2771   }
2772 }
2773 
2774 void
2775 PSParallelCompact::revisit_weak_klass_link(ParCompactionManager* cm, Klass* k) {
2776   cm->revisit_klass_stack()->push(k);
2777 }
2778 
2779 void PSParallelCompact::revisit_mdo(ParCompactionManager* cm, DataLayout* p) {
2780   cm->revisit_mdo_stack()->push(p);
2781 }
2782 
2783 void PSParallelCompact::follow_mdo_weak_refs() {
2784   // All strongly reachable oops have been marked at this point;
2785   // we can visit and clear any weak references from MDO's which
2786   // we memoized during the strong marking phase.
2787   if (PrintRevisitStats) {
2788     gclog_or_tty->print_cr("#classes in system dictionary = %d",
2789                            SystemDictionary::number_of_classes());
2790   }
2791   for (uint i = 0; i < ParallelGCThreads + 1; i++) {
2792     ParCompactionManager* cm = ParCompactionManager::manager_array(i);
2793     Stack<DataLayout*>* rms = cm->revisit_mdo_stack();
2794     if (PrintRevisitStats) {
2795       gclog_or_tty->print_cr("Revisit MDO stack[%u] size = " SIZE_FORMAT,
2796                              i, rms->size());
2797     }
2798     while (!rms->is_empty()) {
2799       rms->pop()->follow_weak_refs(is_alive_closure());
2800     }
2801 
2802     cm->follow_marking_stacks();
2803   }
2804 }
2805 
2806 
2807 #ifdef VALIDATE_MARK_SWEEP
2808 
2809 void PSParallelCompact::track_adjusted_pointer(void* p, bool isroot) {
2810   if (!ValidateMarkSweep)
2811     return;
2812 
2813   if (!isroot) {
2814     if (_pointer_tracking) {
2815       guarantee(_adjusted_pointers->contains(p), "should have seen this pointer");
2816       _adjusted_pointers->remove(p);
2817     }
2818   } else {
2819     ptrdiff_t index = _root_refs_stack->find(p);
2820     if (index != -1) {
2821       int l = _root_refs_stack->length();
2822       if (l > 0 && l - 1 != index) {
2823         void* last = _root_refs_stack->pop();
2824         assert(last != p, "should be different");
2825         _root_refs_stack->at_put(index, last);
2826       } else {
2827         _root_refs_stack->remove(p);
2828       }
2829     }
2830   }
2831 }
2832 
2833 
2834 void PSParallelCompact::check_adjust_pointer(void* p) {
2835   _adjusted_pointers->push(p);
2836 }
2837 
2838 
2839 class AdjusterTracker: public OopClosure {
2840  public:
2841   AdjusterTracker() {};
2842   void do_oop(oop* o)         { PSParallelCompact::check_adjust_pointer(o); }
2843   void do_oop(narrowOop* o)   { PSParallelCompact::check_adjust_pointer(o); }
2844 };
2845 
2846 
2847 void PSParallelCompact::track_interior_pointers(oop obj) {
2848   if (ValidateMarkSweep) {
2849     _adjusted_pointers->clear();
2850     _pointer_tracking = true;
2851 
2852     AdjusterTracker checker;
2853     obj->oop_iterate(&checker);
2854   }
2855 }
2856 
2857 
2858 void PSParallelCompact::check_interior_pointers() {
2859   if (ValidateMarkSweep) {
2860     _pointer_tracking = false;
2861     guarantee(_adjusted_pointers->length() == 0, "should have processed the same pointers");
2862   }
2863 }
2864 
2865 
2866 void PSParallelCompact::reset_live_oop_tracking(bool at_perm) {
2867   if (ValidateMarkSweep) {
2868     guarantee((size_t)_live_oops->length() == _live_oops_index, "should be at end of live oops");
2869     _live_oops_index = at_perm ? _live_oops_index_at_perm : 0;
2870   }
2871 }
2872 
2873 
2874 void PSParallelCompact::register_live_oop(oop p, size_t size) {
2875   if (ValidateMarkSweep) {
2876     _live_oops->push(p);
2877     _live_oops_size->push(size);
2878     _live_oops_index++;
2879   }
2880 }
2881 
2882 void PSParallelCompact::validate_live_oop(oop p, size_t size) {
2883   if (ValidateMarkSweep) {
2884     oop obj = _live_oops->at((int)_live_oops_index);
2885     guarantee(obj == p, "should be the same object");
2886     guarantee(_live_oops_size->at((int)_live_oops_index) == size, "should be the same size");
2887     _live_oops_index++;
2888   }
2889 }
2890 
2891 void PSParallelCompact::live_oop_moved_to(HeapWord* q, size_t size,
2892                                   HeapWord* compaction_top) {
2893   assert(oop(q)->forwardee() == NULL || oop(q)->forwardee() == oop(compaction_top),
2894          "should be moved to forwarded location");
2895   if (ValidateMarkSweep) {
2896     PSParallelCompact::validate_live_oop(oop(q), size);
2897     _live_oops_moved_to->push(oop(compaction_top));
2898   }
2899   if (RecordMarkSweepCompaction) {
2900     _cur_gc_live_oops->push(q);
2901     _cur_gc_live_oops_moved_to->push(compaction_top);
2902     _cur_gc_live_oops_size->push(size);
2903   }
2904 }
2905 
2906 
2907 void PSParallelCompact::compaction_complete() {
2908   if (RecordMarkSweepCompaction) {
2909     GrowableArray<HeapWord*>* _tmp_live_oops          = _cur_gc_live_oops;
2910     GrowableArray<HeapWord*>* _tmp_live_oops_moved_to = _cur_gc_live_oops_moved_to;
2911     GrowableArray<size_t>   * _tmp_live_oops_size     = _cur_gc_live_oops_size;
2912 
2913     _cur_gc_live_oops           = _last_gc_live_oops;
2914     _cur_gc_live_oops_moved_to  = _last_gc_live_oops_moved_to;
2915     _cur_gc_live_oops_size      = _last_gc_live_oops_size;
2916     _last_gc_live_oops          = _tmp_live_oops;
2917     _last_gc_live_oops_moved_to = _tmp_live_oops_moved_to;
2918     _last_gc_live_oops_size     = _tmp_live_oops_size;
2919   }
2920 }
2921 
2922 
2923 void PSParallelCompact::print_new_location_of_heap_address(HeapWord* q) {
2924   if (!RecordMarkSweepCompaction) {
2925     tty->print_cr("Requires RecordMarkSweepCompaction to be enabled");
2926     return;
2927   }
2928 
2929   if (_last_gc_live_oops == NULL) {
2930     tty->print_cr("No compaction information gathered yet");
2931     return;
2932   }
2933 
2934   for (int i = 0; i < _last_gc_live_oops->length(); i++) {
2935     HeapWord* old_oop = _last_gc_live_oops->at(i);
2936     size_t    sz      = _last_gc_live_oops_size->at(i);
2937     if (old_oop <= q && q < (old_oop + sz)) {
2938       HeapWord* new_oop = _last_gc_live_oops_moved_to->at(i);
2939       size_t offset = (q - old_oop);
2940       tty->print_cr("Address " PTR_FORMAT, q);
2941       tty->print_cr(" Was in oop " PTR_FORMAT ", size %d, at offset %d", old_oop, sz, offset);
2942       tty->print_cr(" Now in oop " PTR_FORMAT ", actual address " PTR_FORMAT, new_oop, new_oop + offset);
2943       return;
2944     }
2945   }
2946 
2947   tty->print_cr("Address " PTR_FORMAT " not found in live oop information from last GC", q);
2948 }
2949 #endif //VALIDATE_MARK_SWEEP
2950 
2951 // Update interior oops in the ranges of regions [beg_region, end_region).
2952 void
2953 PSParallelCompact::update_and_deadwood_in_dense_prefix(ParCompactionManager* cm,
2954                                                        SpaceId space_id,
2955                                                        size_t beg_region,
2956                                                        size_t end_region) {
2957   ParallelCompactData& sd = summary_data();
2958   ParMarkBitMap* const mbm = mark_bitmap();
2959 
2960   HeapWord* beg_addr = sd.region_to_addr(beg_region);
2961   HeapWord* const end_addr = sd.region_to_addr(end_region);
2962   assert(beg_region <= end_region, "bad region range");
2963   assert(end_addr <= dense_prefix(space_id), "not in the dense prefix");
2964 
2965 #ifdef  ASSERT
2966   // Claim the regions to avoid triggering an assert when they are marked as
2967   // filled.
2968   for (size_t claim_region = beg_region; claim_region < end_region; ++claim_region) {
2969     assert(sd.region(claim_region)->claim_unsafe(), "claim() failed");
2970   }
2971 #endif  // #ifdef ASSERT
2972 
2973   if (beg_addr != space(space_id)->bottom()) {
2974     // Find the first live object or block of dead space that *starts* in this
2975     // range of regions.  If a partial object crosses onto the region, skip it;
2976     // it will be marked for 'deferred update' when the object head is
2977     // processed.  If dead space crosses onto the region, it is also skipped; it
2978     // will be filled when the prior region is processed.  If neither of those
2979     // apply, the first word in the region is the start of a live object or dead
2980     // space.
2981     assert(beg_addr > space(space_id)->bottom(), "sanity");
2982     const RegionData* const cp = sd.region(beg_region);
2983     if (cp->partial_obj_size() != 0) {
2984       beg_addr = sd.partial_obj_end(beg_region);
2985     } else if (dead_space_crosses_boundary(cp, mbm->addr_to_bit(beg_addr))) {
2986       beg_addr = mbm->find_obj_beg(beg_addr, end_addr);
2987     }
2988   }
2989 
2990   if (beg_addr < end_addr) {
2991     // A live object or block of dead space starts in this range of Regions.
2992      HeapWord* const dense_prefix_end = dense_prefix(space_id);
2993 
2994     // Create closures and iterate.
2995     UpdateOnlyClosure update_closure(mbm, cm, space_id);
2996     FillClosure fill_closure(cm, space_id);
2997     ParMarkBitMap::IterationStatus status;
2998     status = mbm->iterate(&update_closure, &fill_closure, beg_addr, end_addr,
2999                           dense_prefix_end);
3000     if (status == ParMarkBitMap::incomplete) {
3001       update_closure.do_addr(update_closure.source());
3002     }
3003   }
3004 
3005   // Mark the regions as filled.
3006   RegionData* const beg_cp = sd.region(beg_region);
3007   RegionData* const end_cp = sd.region(end_region);
3008   for (RegionData* cp = beg_cp; cp < end_cp; ++cp) {
3009     cp->set_completed();
3010   }
3011 }
3012 
3013 // Return the SpaceId for the space containing addr.  If addr is not in the
3014 // heap, last_space_id is returned.  In debug mode it expects the address to be
3015 // in the heap and asserts such.
3016 PSParallelCompact::SpaceId PSParallelCompact::space_id(HeapWord* addr) {
3017   assert(Universe::heap()->is_in_reserved(addr), "addr not in the heap");
3018 
3019   for (unsigned int id = perm_space_id; id < last_space_id; ++id) {
3020     if (_space_info[id].space()->contains(addr)) {
3021       return SpaceId(id);
3022     }
3023   }
3024 
3025   assert(false, "no space contains the addr");
3026   return last_space_id;
3027 }
3028 
3029 void PSParallelCompact::update_deferred_objects(ParCompactionManager* cm,
3030                                                 SpaceId id) {
3031   assert(id < last_space_id, "bad space id");
3032 
3033   ParallelCompactData& sd = summary_data();
3034   const SpaceInfo* const space_info = _space_info + id;
3035   ObjectStartArray* const start_array = space_info->start_array();
3036 
3037   const MutableSpace* const space = space_info->space();
3038   assert(space_info->dense_prefix() >= space->bottom(), "dense_prefix not set");
3039   HeapWord* const beg_addr = space_info->dense_prefix();
3040   HeapWord* const end_addr = sd.region_align_up(space_info->new_top());
3041 
3042   const RegionData* const beg_region = sd.addr_to_region_ptr(beg_addr);
3043   const RegionData* const end_region = sd.addr_to_region_ptr(end_addr);
3044   const RegionData* cur_region;
3045   for (cur_region = beg_region; cur_region < end_region; ++cur_region) {
3046     HeapWord* const addr = cur_region->deferred_obj_addr();
3047     if (addr != NULL) {
3048       if (start_array != NULL) {
3049         start_array->allocate_block(addr);
3050       }
3051       oop(addr)->update_contents(cm);
3052       assert(oop(addr)->is_oop_or_null(), "should be an oop now");
3053     }
3054   }
3055 }
3056 
3057 // Skip over count live words starting from beg, and return the address of the
3058 // next live word.  Unless marked, the word corresponding to beg is assumed to
3059 // be dead.  Callers must either ensure beg does not correspond to the middle of
3060 // an object, or account for those live words in some other way.  Callers must
3061 // also ensure that there are enough live words in the range [beg, end) to skip.
3062 HeapWord*
3063 PSParallelCompact::skip_live_words(HeapWord* beg, HeapWord* end, size_t count)
3064 {
3065   assert(count > 0, "sanity");
3066 
3067   ParMarkBitMap* m = mark_bitmap();
3068   idx_t bits_to_skip = m->words_to_bits(count);
3069   idx_t cur_beg = m->addr_to_bit(beg);
3070   const idx_t search_end = BitMap::word_align_up(m->addr_to_bit(end));
3071 
3072   do {
3073     cur_beg = m->find_obj_beg(cur_beg, search_end);
3074     idx_t cur_end = m->find_obj_end(cur_beg, search_end);
3075     const size_t obj_bits = cur_end - cur_beg + 1;
3076     if (obj_bits > bits_to_skip) {
3077       return m->bit_to_addr(cur_beg + bits_to_skip);
3078     }
3079     bits_to_skip -= obj_bits;
3080     cur_beg = cur_end + 1;
3081   } while (bits_to_skip > 0);
3082 
3083   // Skipping the desired number of words landed just past the end of an object.
3084   // Find the start of the next object.
3085   cur_beg = m->find_obj_beg(cur_beg, search_end);
3086   assert(cur_beg < m->addr_to_bit(end), "not enough live words to skip");
3087   return m->bit_to_addr(cur_beg);
3088 }
3089 
3090 HeapWord* PSParallelCompact::first_src_addr(HeapWord* const dest_addr,
3091                                             SpaceId src_space_id,
3092                                             size_t src_region_idx)
3093 {
3094   assert(summary_data().is_region_aligned(dest_addr), "not aligned");
3095 
3096   const SplitInfo& split_info = _space_info[src_space_id].split_info();
3097   if (split_info.dest_region_addr() == dest_addr) {
3098     // The partial object ending at the split point contains the first word to
3099     // be copied to dest_addr.
3100     return split_info.first_src_addr();
3101   }
3102 
3103   const ParallelCompactData& sd = summary_data();
3104   ParMarkBitMap* const bitmap = mark_bitmap();
3105   const size_t RegionSize = ParallelCompactData::RegionSize;
3106 
3107   assert(sd.is_region_aligned(dest_addr), "not aligned");
3108   const RegionData* const src_region_ptr = sd.region(src_region_idx);
3109   const size_t partial_obj_size = src_region_ptr->partial_obj_size();
3110   HeapWord* const src_region_destination = src_region_ptr->destination();
3111 
3112   assert(dest_addr >= src_region_destination, "wrong src region");
3113   assert(src_region_ptr->data_size() > 0, "src region cannot be empty");
3114 
3115   HeapWord* const src_region_beg = sd.region_to_addr(src_region_idx);
3116   HeapWord* const src_region_end = src_region_beg + RegionSize;
3117 
3118   HeapWord* addr = src_region_beg;
3119   if (dest_addr == src_region_destination) {
3120     // Return the first live word in the source region.
3121     if (partial_obj_size == 0) {
3122       addr = bitmap->find_obj_beg(addr, src_region_end);
3123       assert(addr < src_region_end, "no objects start in src region");
3124     }
3125     return addr;
3126   }
3127 
3128   // Must skip some live data.
3129   size_t words_to_skip = dest_addr - src_region_destination;
3130   assert(src_region_ptr->data_size() > words_to_skip, "wrong src region");
3131 
3132   if (partial_obj_size >= words_to_skip) {
3133     // All the live words to skip are part of the partial object.
3134     addr += words_to_skip;
3135     if (partial_obj_size == words_to_skip) {
3136       // Find the first live word past the partial object.
3137       addr = bitmap->find_obj_beg(addr, src_region_end);
3138       assert(addr < src_region_end, "wrong src region");
3139     }
3140     return addr;
3141   }
3142 
3143   // Skip over the partial object (if any).
3144   if (partial_obj_size != 0) {
3145     words_to_skip -= partial_obj_size;
3146     addr += partial_obj_size;
3147   }
3148 
3149   // Skip over live words due to objects that start in the region.
3150   addr = skip_live_words(addr, src_region_end, words_to_skip);
3151   assert(addr < src_region_end, "wrong src region");
3152   return addr;
3153 }
3154 
3155 void PSParallelCompact::decrement_destination_counts(ParCompactionManager* cm,
3156                                                      SpaceId src_space_id,
3157                                                      size_t beg_region,
3158                                                      HeapWord* end_addr)
3159 {
3160   ParallelCompactData& sd = summary_data();
3161 
3162 #ifdef ASSERT
3163   MutableSpace* const src_space = _space_info[src_space_id].space();
3164   HeapWord* const beg_addr = sd.region_to_addr(beg_region);
3165   assert(src_space->contains(beg_addr) || beg_addr == src_space->end(),
3166          "src_space_id does not match beg_addr");
3167   assert(src_space->contains(end_addr) || end_addr == src_space->end(),
3168          "src_space_id does not match end_addr");
3169 #endif // #ifdef ASSERT
3170 
3171   RegionData* const beg = sd.region(beg_region);
3172   RegionData* const end = sd.addr_to_region_ptr(sd.region_align_up(end_addr));
3173 
3174   // Regions up to new_top() are enqueued if they become available.
3175   HeapWord* const new_top = _space_info[src_space_id].new_top();
3176   RegionData* const enqueue_end =
3177     sd.addr_to_region_ptr(sd.region_align_up(new_top));
3178 
3179   for (RegionData* cur = beg; cur < end; ++cur) {
3180     assert(cur->data_size() > 0, "region must have live data");
3181     cur->decrement_destination_count();
3182     if (cur < enqueue_end && cur->available() && cur->claim()) {
3183       cm->push_region(sd.region(cur));
3184     }
3185   }
3186 }
3187 
3188 size_t PSParallelCompact::next_src_region(MoveAndUpdateClosure& closure,
3189                                           SpaceId& src_space_id,
3190                                           HeapWord*& src_space_top,
3191                                           HeapWord* end_addr)
3192 {
3193   typedef ParallelCompactData::RegionData RegionData;
3194 
3195   ParallelCompactData& sd = PSParallelCompact::summary_data();
3196   const size_t region_size = ParallelCompactData::RegionSize;
3197 
3198   size_t src_region_idx = 0;
3199 
3200   // Skip empty regions (if any) up to the top of the space.
3201   HeapWord* const src_aligned_up = sd.region_align_up(end_addr);
3202   RegionData* src_region_ptr = sd.addr_to_region_ptr(src_aligned_up);
3203   HeapWord* const top_aligned_up = sd.region_align_up(src_space_top);
3204   const RegionData* const top_region_ptr =
3205     sd.addr_to_region_ptr(top_aligned_up);
3206   while (src_region_ptr < top_region_ptr && src_region_ptr->data_size() == 0) {
3207     ++src_region_ptr;
3208   }
3209 
3210   if (src_region_ptr < top_region_ptr) {
3211     // The next source region is in the current space.  Update src_region_idx
3212     // and the source address to match src_region_ptr.
3213     src_region_idx = sd.region(src_region_ptr);
3214     HeapWord* const src_region_addr = sd.region_to_addr(src_region_idx);
3215     if (src_region_addr > closure.source()) {
3216       closure.set_source(src_region_addr);
3217     }
3218     return src_region_idx;
3219   }
3220 
3221   // Switch to a new source space and find the first non-empty region.
3222   unsigned int space_id = src_space_id + 1;
3223   assert(space_id < last_space_id, "not enough spaces");
3224 
3225   HeapWord* const destination = closure.destination();
3226 
3227   do {
3228     MutableSpace* space = _space_info[space_id].space();
3229     HeapWord* const bottom = space->bottom();
3230     const RegionData* const bottom_cp = sd.addr_to_region_ptr(bottom);
3231 
3232     // Iterate over the spaces that do not compact into themselves.
3233     if (bottom_cp->destination() != bottom) {
3234       HeapWord* const top_aligned_up = sd.region_align_up(space->top());
3235       const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up);
3236 
3237       for (const RegionData* src_cp = bottom_cp; src_cp < top_cp; ++src_cp) {
3238         if (src_cp->live_obj_size() > 0) {
3239           // Found it.
3240           assert(src_cp->destination() == destination,
3241                  "first live obj in the space must match the destination");
3242           assert(src_cp->partial_obj_size() == 0,
3243                  "a space cannot begin with a partial obj");
3244 
3245           src_space_id = SpaceId(space_id);
3246           src_space_top = space->top();
3247           const size_t src_region_idx = sd.region(src_cp);
3248           closure.set_source(sd.region_to_addr(src_region_idx));
3249           return src_region_idx;
3250         } else {
3251           assert(src_cp->data_size() == 0, "sanity");
3252         }
3253       }
3254     }
3255   } while (++space_id < last_space_id);
3256 
3257   assert(false, "no source region was found");
3258   return 0;
3259 }
3260 
3261 void PSParallelCompact::fill_region(ParCompactionManager* cm, size_t region_idx)
3262 {
3263   typedef ParMarkBitMap::IterationStatus IterationStatus;
3264   const size_t RegionSize = ParallelCompactData::RegionSize;
3265   ParMarkBitMap* const bitmap = mark_bitmap();
3266   ParallelCompactData& sd = summary_data();
3267   RegionData* const region_ptr = sd.region(region_idx);
3268 
3269   // Get the items needed to construct the closure.
3270   HeapWord* dest_addr = sd.region_to_addr(region_idx);
3271   SpaceId dest_space_id = space_id(dest_addr);
3272   ObjectStartArray* start_array = _space_info[dest_space_id].start_array();
3273   HeapWord* new_top = _space_info[dest_space_id].new_top();
3274   assert(dest_addr < new_top, "sanity");
3275   const size_t words = MIN2(pointer_delta(new_top, dest_addr), RegionSize);
3276 
3277   // Get the source region and related info.
3278   size_t src_region_idx = region_ptr->source_region();
3279   SpaceId src_space_id = space_id(sd.region_to_addr(src_region_idx));
3280   HeapWord* src_space_top = _space_info[src_space_id].space()->top();
3281 
3282   MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words);
3283   closure.set_source(first_src_addr(dest_addr, src_space_id, src_region_idx));
3284 
3285   // Adjust src_region_idx to prepare for decrementing destination counts (the
3286   // destination count is not decremented when a region is copied to itself).
3287   if (src_region_idx == region_idx) {
3288     src_region_idx += 1;
3289   }
3290 
3291   if (bitmap->is_unmarked(closure.source())) {
3292     // The first source word is in the middle of an object; copy the remainder
3293     // of the object or as much as will fit.  The fact that pointer updates were
3294     // deferred will be noted when the object header is processed.
3295     HeapWord* const old_src_addr = closure.source();
3296     closure.copy_partial_obj();
3297     if (closure.is_full()) {
3298       decrement_destination_counts(cm, src_space_id, src_region_idx,
3299                                    closure.source());
3300       region_ptr->set_deferred_obj_addr(NULL);
3301       region_ptr->set_completed();
3302       return;
3303     }
3304 
3305     HeapWord* const end_addr = sd.region_align_down(closure.source());
3306     if (sd.region_align_down(old_src_addr) != end_addr) {
3307       // The partial object was copied from more than one source region.
3308       decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr);
3309 
3310       // Move to the next source region, possibly switching spaces as well.  All
3311       // args except end_addr may be modified.
3312       src_region_idx = next_src_region(closure, src_space_id, src_space_top,
3313                                        end_addr);
3314     }
3315   }
3316 
3317   do {
3318     HeapWord* const cur_addr = closure.source();
3319     HeapWord* const end_addr = MIN2(sd.region_align_up(cur_addr + 1),
3320                                     src_space_top);
3321     IterationStatus status = bitmap->iterate(&closure, cur_addr, end_addr);
3322 
3323     if (status == ParMarkBitMap::incomplete) {
3324       // The last obj that starts in the source region does not end in the
3325       // region.
3326       assert(closure.source() < end_addr, "sanity");
3327       HeapWord* const obj_beg = closure.source();
3328       HeapWord* const range_end = MIN2(obj_beg + closure.words_remaining(),
3329                                        src_space_top);
3330       HeapWord* const obj_end = bitmap->find_obj_end(obj_beg, range_end);
3331       if (obj_end < range_end) {
3332         // The end was found; the entire object will fit.
3333         status = closure.do_addr(obj_beg, bitmap->obj_size(obj_beg, obj_end));
3334         assert(status != ParMarkBitMap::would_overflow, "sanity");
3335       } else {
3336         // The end was not found; the object will not fit.
3337         assert(range_end < src_space_top, "obj cannot cross space boundary");
3338         status = ParMarkBitMap::would_overflow;
3339       }
3340     }
3341 
3342     if (status == ParMarkBitMap::would_overflow) {
3343       // The last object did not fit.  Note that interior oop updates were
3344       // deferred, then copy enough of the object to fill the region.
3345       region_ptr->set_deferred_obj_addr(closure.destination());
3346       status = closure.copy_until_full(); // copies from closure.source()
3347 
3348       decrement_destination_counts(cm, src_space_id, src_region_idx,
3349                                    closure.source());
3350       region_ptr->set_completed();
3351       return;
3352     }
3353 
3354     if (status == ParMarkBitMap::full) {
3355       decrement_destination_counts(cm, src_space_id, src_region_idx,
3356                                    closure.source());
3357       region_ptr->set_deferred_obj_addr(NULL);
3358       region_ptr->set_completed();
3359       return;
3360     }
3361 
3362     decrement_destination_counts(cm, src_space_id, src_region_idx, end_addr);
3363 
3364     // Move to the next source region, possibly switching spaces as well.  All
3365     // args except end_addr may be modified.
3366     src_region_idx = next_src_region(closure, src_space_id, src_space_top,
3367                                      end_addr);
3368   } while (true);
3369 }
3370 
3371 void
3372 PSParallelCompact::move_and_update(ParCompactionManager* cm, SpaceId space_id) {
3373   const MutableSpace* sp = space(space_id);
3374   if (sp->is_empty()) {
3375     return;
3376   }
3377 
3378   ParallelCompactData& sd = PSParallelCompact::summary_data();
3379   ParMarkBitMap* const bitmap = mark_bitmap();
3380   HeapWord* const dp_addr = dense_prefix(space_id);
3381   HeapWord* beg_addr = sp->bottom();
3382   HeapWord* end_addr = sp->top();
3383 
3384 #ifdef ASSERT
3385   assert(beg_addr <= dp_addr && dp_addr <= end_addr, "bad dense prefix");
3386   if (cm->should_verify_only()) {
3387     VerifyUpdateClosure verify_update(cm, sp);
3388     bitmap->iterate(&verify_update, beg_addr, end_addr);
3389     return;
3390   }
3391 
3392   if (cm->should_reset_only()) {
3393     ResetObjectsClosure reset_objects(cm);
3394     bitmap->iterate(&reset_objects, beg_addr, end_addr);
3395     return;
3396   }
3397 #endif
3398 
3399   const size_t beg_region = sd.addr_to_region_idx(beg_addr);
3400   const size_t dp_region = sd.addr_to_region_idx(dp_addr);
3401   if (beg_region < dp_region) {
3402     update_and_deadwood_in_dense_prefix(cm, space_id, beg_region, dp_region);
3403   }
3404 
3405   // The destination of the first live object that starts in the region is one
3406   // past the end of the partial object entering the region (if any).
3407   HeapWord* const dest_addr = sd.partial_obj_end(dp_region);
3408   HeapWord* const new_top = _space_info[space_id].new_top();
3409   assert(new_top >= dest_addr, "bad new_top value");
3410   const size_t words = pointer_delta(new_top, dest_addr);
3411 
3412   if (words > 0) {
3413     ObjectStartArray* start_array = _space_info[space_id].start_array();
3414     MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words);
3415 
3416     ParMarkBitMap::IterationStatus status;
3417     status = bitmap->iterate(&closure, dest_addr, end_addr);
3418     assert(status == ParMarkBitMap::full, "iteration not complete");
3419     assert(bitmap->find_obj_beg(closure.source(), end_addr) == end_addr,
3420            "live objects skipped because closure is full");
3421   }
3422 }
3423 
3424 jlong PSParallelCompact::millis_since_last_gc() {
3425   jlong ret_val = os::javaTimeMillis() - _time_of_last_gc;
3426   // XXX See note in genCollectedHeap::millis_since_last_gc().
3427   if (ret_val < 0) {
3428     NOT_PRODUCT(warning("time warp: %d", ret_val);)
3429     return 0;
3430   }
3431   return ret_val;
3432 }
3433 
3434 void PSParallelCompact::reset_millis_since_last_gc() {
3435   _time_of_last_gc = os::javaTimeMillis();
3436 }
3437 
3438 ParMarkBitMap::IterationStatus MoveAndUpdateClosure::copy_until_full()
3439 {
3440   if (source() != destination()) {
3441     DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3442     Copy::aligned_conjoint_words(source(), destination(), words_remaining());
3443   }
3444   update_state(words_remaining());
3445   assert(is_full(), "sanity");
3446   return ParMarkBitMap::full;
3447 }
3448 
3449 void MoveAndUpdateClosure::copy_partial_obj()
3450 {
3451   size_t words = words_remaining();
3452 
3453   HeapWord* const range_end = MIN2(source() + words, bitmap()->region_end());
3454   HeapWord* const end_addr = bitmap()->find_obj_end(source(), range_end);
3455   if (end_addr < range_end) {
3456     words = bitmap()->obj_size(source(), end_addr);
3457   }
3458 
3459   // This test is necessary; if omitted, the pointer updates to a partial object
3460   // that crosses the dense prefix boundary could be overwritten.
3461   if (source() != destination()) {
3462     DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3463     Copy::aligned_conjoint_words(source(), destination(), words);
3464   }
3465   update_state(words);
3466 }
3467 
3468 ParMarkBitMapClosure::IterationStatus
3469 MoveAndUpdateClosure::do_addr(HeapWord* addr, size_t words) {
3470   assert(destination() != NULL, "sanity");
3471   assert(bitmap()->obj_size(addr) == words, "bad size");
3472 
3473   _source = addr;
3474   assert(PSParallelCompact::summary_data().calc_new_pointer(source()) ==
3475          destination(), "wrong destination");
3476 
3477   if (words > words_remaining()) {
3478     return ParMarkBitMap::would_overflow;
3479   }
3480 
3481   // The start_array must be updated even if the object is not moving.
3482   if (_start_array != NULL) {
3483     _start_array->allocate_block(destination());
3484   }
3485 
3486   if (destination() != source()) {
3487     DEBUG_ONLY(PSParallelCompact::check_new_location(source(), destination());)
3488     Copy::aligned_conjoint_words(source(), destination(), words);
3489   }
3490 
3491   oop moved_oop = (oop) destination();
3492   moved_oop->update_contents(compaction_manager());
3493   assert(moved_oop->is_oop_or_null(), "Object should be whole at this point");
3494 
3495   update_state(words);
3496   assert(destination() == (HeapWord*)moved_oop + moved_oop->size(), "sanity");
3497   return is_full() ? ParMarkBitMap::full : ParMarkBitMap::incomplete;
3498 }
3499 
3500 UpdateOnlyClosure::UpdateOnlyClosure(ParMarkBitMap* mbm,
3501                                      ParCompactionManager* cm,
3502                                      PSParallelCompact::SpaceId space_id) :
3503   ParMarkBitMapClosure(mbm, cm),
3504   _space_id(space_id),
3505   _start_array(PSParallelCompact::start_array(space_id))
3506 {
3507 }
3508 
3509 // Updates the references in the object to their new values.
3510 ParMarkBitMapClosure::IterationStatus
3511 UpdateOnlyClosure::do_addr(HeapWord* addr, size_t words) {
3512   do_addr(addr);
3513   return ParMarkBitMap::incomplete;
3514 }
3515 
3516 // Verify the new location using the forwarding pointer
3517 // from MarkSweep::mark_sweep_phase2().  Set the mark_word
3518 // to the initial value.
3519 ParMarkBitMapClosure::IterationStatus
3520 PSParallelCompact::VerifyUpdateClosure::do_addr(HeapWord* addr, size_t words) {
3521   // The second arg (words) is not used.
3522   oop obj = (oop) addr;
3523   HeapWord* forwarding_ptr = (HeapWord*) obj->mark()->decode_pointer();
3524   HeapWord* new_pointer = summary_data().calc_new_pointer(obj);
3525   if (forwarding_ptr == NULL) {
3526     // The object is dead or not moving.
3527     assert(bitmap()->is_unmarked(obj) || (new_pointer == (HeapWord*) obj),
3528            "Object liveness is wrong.");
3529     return ParMarkBitMap::incomplete;
3530   }
3531   assert(UseParallelOldGCDensePrefix ||
3532          (HeapMaximumCompactionInterval > 1) ||
3533          (MarkSweepAlwaysCompactCount > 1) ||
3534          (forwarding_ptr == new_pointer),
3535     "Calculation of new location is incorrect");
3536   return ParMarkBitMap::incomplete;
3537 }
3538 
3539 // Reset objects modified for debug checking.
3540 ParMarkBitMapClosure::IterationStatus
3541 PSParallelCompact::ResetObjectsClosure::do_addr(HeapWord* addr, size_t words) {
3542   // The second arg (words) is not used.
3543   oop obj = (oop) addr;
3544   obj->init_mark();
3545   return ParMarkBitMap::incomplete;
3546 }
3547 
3548 // Prepare for compaction.  This method is executed once
3549 // (i.e., by a single thread) before compaction.
3550 // Save the updated location of the intArrayKlassObj for
3551 // filling holes in the dense prefix.
3552 void PSParallelCompact::compact_prologue() {
3553   _updated_int_array_klass_obj = (klassOop)
3554     summary_data().calc_new_pointer(Universe::intArrayKlassObj());
3555 }
3556