1 /*
   2  * Copyright 2005-2008 Sun Microsystems, Inc.  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 Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  20  * CA 95054 USA or visit www.sun.com if you need additional information or
  21  * have any questions.
  22  *
  23  */
  24 
  25 #include "incls/_precompiled.incl"
  26 #include "incls/_psParallelCompact.cpp.incl"
  27 
  28 #include <math.h>
  29 
  30 // All sizes are in HeapWords.
  31 const size_t ParallelCompactData::Log2RegionSize  = 9; // 512 words
  32 const size_t ParallelCompactData::RegionSize      = (size_t)1 << Log2RegionSize;
  33 const size_t ParallelCompactData::RegionSizeBytes =
  34   RegionSize << LogHeapWordSize;
  35 const size_t ParallelCompactData::RegionSizeOffsetMask = RegionSize - 1;
  36 const size_t ParallelCompactData::RegionAddrOffsetMask = RegionSizeBytes - 1;
  37 const size_t ParallelCompactData::RegionAddrMask  = ~RegionAddrOffsetMask;
  38 
  39 const ParallelCompactData::RegionData::region_sz_t
  40 ParallelCompactData::RegionData::dc_shift = 27;
  41 
  42 const ParallelCompactData::RegionData::region_sz_t
  43 ParallelCompactData::RegionData::dc_mask = ~0U << dc_shift;
  44 
  45 const ParallelCompactData::RegionData::region_sz_t
  46 ParallelCompactData::RegionData::dc_one = 0x1U << dc_shift;
  47 
  48 const ParallelCompactData::RegionData::region_sz_t
  49 ParallelCompactData::RegionData::los_mask = ~dc_mask;
  50 
  51 const ParallelCompactData::RegionData::region_sz_t
  52 ParallelCompactData::RegionData::dc_claimed = 0x8U << dc_shift;
  53 
  54 const ParallelCompactData::RegionData::region_sz_t
  55 ParallelCompactData::RegionData::dc_completed = 0xcU << dc_shift;
  56 
  57 SpaceInfo PSParallelCompact::_space_info[PSParallelCompact::last_space_id];
  58 bool      PSParallelCompact::_print_phases = false;
  59 
  60 ReferenceProcessor* PSParallelCompact::_ref_processor = NULL;
  61 klassOop            PSParallelCompact::_updated_int_array_klass_obj = NULL;
  62 
  63 double PSParallelCompact::_dwl_mean;
  64 double PSParallelCompact::_dwl_std_dev;
  65 double PSParallelCompact::_dwl_first_term;
  66 double PSParallelCompact::_dwl_adjustment;
  67 #ifdef  ASSERT
  68 bool   PSParallelCompact::_dwl_initialized = false;
  69 #endif  // #ifdef ASSERT
  70 
  71 #ifdef VALIDATE_MARK_SWEEP
  72 GrowableArray<void*>*   PSParallelCompact::_root_refs_stack = NULL;
  73 GrowableArray<oop> *    PSParallelCompact::_live_oops = NULL;
  74 GrowableArray<oop> *    PSParallelCompact::_live_oops_moved_to = NULL;
  75 GrowableArray<size_t>*  PSParallelCompact::_live_oops_size = NULL;
  76 size_t                  PSParallelCompact::_live_oops_index = 0;
  77 size_t                  PSParallelCompact::_live_oops_index_at_perm = 0;
  78 GrowableArray<void*>*   PSParallelCompact::_other_refs_stack = NULL;
  79 GrowableArray<void*>*   PSParallelCompact::_adjusted_pointers = NULL;
  80 bool                    PSParallelCompact::_pointer_tracking = false;
  81 bool                    PSParallelCompact::_root_tracking = true;
  82 
  83 GrowableArray<HeapWord*>* PSParallelCompact::_cur_gc_live_oops = NULL;
  84 GrowableArray<HeapWord*>* PSParallelCompact::_cur_gc_live_oops_moved_to = NULL;
  85 GrowableArray<size_t>   * PSParallelCompact::_cur_gc_live_oops_size = NULL;
  86 GrowableArray<HeapWord*>* PSParallelCompact::_last_gc_live_oops = NULL;
  87 GrowableArray<HeapWord*>* PSParallelCompact::_last_gc_live_oops_moved_to = NULL;
  88 GrowableArray<size_t>   * PSParallelCompact::_last_gc_live_oops_size = NULL;
  89 #endif
  90 
  91 #ifndef PRODUCT
  92 const char* PSParallelCompact::space_names[] = {
  93   "perm", "old ", "eden", "from", "to  "
  94 };
  95 
  96 void PSParallelCompact::print_region_ranges()
  97 {
  98   tty->print_cr("space  bottom     top        end        new_top");
  99   tty->print_cr("------ ---------- ---------- ---------- ----------");
 100 
 101   for (unsigned int id = 0; id < last_space_id; ++id) {
 102     const MutableSpace* space = _space_info[id].space();
 103     tty->print_cr("%u %s "
 104                   SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " "
 105                   SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " ",
 106                   id, space_names[id],
 107                   summary_data().addr_to_region_idx(space->bottom()),
 108                   summary_data().addr_to_region_idx(space->top()),
 109                   summary_data().addr_to_region_idx(space->end()),
 110                   summary_data().addr_to_region_idx(_space_info[id].new_top()));
 111   }
 112 }
 113 
 114 void
 115 print_generic_summary_region(size_t i, const ParallelCompactData::RegionData* c)
 116 {
 117 #define REGION_IDX_FORMAT        SIZE_FORMAT_W(7)
 118 #define REGION_DATA_FORMAT       SIZE_FORMAT_W(5)
 119 
 120   ParallelCompactData& sd = PSParallelCompact::summary_data();
 121   size_t dci = c->destination() ? sd.addr_to_region_idx(c->destination()) : 0;
 122   tty->print_cr(REGION_IDX_FORMAT " " PTR_FORMAT " "
 123                 REGION_IDX_FORMAT " " PTR_FORMAT " "
 124                 REGION_DATA_FORMAT " " REGION_DATA_FORMAT " "
 125                 REGION_DATA_FORMAT " " REGION_IDX_FORMAT " %d",
 126                 i, c->data_location(), dci, c->destination(),
 127                 c->partial_obj_size(), c->live_obj_size(),
 128                 c->data_size(), c->source_region(), c->destination_count());
 129 
 130 #undef  REGION_IDX_FORMAT
 131 #undef  REGION_DATA_FORMAT
 132 }
 133 
 134 void
 135 print_generic_summary_data(ParallelCompactData& summary_data,
 136                            HeapWord* const beg_addr,
 137                            HeapWord* const end_addr)
 138 {
 139   size_t total_words = 0;
 140   size_t i = summary_data.addr_to_region_idx(beg_addr);
 141   const size_t last = summary_data.addr_to_region_idx(end_addr);
 142   HeapWord* pdest = 0;
 143 
 144   while (i <= last) {
 145     ParallelCompactData::RegionData* c = summary_data.region(i);
 146     if (c->data_size() != 0 || c->destination() != pdest) {
 147       print_generic_summary_region(i, c);
 148       total_words += c->data_size();
 149       pdest = c->destination();
 150     }
 151     ++i;
 152   }
 153 
 154   tty->print_cr("summary_data_bytes=" SIZE_FORMAT, total_words * HeapWordSize);
 155 }
 156 
 157 void
 158 print_generic_summary_data(ParallelCompactData& summary_data,
 159                            SpaceInfo* space_info)
 160 {
 161   for (unsigned int id = 0; id < PSParallelCompact::last_space_id; ++id) {
 162     const MutableSpace* space = space_info[id].space();
 163     print_generic_summary_data(summary_data, space->bottom(),
 164                                MAX2(space->top(), space_info[id].new_top()));
 165   }
 166 }
 167 
 168 void
 169 print_initial_summary_region(size_t i,
 170                              const ParallelCompactData::RegionData* c,
 171                              bool newline = true)
 172 {
 173   tty->print(SIZE_FORMAT_W(5) " " PTR_FORMAT " "
 174              SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " "
 175              SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d",
 176              i, c->destination(),
 177              c->partial_obj_size(), c->live_obj_size(),
 178              c->data_size(), c->source_region(), c->destination_count());
 179   if (newline) tty->cr();
 180 }
 181 
 182 void
 183 print_initial_summary_data(ParallelCompactData& summary_data,
 184                            const MutableSpace* space) {
 185   if (space->top() == space->bottom()) {
 186     return;
 187   }
 188 
 189   const size_t region_size = ParallelCompactData::RegionSize;
 190   typedef ParallelCompactData::RegionData RegionData;
 191   HeapWord* const top_aligned_up = summary_data.region_align_up(space->top());
 192   const size_t end_region = summary_data.addr_to_region_idx(top_aligned_up);
 193   const RegionData* c = summary_data.region(end_region - 1);
 194   HeapWord* end_addr = c->destination() + c->data_size();
 195   const size_t live_in_space = pointer_delta(end_addr, space->bottom());
 196 
 197   // Print (and count) the full regions at the beginning of the space.
 198   size_t full_region_count = 0;
 199   size_t i = summary_data.addr_to_region_idx(space->bottom());
 200   while (i < end_region && summary_data.region(i)->data_size() == region_size) {
 201     print_initial_summary_region(i, summary_data.region(i));
 202     ++full_region_count;
 203     ++i;
 204   }
 205 
 206   size_t live_to_right = live_in_space - full_region_count * region_size;
 207 
 208   double max_reclaimed_ratio = 0.0;
 209   size_t max_reclaimed_ratio_region = 0;
 210   size_t max_dead_to_right = 0;
 211   size_t max_live_to_right = 0;
 212 
 213   // Print the 'reclaimed ratio' for regions while there is something live in
 214   // the region or to the right of it.  The remaining regions are empty (and
 215   // uninteresting), and computing the ratio will result in division by 0.
 216   while (i < end_region && live_to_right > 0) {
 217     c = summary_data.region(i);
 218     HeapWord* const region_addr = summary_data.region_to_addr(i);
 219     const size_t used_to_right = pointer_delta(space->top(), region_addr);
 220     const size_t dead_to_right = used_to_right - live_to_right;
 221     const double reclaimed_ratio = double(dead_to_right) / live_to_right;
 222 
 223     if (reclaimed_ratio > max_reclaimed_ratio) {
 224             max_reclaimed_ratio = reclaimed_ratio;
 225             max_reclaimed_ratio_region = i;
 226             max_dead_to_right = dead_to_right;
 227             max_live_to_right = live_to_right;
 228     }
 229 
 230     print_initial_summary_region(i, c, false);
 231     tty->print_cr(" %12.10f " SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10),
 232                   reclaimed_ratio, dead_to_right, live_to_right);
 233 
 234     live_to_right -= c->data_size();
 235     ++i;
 236   }
 237 
 238   // Any remaining regions are empty.  Print one more if there is one.
 239   if (i < end_region) {
 240     print_initial_summary_region(i, summary_data.region(i));
 241   }
 242 
 243   tty->print_cr("max:  " SIZE_FORMAT_W(4) " d2r=" SIZE_FORMAT_W(10) " "
 244                 "l2r=" SIZE_FORMAT_W(10) " max_ratio=%14.12f",
 245                 max_reclaimed_ratio_region, max_dead_to_right,
 246                 max_live_to_right, max_reclaimed_ratio);
 247 }
 248 
 249 void
 250 print_initial_summary_data(ParallelCompactData& summary_data,
 251                            SpaceInfo* space_info) {
 252   unsigned int id = PSParallelCompact::perm_space_id;
 253   const MutableSpace* space;
 254   do {
 255     space = space_info[id].space();
 256     print_initial_summary_data(summary_data, space);
 257   } while (++id < PSParallelCompact::eden_space_id);
 258 
 259   do {
 260     space = space_info[id].space();
 261     print_generic_summary_data(summary_data, space->bottom(), space->top());
 262   } while (++id < PSParallelCompact::last_space_id);
 263 }
 264 #endif  // #ifndef PRODUCT
 265 
 266 #ifdef  ASSERT
 267 size_t add_obj_count;
 268 size_t add_obj_size;
 269 size_t mark_bitmap_count;
 270 size_t mark_bitmap_size;
 271 #endif  // #ifdef ASSERT
 272 
 273 ParallelCompactData::ParallelCompactData()
 274 {
 275   _region_start = 0;
 276 
 277   _region_vspace = 0;
 278   _region_data = 0;
 279   _region_count = 0;
 280 }
 281 
 282 bool ParallelCompactData::initialize(MemRegion covered_region)
 283 {
 284   _region_start = covered_region.start();
 285   const size_t region_size = covered_region.word_size();
 286   DEBUG_ONLY(_region_end = _region_start + region_size;)
 287 
 288   assert(region_align_down(_region_start) == _region_start,
 289          "region start not aligned");
 290   assert((region_size & RegionSizeOffsetMask) == 0,
 291          "region size not a multiple of RegionSize");
 292 
 293   bool result = initialize_region_data(region_size);
 294 
 295   return result;
 296 }
 297 
 298 PSVirtualSpace*
 299 ParallelCompactData::create_vspace(size_t count, size_t element_size)
 300 {
 301   const size_t raw_bytes = count * element_size;
 302   const size_t page_sz = os::page_size_for_region(raw_bytes, raw_bytes, 10);
 303   const size_t granularity = os::vm_allocation_granularity();
 304   const size_t bytes = align_size_up(raw_bytes, MAX2(page_sz, granularity));
 305 
 306   const size_t rs_align = page_sz == (size_t) os::vm_page_size() ? 0 :
 307     MAX2(page_sz, granularity);
 308   ReservedSpace rs(bytes, rs_align, rs_align > 0);
 309   os::trace_page_sizes("par compact", raw_bytes, raw_bytes, page_sz, rs.base(),
 310                        rs.size());
 311   PSVirtualSpace* vspace = new PSVirtualSpace(rs, page_sz);
 312   if (vspace != 0) {
 313     if (vspace->expand_by(bytes)) {
 314       return vspace;
 315     }
 316     delete vspace;
 317     // Release memory reserved in the space.
 318     rs.release();
 319   }
 320 
 321   return 0;
 322 }
 323 
 324 bool ParallelCompactData::initialize_region_data(size_t region_size)
 325 {
 326   const size_t count = (region_size + RegionSizeOffsetMask) >> Log2RegionSize;
 327   _region_vspace = create_vspace(count, sizeof(RegionData));
 328   if (_region_vspace != 0) {
 329     _region_data = (RegionData*)_region_vspace->reserved_low_addr();
 330     _region_count = count;
 331     return true;
 332   }
 333   return false;
 334 }
 335 
 336 void ParallelCompactData::clear()
 337 {
 338   memset(_region_data, 0, _region_vspace->committed_size());
 339 }
 340 
 341 void ParallelCompactData::clear_range(size_t beg_region, size_t end_region) {
 342   assert(beg_region <= _region_count, "beg_region out of range");
 343   assert(end_region <= _region_count, "end_region out of range");
 344 
 345   const size_t region_cnt = end_region - beg_region;
 346   memset(_region_data + beg_region, 0, region_cnt * sizeof(RegionData));
 347 }
 348 
 349 HeapWord* ParallelCompactData::partial_obj_end(size_t region_idx) const
 350 {
 351   const RegionData* cur_cp = region(region_idx);
 352   const RegionData* const end_cp = region(region_count() - 1);
 353 
 354   HeapWord* result = region_to_addr(region_idx);
 355   if (cur_cp < end_cp) {
 356     do {
 357       result += cur_cp->partial_obj_size();
 358     } while (cur_cp->partial_obj_size() == RegionSize && ++cur_cp < end_cp);
 359   }
 360   return result;
 361 }
 362 
 363 void ParallelCompactData::add_obj(HeapWord* addr, size_t len)
 364 {
 365   const size_t obj_ofs = pointer_delta(addr, _region_start);
 366   const size_t beg_region = obj_ofs >> Log2RegionSize;
 367   const size_t end_region = (obj_ofs + len - 1) >> Log2RegionSize;
 368 
 369   DEBUG_ONLY(Atomic::inc_ptr(&add_obj_count);)
 370   DEBUG_ONLY(Atomic::add_ptr(len, &add_obj_size);)
 371 
 372   if (beg_region == end_region) {
 373     // All in one region.
 374     _region_data[beg_region].add_live_obj(len);
 375     return;
 376   }
 377 
 378   // First region.
 379   const size_t beg_ofs = region_offset(addr);
 380   _region_data[beg_region].add_live_obj(RegionSize - beg_ofs);
 381 
 382   klassOop klass = ((oop)addr)->klass();
 383   // Middle regions--completely spanned by this object.
 384   for (size_t region = beg_region + 1; region < end_region; ++region) {
 385     _region_data[region].set_partial_obj_size(RegionSize);
 386     _region_data[region].set_partial_obj_addr(addr);
 387   }
 388 
 389   // Last region.
 390   const size_t end_ofs = region_offset(addr + len - 1);
 391   _region_data[end_region].set_partial_obj_size(end_ofs + 1);
 392   _region_data[end_region].set_partial_obj_addr(addr);
 393 }
 394 
 395 void
 396 ParallelCompactData::summarize_dense_prefix(HeapWord* beg, HeapWord* end)
 397 {
 398   assert(region_offset(beg) == 0, "not RegionSize aligned");
 399   assert(region_offset(end) == 0, "not RegionSize aligned");
 400 
 401   size_t cur_region = addr_to_region_idx(beg);
 402   const size_t end_region = addr_to_region_idx(end);
 403   HeapWord* addr = beg;
 404   while (cur_region < end_region) {
 405     _region_data[cur_region].set_destination(addr);
 406     _region_data[cur_region].set_destination_count(0);
 407     _region_data[cur_region].set_source_region(cur_region);
 408     _region_data[cur_region].set_data_location(addr);
 409 
 410     // Update live_obj_size so the region appears completely full.
 411     size_t live_size = RegionSize - _region_data[cur_region].partial_obj_size();
 412     _region_data[cur_region].set_live_obj_size(live_size);
 413 
 414     ++cur_region;
 415     addr += RegionSize;
 416   }
 417 }
 418 
 419 bool ParallelCompactData::summarize(HeapWord* target_beg, HeapWord* target_end,
 420                                     HeapWord* source_beg, HeapWord* source_end,
 421                                     HeapWord** target_next,
 422                                     HeapWord** source_next) {
 423   // This is too strict.
 424   // assert(region_offset(source_beg) == 0, "not RegionSize aligned");
 425 
 426   if (TraceParallelOldGCSummaryPhase) {
 427     tty->print_cr("tb=" PTR_FORMAT " te=" PTR_FORMAT " "
 428                   "sb=" PTR_FORMAT " se=" PTR_FORMAT " "
 429                   "tn=" PTR_FORMAT " sn=" PTR_FORMAT,
 430                   target_beg, target_end,
 431                   source_beg, source_end,
 432                   target_next != 0 ? *target_next : (HeapWord*) 0,
 433                   source_next != 0 ? *source_next : (HeapWord*) 0);
 434   }
 435 
 436   size_t cur_region = addr_to_region_idx(source_beg);
 437   const size_t end_region = addr_to_region_idx(region_align_up(source_end));
 438 
 439   HeapWord *dest_addr = target_beg;
 440   while (cur_region < end_region) {
 441     size_t words = _region_data[cur_region].data_size();
 442 
 443 #if     1
 444     assert(pointer_delta(target_end, dest_addr) >= words,
 445            "source region does not fit into target region");
 446 #else
 447     // XXX - need some work on the corner cases here.  If the region does not
 448     // fit, then must either make sure any partial_obj from the region fits, or
 449     // "undo" the initial part of the partial_obj that is in the previous
 450     // region.
 451     if (dest_addr + words >= target_end) {
 452       // Let the caller know where to continue.
 453       *target_next = dest_addr;
 454       *source_next = region_to_addr(cur_region);
 455       return false;
 456     }
 457 #endif  // #if 1
 458 
 459     _region_data[cur_region].set_destination(dest_addr);
 460 
 461     // Set the destination_count for cur_region, and if necessary, update
 462     // source_region for a destination region.  The source_region field is
 463     // updated if cur_region is the first (left-most) region to be copied to a
 464     // destination region.
 465     //
 466     // The destination_count calculation is a bit subtle.  A region that has
 467     // data that compacts into itself does not count itself as a destination.
 468     // This maintains the invariant that a zero count means the region is
 469     // available and can be claimed and then filled.
 470     if (words > 0) {
 471       HeapWord* const last_addr = dest_addr + words - 1;
 472       const size_t dest_region_1 = addr_to_region_idx(dest_addr);
 473       const size_t dest_region_2 = addr_to_region_idx(last_addr);
 474 #if     0
 475       // Initially assume that the destination regions will be the same and
 476       // adjust the value below if necessary.  Under this assumption, if
 477       // cur_region == dest_region_2, then cur_region will be compacted
 478       // completely into itself.
 479       uint destination_count = cur_region == dest_region_2 ? 0 : 1;
 480       if (dest_region_1 != dest_region_2) {
 481         // Destination regions differ; adjust destination_count.
 482         destination_count += 1;
 483         // Data from cur_region will be copied to the start of dest_region_2.
 484         _region_data[dest_region_2].set_source_region(cur_region);
 485       } else if (region_offset(dest_addr) == 0) {
 486         // Data from cur_region will be copied to the start of the destination
 487         // region.
 488         _region_data[dest_region_1].set_source_region(cur_region);
 489       }
 490 #else
 491       // Initially assume that the destination regions will be different and
 492       // adjust the value below if necessary.  Under this assumption, if
 493       // cur_region == dest_region2, then cur_region will be compacted partially
 494       // into dest_region_1 and partially into itself.
 495       uint destination_count = cur_region == dest_region_2 ? 1 : 2;
 496       if (dest_region_1 != dest_region_2) {
 497         // Data from cur_region will be copied to the start of dest_region_2.
 498         _region_data[dest_region_2].set_source_region(cur_region);
 499       } else {
 500         // Destination regions are the same; adjust destination_count.
 501         destination_count -= 1;
 502         if (region_offset(dest_addr) == 0) {
 503           // Data from cur_region will be copied to the start of the destination
 504           // region.
 505           _region_data[dest_region_1].set_source_region(cur_region);
 506         }
 507       }
 508 #endif  // #if 0
 509 
 510       _region_data[cur_region].set_destination_count(destination_count);
 511       _region_data[cur_region].set_data_location(region_to_addr(cur_region));
 512       dest_addr += words;
 513     }
 514 
 515     ++cur_region;
 516   }
 517 
 518   *target_next = dest_addr;
 519   return true;
 520 }
 521 
 522 HeapWord* ParallelCompactData::calc_new_pointer(HeapWord* addr) {
 523   assert(addr != NULL, "Should detect NULL oop earlier");
 524   assert(PSParallelCompact::gc_heap()->is_in(addr), "addr not in heap");
 525 #ifdef ASSERT
 526   if (PSParallelCompact::mark_bitmap()->is_unmarked(addr)) {
 527     gclog_or_tty->print_cr("calc_new_pointer:: addr " PTR_FORMAT, addr);
 528   }
 529 #endif
 530   assert(PSParallelCompact::mark_bitmap()->is_marked(addr), "obj not marked");
 531 
 532   // Region covering the object.
 533   size_t region_index = addr_to_region_idx(addr);
 534   const RegionData* const region_ptr = region(region_index);
 535   HeapWord* const region_addr = region_align_down(addr);
 536 
 537   assert(addr < region_addr + RegionSize, "Region does not cover object");
 538   assert(addr_to_region_ptr(region_addr) == region_ptr, "sanity check");
 539 
 540   HeapWord* result = region_ptr->destination();
 541 
 542   // If all the data in the region is live, then the new location of the object
 543   // can be calculated from the destination of the region plus the offset of the
 544   // object in the region.
 545   if (region_ptr->data_size() == RegionSize) {
 546     result += pointer_delta(addr, region_addr);
 547     return result;
 548   }
 549 
 550   // The new location of the object is
 551   //    region destination +
 552   //    size of the partial object extending onto the region +
 553   //    sizes of the live objects in the Region that are to the left of addr
 554   const size_t partial_obj_size = region_ptr->partial_obj_size();
 555   HeapWord* const search_start = region_addr + partial_obj_size;
 556 
 557   const ParMarkBitMap* bitmap = PSParallelCompact::mark_bitmap();
 558   size_t live_to_left = bitmap->live_words_in_range(search_start, oop(addr));
 559 
 560   result += partial_obj_size + live_to_left;
 561   assert(result <= addr, "object cannot move to the right");
 562   return result;
 563 }
 564 
 565 klassOop ParallelCompactData::calc_new_klass(klassOop old_klass) {
 566   klassOop updated_klass;
 567   if (PSParallelCompact::should_update_klass(old_klass)) {
 568     updated_klass = (klassOop) calc_new_pointer(old_klass);
 569   } else {
 570     updated_klass = old_klass;
 571   }
 572 
 573   return updated_klass;
 574 }
 575 
 576 #ifdef  ASSERT
 577 void ParallelCompactData::verify_clear(const PSVirtualSpace* vspace)
 578 {
 579   const size_t* const beg = (const size_t*)vspace->committed_low_addr();
 580   const size_t* const end = (const size_t*)vspace->committed_high_addr();
 581   for (const size_t* p = beg; p < end; ++p) {
 582     assert(*p == 0, "not zero");
 583   }
 584 }
 585 
 586 void ParallelCompactData::verify_clear()
 587 {
 588   verify_clear(_region_vspace);
 589 }
 590 #endif  // #ifdef ASSERT
 591 
 592 #ifdef NOT_PRODUCT
 593 ParallelCompactData::RegionData* debug_region(size_t region_index) {
 594   ParallelCompactData& sd = PSParallelCompact::summary_data();
 595   return sd.region(region_index);
 596 }
 597 #endif
 598 
 599 elapsedTimer        PSParallelCompact::_accumulated_time;
 600 unsigned int        PSParallelCompact::_total_invocations = 0;
 601 unsigned int        PSParallelCompact::_maximum_compaction_gc_num = 0;
 602 jlong               PSParallelCompact::_time_of_last_gc = 0;
 603 CollectorCounters*  PSParallelCompact::_counters = NULL;
 604 ParMarkBitMap       PSParallelCompact::_mark_bitmap;
 605 ParallelCompactData PSParallelCompact::_summary_data;
 606 
 607 PSParallelCompact::IsAliveClosure PSParallelCompact::_is_alive_closure;
 608 
 609 void PSParallelCompact::IsAliveClosure::do_object(oop p)   { ShouldNotReachHere(); }
 610 bool PSParallelCompact::IsAliveClosure::do_object_b(oop p) { return mark_bitmap()->is_marked(p); }
 611 
 612 void PSParallelCompact::KeepAliveClosure::do_oop(oop* p)       { PSParallelCompact::KeepAliveClosure::do_oop_work(p); }
 613 void PSParallelCompact::KeepAliveClosure::do_oop(narrowOop* p) { PSParallelCompact::KeepAliveClosure::do_oop_work(p); }
 614 
 615 PSParallelCompact::AdjustPointerClosure PSParallelCompact::_adjust_root_pointer_closure(true);
 616 PSParallelCompact::AdjustPointerClosure PSParallelCompact::_adjust_pointer_closure(false);
 617 
 618 void PSParallelCompact::AdjustPointerClosure::do_oop(oop* p)       { adjust_pointer(p, _is_root); }
 619 void PSParallelCompact::AdjustPointerClosure::do_oop(narrowOop* p) { adjust_pointer(p, _is_root); }
 620 
 621 void PSParallelCompact::FollowStackClosure::do_void() { follow_stack(_compaction_manager); }
 622 
 623 void PSParallelCompact::MarkAndPushClosure::do_oop(oop* p)       { mark_and_push(_compaction_manager, p); }
 624 void PSParallelCompact::MarkAndPushClosure::do_oop(narrowOop* p) { mark_and_push(_compaction_manager, p); }
 625 
 626 void PSParallelCompact::post_initialize() {
 627   ParallelScavengeHeap* heap = gc_heap();
 628   assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity");
 629 
 630   MemRegion mr = heap->reserved_region();
 631   _ref_processor = ReferenceProcessor::create_ref_processor(
 632     mr,                         // span
 633     true,                       // atomic_discovery
 634     true,                       // mt_discovery
 635     &_is_alive_closure,
 636     ParallelGCThreads,
 637     ParallelRefProcEnabled);
 638   _counters = new CollectorCounters("PSParallelCompact", 1);
 639 
 640   // Initialize static fields in ParCompactionManager.
 641   ParCompactionManager::initialize(mark_bitmap());
 642 }
 643 
 644 bool PSParallelCompact::initialize() {
 645   ParallelScavengeHeap* heap = gc_heap();
 646   assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity");
 647   MemRegion mr = heap->reserved_region();
 648 
 649   // Was the old gen get allocated successfully?
 650   if (!heap->old_gen()->is_allocated()) {
 651     return false;
 652   }
 653 
 654   initialize_space_info();
 655   initialize_dead_wood_limiter();
 656 
 657   if (!_mark_bitmap.initialize(mr)) {
 658     vm_shutdown_during_initialization("Unable to allocate bit map for "
 659       "parallel garbage collection for the requested heap size.");
 660     return false;
 661   }
 662 
 663   if (!_summary_data.initialize(mr)) {
 664     vm_shutdown_during_initialization("Unable to allocate tables for "
 665       "parallel garbage collection for the requested heap size.");
 666     return false;
 667   }
 668 
 669   return true;
 670 }
 671 
 672 void PSParallelCompact::initialize_space_info()
 673 {
 674   memset(&_space_info, 0, sizeof(_space_info));
 675 
 676   ParallelScavengeHeap* heap = gc_heap();
 677   PSYoungGen* young_gen = heap->young_gen();
 678   MutableSpace* perm_space = heap->perm_gen()->object_space();
 679 
 680   _space_info[perm_space_id].set_space(perm_space);
 681   _space_info[old_space_id].set_space(heap->old_gen()->object_space());
 682   _space_info[eden_space_id].set_space(young_gen->eden_space());
 683   _space_info[from_space_id].set_space(young_gen->from_space());
 684   _space_info[to_space_id].set_space(young_gen->to_space());
 685 
 686   _space_info[perm_space_id].set_start_array(heap->perm_gen()->start_array());
 687   _space_info[old_space_id].set_start_array(heap->old_gen()->start_array());
 688 
 689   _space_info[perm_space_id].set_min_dense_prefix(perm_space->top());
 690   if (TraceParallelOldGCDensePrefix) {
 691     tty->print_cr("perm min_dense_prefix=" PTR_FORMAT,
 692                   _space_info[perm_space_id].min_dense_prefix());
 693   }
 694 }
 695 
 696 void PSParallelCompact::initialize_dead_wood_limiter()
 697 {
 698   const size_t max = 100;
 699   _dwl_mean = double(MIN2(ParallelOldDeadWoodLimiterMean, max)) / 100.0;
 700   _dwl_std_dev = double(MIN2(ParallelOldDeadWoodLimiterStdDev, max)) / 100.0;
 701   _dwl_first_term = 1.0 / (sqrt(2.0 * M_PI) * _dwl_std_dev);
 702   DEBUG_ONLY(_dwl_initialized = true;)
 703   _dwl_adjustment = normal_distribution(1.0);
 704 }
 705 
 706 // Simple class for storing info about the heap at the start of GC, to be used
 707 // after GC for comparison/printing.
 708 class PreGCValues {
 709 public:
 710   PreGCValues() { }
 711   PreGCValues(ParallelScavengeHeap* heap) { fill(heap); }
 712 
 713   void fill(ParallelScavengeHeap* heap) {
 714     _heap_used      = heap->used();
 715     _young_gen_used = heap->young_gen()->used_in_bytes();
 716     _old_gen_used   = heap->old_gen()->used_in_bytes();
 717     _perm_gen_used  = heap->perm_gen()->used_in_bytes();
 718   };
 719 
 720   size_t heap_used() const      { return _heap_used; }
 721   size_t young_gen_used() const { return _young_gen_used; }
 722   size_t old_gen_used() const   { return _old_gen_used; }
 723   size_t perm_gen_used() const  { return _perm_gen_used; }
 724 
 725 private:
 726   size_t _heap_used;
 727   size_t _young_gen_used;
 728   size_t _old_gen_used;
 729   size_t _perm_gen_used;
 730 };
 731 
 732 void
 733 PSParallelCompact::clear_data_covering_space(SpaceId id)
 734 {
 735   // At this point, top is the value before GC, new_top() is the value that will
 736   // be set at the end of GC.  The marking bitmap is cleared to top; nothing
 737   // should be marked above top.  The summary data is cleared to the larger of
 738   // top & new_top.
 739   MutableSpace* const space = _space_info[id].space();
 740   HeapWord* const bot = space->bottom();
 741   HeapWord* const top = space->top();
 742   HeapWord* const max_top = MAX2(top, _space_info[id].new_top());
 743 
 744   const idx_t beg_bit = _mark_bitmap.addr_to_bit(bot);
 745   const idx_t end_bit = BitMap::word_align_up(_mark_bitmap.addr_to_bit(top));
 746   _mark_bitmap.clear_range(beg_bit, end_bit);
 747 
 748   const size_t beg_region = _summary_data.addr_to_region_idx(bot);
 749   const size_t end_region =
 750     _summary_data.addr_to_region_idx(_summary_data.region_align_up(max_top));
 751   _summary_data.clear_range(beg_region, end_region);
 752 }
 753 
 754 void PSParallelCompact::pre_compact(PreGCValues* pre_gc_values)
 755 {
 756   // Update the from & to space pointers in space_info, since they are swapped
 757   // at each young gen gc.  Do the update unconditionally (even though a
 758   // promotion failure does not swap spaces) because an unknown number of minor
 759   // collections will have swapped the spaces an unknown number of times.
 760   TraceTime tm("pre compact", print_phases(), true, gclog_or_tty);
 761   ParallelScavengeHeap* heap = gc_heap();
 762   _space_info[from_space_id].set_space(heap->young_gen()->from_space());
 763   _space_info[to_space_id].set_space(heap->young_gen()->to_space());
 764 
 765   pre_gc_values->fill(heap);
 766 
 767   ParCompactionManager::reset();
 768   NOT_PRODUCT(_mark_bitmap.reset_counters());
 769   DEBUG_ONLY(add_obj_count = add_obj_size = 0;)
 770   DEBUG_ONLY(mark_bitmap_count = mark_bitmap_size = 0;)
 771 
 772   // Increment the invocation count
 773   heap->increment_total_collections(true);
 774 
 775   // We need to track unique mark sweep invocations as well.
 776   _total_invocations++;
 777 
 778   if (PrintHeapAtGC) {
 779     Universe::print_heap_before_gc();
 780   }
 781 
 782   // Fill in TLABs
 783   heap->accumulate_statistics_all_tlabs();
 784   heap->ensure_parsability(true);  // retire TLABs
 785 
 786   if (VerifyBeforeGC && heap->total_collections() >= VerifyGCStartAt) {
 787     HandleMark hm;  // Discard invalid handles created during verification
 788     gclog_or_tty->print(" VerifyBeforeGC:");
 789     Universe::verify(true);
 790   }
 791 
 792   // Verify object start arrays
 793   if (VerifyObjectStartArray &&
 794       VerifyBeforeGC) {
 795     heap->old_gen()->verify_object_start_array();
 796     heap->perm_gen()->verify_object_start_array();
 797   }
 798 
 799   DEBUG_ONLY(mark_bitmap()->verify_clear();)
 800   DEBUG_ONLY(summary_data().verify_clear();)
 801 
 802   // Have worker threads release resources the next time they run a task.
 803   gc_task_manager()->release_all_resources();
 804 }
 805 
 806 void PSParallelCompact::post_compact()
 807 {
 808   TraceTime tm("post compact", print_phases(), true, gclog_or_tty);
 809 
 810   // Clear the marking bitmap and summary data and update top() in each space.
 811   for (unsigned int id = perm_space_id; id < last_space_id; ++id) {
 812     clear_data_covering_space(SpaceId(id));
 813     _space_info[id].space()->set_top(_space_info[id].new_top());
 814   }
 815 
 816   MutableSpace* const eden_space = _space_info[eden_space_id].space();
 817   MutableSpace* const from_space = _space_info[from_space_id].space();
 818   MutableSpace* const to_space   = _space_info[to_space_id].space();
 819 
 820   ParallelScavengeHeap* heap = gc_heap();
 821   bool eden_empty = eden_space->is_empty();
 822   if (!eden_empty) {
 823     eden_empty = absorb_live_data_from_eden(heap->size_policy(),
 824                                             heap->young_gen(), heap->old_gen());
 825   }
 826 
 827   // Update heap occupancy information which is used as input to the soft ref
 828   // clearing policy at the next gc.
 829   Universe::update_heap_info_at_gc();
 830 
 831   bool young_gen_empty = eden_empty && from_space->is_empty() &&
 832     to_space->is_empty();
 833 
 834   BarrierSet* bs = heap->barrier_set();
 835   if (bs->is_a(BarrierSet::ModRef)) {
 836     ModRefBarrierSet* modBS = (ModRefBarrierSet*)bs;
 837     MemRegion old_mr = heap->old_gen()->reserved();
 838     MemRegion perm_mr = heap->perm_gen()->reserved();
 839     assert(perm_mr.end() <= old_mr.start(), "Generations out of order");
 840 
 841     if (young_gen_empty) {
 842       modBS->clear(MemRegion(perm_mr.start(), old_mr.end()));
 843     } else {
 844       modBS->invalidate(MemRegion(perm_mr.start(), old_mr.end()));
 845     }
 846   }
 847 
 848   Threads::gc_epilogue();
 849   CodeCache::gc_epilogue();
 850 
 851   COMPILER2_PRESENT(DerivedPointerTable::update_pointers());
 852 
 853   ref_processor()->enqueue_discovered_references(NULL);
 854 
 855   if (ZapUnusedHeapArea) {
 856     heap->gen_mangle_unused_area();
 857   }
 858 
 859   // Update time of last GC
 860   reset_millis_since_last_gc();
 861 }
 862 
 863 HeapWord*
 864 PSParallelCompact::compute_dense_prefix_via_density(const SpaceId id,
 865                                                     bool maximum_compaction)
 866 {
 867   const size_t region_size = ParallelCompactData::RegionSize;
 868   const ParallelCompactData& sd = summary_data();
 869 
 870   const MutableSpace* const space = _space_info[id].space();
 871   HeapWord* const top_aligned_up = sd.region_align_up(space->top());
 872   const RegionData* const beg_cp = sd.addr_to_region_ptr(space->bottom());
 873   const RegionData* const end_cp = sd.addr_to_region_ptr(top_aligned_up);
 874 
 875   // Skip full regions at the beginning of the space--they are necessarily part
 876   // of the dense prefix.
 877   size_t full_count = 0;
 878   const RegionData* cp;
 879   for (cp = beg_cp; cp < end_cp && cp->data_size() == region_size; ++cp) {
 880     ++full_count;
 881   }
 882 
 883   assert(total_invocations() >= _maximum_compaction_gc_num, "sanity");
 884   const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num;
 885   const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval;
 886   if (maximum_compaction || cp == end_cp || interval_ended) {
 887     _maximum_compaction_gc_num = total_invocations();
 888     return sd.region_to_addr(cp);
 889   }
 890 
 891   HeapWord* const new_top = _space_info[id].new_top();
 892   const size_t space_live = pointer_delta(new_top, space->bottom());
 893   const size_t space_used = space->used_in_words();
 894   const size_t space_capacity = space->capacity_in_words();
 895 
 896   const double cur_density = double(space_live) / space_capacity;
 897   const double deadwood_density =
 898     (1.0 - cur_density) * (1.0 - cur_density) * cur_density * cur_density;
 899   const size_t deadwood_goal = size_t(space_capacity * deadwood_density);
 900 
 901   if (TraceParallelOldGCDensePrefix) {
 902     tty->print_cr("cur_dens=%5.3f dw_dens=%5.3f dw_goal=" SIZE_FORMAT,
 903                   cur_density, deadwood_density, deadwood_goal);
 904     tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " "
 905                   "space_cap=" SIZE_FORMAT,
 906                   space_live, space_used,
 907                   space_capacity);
 908   }
 909 
 910   // XXX - Use binary search?
 911   HeapWord* dense_prefix = sd.region_to_addr(cp);
 912   const RegionData* full_cp = cp;
 913   const RegionData* const top_cp = sd.addr_to_region_ptr(space->top() - 1);
 914   while (cp < end_cp) {
 915     HeapWord* region_destination = cp->destination();
 916     const size_t cur_deadwood = pointer_delta(dense_prefix, region_destination);
 917     if (TraceParallelOldGCDensePrefix && Verbose) {
 918       tty->print_cr("c#=" SIZE_FORMAT_W(4) " dst=" PTR_FORMAT " "
 919                     "dp=" SIZE_FORMAT_W(8) " " "cdw=" SIZE_FORMAT_W(8),
 920                     sd.region(cp), region_destination,
 921                     dense_prefix, cur_deadwood);
 922     }
 923 
 924     if (cur_deadwood >= deadwood_goal) {
 925       // Found the region that has the correct amount of deadwood to the left.
 926       // This typically occurs after crossing a fairly sparse set of regions, so
 927       // iterate backwards over those sparse regions, looking for the region
 928       // that has the lowest density of live objects 'to the right.'
 929       size_t space_to_left = sd.region(cp) * region_size;
 930       size_t live_to_left = space_to_left - cur_deadwood;
 931       size_t space_to_right = space_capacity - space_to_left;
 932       size_t live_to_right = space_live - live_to_left;
 933       double density_to_right = double(live_to_right) / space_to_right;
 934       while (cp > full_cp) {
 935         --cp;
 936         const size_t prev_region_live_to_right = live_to_right -
 937           cp->data_size();
 938         const size_t prev_region_space_to_right = space_to_right + region_size;
 939         double prev_region_density_to_right =
 940           double(prev_region_live_to_right) / prev_region_space_to_right;
 941         if (density_to_right <= prev_region_density_to_right) {
 942           return dense_prefix;
 943         }
 944         if (TraceParallelOldGCDensePrefix && Verbose) {
 945           tty->print_cr("backing up from c=" SIZE_FORMAT_W(4) " d2r=%10.8f "
 946                         "pc_d2r=%10.8f", sd.region(cp), density_to_right,
 947                         prev_region_density_to_right);
 948         }
 949         dense_prefix -= region_size;
 950         live_to_right = prev_region_live_to_right;
 951         space_to_right = prev_region_space_to_right;
 952         density_to_right = prev_region_density_to_right;
 953       }
 954       return dense_prefix;
 955     }
 956 
 957     dense_prefix += region_size;
 958     ++cp;
 959   }
 960 
 961   return dense_prefix;
 962 }
 963 
 964 #ifndef PRODUCT
 965 void PSParallelCompact::print_dense_prefix_stats(const char* const algorithm,
 966                                                  const SpaceId id,
 967                                                  const bool maximum_compaction,
 968                                                  HeapWord* const addr)
 969 {
 970   const size_t region_idx = summary_data().addr_to_region_idx(addr);
 971   RegionData* const cp = summary_data().region(region_idx);
 972   const MutableSpace* const space = _space_info[id].space();
 973   HeapWord* const new_top = _space_info[id].new_top();
 974 
 975   const size_t space_live = pointer_delta(new_top, space->bottom());
 976   const size_t dead_to_left = pointer_delta(addr, cp->destination());
 977   const size_t space_cap = space->capacity_in_words();
 978   const double dead_to_left_pct = double(dead_to_left) / space_cap;
 979   const size_t live_to_right = new_top - cp->destination();
 980   const size_t dead_to_right = space->top() - addr - live_to_right;
 981 
 982   tty->print_cr("%s=" PTR_FORMAT " dpc=" SIZE_FORMAT_W(5) " "
 983                 "spl=" SIZE_FORMAT " "
 984                 "d2l=" SIZE_FORMAT " d2l%%=%6.4f "
 985                 "d2r=" SIZE_FORMAT " l2r=" SIZE_FORMAT
 986                 " ratio=%10.8f",
 987                 algorithm, addr, region_idx,
 988                 space_live,
 989                 dead_to_left, dead_to_left_pct,
 990                 dead_to_right, live_to_right,
 991                 double(dead_to_right) / live_to_right);
 992 }
 993 #endif  // #ifndef PRODUCT
 994 
 995 // Return a fraction indicating how much of the generation can be treated as
 996 // "dead wood" (i.e., not reclaimed).  The function uses a normal distribution
 997 // based on the density of live objects in the generation to determine a limit,
 998 // which is then adjusted so the return value is min_percent when the density is
 999 // 1.
1000 //
1001 // The following table shows some return values for a different values of the
1002 // standard deviation (ParallelOldDeadWoodLimiterStdDev); the mean is 0.5 and
1003 // min_percent is 1.
1004 //
1005 //                          fraction allowed as dead wood
1006 //         -----------------------------------------------------------------
1007 // density std_dev=70 std_dev=75 std_dev=80 std_dev=85 std_dev=90 std_dev=95
1008 // ------- ---------- ---------- ---------- ---------- ---------- ----------
1009 // 0.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000
1010 // 0.05000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941
1011 // 0.10000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272
1012 // 0.15000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066
1013 // 0.20000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975
1014 // 0.25000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313
1015 // 0.30000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132
1016 // 0.35000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289
1017 // 0.40000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500
1018 // 0.45000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386
1019 // 0.50000 0.13832410 0.11599237 0.09847664 0.08456518 0.07338887 0.06431510
1020 // 0.55000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386
1021 // 0.60000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500
1022 // 0.65000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289
1023 // 0.70000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132
1024 // 0.75000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313
1025 // 0.80000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975
1026 // 0.85000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066
1027 // 0.90000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272
1028 // 0.95000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941
1029 // 1.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000
1030 
1031 double PSParallelCompact::dead_wood_limiter(double density, size_t min_percent)
1032 {
1033   assert(_dwl_initialized, "uninitialized");
1034 
1035   // The raw limit is the value of the normal distribution at x = density.
1036   const double raw_limit = normal_distribution(density);
1037 
1038   // Adjust the raw limit so it becomes the minimum when the density is 1.
1039   //
1040   // First subtract the adjustment value (which is simply the precomputed value
1041   // normal_distribution(1.0)); this yields a value of 0 when the density is 1.
1042   // Then add the minimum value, so the minimum is returned when the density is
1043   // 1.  Finally, prevent negative values, which occur when the mean is not 0.5.
1044   const double min = double(min_percent) / 100.0;
1045   const double limit = raw_limit - _dwl_adjustment + min;
1046   return MAX2(limit, 0.0);
1047 }
1048 
1049 ParallelCompactData::RegionData*
1050 PSParallelCompact::first_dead_space_region(const RegionData* beg,
1051                                            const RegionData* end)
1052 {
1053   const size_t region_size = ParallelCompactData::RegionSize;
1054   ParallelCompactData& sd = summary_data();
1055   size_t left = sd.region(beg);
1056   size_t right = end > beg ? sd.region(end) - 1 : left;
1057 
1058   // Binary search.
1059   while (left < right) {
1060     // Equivalent to (left + right) / 2, but does not overflow.
1061     const size_t middle = left + (right - left) / 2;
1062     RegionData* const middle_ptr = sd.region(middle);
1063     HeapWord* const dest = middle_ptr->destination();
1064     HeapWord* const addr = sd.region_to_addr(middle);
1065     assert(dest != NULL, "sanity");
1066     assert(dest <= addr, "must move left");
1067 
1068     if (middle > left && dest < addr) {
1069       right = middle - 1;
1070     } else if (middle < right && middle_ptr->data_size() == region_size) {
1071       left = middle + 1;
1072     } else {
1073       return middle_ptr;
1074     }
1075   }
1076   return sd.region(left);
1077 }
1078 
1079 ParallelCompactData::RegionData*
1080 PSParallelCompact::dead_wood_limit_region(const RegionData* beg,
1081                                           const RegionData* end,
1082                                           size_t dead_words)
1083 {
1084   ParallelCompactData& sd = summary_data();
1085   size_t left = sd.region(beg);
1086   size_t right = end > beg ? sd.region(end) - 1 : left;
1087 
1088   // Binary search.
1089   while (left < right) {
1090     // Equivalent to (left + right) / 2, but does not overflow.
1091     const size_t middle = left + (right - left) / 2;
1092     RegionData* const middle_ptr = sd.region(middle);
1093     HeapWord* const dest = middle_ptr->destination();
1094     HeapWord* const addr = sd.region_to_addr(middle);
1095     assert(dest != NULL, "sanity");
1096     assert(dest <= addr, "must move left");
1097 
1098     const size_t dead_to_left = pointer_delta(addr, dest);
1099     if (middle > left && dead_to_left > dead_words) {
1100       right = middle - 1;
1101     } else if (middle < right && dead_to_left < dead_words) {
1102       left = middle + 1;
1103     } else {
1104       return middle_ptr;
1105     }
1106   }
1107   return sd.region(left);
1108 }
1109 
1110 // The result is valid during the summary phase, after the initial summarization
1111 // of each space into itself, and before final summarization.
1112 inline double
1113 PSParallelCompact::reclaimed_ratio(const RegionData* const cp,
1114                                    HeapWord* const bottom,
1115                                    HeapWord* const top,
1116                                    HeapWord* const new_top)
1117 {
1118   ParallelCompactData& sd = summary_data();
1119 
1120   assert(cp != NULL, "sanity");
1121   assert(bottom != NULL, "sanity");
1122   assert(top != NULL, "sanity");
1123   assert(new_top != NULL, "sanity");
1124   assert(top >= new_top, "summary data problem?");
1125   assert(new_top > bottom, "space is empty; should not be here");
1126   assert(new_top >= cp->destination(), "sanity");
1127   assert(top >= sd.region_to_addr(cp), "sanity");
1128 
1129   HeapWord* const destination = cp->destination();
1130   const size_t dense_prefix_live  = pointer_delta(destination, bottom);
1131   const size_t compacted_region_live = pointer_delta(new_top, destination);
1132   const size_t compacted_region_used = pointer_delta(top,
1133                                                      sd.region_to_addr(cp));
1134   const size_t reclaimable = compacted_region_used - compacted_region_live;
1135 
1136   const double divisor = dense_prefix_live + 1.25 * compacted_region_live;
1137   return double(reclaimable) / divisor;
1138 }
1139 
1140 // Return the address of the end of the dense prefix, a.k.a. the start of the
1141 // compacted region.  The address is always on a region boundary.
1142 //
1143 // Completely full regions at the left are skipped, since no compaction can
1144 // occur in those regions.  Then the maximum amount of dead wood to allow is
1145 // computed, based on the density (amount live / capacity) of the generation;
1146 // the region with approximately that amount of dead space to the left is
1147 // identified as the limit region.  Regions between the last completely full
1148 // region and the limit region are scanned and the one that has the best
1149 // (maximum) reclaimed_ratio() is selected.
1150 HeapWord*
1151 PSParallelCompact::compute_dense_prefix(const SpaceId id,
1152                                         bool maximum_compaction)
1153 {
1154   const size_t region_size = ParallelCompactData::RegionSize;
1155   const ParallelCompactData& sd = summary_data();
1156 
1157   const MutableSpace* const space = _space_info[id].space();
1158   HeapWord* const top = space->top();
1159   HeapWord* const top_aligned_up = sd.region_align_up(top);
1160   HeapWord* const new_top = _space_info[id].new_top();
1161   HeapWord* const new_top_aligned_up = sd.region_align_up(new_top);
1162   HeapWord* const bottom = space->bottom();
1163   const RegionData* const beg_cp = sd.addr_to_region_ptr(bottom);
1164   const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up);
1165   const RegionData* const new_top_cp =
1166     sd.addr_to_region_ptr(new_top_aligned_up);
1167 
1168   // Skip full regions at the beginning of the space--they are necessarily part
1169   // of the dense prefix.
1170   const RegionData* const full_cp = first_dead_space_region(beg_cp, new_top_cp);
1171   assert(full_cp->destination() == sd.region_to_addr(full_cp) ||
1172          space->is_empty(), "no dead space allowed to the left");
1173   assert(full_cp->data_size() < region_size || full_cp == new_top_cp - 1,
1174          "region must have dead space");
1175 
1176   // The gc number is saved whenever a maximum compaction is done, and used to
1177   // determine when the maximum compaction interval has expired.  This avoids
1178   // successive max compactions for different reasons.
1179   assert(total_invocations() >= _maximum_compaction_gc_num, "sanity");
1180   const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num;
1181   const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval ||
1182     total_invocations() == HeapFirstMaximumCompactionCount;
1183   if (maximum_compaction || full_cp == top_cp || interval_ended) {
1184     _maximum_compaction_gc_num = total_invocations();
1185     return sd.region_to_addr(full_cp);
1186   }
1187 
1188   const size_t space_live = pointer_delta(new_top, bottom);
1189   const size_t space_used = space->used_in_words();
1190   const size_t space_capacity = space->capacity_in_words();
1191 
1192   const double density = double(space_live) / double(space_capacity);
1193   const size_t min_percent_free =
1194           id == perm_space_id ? PermMarkSweepDeadRatio : MarkSweepDeadRatio;
1195   const double limiter = dead_wood_limiter(density, min_percent_free);
1196   const size_t dead_wood_max = space_used - space_live;
1197   const size_t dead_wood_limit = MIN2(size_t(space_capacity * limiter),
1198                                       dead_wood_max);
1199 
1200   if (TraceParallelOldGCDensePrefix) {
1201     tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " "
1202                   "space_cap=" SIZE_FORMAT,
1203                   space_live, space_used,
1204                   space_capacity);
1205     tty->print_cr("dead_wood_limiter(%6.4f, %d)=%6.4f "
1206                   "dead_wood_max=" SIZE_FORMAT " dead_wood_limit=" SIZE_FORMAT,
1207                   density, min_percent_free, limiter,
1208                   dead_wood_max, dead_wood_limit);
1209   }
1210 
1211   // Locate the region with the desired amount of dead space to the left.
1212   const RegionData* const limit_cp =
1213     dead_wood_limit_region(full_cp, top_cp, dead_wood_limit);
1214 
1215   // Scan from the first region with dead space to the limit region and find the
1216   // one with the best (largest) reclaimed ratio.
1217   double best_ratio = 0.0;
1218   const RegionData* best_cp = full_cp;
1219   for (const RegionData* cp = full_cp; cp < limit_cp; ++cp) {
1220     double tmp_ratio = reclaimed_ratio(cp, bottom, top, new_top);
1221     if (tmp_ratio > best_ratio) {
1222       best_cp = cp;
1223       best_ratio = tmp_ratio;
1224     }
1225   }
1226 
1227 #if     0
1228   // Something to consider:  if the region with the best ratio is 'close to' the
1229   // first region w/free space, choose the first region with free space
1230   // ("first-free").  The first-free region is usually near the start of the
1231   // heap, which means we are copying most of the heap already, so copy a bit
1232   // more to get complete compaction.
1233   if (pointer_delta(best_cp, full_cp, sizeof(RegionData)) < 4) {
1234     _maximum_compaction_gc_num = total_invocations();
1235     best_cp = full_cp;
1236   }
1237 #endif  // #if 0
1238 
1239   return sd.region_to_addr(best_cp);
1240 }
1241 
1242 void PSParallelCompact::summarize_spaces_quick()
1243 {
1244   for (unsigned int i = 0; i < last_space_id; ++i) {
1245     const MutableSpace* space = _space_info[i].space();
1246     bool result = _summary_data.summarize(space->bottom(), space->end(),
1247                                           space->bottom(), space->top(),
1248                                           _space_info[i].new_top_addr());
1249     assert(result, "should never fail");
1250     _space_info[i].set_dense_prefix(space->bottom());
1251   }
1252 }
1253 
1254 void PSParallelCompact::fill_dense_prefix_end(SpaceId id)
1255 {
1256   HeapWord* const dense_prefix_end = dense_prefix(id);
1257   const RegionData* region = _summary_data.addr_to_region_ptr(dense_prefix_end);
1258   const idx_t dense_prefix_bit = _mark_bitmap.addr_to_bit(dense_prefix_end);
1259   if (dead_space_crosses_boundary(region, dense_prefix_bit)) {
1260     // Only enough dead space is filled so that any remaining dead space to the
1261     // left is larger than the minimum filler object.  (The remainder is filled
1262     // during the copy/update phase.)
1263     //
1264     // The size of the dead space to the right of the boundary is not a
1265     // concern, since compaction will be able to use whatever space is
1266     // available.
1267     //
1268     // Here '||' is the boundary, 'x' represents a don't care bit and a box
1269     // surrounds the space to be filled with an object.
1270     //
1271     // In the 32-bit VM, each bit represents two 32-bit words:
1272     //                              +---+
1273     // a) beg_bits:  ...  x   x   x | 0 | ||   0   x  x  ...
1274     //    end_bits:  ...  x   x   x | 0 | ||   0   x  x  ...
1275     //                              +---+
1276     //
1277     // In the 64-bit VM, each bit represents one 64-bit word:
1278     //                              +------------+
1279     // b) beg_bits:  ...  x   x   x | 0   ||   0 | x  x  ...
1280     //    end_bits:  ...  x   x   1 | 0   ||   0 | x  x  ...
1281     //                              +------------+
1282     //                          +-------+
1283     // c) beg_bits:  ...  x   x | 0   0 | ||   0   x  x  ...
1284     //    end_bits:  ...  x   1 | 0   0 | ||   0   x  x  ...
1285     //                          +-------+
1286     //                      +-----------+
1287     // d) beg_bits:  ...  x | 0   0   0 | ||   0   x  x  ...
1288     //    end_bits:  ...  1 | 0   0   0 | ||   0   x  x  ...
1289     //                      +-----------+
1290     //                          +-------+
1291     // e) beg_bits:  ...  0   0 | 0   0 | ||   0   x  x  ...
1292     //    end_bits:  ...  0   0 | 0   0 | ||   0   x  x  ...
1293     //                          +-------+
1294 
1295     // Initially assume case a, c or e will apply.
1296     size_t obj_len = (size_t)oopDesc::header_size();
1297     HeapWord* obj_beg = dense_prefix_end - obj_len;
1298 
1299 #ifdef  _LP64
1300     if (_mark_bitmap.is_obj_end(dense_prefix_bit - 2)) {
1301       // Case b above.
1302       obj_beg = dense_prefix_end - 1;
1303     } else if (!_mark_bitmap.is_obj_end(dense_prefix_bit - 3) &&
1304                _mark_bitmap.is_obj_end(dense_prefix_bit - 4)) {
1305       // Case d above.
1306       obj_beg = dense_prefix_end - 3;
1307       obj_len = 3;
1308     }
1309 #endif  // #ifdef _LP64
1310 
1311     MemRegion region(obj_beg, obj_len);
1312     SharedHeap::fill_region_with_object(region);
1313     _mark_bitmap.mark_obj(obj_beg, obj_len);
1314     _summary_data.add_obj(obj_beg, obj_len);
1315     assert(start_array(id) != NULL, "sanity");
1316     start_array(id)->allocate_block(obj_beg);
1317   }
1318 }
1319 
1320 void
1321 PSParallelCompact::summarize_space(SpaceId id, bool maximum_compaction)
1322 {
1323   assert(id < last_space_id, "id out of range");
1324   assert(_space_info[id].dense_prefix() == _space_info[id].space()->bottom(),
1325          "should have been set in summarize_spaces_quick()");
1326 
1327   const MutableSpace* space = _space_info[id].space();
1328   if (_space_info[id].new_top() != space->bottom()) {
1329     HeapWord* dense_prefix_end = compute_dense_prefix(id, maximum_compaction);
1330     _space_info[id].set_dense_prefix(dense_prefix_end);
1331 
1332 #ifndef PRODUCT
1333     if (TraceParallelOldGCDensePrefix) {
1334       print_dense_prefix_stats("ratio", id, maximum_compaction,
1335                                dense_prefix_end);
1336       HeapWord* addr = compute_dense_prefix_via_density(id, maximum_compaction);
1337       print_dense_prefix_stats("density", id, maximum_compaction, addr);
1338     }
1339 #endif  // #ifndef PRODUCT
1340 
1341     // If dead space crosses the dense prefix boundary, it is (at least
1342     // partially) filled with a dummy object, marked live and added to the
1343     // summary data.  This simplifies the copy/update phase and must be done
1344     // before the final locations of objects are determined, to prevent leaving
1345     // a fragment of dead space that is too small to fill with an object.
1346     if (!maximum_compaction && dense_prefix_end != space->bottom()) {
1347       fill_dense_prefix_end(id);
1348     }
1349 
1350     // Compute the destination of each Region, and thus each object.
1351     _summary_data.summarize_dense_prefix(space->bottom(), dense_prefix_end);
1352     _summary_data.summarize(dense_prefix_end, space->end(),
1353                             dense_prefix_end, space->top(),
1354                             _space_info[id].new_top_addr());
1355   }
1356 
1357   if (TraceParallelOldGCSummaryPhase) {
1358     const size_t region_size = ParallelCompactData::RegionSize;
1359     HeapWord* const dense_prefix_end = _space_info[id].dense_prefix();
1360     const size_t dp_region = _summary_data.addr_to_region_idx(dense_prefix_end);
1361     const size_t dp_words = pointer_delta(dense_prefix_end, space->bottom());
1362     HeapWord* const new_top = _space_info[id].new_top();
1363     const HeapWord* nt_aligned_up = _summary_data.region_align_up(new_top);
1364     const size_t cr_words = pointer_delta(nt_aligned_up, dense_prefix_end);
1365     tty->print_cr("id=%d cap=" SIZE_FORMAT " dp=" PTR_FORMAT " "
1366                   "dp_region=" SIZE_FORMAT " " "dp_count=" SIZE_FORMAT " "
1367                   "cr_count=" SIZE_FORMAT " " "nt=" PTR_FORMAT,
1368                   id, space->capacity_in_words(), dense_prefix_end,
1369                   dp_region, dp_words / region_size,
1370                   cr_words / region_size, new_top);
1371   }
1372 }
1373 
1374 void PSParallelCompact::summary_phase(ParCompactionManager* cm,
1375                                       bool maximum_compaction)
1376 {
1377   EventMark m("2 summarize");
1378   TraceTime tm("summary phase", print_phases(), true, gclog_or_tty);
1379   // trace("2");
1380 
1381 #ifdef  ASSERT
1382   if (TraceParallelOldGCMarkingPhase) {
1383     tty->print_cr("add_obj_count=" SIZE_FORMAT " "
1384                   "add_obj_bytes=" SIZE_FORMAT,
1385                   add_obj_count, add_obj_size * HeapWordSize);
1386     tty->print_cr("mark_bitmap_count=" SIZE_FORMAT " "
1387                   "mark_bitmap_bytes=" SIZE_FORMAT,
1388                   mark_bitmap_count, mark_bitmap_size * HeapWordSize);
1389   }
1390 #endif  // #ifdef ASSERT
1391 
1392   // Quick summarization of each space into itself, to see how much is live.
1393   summarize_spaces_quick();
1394 
1395   if (TraceParallelOldGCSummaryPhase) {
1396     tty->print_cr("summary_phase:  after summarizing each space to self");
1397     Universe::print();
1398     NOT_PRODUCT(print_region_ranges());
1399     if (Verbose) {
1400       NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info));
1401     }
1402   }
1403 
1404   // The amount of live data that will end up in old space (assuming it fits).
1405   size_t old_space_total_live = 0;
1406   unsigned int id;
1407   for (id = old_space_id; id < last_space_id; ++id) {
1408     old_space_total_live += pointer_delta(_space_info[id].new_top(),
1409                                           _space_info[id].space()->bottom());
1410   }
1411 
1412   const MutableSpace* old_space = _space_info[old_space_id].space();
1413   if (old_space_total_live > old_space->capacity_in_words()) {
1414     // XXX - should also try to expand
1415     maximum_compaction = true;
1416   } else if (!UseParallelOldGCDensePrefix) {
1417     maximum_compaction = true;
1418   }
1419 
1420   // Permanent and Old generations.
1421   summarize_space(perm_space_id, maximum_compaction);
1422   summarize_space(old_space_id, maximum_compaction);
1423 
1424   // Summarize the remaining spaces (those in the young gen) into old space.  If
1425   // the live data from a space doesn't fit, the existing summarization is left
1426   // intact, so the data is compacted down within the space itself.
1427   HeapWord** new_top_addr = _space_info[old_space_id].new_top_addr();
1428   HeapWord* const target_space_end = old_space->end();
1429   for (id = eden_space_id; id < last_space_id; ++id) {
1430     const MutableSpace* space = _space_info[id].space();
1431     const size_t live = pointer_delta(_space_info[id].new_top(),
1432                                       space->bottom());
1433     const size_t available = pointer_delta(target_space_end, *new_top_addr);
1434     if (live > 0 && live <= available) {
1435       // All the live data will fit.
1436       if (TraceParallelOldGCSummaryPhase) {
1437         tty->print_cr("summarizing %d into old_space @ " PTR_FORMAT,
1438                       id, *new_top_addr);
1439       }
1440       _summary_data.summarize(*new_top_addr, target_space_end,
1441                               space->bottom(), space->top(),
1442                               new_top_addr);
1443 
1444       // Clear the source_region field for each region in the space.
1445       HeapWord* const new_top = _space_info[id].new_top();
1446       HeapWord* const clear_end = _summary_data.region_align_up(new_top);
1447       RegionData* beg_region =
1448         _summary_data.addr_to_region_ptr(space->bottom());
1449       RegionData* end_region = _summary_data.addr_to_region_ptr(clear_end);
1450       while (beg_region < end_region) {
1451         beg_region->set_source_region(0);
1452         ++beg_region;
1453       }
1454 
1455       // Reset the new_top value for the space.
1456       _space_info[id].set_new_top(space->bottom());
1457     }
1458   }
1459 
1460   if (TraceParallelOldGCSummaryPhase) {
1461     tty->print_cr("summary_phase:  after final summarization");
1462     Universe::print();
1463     NOT_PRODUCT(print_region_ranges());
1464     if (Verbose) {
1465       NOT_PRODUCT(print_generic_summary_data(_summary_data, _space_info));
1466     }
1467   }
1468 }
1469 
1470 // This method should contain all heap-specific policy for invoking a full
1471 // collection.  invoke_no_policy() will only attempt to compact the heap; it
1472 // will do nothing further.  If we need to bail out for policy reasons, scavenge
1473 // before full gc, or any other specialized behavior, it needs to be added here.
1474 //
1475 // Note that this method should only be called from the vm_thread while at a
1476 // safepoint.
1477 void PSParallelCompact::invoke(bool maximum_heap_compaction) {
1478   assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
1479   assert(Thread::current() == (Thread*)VMThread::vm_thread(),
1480          "should be in vm thread");
1481   ParallelScavengeHeap* heap = gc_heap();
1482   GCCause::Cause gc_cause = heap->gc_cause();
1483   assert(!heap->is_gc_active(), "not reentrant");
1484 
1485   PSAdaptiveSizePolicy* policy = heap->size_policy();
1486 
1487   // Before each allocation/collection attempt, find out from the
1488   // policy object if GCs are, on the whole, taking too long. If so,
1489   // bail out without attempting a collection.  The exceptions are
1490   // for explicitly requested GC's.
1491   if (!policy->gc_time_limit_exceeded() ||
1492       GCCause::is_user_requested_gc(gc_cause) ||
1493       GCCause::is_serviceability_requested_gc(gc_cause)) {
1494     IsGCActiveMark mark;
1495 
1496     if (ScavengeBeforeFullGC) {
1497       PSScavenge::invoke_no_policy();
1498     }
1499 
1500     PSParallelCompact::invoke_no_policy(maximum_heap_compaction);
1501   }
1502 }
1503 
1504 bool ParallelCompactData::region_contains(size_t region_index, HeapWord* addr) {
1505   size_t addr_region_index = addr_to_region_idx(addr);
1506   return region_index == addr_region_index;
1507 }
1508 
1509 // This method contains no policy. You should probably
1510 // be calling invoke() instead.
1511 void PSParallelCompact::invoke_no_policy(bool maximum_heap_compaction) {
1512   assert(SafepointSynchronize::is_at_safepoint(), "must be at a safepoint");
1513   assert(ref_processor() != NULL, "Sanity");
1514 
1515   if (GC_locker::check_active_before_gc()) {
1516     return;
1517   }
1518 
1519   TimeStamp marking_start;
1520   TimeStamp compaction_start;
1521   TimeStamp collection_exit;
1522 
1523   ParallelScavengeHeap* heap = gc_heap();
1524   GCCause::Cause gc_cause = heap->gc_cause();
1525   PSYoungGen* young_gen = heap->young_gen();
1526   PSOldGen* old_gen = heap->old_gen();
1527   PSPermGen* perm_gen = heap->perm_gen();
1528   PSAdaptiveSizePolicy* size_policy = heap->size_policy();
1529 
1530   if (ZapUnusedHeapArea) {
1531     // Save information needed to minimize mangling
1532     heap->record_gen_tops_before_GC();
1533   }
1534 
1535   _print_phases = PrintGCDetails && PrintParallelOldGCPhaseTimes;
1536 
1537   // Make sure data structures are sane, make the heap parsable, and do other
1538   // miscellaneous bookkeeping.
1539   PreGCValues pre_gc_values;
1540   pre_compact(&pre_gc_values);
1541 
1542   // Get the compaction manager reserved for the VM thread.
1543   ParCompactionManager* const vmthread_cm =
1544     ParCompactionManager::manager_array(gc_task_manager()->workers());
1545 
1546   // Place after pre_compact() where the number of invocations is incremented.
1547   AdaptiveSizePolicyOutput(size_policy, heap->total_collections());
1548 
1549   {
1550     ResourceMark rm;
1551     HandleMark hm;
1552 
1553     const bool is_system_gc = gc_cause == GCCause::_java_lang_system_gc;
1554 
1555     // This is useful for debugging but don't change the output the
1556     // the customer sees.
1557     const char* gc_cause_str = "Full GC";
1558     if (is_system_gc && PrintGCDetails) {
1559       gc_cause_str = "Full GC (System)";
1560     }
1561     gclog_or_tty->date_stamp(PrintGC && PrintGCDateStamps);
1562     TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
1563     TraceTime t1(gc_cause_str, PrintGC, !PrintGCDetails, gclog_or_tty);
1564     TraceCollectorStats tcs(counters());
1565     TraceMemoryManagerStats tms(true /* Full GC */);
1566 
1567     if (TraceGen1Time) accumulated_time()->start();
1568 
1569     // Let the size policy know we're starting
1570     size_policy->major_collection_begin();
1571 
1572     // When collecting the permanent generation methodOops may be moving,
1573     // so we either have to flush all bcp data or convert it into bci.
1574     CodeCache::gc_prologue();
1575     Threads::gc_prologue();
1576 
1577     NOT_PRODUCT(ref_processor()->verify_no_references_recorded());
1578     COMPILER2_PRESENT(DerivedPointerTable::clear());
1579 
1580     ref_processor()->enable_discovery();
1581     ref_processor()->snap_policy(maximum_heap_compaction);
1582 
1583     bool marked_for_unloading = false;
1584 
1585     marking_start.update();
1586     marking_phase(vmthread_cm, maximum_heap_compaction);
1587 
1588 #ifndef PRODUCT
1589     if (TraceParallelOldGCMarkingPhase) {
1590       gclog_or_tty->print_cr("marking_phase: cas_tries %d  cas_retries %d "
1591         "cas_by_another %d",
1592         mark_bitmap()->cas_tries(), mark_bitmap()->cas_retries(),
1593         mark_bitmap()->cas_by_another());
1594     }
1595 #endif  // #ifndef PRODUCT
1596 
1597     bool max_on_system_gc = UseMaximumCompactionOnSystemGC && is_system_gc;
1598     summary_phase(vmthread_cm, maximum_heap_compaction || max_on_system_gc);
1599 
1600     COMPILER2_PRESENT(assert(DerivedPointerTable::is_active(), "Sanity"));
1601     COMPILER2_PRESENT(DerivedPointerTable::set_active(false));
1602 
1603     // adjust_roots() updates Universe::_intArrayKlassObj which is
1604     // needed by the compaction for filling holes in the dense prefix.
1605     adjust_roots();
1606 
1607     compaction_start.update();
1608     // Does the perm gen always have to be done serially because
1609     // klasses are used in the update of an object?
1610     compact_perm(vmthread_cm);
1611 
1612     if (UseParallelOldGCCompacting) {
1613       compact();
1614     } else {
1615       compact_serial(vmthread_cm);
1616     }
1617 
1618     // Reset the mark bitmap, summary data, and do other bookkeeping.  Must be
1619     // done before resizing.
1620     post_compact();
1621 
1622     // Let the size policy know we're done
1623     size_policy->major_collection_end(old_gen->used_in_bytes(), gc_cause);
1624 
1625     if (UseAdaptiveSizePolicy) {
1626       if (PrintAdaptiveSizePolicy) {
1627         gclog_or_tty->print("AdaptiveSizeStart: ");
1628         gclog_or_tty->stamp();
1629         gclog_or_tty->print_cr(" collection: %d ",
1630                        heap->total_collections());
1631         if (Verbose) {
1632           gclog_or_tty->print("old_gen_capacity: %d young_gen_capacity: %d"
1633             " perm_gen_capacity: %d ",
1634             old_gen->capacity_in_bytes(), young_gen->capacity_in_bytes(),
1635             perm_gen->capacity_in_bytes());
1636         }
1637       }
1638 
1639       // Don't check if the size_policy is ready here.  Let
1640       // the size_policy check that internally.
1641       if (UseAdaptiveGenerationSizePolicyAtMajorCollection &&
1642           ((gc_cause != GCCause::_java_lang_system_gc) ||
1643             UseAdaptiveSizePolicyWithSystemGC)) {
1644         // Calculate optimal free space amounts
1645         assert(young_gen->max_size() >
1646           young_gen->from_space()->capacity_in_bytes() +
1647           young_gen->to_space()->capacity_in_bytes(),
1648           "Sizes of space in young gen are out-of-bounds");
1649         size_t max_eden_size = young_gen->max_size() -
1650           young_gen->from_space()->capacity_in_bytes() -
1651           young_gen->to_space()->capacity_in_bytes();
1652         size_policy->compute_generation_free_space(
1653                               young_gen->used_in_bytes(),
1654                               young_gen->eden_space()->used_in_bytes(),
1655                               old_gen->used_in_bytes(),
1656                               perm_gen->used_in_bytes(),
1657                               young_gen->eden_space()->capacity_in_bytes(),
1658                               old_gen->max_gen_size(),
1659                               max_eden_size,
1660                               true /* full gc*/,
1661                               gc_cause);
1662 
1663         heap->resize_old_gen(
1664           size_policy->calculated_old_free_size_in_bytes());
1665 
1666         // Don't resize the young generation at an major collection.  A
1667         // desired young generation size may have been calculated but
1668         // resizing the young generation complicates the code because the
1669         // resizing of the old generation may have moved the boundary
1670         // between the young generation and the old generation.  Let the
1671         // young generation resizing happen at the minor collections.
1672       }
1673       if (PrintAdaptiveSizePolicy) {
1674         gclog_or_tty->print_cr("AdaptiveSizeStop: collection: %d ",
1675                        heap->total_collections());
1676       }
1677     }
1678 
1679     if (UsePerfData) {
1680       PSGCAdaptivePolicyCounters* const counters = heap->gc_policy_counters();
1681       counters->update_counters();
1682       counters->update_old_capacity(old_gen->capacity_in_bytes());
1683       counters->update_young_capacity(young_gen->capacity_in_bytes());
1684     }
1685 
1686     heap->resize_all_tlabs();
1687 
1688     // We collected the perm gen, so we'll resize it here.
1689     perm_gen->compute_new_size(pre_gc_values.perm_gen_used());
1690 
1691     if (TraceGen1Time) accumulated_time()->stop();
1692 
1693     if (PrintGC) {
1694       if (PrintGCDetails) {
1695         // No GC timestamp here.  This is after GC so it would be confusing.
1696         young_gen->print_used_change(pre_gc_values.young_gen_used());
1697         old_gen->print_used_change(pre_gc_values.old_gen_used());
1698         heap->print_heap_change(pre_gc_values.heap_used());
1699         // Print perm gen last (print_heap_change() excludes the perm gen).
1700         perm_gen->print_used_change(pre_gc_values.perm_gen_used());
1701       } else {
1702         heap->print_heap_change(pre_gc_values.heap_used());
1703       }
1704     }
1705 
1706     // Track memory usage and detect low memory
1707     MemoryService::track_memory_usage();
1708     heap->update_counters();
1709 
1710     if (PrintGCDetails) {
1711       if (size_policy->print_gc_time_limit_would_be_exceeded()) {
1712         if (size_policy->gc_time_limit_exceeded()) {
1713           gclog_or_tty->print_cr("      GC time is exceeding GCTimeLimit "
1714             "of %d%%", GCTimeLimit);
1715         } else {
1716           gclog_or_tty->print_cr("      GC time would exceed GCTimeLimit "
1717             "of %d%%", GCTimeLimit);
1718         }
1719       }
1720       size_policy->set_print_gc_time_limit_would_be_exceeded(false);
1721     }
1722   }
1723 
1724   if (VerifyAfterGC && heap->total_collections() >= VerifyGCStartAt) {
1725     HandleMark hm;  // Discard invalid handles created during verification
1726     gclog_or_tty->print(" VerifyAfterGC:");
1727     Universe::verify(false);
1728   }
1729 
1730   // Re-verify object start arrays
1731   if (VerifyObjectStartArray &&
1732       VerifyAfterGC) {
1733     old_gen->verify_object_start_array();
1734     perm_gen->verify_object_start_array();
1735   }
1736 
1737   if (ZapUnusedHeapArea) {
1738     old_gen->object_space()->check_mangled_unused_area_complete();
1739     perm_gen->object_space()->check_mangled_unused_area_complete();
1740   }
1741 
1742   NOT_PRODUCT(ref_processor()->verify_no_references_recorded());
1743 
1744   collection_exit.update();
1745 
1746   if (PrintHeapAtGC) {
1747     Universe::print_heap_after_gc();
1748   }
1749   if (PrintGCTaskTimeStamps) {
1750     gclog_or_tty->print_cr("VM-Thread " INT64_FORMAT " " INT64_FORMAT " "
1751                            INT64_FORMAT,
1752                            marking_start.ticks(), compaction_start.ticks(),
1753                            collection_exit.ticks());
1754     gc_task_manager()->print_task_time_stamps();
1755   }
1756 }
1757 
1758 bool PSParallelCompact::absorb_live_data_from_eden(PSAdaptiveSizePolicy* size_policy,
1759                                              PSYoungGen* young_gen,
1760                                              PSOldGen* old_gen) {
1761   MutableSpace* const eden_space = young_gen->eden_space();
1762   assert(!eden_space->is_empty(), "eden must be non-empty");
1763   assert(young_gen->virtual_space()->alignment() ==
1764          old_gen->virtual_space()->alignment(), "alignments do not match");
1765 
1766   if (!(UseAdaptiveSizePolicy && UseAdaptiveGCBoundary)) {
1767     return false;
1768   }
1769 
1770   // Both generations must be completely committed.
1771   if (young_gen->virtual_space()->uncommitted_size() != 0) {
1772     return false;
1773   }
1774   if (old_gen->virtual_space()->uncommitted_size() != 0) {
1775     return false;
1776   }
1777 
1778   // Figure out how much to take from eden.  Include the average amount promoted
1779   // in the total; otherwise the next young gen GC will simply bail out to a
1780   // full GC.
1781   const size_t alignment = old_gen->virtual_space()->alignment();
1782   const size_t eden_used = eden_space->used_in_bytes();
1783   const size_t promoted = (size_t)size_policy->avg_promoted()->padded_average();
1784   const size_t absorb_size = align_size_up(eden_used + promoted, alignment);
1785   const size_t eden_capacity = eden_space->capacity_in_bytes();
1786 
1787   if (absorb_size >= eden_capacity) {
1788     return false; // Must leave some space in eden.
1789   }
1790 
1791   const size_t new_young_size = young_gen->capacity_in_bytes() - absorb_size;
1792   if (new_young_size < young_gen->min_gen_size()) {
1793     return false; // Respect young gen minimum size.
1794   }
1795 
1796   if (TraceAdaptiveGCBoundary && Verbose) {
1797     gclog_or_tty->print(" absorbing " SIZE_FORMAT "K:  "
1798                         "eden " SIZE_FORMAT "K->" SIZE_FORMAT "K "
1799                         "from " SIZE_FORMAT "K, to " SIZE_FORMAT "K "
1800                         "young_gen " SIZE_FORMAT "K->" SIZE_FORMAT "K ",
1801                         absorb_size / K,
1802                         eden_capacity / K, (eden_capacity - absorb_size) / K,
1803                         young_gen->from_space()->used_in_bytes() / K,
1804                         young_gen->to_space()->used_in_bytes() / K,
1805                         young_gen->capacity_in_bytes() / K, new_young_size / K);
1806   }
1807 
1808   // Fill the unused part of the old gen.
1809   MutableSpace* const old_space = old_gen->object_space();
1810   MemRegion old_gen_unused(old_space->top(), old_space->end());
1811   if (!old_gen_unused.is_empty()) {
1812     SharedHeap::fill_region_with_object(old_gen_unused);
1813   }
1814 
1815   // Take the live data from eden and set both top and end in the old gen to
1816   // eden top.  (Need to set end because reset_after_change() mangles the region
1817   // from end to virtual_space->high() in debug builds).
1818   HeapWord* const new_top = eden_space->top();
1819   old_gen->virtual_space()->expand_into(young_gen->virtual_space(),
1820                                         absorb_size);
1821   young_gen->reset_after_change();
1822   old_space->set_top(new_top);
1823   old_space->set_end(new_top);
1824   old_gen->reset_after_change();
1825 
1826   // Update the object start array for the filler object and the data from eden.
1827   ObjectStartArray* const start_array = old_gen->start_array();
1828   HeapWord* const start = old_gen_unused.start();
1829   for (HeapWord* addr = start; addr < new_top; addr += oop(addr)->size()) {
1830     start_array->allocate_block(addr);
1831   }
1832 
1833   // Could update the promoted average here, but it is not typically updated at
1834   // full GCs and the value to use is unclear.  Something like
1835   //
1836   // cur_promoted_avg + absorb_size / number_of_scavenges_since_last_full_gc.
1837 
1838   size_policy->set_bytes_absorbed_from_eden(absorb_size);
1839   return true;
1840 }
1841 
1842 GCTaskManager* const PSParallelCompact::gc_task_manager() {
1843   assert(ParallelScavengeHeap::gc_task_manager() != NULL,
1844     "shouldn't return NULL");
1845   return ParallelScavengeHeap::gc_task_manager();
1846 }
1847 
1848 void PSParallelCompact::marking_phase(ParCompactionManager* cm,
1849                                       bool maximum_heap_compaction) {
1850   // Recursively traverse all live objects and mark them
1851   EventMark m("1 mark object");
1852   TraceTime tm("marking phase", print_phases(), true, gclog_or_tty);
1853 
1854   ParallelScavengeHeap* heap = gc_heap();
1855   uint parallel_gc_threads = heap->gc_task_manager()->workers();
1856   TaskQueueSetSuper* qset = ParCompactionManager::region_array();
1857   ParallelTaskTerminator terminator(parallel_gc_threads, qset);
1858 
1859   PSParallelCompact::MarkAndPushClosure mark_and_push_closure(cm);
1860   PSParallelCompact::FollowStackClosure follow_stack_closure(cm);
1861 
1862   {
1863     TraceTime tm_m("par mark", print_phases(), true, gclog_or_tty);
1864 
1865     GCTaskQueue* q = GCTaskQueue::create();
1866 
1867     q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::universe));
1868     q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jni_handles));
1869     // We scan the thread roots in parallel
1870     Threads::create_thread_roots_marking_tasks(q);
1871     q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::object_synchronizer));
1872     q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::flat_profiler));
1873     q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::management));
1874     q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::system_dictionary));
1875     q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jvmti));
1876     q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::vm_symbols));
1877 
1878     if (parallel_gc_threads > 1) {
1879       for (uint j = 0; j < parallel_gc_threads; j++) {
1880         q->enqueue(new StealMarkingTask(&terminator));
1881       }
1882     }
1883 
1884     WaitForBarrierGCTask* fin = WaitForBarrierGCTask::create();
1885     q->enqueue(fin);
1886 
1887     gc_task_manager()->add_list(q);
1888 
1889     fin->wait_for();
1890 
1891     // We have to release the barrier tasks!
1892     WaitForBarrierGCTask::destroy(fin);
1893   }
1894 
1895   // Process reference objects found during marking
1896   {
1897     TraceTime tm_r("reference processing", print_phases(), true, gclog_or_tty);
1898     if (ref_processor()->processing_is_mt()) {
1899       RefProcTaskExecutor task_executor;
1900       ref_processor()->process_discovered_references(
1901         is_alive_closure(), &mark_and_push_closure, &follow_stack_closure, &task_executor);
1902     } else {
1903       ref_processor()->process_discovered_references(
1904         is_alive_closure(), &mark_and_push_closure, &follow_stack_closure, NULL);
1905     }
1906   }
1907 
1908   TraceTime tm_c("class unloading", print_phases(), true, gclog_or_tty);
1909   // Follow system dictionary roots and unload classes.
1910   bool purged_class = SystemDictionary::do_unloading(is_alive_closure());
1911 
1912   // Follow code cache roots.
1913   CodeCache::do_unloading(is_alive_closure(), &mark_and_push_closure,
1914                           purged_class);
1915   follow_stack(cm); // Flush marking stack.
1916 
1917   // Update subklass/sibling/implementor links of live klasses
1918   // revisit_klass_stack is used in follow_weak_klass_links().
1919   follow_weak_klass_links(cm);
1920 
1921   // Visit symbol and interned string tables and delete unmarked oops
1922   SymbolTable::unlink(is_alive_closure());
1923   StringTable::unlink(is_alive_closure());
1924 
1925   assert(cm->marking_stack()->size() == 0, "stack should be empty by now");
1926   assert(cm->overflow_stack()->is_empty(), "stack should be empty by now");
1927 }
1928 
1929 // This should be moved to the shared markSweep code!
1930 class PSAlwaysTrueClosure: public BoolObjectClosure {
1931 public:
1932   void do_object(oop p) { ShouldNotReachHere(); }
1933   bool do_object_b(oop p) { return true; }
1934 };
1935 static PSAlwaysTrueClosure always_true;
1936 
1937 void PSParallelCompact::adjust_roots() {
1938   // Adjust the pointers to reflect the new locations
1939   EventMark m("3 adjust roots");
1940   TraceTime tm("adjust roots", print_phases(), true, gclog_or_tty);
1941 
1942   // General strong roots.
1943   Universe::oops_do(adjust_root_pointer_closure());
1944   ReferenceProcessor::oops_do(adjust_root_pointer_closure());
1945   JNIHandles::oops_do(adjust_root_pointer_closure());   // Global (strong) JNI handles
1946   Threads::oops_do(adjust_root_pointer_closure());
1947   ObjectSynchronizer::oops_do(adjust_root_pointer_closure());
1948   FlatProfiler::oops_do(adjust_root_pointer_closure());
1949   Management::oops_do(adjust_root_pointer_closure());
1950   JvmtiExport::oops_do(adjust_root_pointer_closure());
1951   // SO_AllClasses
1952   SystemDictionary::oops_do(adjust_root_pointer_closure());
1953   vmSymbols::oops_do(adjust_root_pointer_closure());
1954 
1955   // Now adjust pointers in remaining weak roots.  (All of which should
1956   // have been cleared if they pointed to non-surviving objects.)
1957   // Global (weak) JNI handles
1958   JNIHandles::weak_oops_do(&always_true, adjust_root_pointer_closure());
1959 
1960   CodeCache::oops_do(adjust_pointer_closure());
1961   SymbolTable::oops_do(adjust_root_pointer_closure());
1962   StringTable::oops_do(adjust_root_pointer_closure());
1963   ref_processor()->weak_oops_do(adjust_root_pointer_closure());
1964   // Roots were visited so references into the young gen in roots
1965   // may have been scanned.  Process them also.
1966   // Should the reference processor have a span that excludes
1967   // young gen objects?
1968   PSScavenge::reference_processor()->weak_oops_do(
1969                                               adjust_root_pointer_closure());
1970 }
1971 
1972 void PSParallelCompact::compact_perm(ParCompactionManager* cm) {
1973   EventMark m("4 compact perm");
1974   TraceTime tm("compact perm gen", print_phases(), true, gclog_or_tty);
1975   // trace("4");
1976 
1977   gc_heap()->perm_gen()->start_array()->reset();
1978   move_and_update(cm, perm_space_id);
1979 }
1980 
1981 void PSParallelCompact::enqueue_region_draining_tasks(GCTaskQueue* q,
1982                                                       uint parallel_gc_threads)
1983 {
1984   TraceTime tm("drain task setup", print_phases(), true, gclog_or_tty);
1985 
1986   const unsigned int task_count = MAX2(parallel_gc_threads, 1U);
1987   for (unsigned int j = 0; j < task_count; j++) {
1988     q->enqueue(new DrainStacksCompactionTask());
1989   }
1990 
1991   // Find all regions that are available (can be filled immediately) and
1992   // distribute them to the thread stacks.  The iteration is done in reverse
1993   // order (high to low) so the regions will be removed in ascending order.
1994 
1995   const ParallelCompactData& sd = PSParallelCompact::summary_data();
1996 
1997   size_t fillable_regions = 0;   // A count for diagnostic purposes.
1998   unsigned int which = 0;       // The worker thread number.
1999 
2000   for (unsigned int id = to_space_id; id > perm_space_id; --id) {
2001     SpaceInfo* const space_info = _space_info + id;
2002     MutableSpace* const space = space_info->space();
2003     HeapWord* const new_top = space_info->new_top();
2004 
2005     const size_t beg_region = sd.addr_to_region_idx(space_info->dense_prefix());
2006     const size_t end_region =
2007       sd.addr_to_region_idx(sd.region_align_up(new_top));
2008     assert(end_region > 0, "perm gen cannot be empty");
2009 
2010     for (size_t cur = end_region - 1; cur >= beg_region; --cur) {
2011       if (sd.region(cur)->claim_unsafe()) {
2012         ParCompactionManager* cm = ParCompactionManager::manager_array(which);
2013         cm->save_for_processing(cur);
2014 
2015         if (TraceParallelOldGCCompactionPhase && Verbose) {
2016           const size_t count_mod_8 = fillable_regions & 7;
2017           if (count_mod_8 == 0) gclog_or_tty->print("fillable: ");
2018           gclog_or_tty->print(" " SIZE_FORMAT_W(7), cur);
2019           if (count_mod_8 == 7) gclog_or_tty->cr();
2020         }
2021 
2022         NOT_PRODUCT(++fillable_regions;)
2023 
2024         // Assign regions to threads in round-robin fashion.
2025         if (++which == task_count) {
2026           which = 0;
2027         }
2028       }
2029     }
2030   }
2031 
2032   if (TraceParallelOldGCCompactionPhase) {
2033     if (Verbose && (fillable_regions & 7) != 0) gclog_or_tty->cr();
2034     gclog_or_tty->print_cr("%u initially fillable regions", fillable_regions);
2035   }
2036 }
2037 
2038 #define PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING 4
2039 
2040 void PSParallelCompact::enqueue_dense_prefix_tasks(GCTaskQueue* q,
2041                                                     uint parallel_gc_threads) {
2042   TraceTime tm("dense prefix task setup", print_phases(), true, gclog_or_tty);
2043 
2044   ParallelCompactData& sd = PSParallelCompact::summary_data();
2045 
2046   // Iterate over all the spaces adding tasks for updating
2047   // regions in the dense prefix.  Assume that 1 gc thread
2048   // will work on opening the gaps and the remaining gc threads
2049   // will work on the dense prefix.
2050   SpaceId space_id = old_space_id;
2051   while (space_id != last_space_id) {
2052     HeapWord* const dense_prefix_end = _space_info[space_id].dense_prefix();
2053     const MutableSpace* const space = _space_info[space_id].space();
2054 
2055     if (dense_prefix_end == space->bottom()) {
2056       // There is no dense prefix for this space.
2057       space_id = next_compaction_space_id(space_id);
2058       continue;
2059     }
2060 
2061     // The dense prefix is before this region.
2062     size_t region_index_end_dense_prefix =
2063         sd.addr_to_region_idx(dense_prefix_end);
2064     RegionData* const dense_prefix_cp =
2065       sd.region(region_index_end_dense_prefix);
2066     assert(dense_prefix_end == space->end() ||
2067            dense_prefix_cp->available() ||
2068            dense_prefix_cp->claimed(),
2069            "The region after the dense prefix should always be ready to fill");
2070 
2071     size_t region_index_start = sd.addr_to_region_idx(space->bottom());
2072 
2073     // Is there dense prefix work?
2074     size_t total_dense_prefix_regions =
2075       region_index_end_dense_prefix - region_index_start;
2076     // How many regions of the dense prefix should be given to
2077     // each thread?
2078     if (total_dense_prefix_regions > 0) {
2079       uint tasks_for_dense_prefix = 1;
2080       if (UseParallelDensePrefixUpdate) {
2081         if (total_dense_prefix_regions <=
2082             (parallel_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING)) {
2083           // Don't over partition.  This assumes that
2084           // PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING is a small integer value
2085           // so there are not many regions to process.
2086           tasks_for_dense_prefix = parallel_gc_threads;
2087         } else {
2088           // Over partition
2089           tasks_for_dense_prefix = parallel_gc_threads *
2090             PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING;
2091         }
2092       }
2093       size_t regions_per_thread = total_dense_prefix_regions /
2094         tasks_for_dense_prefix;
2095       // Give each thread at least 1 region.
2096       if (regions_per_thread == 0) {
2097         regions_per_thread = 1;
2098       }
2099 
2100       for (uint k = 0; k < tasks_for_dense_prefix; k++) {
2101         if (region_index_start >= region_index_end_dense_prefix) {
2102           break;
2103         }
2104         // region_index_end is not processed
2105         size_t region_index_end = MIN2(region_index_start + regions_per_thread,
2106                                        region_index_end_dense_prefix);
2107         q->enqueue(new UpdateDensePrefixTask(
2108                                  space_id,
2109                                  region_index_start,
2110                                  region_index_end));
2111         region_index_start = region_index_end;
2112       }
2113     }
2114     // This gets any part of the dense prefix that did not
2115     // fit evenly.
2116     if (region_index_start < region_index_end_dense_prefix) {
2117       q->enqueue(new UpdateDensePrefixTask(
2118                                  space_id,
2119                                  region_index_start,
2120                                  region_index_end_dense_prefix));
2121     }
2122     space_id = next_compaction_space_id(space_id);
2123   }  // End tasks for dense prefix
2124 }
2125 
2126 void PSParallelCompact::enqueue_region_stealing_tasks(
2127                                      GCTaskQueue* q,
2128                                      ParallelTaskTerminator* terminator_ptr,
2129                                      uint parallel_gc_threads) {
2130   TraceTime tm("steal task setup", print_phases(), true, gclog_or_tty);
2131 
2132   // Once a thread has drained it's stack, it should try to steal regions from
2133   // other threads.
2134   if (parallel_gc_threads > 1) {
2135     for (uint j = 0; j < parallel_gc_threads; j++) {
2136       q->enqueue(new StealRegionCompactionTask(terminator_ptr));
2137     }
2138   }
2139 }
2140 
2141 void PSParallelCompact::compact() {
2142   EventMark m("5 compact");
2143   // trace("5");
2144   TraceTime tm("compaction phase", print_phases(), true, gclog_or_tty);
2145 
2146   ParallelScavengeHeap* heap = (ParallelScavengeHeap*)Universe::heap();
2147   assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity");
2148   PSOldGen* old_gen = heap->old_gen();
2149   old_gen->start_array()->reset();
2150   uint parallel_gc_threads = heap->gc_task_manager()->workers();
2151   TaskQueueSetSuper* qset = ParCompactionManager::region_array();
2152   ParallelTaskTerminator terminator(parallel_gc_threads, qset);
2153 
2154   GCTaskQueue* q = GCTaskQueue::create();
2155   enqueue_region_draining_tasks(q, parallel_gc_threads);
2156   enqueue_dense_prefix_tasks(q, parallel_gc_threads);
2157   enqueue_region_stealing_tasks(q, &terminator, parallel_gc_threads);
2158 
2159   {
2160     TraceTime tm_pc("par compact", print_phases(), true, gclog_or_tty);
2161 
2162     WaitForBarrierGCTask* fin = WaitForBarrierGCTask::create();
2163     q->enqueue(fin);
2164 
2165     gc_task_manager()->add_list(q);
2166 
2167     fin->wait_for();
2168 
2169     // We have to release the barrier tasks!
2170     WaitForBarrierGCTask::destroy(fin);
2171 
2172 #ifdef  ASSERT
2173     // Verify that all regions have been processed before the deferred updates.
2174     // Note that perm_space_id is skipped; this type of verification is not
2175     // valid until the perm gen is compacted by regions.
2176     for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2177       verify_complete(SpaceId(id));
2178     }
2179 #endif
2180   }
2181 
2182   {
2183     // Update the deferred objects, if any.  Any compaction manager can be used.
2184     TraceTime tm_du("deferred updates", print_phases(), true, gclog_or_tty);
2185     ParCompactionManager* cm = ParCompactionManager::manager_array(0);
2186     for (unsigned int id = old_space_id; id < last_space_id; ++id) {
2187       update_deferred_objects(cm, SpaceId(id));
2188     }
2189   }
2190 }
2191 
2192 #ifdef  ASSERT
2193 void PSParallelCompact::verify_complete(SpaceId space_id) {
2194   // All Regions between space bottom() to new_top() should be marked as filled
2195   // and all Regions between new_top() and top() should be available (i.e.,
2196   // should have been emptied).
2197   ParallelCompactData& sd = summary_data();
2198   SpaceInfo si = _space_info[space_id];
2199   HeapWord* new_top_addr = sd.region_align_up(si.new_top());
2200   HeapWord* old_top_addr = sd.region_align_up(si.space()->top());
2201   const size_t beg_region = sd.addr_to_region_idx(si.space()->bottom());
2202   const size_t new_top_region = sd.addr_to_region_idx(new_top_addr);
2203   const size_t old_top_region = sd.addr_to_region_idx(old_top_addr);
2204 
2205   bool issued_a_warning = false;
2206 
2207   size_t cur_region;
2208   for (cur_region = beg_region; cur_region < new_top_region; ++cur_region) {
2209     const RegionData* const c = sd.region(cur_region);
2210     if (!c->completed()) {
2211       warning("region " SIZE_FORMAT " not filled:  "
2212               "destination_count=" SIZE_FORMAT,
2213               cur_region, c->destination_count());
2214       issued_a_warning = true;
2215     }
2216   }
2217 
2218   for (cur_region = new_top_region; cur_region < old_top_region; ++cur_region) {
2219     const RegionData* const c = sd.region(cur_region);
2220     if (!c->available()) {
2221       warning("region " SIZE_FORMAT " not empty:   "
2222               "destination_count=" SIZE_FORMAT,
2223               cur_region, c->destination_count());
2224       issued_a_warning = true;
2225     }
2226   }
2227 
2228   if (issued_a_warning) {
2229     print_region_ranges();
2230   }
2231 }
2232 #endif  // #ifdef ASSERT
2233 
2234 void PSParallelCompact::compact_serial(ParCompactionManager* cm) {
2235   EventMark m("5 compact serial");
2236   TraceTime tm("compact serial", print_phases(), true, gclog_or_tty);
2237 
2238   ParallelScavengeHeap* heap = (ParallelScavengeHeap*)Universe::heap();
2239   assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity");
2240 
2241   PSYoungGen* young_gen = heap->young_gen();
2242   PSOldGen* old_gen = heap->old_gen();
2243 
2244   old_gen->start_array()->reset();
2245   old_gen->move_and_update(cm);
2246   young_gen->move_and_update(cm);
2247 }
2248 
2249 
2250 void PSParallelCompact::follow_stack(ParCompactionManager* cm) {
2251   while(!cm->overflow_stack()->is_empty()) {
2252     oop obj = cm->overflow_stack()->pop();
2253     obj->follow_contents(cm);
2254   }
2255 
2256   oop obj;
2257   // obj is a reference!!!
2258   while (cm->marking_stack()->pop_local(obj)) {
2259     // It would be nice to assert about the type of objects we might
2260     // pop, but they can come from anywhere, unfortunately.
2261     obj->follow_contents(cm);
2262   }
2263 }
2264 
2265 void
2266 PSParallelCompact::follow_weak_klass_links(ParCompactionManager* serial_cm) {
2267   // All klasses on the revisit stack are marked at this point.
2268   // Update and follow all subklass, sibling and implementor links.
2269   for (uint i = 0; i < ParallelGCThreads+1; i++) {
2270     ParCompactionManager* cm = ParCompactionManager::manager_array(i);
2271     KeepAliveClosure keep_alive_closure(cm);
2272     for (int i = 0; i < cm->revisit_klass_stack()->length(); i++) {
2273       cm->revisit_klass_stack()->at(i)->follow_weak_klass_links(
2274         is_alive_closure(),
2275         &keep_alive_closure);
2276     }
2277     follow_stack(cm);
2278   }
2279 }
2280 
2281 void
2282 PSParallelCompact::revisit_weak_klass_link(ParCompactionManager* cm, Klass* k) {
2283   cm->revisit_klass_stack()->push(k);
2284 }
2285 
2286 #ifdef VALIDATE_MARK_SWEEP
2287 
2288 void PSParallelCompact::track_adjusted_pointer(void* p, bool isroot) {
2289   if (!ValidateMarkSweep)
2290     return;
2291 
2292   if (!isroot) {
2293     if (_pointer_tracking) {
2294       guarantee(_adjusted_pointers->contains(p), "should have seen this pointer");
2295       _adjusted_pointers->remove(p);
2296     }
2297   } else {
2298     ptrdiff_t index = _root_refs_stack->find(p);
2299     if (index != -1) {
2300       int l = _root_refs_stack->length();
2301       if (l > 0 && l - 1 != index) {
2302         void* last = _root_refs_stack->pop();
2303         assert(last != p, "should be different");
2304         _root_refs_stack->at_put(index, last);
2305       } else {
2306         _root_refs_stack->remove(p);
2307       }
2308     }
2309   }
2310 }
2311 
2312 
2313 void PSParallelCompact::check_adjust_pointer(void* p) {
2314   _adjusted_pointers->push(p);
2315 }
2316 
2317 
2318 class AdjusterTracker: public OopClosure {
2319  public:
2320   AdjusterTracker() {};
2321   void do_oop(oop* o)         { PSParallelCompact::check_adjust_pointer(o); }
2322   void do_oop(narrowOop* o)   { PSParallelCompact::check_adjust_pointer(o); }
2323 };
2324 
2325 
2326 void PSParallelCompact::track_interior_pointers(oop obj) {
2327   if (ValidateMarkSweep) {
2328     _adjusted_pointers->clear();
2329     _pointer_tracking = true;
2330 
2331     AdjusterTracker checker;
2332     obj->oop_iterate(&checker);
2333   }
2334 }
2335 
2336 
2337 void PSParallelCompact::check_interior_pointers() {
2338   if (ValidateMarkSweep) {
2339     _pointer_tracking = false;
2340     guarantee(_adjusted_pointers->length() == 0, "should have processed the same pointers");
2341   }
2342 }
2343 
2344 
2345 void PSParallelCompact::reset_live_oop_tracking(bool at_perm) {
2346   if (ValidateMarkSweep) {
2347     guarantee((size_t)_live_oops->length() == _live_oops_index, "should be at end of live oops");
2348     _live_oops_index = at_perm ? _live_oops_index_at_perm : 0;
2349   }
2350 }
2351 
2352 
2353 void PSParallelCompact::register_live_oop(oop p, size_t size) {
2354   if (ValidateMarkSweep) {
2355     _live_oops->push(p);
2356     _live_oops_size->push(size);
2357     _live_oops_index++;
2358   }
2359 }
2360 
2361 void PSParallelCompact::validate_live_oop(oop p, size_t size) {
2362   if (ValidateMarkSweep) {
2363     oop obj = _live_oops->at((int)_live_oops_index);
2364     guarantee(obj == p, "should be the same object");
2365     guarantee(_live_oops_size->at((int)_live_oops_index) == size, "should be the same size");
2366     _live_oops_index++;
2367   }
2368 }
2369 
2370 void PSParallelCompact::live_oop_moved_to(HeapWord* q, size_t size,
2371                                   HeapWord* compaction_top) {
2372   assert(oop(q)->forwardee() == NULL || oop(q)->forwardee() == oop(compaction_top),
2373          "should be moved to forwarded location");
2374   if (ValidateMarkSweep) {
2375     PSParallelCompact::validate_live_oop(oop(q), size);
2376     _live_oops_moved_to->push(oop(compaction_top));
2377   }
2378   if (RecordMarkSweepCompaction) {
2379     _cur_gc_live_oops->push(q);
2380     _cur_gc_live_oops_moved_to->push(compaction_top);
2381     _cur_gc_live_oops_size->push(size);
2382   }
2383 }
2384 
2385 
2386 void PSParallelCompact::compaction_complete() {
2387   if (RecordMarkSweepCompaction) {
2388     GrowableArray<HeapWord*>* _tmp_live_oops          = _cur_gc_live_oops;
2389     GrowableArray<HeapWord*>* _tmp_live_oops_moved_to = _cur_gc_live_oops_moved_to;
2390     GrowableArray<size_t>   * _tmp_live_oops_size     = _cur_gc_live_oops_size;
2391 
2392     _cur_gc_live_oops           = _last_gc_live_oops;
2393     _cur_gc_live_oops_moved_to  = _last_gc_live_oops_moved_to;
2394     _cur_gc_live_oops_size      = _last_gc_live_oops_size;
2395     _last_gc_live_oops          = _tmp_live_oops;
2396     _last_gc_live_oops_moved_to = _tmp_live_oops_moved_to;
2397     _last_gc_live_oops_size     = _tmp_live_oops_size;
2398   }
2399 }
2400 
2401 
2402 void PSParallelCompact::print_new_location_of_heap_address(HeapWord* q) {
2403   if (!RecordMarkSweepCompaction) {
2404     tty->print_cr("Requires RecordMarkSweepCompaction to be enabled");
2405     return;
2406   }
2407 
2408   if (_last_gc_live_oops == NULL) {
2409     tty->print_cr("No compaction information gathered yet");
2410     return;
2411   }
2412 
2413   for (int i = 0; i < _last_gc_live_oops->length(); i++) {
2414     HeapWord* old_oop = _last_gc_live_oops->at(i);
2415     size_t    sz      = _last_gc_live_oops_size->at(i);
2416     if (old_oop <= q && q < (old_oop + sz)) {
2417       HeapWord* new_oop = _last_gc_live_oops_moved_to->at(i);
2418       size_t offset = (q - old_oop);
2419       tty->print_cr("Address " PTR_FORMAT, q);
2420       tty->print_cr(" Was in oop " PTR_FORMAT ", size %d, at offset %d", old_oop, sz, offset);
2421       tty->print_cr(" Now in oop " PTR_FORMAT ", actual address " PTR_FORMAT, new_oop, new_oop + offset);
2422       return;
2423     }
2424   }
2425 
2426   tty->print_cr("Address " PTR_FORMAT " not found in live oop information from last GC", q);
2427 }
2428 #endif //VALIDATE_MARK_SWEEP
2429 
2430 // Update interior oops in the ranges of regions [beg_region, end_region).
2431 void
2432 PSParallelCompact::update_and_deadwood_in_dense_prefix(ParCompactionManager* cm,
2433                                                        SpaceId space_id,
2434                                                        size_t beg_region,
2435                                                        size_t end_region) {
2436   ParallelCompactData& sd = summary_data();
2437   ParMarkBitMap* const mbm = mark_bitmap();
2438 
2439   HeapWord* beg_addr = sd.region_to_addr(beg_region);
2440   HeapWord* const end_addr = sd.region_to_addr(end_region);
2441   assert(beg_region <= end_region, "bad region range");
2442   assert(end_addr <= dense_prefix(space_id), "not in the dense prefix");
2443 
2444 #ifdef  ASSERT
2445   // Claim the regions to avoid triggering an assert when they are marked as
2446   // filled.
2447   for (size_t claim_region = beg_region; claim_region < end_region; ++claim_region) {
2448     assert(sd.region(claim_region)->claim_unsafe(), "claim() failed");
2449   }
2450 #endif  // #ifdef ASSERT
2451 
2452   if (beg_addr != space(space_id)->bottom()) {
2453     // Find the first live object or block of dead space that *starts* in this
2454     // range of regions.  If a partial object crosses onto the region, skip it;
2455     // it will be marked for 'deferred update' when the object head is
2456     // processed.  If dead space crosses onto the region, it is also skipped; it
2457     // will be filled when the prior region is processed.  If neither of those
2458     // apply, the first word in the region is the start of a live object or dead
2459     // space.
2460     assert(beg_addr > space(space_id)->bottom(), "sanity");
2461     const RegionData* const cp = sd.region(beg_region);
2462     if (cp->partial_obj_size() != 0) {
2463       beg_addr = sd.partial_obj_end(beg_region);
2464     } else if (dead_space_crosses_boundary(cp, mbm->addr_to_bit(beg_addr))) {
2465       beg_addr = mbm->find_obj_beg(beg_addr, end_addr);
2466     }
2467   }
2468 
2469   if (beg_addr < end_addr) {
2470     // A live object or block of dead space starts in this range of Regions.
2471      HeapWord* const dense_prefix_end = dense_prefix(space_id);
2472 
2473     // Create closures and iterate.
2474     UpdateOnlyClosure update_closure(mbm, cm, space_id);
2475     FillClosure fill_closure(cm, space_id);
2476     ParMarkBitMap::IterationStatus status;
2477     status = mbm->iterate(&update_closure, &fill_closure, beg_addr, end_addr,
2478                           dense_prefix_end);
2479     if (status == ParMarkBitMap::incomplete) {
2480       update_closure.do_addr(update_closure.source());
2481     }
2482   }
2483 
2484   // Mark the regions as filled.
2485   RegionData* const beg_cp = sd.region(beg_region);
2486   RegionData* const end_cp = sd.region(end_region);
2487   for (RegionData* cp = beg_cp; cp < end_cp; ++cp) {
2488     cp->set_completed();
2489   }
2490 }
2491 
2492 // Return the SpaceId for the space containing addr.  If addr is not in the
2493 // heap, last_space_id is returned.  In debug mode it expects the address to be
2494 // in the heap and asserts such.
2495 PSParallelCompact::SpaceId PSParallelCompact::space_id(HeapWord* addr) {
2496   assert(Universe::heap()->is_in_reserved(addr), "addr not in the heap");
2497 
2498   for (unsigned int id = perm_space_id; id < last_space_id; ++id) {
2499     if (_space_info[id].space()->contains(addr)) {
2500       return SpaceId(id);
2501     }
2502   }
2503 
2504   assert(false, "no space contains the addr");
2505   return last_space_id;
2506 }
2507 
2508 void PSParallelCompact::update_deferred_objects(ParCompactionManager* cm,
2509                                                 SpaceId id) {
2510   assert(id < last_space_id, "bad space id");
2511 
2512   ParallelCompactData& sd = summary_data();
2513   const SpaceInfo* const space_info = _space_info + id;
2514   ObjectStartArray* const start_array = space_info->start_array();
2515 
2516   const MutableSpace* const space = space_info->space();
2517   assert(space_info->dense_prefix() >= space->bottom(), "dense_prefix not set");
2518   HeapWord* const beg_addr = space_info->dense_prefix();
2519   HeapWord* const end_addr = sd.region_align_up(space_info->new_top());
2520 
2521   const RegionData* const beg_region = sd.addr_to_region_ptr(beg_addr);
2522   const RegionData* const end_region = sd.addr_to_region_ptr(end_addr);
2523   const RegionData* cur_region;
2524   for (cur_region = beg_region; cur_region < end_region; ++cur_region) {
2525     HeapWord* const addr = cur_region->deferred_obj_addr();
2526     if (addr != NULL) {
2527       if (start_array != NULL) {
2528         start_array->allocate_block(addr);
2529       }
2530       oop(addr)->update_contents(cm);
2531       assert(oop(addr)->is_oop_or_null(), "should be an oop now");
2532     }
2533   }
2534 }
2535 
2536 // Skip over count live words starting from beg, and return the address of the
2537 // next live word.  Unless marked, the word corresponding to beg is assumed to
2538 // be dead.  Callers must either ensure beg does not correspond to the middle of
2539 // an object, or account for those live words in some other way.  Callers must
2540 // also ensure that there are enough live words in the range [beg, end) to skip.
2541 HeapWord*
2542 PSParallelCompact::skip_live_words(HeapWord* beg, HeapWord* end, size_t count)
2543 {
2544   assert(count > 0, "sanity");
2545 
2546   ParMarkBitMap* m = mark_bitmap();
2547   idx_t bits_to_skip = m->words_to_bits(count);
2548   idx_t cur_beg = m->addr_to_bit(beg);
2549   const idx_t search_end = BitMap::word_align_up(m->addr_to_bit(end));
2550 
2551   do {
2552     cur_beg = m->find_obj_beg(cur_beg, search_end);
2553     idx_t cur_end = m->find_obj_end(cur_beg, search_end);
2554     const size_t obj_bits = cur_end - cur_beg + 1;
2555     if (obj_bits > bits_to_skip) {
2556       return m->bit_to_addr(cur_beg + bits_to_skip);
2557     }
2558     bits_to_skip -= obj_bits;
2559     cur_beg = cur_end + 1;
2560   } while (bits_to_skip > 0);
2561 
2562   // Skipping the desired number of words landed just past the end of an object.
2563   // Find the start of the next object.
2564   cur_beg = m->find_obj_beg(cur_beg, search_end);
2565   assert(cur_beg < m->addr_to_bit(end), "not enough live words to skip");
2566   return m->bit_to_addr(cur_beg);
2567 }
2568 
2569 HeapWord*
2570 PSParallelCompact::first_src_addr(HeapWord* const dest_addr,
2571                                  size_t src_region_idx)
2572 {
2573   ParMarkBitMap* const bitmap = mark_bitmap();
2574   const ParallelCompactData& sd = summary_data();
2575   const size_t RegionSize = ParallelCompactData::RegionSize;
2576 
2577   assert(sd.is_region_aligned(dest_addr), "not aligned");
2578 
2579   const RegionData* const src_region_ptr = sd.region(src_region_idx);
2580   const size_t partial_obj_size = src_region_ptr->partial_obj_size();
2581   HeapWord* const src_region_destination = src_region_ptr->destination();
2582 
2583   assert(dest_addr >= src_region_destination, "wrong src region");
2584   assert(src_region_ptr->data_size() > 0, "src region cannot be empty");
2585 
2586   HeapWord* const src_region_beg = sd.region_to_addr(src_region_idx);
2587   HeapWord* const src_region_end = src_region_beg + RegionSize;
2588 
2589   HeapWord* addr = src_region_beg;
2590   if (dest_addr == src_region_destination) {
2591     // Return the first live word in the source region.
2592     if (partial_obj_size == 0) {
2593       addr = bitmap->find_obj_beg(addr, src_region_end);
2594       assert(addr < src_region_end, "no objects start in src region");
2595     }
2596     return addr;
2597   }
2598 
2599   // Must skip some live data.
2600   size_t words_to_skip = dest_addr - src_region_destination;
2601   assert(src_region_ptr->data_size() > words_to_skip, "wrong src region");
2602 
2603   if (partial_obj_size >= words_to_skip) {
2604     // All the live words to skip are part of the partial object.
2605     addr += words_to_skip;
2606     if (partial_obj_size == words_to_skip) {
2607       // Find the first live word past the partial object.
2608       addr = bitmap->find_obj_beg(addr, src_region_end);
2609       assert(addr < src_region_end, "wrong src region");
2610     }
2611     return addr;
2612   }
2613 
2614   // Skip over the partial object (if any).
2615   if (partial_obj_size != 0) {
2616     words_to_skip -= partial_obj_size;
2617     addr += partial_obj_size;
2618   }
2619 
2620   // Skip over live words due to objects that start in the region.
2621   addr = skip_live_words(addr, src_region_end, words_to_skip);
2622   assert(addr < src_region_end, "wrong src region");
2623   return addr;
2624 }
2625 
2626 void PSParallelCompact::decrement_destination_counts(ParCompactionManager* cm,
2627                                                      size_t beg_region,
2628                                                      HeapWord* end_addr)
2629 {
2630   ParallelCompactData& sd = summary_data();
2631   RegionData* const beg = sd.region(beg_region);
2632   HeapWord* const end_addr_aligned_up = sd.region_align_up(end_addr);
2633   RegionData* const end = sd.addr_to_region_ptr(end_addr_aligned_up);
2634   size_t cur_idx = beg_region;
2635   for (RegionData* cur = beg; cur < end; ++cur, ++cur_idx) {
2636     assert(cur->data_size() > 0, "region must have live data");
2637     cur->decrement_destination_count();
2638     if (cur_idx <= cur->source_region() && cur->available() && cur->claim()) {
2639       cm->save_for_processing(cur_idx);
2640     }
2641   }
2642 }
2643 
2644 size_t PSParallelCompact::next_src_region(MoveAndUpdateClosure& closure,
2645                                           SpaceId& src_space_id,
2646                                           HeapWord*& src_space_top,
2647                                           HeapWord* end_addr)
2648 {
2649   typedef ParallelCompactData::RegionData RegionData;
2650 
2651   ParallelCompactData& sd = PSParallelCompact::summary_data();
2652   const size_t region_size = ParallelCompactData::RegionSize;
2653 
2654   size_t src_region_idx = 0;
2655 
2656   // Skip empty regions (if any) up to the top of the space.
2657   HeapWord* const src_aligned_up = sd.region_align_up(end_addr);
2658   RegionData* src_region_ptr = sd.addr_to_region_ptr(src_aligned_up);
2659   HeapWord* const top_aligned_up = sd.region_align_up(src_space_top);
2660   const RegionData* const top_region_ptr =
2661     sd.addr_to_region_ptr(top_aligned_up);
2662   while (src_region_ptr < top_region_ptr && src_region_ptr->data_size() == 0) {
2663     ++src_region_ptr;
2664   }
2665 
2666   if (src_region_ptr < top_region_ptr) {
2667     // The next source region is in the current space.  Update src_region_idx
2668     // and the source address to match src_region_ptr.
2669     src_region_idx = sd.region(src_region_ptr);
2670     HeapWord* const src_region_addr = sd.region_to_addr(src_region_idx);
2671     if (src_region_addr > closure.source()) {
2672       closure.set_source(src_region_addr);
2673     }
2674     return src_region_idx;
2675   }
2676 
2677   // Switch to a new source space and find the first non-empty region.
2678   unsigned int space_id = src_space_id + 1;
2679   assert(space_id < last_space_id, "not enough spaces");
2680 
2681   HeapWord* const destination = closure.destination();
2682 
2683   do {
2684     MutableSpace* space = _space_info[space_id].space();
2685     HeapWord* const bottom = space->bottom();
2686     const RegionData* const bottom_cp = sd.addr_to_region_ptr(bottom);
2687 
2688     // Iterate over the spaces that do not compact into themselves.
2689     if (bottom_cp->destination() != bottom) {
2690       HeapWord* const top_aligned_up = sd.region_align_up(space->top());
2691       const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up);
2692 
2693       for (const RegionData* src_cp = bottom_cp; src_cp < top_cp; ++src_cp) {
2694         if (src_cp->live_obj_size() > 0) {
2695           // Found it.
2696           assert(src_cp->destination() == destination,
2697                  "first live obj in the space must match the destination");
2698           assert(src_cp->partial_obj_size() == 0,
2699                  "a space cannot begin with a partial obj");
2700 
2701           src_space_id = SpaceId(space_id);
2702           src_space_top = space->top();
2703           const size_t src_region_idx = sd.region(src_cp);
2704           closure.set_source(sd.region_to_addr(src_region_idx));
2705           return src_region_idx;
2706         } else {
2707           assert(src_cp->data_size() == 0, "sanity");
2708         }
2709       }
2710     }
2711   } while (++space_id < last_space_id);
2712 
2713   assert(false, "no source region was found");
2714   return 0;
2715 }
2716 
2717 void PSParallelCompact::fill_region(ParCompactionManager* cm, size_t region_idx)
2718 {
2719   typedef ParMarkBitMap::IterationStatus IterationStatus;
2720   const size_t RegionSize = ParallelCompactData::RegionSize;
2721   ParMarkBitMap* const bitmap = mark_bitmap();
2722   ParallelCompactData& sd = summary_data();
2723   RegionData* const region_ptr = sd.region(region_idx);
2724 
2725   // Get the items needed to construct the closure.
2726   HeapWord* dest_addr = sd.region_to_addr(region_idx);
2727   SpaceId dest_space_id = space_id(dest_addr);
2728   ObjectStartArray* start_array = _space_info[dest_space_id].start_array();
2729   HeapWord* new_top = _space_info[dest_space_id].new_top();
2730   assert(dest_addr < new_top, "sanity");
2731   const size_t words = MIN2(pointer_delta(new_top, dest_addr), RegionSize);
2732 
2733   // Get the source region and related info.
2734   size_t src_region_idx = region_ptr->source_region();
2735   SpaceId src_space_id = space_id(sd.region_to_addr(src_region_idx));
2736   HeapWord* src_space_top = _space_info[src_space_id].space()->top();
2737 
2738   MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words);
2739   closure.set_source(first_src_addr(dest_addr, src_region_idx));
2740 
2741   // Adjust src_region_idx to prepare for decrementing destination counts (the
2742   // destination count is not decremented when a region is copied to itself).
2743   if (src_region_idx == region_idx) {
2744     src_region_idx += 1;
2745   }
2746 
2747   if (bitmap->is_unmarked(closure.source())) {
2748     // The first source word is in the middle of an object; copy the remainder
2749     // of the object or as much as will fit.  The fact that pointer updates were
2750     // deferred will be noted when the object header is processed.
2751     HeapWord* const old_src_addr = closure.source();
2752     closure.copy_partial_obj();
2753     if (closure.is_full()) {
2754       decrement_destination_counts(cm, src_region_idx, closure.source());
2755       region_ptr->set_deferred_obj_addr(NULL);
2756       region_ptr->set_completed();
2757       return;
2758     }
2759 
2760     HeapWord* const end_addr = sd.region_align_down(closure.source());
2761     if (sd.region_align_down(old_src_addr) != end_addr) {
2762       // The partial object was copied from more than one source region.
2763       decrement_destination_counts(cm, src_region_idx, end_addr);
2764 
2765       // Move to the next source region, possibly switching spaces as well.  All
2766       // args except end_addr may be modified.
2767       src_region_idx = next_src_region(closure, src_space_id, src_space_top,
2768                                        end_addr);
2769     }
2770   }
2771 
2772   do {
2773     HeapWord* const cur_addr = closure.source();
2774     HeapWord* const end_addr = MIN2(sd.region_align_up(cur_addr + 1),
2775                                     src_space_top);
2776     IterationStatus status = bitmap->iterate(&closure, cur_addr, end_addr);
2777 
2778     if (status == ParMarkBitMap::incomplete) {
2779       // The last obj that starts in the source region does not end in the
2780       // region.
2781       assert(closure.source() < end_addr, "sanity")
2782       HeapWord* const obj_beg = closure.source();
2783       HeapWord* const range_end = MIN2(obj_beg + closure.words_remaining(),
2784                                        src_space_top);
2785       HeapWord* const obj_end = bitmap->find_obj_end(obj_beg, range_end);
2786       if (obj_end < range_end) {
2787         // The end was found; the entire object will fit.
2788         status = closure.do_addr(obj_beg, bitmap->obj_size(obj_beg, obj_end));
2789         assert(status != ParMarkBitMap::would_overflow, "sanity");
2790       } else {
2791         // The end was not found; the object will not fit.
2792         assert(range_end < src_space_top, "obj cannot cross space boundary");
2793         status = ParMarkBitMap::would_overflow;
2794       }
2795     }
2796 
2797     if (status == ParMarkBitMap::would_overflow) {
2798       // The last object did not fit.  Note that interior oop updates were
2799       // deferred, then copy enough of the object to fill the region.
2800       region_ptr->set_deferred_obj_addr(closure.destination());
2801       status = closure.copy_until_full(); // copies from closure.source()
2802 
2803       decrement_destination_counts(cm, src_region_idx, closure.source());
2804       region_ptr->set_completed();
2805       return;
2806     }
2807 
2808     if (status == ParMarkBitMap::full) {
2809       decrement_destination_counts(cm, src_region_idx, closure.source());
2810       region_ptr->set_deferred_obj_addr(NULL);
2811       region_ptr->set_completed();
2812       return;
2813     }
2814 
2815     decrement_destination_counts(cm, src_region_idx, end_addr);
2816 
2817     // Move to the next source region, possibly switching spaces as well.  All
2818     // args except end_addr may be modified.
2819     src_region_idx = next_src_region(closure, src_space_id, src_space_top,
2820                                      end_addr);
2821   } while (true);
2822 }
2823 
2824 void
2825 PSParallelCompact::move_and_update(ParCompactionManager* cm, SpaceId space_id) {
2826   const MutableSpace* sp = space(space_id);
2827   if (sp->is_empty()) {
2828     return;
2829   }
2830 
2831   ParallelCompactData& sd = PSParallelCompact::summary_data();
2832   ParMarkBitMap* const bitmap = mark_bitmap();
2833   HeapWord* const dp_addr = dense_prefix(space_id);
2834   HeapWord* beg_addr = sp->bottom();
2835   HeapWord* end_addr = sp->top();
2836 
2837 #ifdef ASSERT
2838   assert(beg_addr <= dp_addr && dp_addr <= end_addr, "bad dense prefix");
2839   if (cm->should_verify_only()) {
2840     VerifyUpdateClosure verify_update(cm, sp);
2841     bitmap->iterate(&verify_update, beg_addr, end_addr);
2842     return;
2843   }
2844 
2845   if (cm->should_reset_only()) {
2846     ResetObjectsClosure reset_objects(cm);
2847     bitmap->iterate(&reset_objects, beg_addr, end_addr);
2848     return;
2849   }
2850 #endif
2851 
2852   const size_t beg_region = sd.addr_to_region_idx(beg_addr);
2853   const size_t dp_region = sd.addr_to_region_idx(dp_addr);
2854   if (beg_region < dp_region) {
2855     update_and_deadwood_in_dense_prefix(cm, space_id, beg_region, dp_region);
2856   }
2857 
2858   // The destination of the first live object that starts in the region is one
2859   // past the end of the partial object entering the region (if any).
2860   HeapWord* const dest_addr = sd.partial_obj_end(dp_region);
2861   HeapWord* const new_top = _space_info[space_id].new_top();
2862   assert(new_top >= dest_addr, "bad new_top value");
2863   const size_t words = pointer_delta(new_top, dest_addr);
2864 
2865   if (words > 0) {
2866     ObjectStartArray* start_array = _space_info[space_id].start_array();
2867     MoveAndUpdateClosure closure(bitmap, cm, start_array, dest_addr, words);
2868 
2869     ParMarkBitMap::IterationStatus status;
2870     status = bitmap->iterate(&closure, dest_addr, end_addr);
2871     assert(status == ParMarkBitMap::full, "iteration not complete");
2872     assert(bitmap->find_obj_beg(closure.source(), end_addr) == end_addr,
2873            "live objects skipped because closure is full");
2874   }
2875 }
2876 
2877 jlong PSParallelCompact::millis_since_last_gc() {
2878   jlong ret_val = os::javaTimeMillis() - _time_of_last_gc;
2879   // XXX See note in genCollectedHeap::millis_since_last_gc().
2880   if (ret_val < 0) {
2881     NOT_PRODUCT(warning("time warp: %d", ret_val);)
2882     return 0;
2883   }
2884   return ret_val;
2885 }
2886 
2887 void PSParallelCompact::reset_millis_since_last_gc() {
2888   _time_of_last_gc = os::javaTimeMillis();
2889 }
2890 
2891 ParMarkBitMap::IterationStatus MoveAndUpdateClosure::copy_until_full()
2892 {
2893   if (source() != destination()) {
2894     assert(source() > destination(), "must copy to the left");
2895     Copy::aligned_conjoint_words(source(), destination(), words_remaining());
2896   }
2897   update_state(words_remaining());
2898   assert(is_full(), "sanity");
2899   return ParMarkBitMap::full;
2900 }
2901 
2902 void MoveAndUpdateClosure::copy_partial_obj()
2903 {
2904   size_t words = words_remaining();
2905 
2906   HeapWord* const range_end = MIN2(source() + words, bitmap()->region_end());
2907   HeapWord* const end_addr = bitmap()->find_obj_end(source(), range_end);
2908   if (end_addr < range_end) {
2909     words = bitmap()->obj_size(source(), end_addr);
2910   }
2911 
2912   // This test is necessary; if omitted, the pointer updates to a partial object
2913   // that crosses the dense prefix boundary could be overwritten.
2914   if (source() != destination()) {
2915     assert(source() > destination(), "must copy to the left");
2916     Copy::aligned_conjoint_words(source(), destination(), words);
2917   }
2918   update_state(words);
2919 }
2920 
2921 ParMarkBitMapClosure::IterationStatus
2922 MoveAndUpdateClosure::do_addr(HeapWord* addr, size_t words) {
2923   assert(destination() != NULL, "sanity");
2924   assert(bitmap()->obj_size(addr) == words, "bad size");
2925 
2926   _source = addr;
2927   assert(PSParallelCompact::summary_data().calc_new_pointer(source()) ==
2928          destination(), "wrong destination");
2929 
2930   if (words > words_remaining()) {
2931     return ParMarkBitMap::would_overflow;
2932   }
2933 
2934   // The start_array must be updated even if the object is not moving.
2935   if (_start_array != NULL) {
2936     _start_array->allocate_block(destination());
2937   }
2938 
2939   if (destination() != source()) {
2940     assert(destination() < source(), "must copy to the left");
2941     Copy::aligned_conjoint_words(source(), destination(), words);
2942   }
2943 
2944   oop moved_oop = (oop) destination();
2945   moved_oop->update_contents(compaction_manager());
2946   assert(moved_oop->is_oop_or_null(), "Object should be whole at this point");
2947 
2948   update_state(words);
2949   assert(destination() == (HeapWord*)moved_oop + moved_oop->size(), "sanity");
2950   return is_full() ? ParMarkBitMap::full : ParMarkBitMap::incomplete;
2951 }
2952 
2953 UpdateOnlyClosure::UpdateOnlyClosure(ParMarkBitMap* mbm,
2954                                      ParCompactionManager* cm,
2955                                      PSParallelCompact::SpaceId space_id) :
2956   ParMarkBitMapClosure(mbm, cm),
2957   _space_id(space_id),
2958   _start_array(PSParallelCompact::start_array(space_id))
2959 {
2960 }
2961 
2962 // Updates the references in the object to their new values.
2963 ParMarkBitMapClosure::IterationStatus
2964 UpdateOnlyClosure::do_addr(HeapWord* addr, size_t words) {
2965   do_addr(addr);
2966   return ParMarkBitMap::incomplete;
2967 }
2968 
2969 // Verify the new location using the forwarding pointer
2970 // from MarkSweep::mark_sweep_phase2().  Set the mark_word
2971 // to the initial value.
2972 ParMarkBitMapClosure::IterationStatus
2973 PSParallelCompact::VerifyUpdateClosure::do_addr(HeapWord* addr, size_t words) {
2974   // The second arg (words) is not used.
2975   oop obj = (oop) addr;
2976   HeapWord* forwarding_ptr = (HeapWord*) obj->mark()->decode_pointer();
2977   HeapWord* new_pointer = summary_data().calc_new_pointer(obj);
2978   if (forwarding_ptr == NULL) {
2979     // The object is dead or not moving.
2980     assert(bitmap()->is_unmarked(obj) || (new_pointer == (HeapWord*) obj),
2981            "Object liveness is wrong.");
2982     return ParMarkBitMap::incomplete;
2983   }
2984   assert(UseParallelOldGCDensePrefix ||
2985          (HeapMaximumCompactionInterval > 1) ||
2986          (MarkSweepAlwaysCompactCount > 1) ||
2987          (forwarding_ptr == new_pointer),
2988     "Calculation of new location is incorrect");
2989   return ParMarkBitMap::incomplete;
2990 }
2991 
2992 // Reset objects modified for debug checking.
2993 ParMarkBitMapClosure::IterationStatus
2994 PSParallelCompact::ResetObjectsClosure::do_addr(HeapWord* addr, size_t words) {
2995   // The second arg (words) is not used.
2996   oop obj = (oop) addr;
2997   obj->init_mark();
2998   return ParMarkBitMap::incomplete;
2999 }
3000 
3001 // Prepare for compaction.  This method is executed once
3002 // (i.e., by a single thread) before compaction.
3003 // Save the updated location of the intArrayKlassObj for
3004 // filling holes in the dense prefix.
3005 void PSParallelCompact::compact_prologue() {
3006   _updated_int_array_klass_obj = (klassOop)
3007     summary_data().calc_new_pointer(Universe::intArrayKlassObj());
3008 }
3009 
3010 // The initial implementation of this method created a field
3011 // _next_compaction_space_id in SpaceInfo and initialized
3012 // that field in SpaceInfo::initialize_space_info().  That
3013 // required that _next_compaction_space_id be declared a
3014 // SpaceId in SpaceInfo and that would have required that
3015 // either SpaceId be declared in a separate class or that
3016 // it be declared in SpaceInfo.  It didn't seem consistent
3017 // to declare it in SpaceInfo (didn't really fit logically).
3018 // Alternatively, defining a separate class to define SpaceId
3019 // seem excessive.  This implementation is simple and localizes
3020 // the knowledge.
3021 
3022 PSParallelCompact::SpaceId
3023 PSParallelCompact::next_compaction_space_id(SpaceId id) {
3024   assert(id < last_space_id, "id out of range");
3025   switch (id) {
3026     case perm_space_id :
3027       return last_space_id;
3028     case old_space_id :
3029       return eden_space_id;
3030     case eden_space_id :
3031       return from_space_id;
3032     case from_space_id :
3033       return to_space_id;
3034     case to_space_id :
3035       return last_space_id;
3036     default:
3037       assert(false, "Bad space id");
3038       return last_space_id;
3039   }
3040 }