Mercurial > jdk8-mips64-public > hotspot / file revision

 /*

  * Copyright (c) 2005, 2014, Oracle and/or its affiliates. All rights reserved.

  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.

  *

  * This code is free software; you can redistribute it and/or modify it

  * under the terms of the GNU General Public License version 2 only, as

  * published by the Free Software Foundation.

  *

  * This code is distributed in the hope that it will be useful, but WITHOUT

  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or

  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License

  * version 2 for more details (a copy is included in the LICENSE file that

  * accompanied this code).

  *

  * You should have received a copy of the GNU General Public License version

  * 2 along with this work; if not, write to the Free Software Foundation,

  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.

  *

  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA

  * or visit www.oracle.com if you need additional information or have any

  * questions.

  *

  */

 /*

  * This file has been modified by Loongson Technology in 2015. These

  * modifications are Copyright (c) 2015 Loongson Technology, and are made

  * available on the same license terms set forth above.

  */

 #include "precompiled.hpp"

 #include "classfile/symbolTable.hpp"

 #include "classfile/systemDictionary.hpp"

 #include "code/codeCache.hpp"

 #include "gc_implementation/parallelScavenge/gcTaskManager.hpp"

 #include "gc_implementation/parallelScavenge/parallelScavengeHeap.inline.hpp"

 #include "gc_implementation/parallelScavenge/pcTasks.hpp"

 #include "gc_implementation/parallelScavenge/psAdaptiveSizePolicy.hpp"

 #include "gc_implementation/parallelScavenge/psCompactionManager.inline.hpp"

 #include "gc_implementation/parallelScavenge/psMarkSweep.hpp"

 #include "gc_implementation/parallelScavenge/psMarkSweepDecorator.hpp"

 #include "gc_implementation/parallelScavenge/psOldGen.hpp"

 #include "gc_implementation/parallelScavenge/psParallelCompact.hpp"

 #include "gc_implementation/parallelScavenge/psPromotionManager.inline.hpp"

 #include "gc_implementation/parallelScavenge/psScavenge.hpp"

 #include "gc_implementation/parallelScavenge/psYoungGen.hpp"

 #include "gc_implementation/shared/gcHeapSummary.hpp"

 #include "gc_implementation/shared/gcTimer.hpp"

 #include "gc_implementation/shared/gcTrace.hpp"

 #include "gc_implementation/shared/gcTraceTime.hpp"

 #include "gc_implementation/shared/isGCActiveMark.hpp"

 #include "gc_interface/gcCause.hpp"

 #include "memory/gcLocker.inline.hpp"

 #include "memory/referencePolicy.hpp"

 #include "memory/referenceProcessor.hpp"

 #include "oops/methodData.hpp"

 #include "oops/oop.inline.hpp"

 #include "oops/oop.pcgc.inline.hpp"

 #include "runtime/fprofiler.hpp"

 #include "runtime/safepoint.hpp"

 #include "runtime/vmThread.hpp"

 #include "services/management.hpp"

 #include "services/memoryService.hpp"

 #include "services/memTracker.hpp"

 #include "utilities/events.hpp"

 #include "utilities/stack.inline.hpp"

 #include <math.h>

 PRAGMA_FORMAT_MUTE_WARNINGS_FOR_GCC

 // All sizes are in HeapWords.

 const size_t ParallelCompactData::Log2RegionSize  = 16; // 64K words

 const size_t ParallelCompactData::RegionSize      = (size_t)1 << Log2RegionSize;

 const size_t ParallelCompactData::RegionSizeBytes =

   RegionSize << LogHeapWordSize;

 const size_t ParallelCompactData::RegionSizeOffsetMask = RegionSize - 1;

 const size_t ParallelCompactData::RegionAddrOffsetMask = RegionSizeBytes - 1;

 const size_t ParallelCompactData::RegionAddrMask       = ~RegionAddrOffsetMask;

 const size_t ParallelCompactData::Log2BlockSize   = 7; // 128 words

 const size_t ParallelCompactData::BlockSize       = (size_t)1 << Log2BlockSize;

 const size_t ParallelCompactData::BlockSizeBytes  =

   BlockSize << LogHeapWordSize;

 const size_t ParallelCompactData::BlockSizeOffsetMask = BlockSize - 1;

 const size_t ParallelCompactData::BlockAddrOffsetMask = BlockSizeBytes - 1;

 const size_t ParallelCompactData::BlockAddrMask       = ~BlockAddrOffsetMask;

 const size_t ParallelCompactData::BlocksPerRegion = RegionSize / BlockSize;

 const size_t ParallelCompactData::Log2BlocksPerRegion =

   Log2RegionSize - Log2BlockSize;

 const ParallelCompactData::RegionData::region_sz_t

 ParallelCompactData::RegionData::dc_shift = 27;

 const ParallelCompactData::RegionData::region_sz_t

 ParallelCompactData::RegionData::dc_mask = ~0U << dc_shift;

 const ParallelCompactData::RegionData::region_sz_t

 ParallelCompactData::RegionData::dc_one = 0x1U << dc_shift;

 const ParallelCompactData::RegionData::region_sz_t

 ParallelCompactData::RegionData::los_mask = ~dc_mask;

 const ParallelCompactData::RegionData::region_sz_t

 ParallelCompactData::RegionData::dc_claimed = 0x8U << dc_shift;

 const ParallelCompactData::RegionData::region_sz_t

 ParallelCompactData::RegionData::dc_completed = 0xcU << dc_shift;

 SpaceInfo PSParallelCompact::_space_info[PSParallelCompact::last_space_id];

 bool      PSParallelCompact::_print_phases = false;

 ReferenceProcessor* PSParallelCompact::_ref_processor = NULL;

 Klass*              PSParallelCompact::_updated_int_array_klass_obj = NULL;

 double PSParallelCompact::_dwl_mean;

 double PSParallelCompact::_dwl_std_dev;

 double PSParallelCompact::_dwl_first_term;

 double PSParallelCompact::_dwl_adjustment;

 #ifdef  ASSERT

 bool   PSParallelCompact::_dwl_initialized = false;

 #endif  // #ifdef ASSERT

 void SplitInfo::record(size_t src_region_idx, size_t partial_obj_size,

                        HeapWord* destination)

 {

   assert(src_region_idx != 0, "invalid src_region_idx");

   assert(partial_obj_size != 0, "invalid partial_obj_size argument");

   assert(destination != NULL, "invalid destination argument");

   _src_region_idx = src_region_idx;

   _partial_obj_size = partial_obj_size;

   _destination = destination;

   // These fields may not be updated below, so make sure they're clear.

   assert(_dest_region_addr == NULL, "should have been cleared");

   assert(_first_src_addr == NULL, "should have been cleared");

   // Determine the number of destination regions for the partial object.

   HeapWord* const last_word = destination + partial_obj_size - 1;

   const ParallelCompactData& sd = PSParallelCompact::summary_data();

   HeapWord* const beg_region_addr = sd.region_align_down(destination);

   HeapWord* const end_region_addr = sd.region_align_down(last_word);

   if (beg_region_addr == end_region_addr) {

     // One destination region.

     _destination_count = 1;

     if (end_region_addr == destination) {

       // The destination falls on a region boundary, thus the first word of the

       // partial object will be the first word copied to the destination region.

       _dest_region_addr = end_region_addr;

       _first_src_addr = sd.region_to_addr(src_region_idx);

     }

   } else {

     // Two destination regions.  When copied, the partial object will cross a

     // destination region boundary, so a word somewhere within the partial

     // object will be the first word copied to the second destination region.

     _destination_count = 2;

     _dest_region_addr = end_region_addr;

     const size_t ofs = pointer_delta(end_region_addr, destination);

     assert(ofs < _partial_obj_size, "sanity");

     _first_src_addr = sd.region_to_addr(src_region_idx) + ofs;

   }

 }

 void SplitInfo::clear()

 {

   _src_region_idx = 0;

   _partial_obj_size = 0;

   _destination = NULL;

   _destination_count = 0;

   _dest_region_addr = NULL;

   _first_src_addr = NULL;

   assert(!is_valid(), "sanity");

 }

 #ifdef  ASSERT

 void SplitInfo::verify_clear()

 {

   assert(_src_region_idx == 0, "not clear");

   assert(_partial_obj_size == 0, "not clear");

   assert(_destination == NULL, "not clear");

   assert(_destination_count == 0, "not clear");

   assert(_dest_region_addr == NULL, "not clear");

   assert(_first_src_addr == NULL, "not clear");

 }

 #endif  // #ifdef ASSERT

 void PSParallelCompact::print_on_error(outputStream* st) {

   _mark_bitmap.print_on_error(st);

 }

 #ifndef PRODUCT

 const char* PSParallelCompact::space_names[] = {

   "old ", "eden", "from", "to  "

 };

 void PSParallelCompact::print_region_ranges()

 {

   tty->print_cr("space  bottom     top        end        new_top");

   tty->print_cr("------ ---------- ---------- ---------- ----------");

   for (unsigned int id = 0; id < last_space_id; ++id) {

     const MutableSpace* space = _space_info[id].space();

     tty->print_cr("%u %s "

                   SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " "

                   SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10) " ",

                   id, space_names[id],

                   summary_data().addr_to_region_idx(space->bottom()),

                   summary_data().addr_to_region_idx(space->top()),

                   summary_data().addr_to_region_idx(space->end()),

                   summary_data().addr_to_region_idx(_space_info[id].new_top()));

   }

 }

 void

 print_generic_summary_region(size_t i, const ParallelCompactData::RegionData* c)

 {

 #define REGION_IDX_FORMAT        SIZE_FORMAT_W(7)

 #define REGION_DATA_FORMAT       SIZE_FORMAT_W(5)

   ParallelCompactData& sd = PSParallelCompact::summary_data();

   size_t dci = c->destination() ? sd.addr_to_region_idx(c->destination()) : 0;

   tty->print_cr(REGION_IDX_FORMAT " " PTR_FORMAT " "

                 REGION_IDX_FORMAT " " PTR_FORMAT " "

                 REGION_DATA_FORMAT " " REGION_DATA_FORMAT " "

                 REGION_DATA_FORMAT " " REGION_IDX_FORMAT " %d",

                 i, c->data_location(), dci, c->destination(),

                 c->partial_obj_size(), c->live_obj_size(),

                 c->data_size(), c->source_region(), c->destination_count());

 #undef  REGION_IDX_FORMAT

 #undef  REGION_DATA_FORMAT

 }

 void

 print_generic_summary_data(ParallelCompactData& summary_data,

                            HeapWord* const beg_addr,

                            HeapWord* const end_addr)

 {

   size_t total_words = 0;

   size_t i = summary_data.addr_to_region_idx(beg_addr);

   const size_t last = summary_data.addr_to_region_idx(end_addr);

   HeapWord* pdest = 0;

   while (i <= last) {

     ParallelCompactData::RegionData* c = summary_data.region(i);

     if (c->data_size() != 0 || c->destination() != pdest) {

       print_generic_summary_region(i, c);

       total_words += c->data_size();

       pdest = c->destination();

     }

     ++i;

   }

   tty->print_cr("summary_data_bytes=" SIZE_FORMAT, total_words * HeapWordSize);

 }

 void

 print_generic_summary_data(ParallelCompactData& summary_data,

                            SpaceInfo* space_info)

 {

   for (unsigned int id = 0; id < PSParallelCompact::last_space_id; ++id) {

     const MutableSpace* space = space_info[id].space();

     print_generic_summary_data(summary_data, space->bottom(),

                                MAX2(space->top(), space_info[id].new_top()));

   }

 }

 void

 print_initial_summary_region(size_t i,

                              const ParallelCompactData::RegionData* c,

                              bool newline = true)

 {

   tty->print(SIZE_FORMAT_W(5) " " PTR_FORMAT " "

              SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " "

              SIZE_FORMAT_W(5) " " SIZE_FORMAT_W(5) " %d",

              i, c->destination(),

              c->partial_obj_size(), c->live_obj_size(),

              c->data_size(), c->source_region(), c->destination_count());

   if (newline) tty->cr();

 }

 void

 print_initial_summary_data(ParallelCompactData& summary_data,

                            const MutableSpace* space) {

   if (space->top() == space->bottom()) {

     return;

   }

   const size_t region_size = ParallelCompactData::RegionSize;

   typedef ParallelCompactData::RegionData RegionData;

   HeapWord* const top_aligned_up = summary_data.region_align_up(space->top());

   const size_t end_region = summary_data.addr_to_region_idx(top_aligned_up);

   const RegionData* c = summary_data.region(end_region - 1);

   HeapWord* end_addr = c->destination() + c->data_size();

   const size_t live_in_space = pointer_delta(end_addr, space->bottom());

   // Print (and count) the full regions at the beginning of the space.

   size_t full_region_count = 0;

   size_t i = summary_data.addr_to_region_idx(space->bottom());

   while (i < end_region && summary_data.region(i)->data_size() == region_size) {

     print_initial_summary_region(i, summary_data.region(i));

     ++full_region_count;

     ++i;

   }

   size_t live_to_right = live_in_space - full_region_count * region_size;

   double max_reclaimed_ratio = 0.0;

   size_t max_reclaimed_ratio_region = 0;

   size_t max_dead_to_right = 0;

   size_t max_live_to_right = 0;

   // Print the 'reclaimed ratio' for regions while there is something live in

   // the region or to the right of it.  The remaining regions are empty (and

   // uninteresting), and computing the ratio will result in division by 0.

   while (i < end_region && live_to_right > 0) {

     c = summary_data.region(i);

     HeapWord* const region_addr = summary_data.region_to_addr(i);

     const size_t used_to_right = pointer_delta(space->top(), region_addr);

     const size_t dead_to_right = used_to_right - live_to_right;

     const double reclaimed_ratio = double(dead_to_right) / live_to_right;

     if (reclaimed_ratio > max_reclaimed_ratio) {

             max_reclaimed_ratio = reclaimed_ratio;

             max_reclaimed_ratio_region = i;

             max_dead_to_right = dead_to_right;

             max_live_to_right = live_to_right;

     }

     print_initial_summary_region(i, c, false);

     tty->print_cr(" %12.10f " SIZE_FORMAT_W(10) " " SIZE_FORMAT_W(10),

                   reclaimed_ratio, dead_to_right, live_to_right);

     live_to_right -= c->data_size();

     ++i;

   }

   // Any remaining regions are empty.  Print one more if there is one.

   if (i < end_region) {

     print_initial_summary_region(i, summary_data.region(i));

   }

   tty->print_cr("max:  " SIZE_FORMAT_W(4) " d2r=" SIZE_FORMAT_W(10) " "

                 "l2r=" SIZE_FORMAT_W(10) " max_ratio=%14.12f",

                 max_reclaimed_ratio_region, max_dead_to_right,

                 max_live_to_right, max_reclaimed_ratio);

 }

 void

 print_initial_summary_data(ParallelCompactData& summary_data,

                            SpaceInfo* space_info) {

   unsigned int id = PSParallelCompact::old_space_id;

   const MutableSpace* space;

   do {

     space = space_info[id].space();

     print_initial_summary_data(summary_data, space);

   } while (++id < PSParallelCompact::eden_space_id);

   do {

     space = space_info[id].space();

     print_generic_summary_data(summary_data, space->bottom(), space->top());

   } while (++id < PSParallelCompact::last_space_id);

 }

 #endif  // #ifndef PRODUCT

 #ifdef  ASSERT

 size_t add_obj_count;

 size_t add_obj_size;

 size_t mark_bitmap_count;

 size_t mark_bitmap_size;

 #endif  // #ifdef ASSERT

 ParallelCompactData::ParallelCompactData()

 {

   _region_start = 0;

   _region_vspace = 0;

   _reserved_byte_size = 0;

   _region_data = 0;

   _region_count = 0;

   _block_vspace = 0;

   _block_data = 0;

   _block_count = 0;

 }

 bool ParallelCompactData::initialize(MemRegion covered_region)

 {

   _region_start = covered_region.start();

   const size_t region_size = covered_region.word_size();

   DEBUG_ONLY(_region_end = _region_start + region_size;)

   assert(region_align_down(_region_start) == _region_start,

          "region start not aligned");

   assert((region_size & RegionSizeOffsetMask) == 0,

          "region size not a multiple of RegionSize");

   bool result = initialize_region_data(region_size) && initialize_block_data();

   return result;

 }

 PSVirtualSpace*

 ParallelCompactData::create_vspace(size_t count, size_t element_size)

 {

   const size_t raw_bytes = count * element_size;

   const size_t page_sz = os::page_size_for_region(raw_bytes, raw_bytes, 10);

   const size_t granularity = os::vm_allocation_granularity();

   _reserved_byte_size = align_size_up(raw_bytes, MAX2(page_sz, granularity));

   const size_t rs_align = page_sz == (size_t) os::vm_page_size() ? 0 :

     MAX2(page_sz, granularity);

   ReservedSpace rs(_reserved_byte_size, rs_align, rs_align > 0);

   os::trace_page_sizes("par compact", raw_bytes, raw_bytes, page_sz, rs.base(),

                        rs.size());

   MemTracker::record_virtual_memory_type((address)rs.base(), mtGC);

   PSVirtualSpace* vspace = new PSVirtualSpace(rs, page_sz);

   if (vspace != 0) {

     if (vspace->expand_by(_reserved_byte_size)) {

       return vspace;

     }

     delete vspace;

     // Release memory reserved in the space.

     rs.release();

   }

   return 0;

 }

 bool ParallelCompactData::initialize_region_data(size_t region_size)

 {

   const size_t count = (region_size + RegionSizeOffsetMask) >> Log2RegionSize;

   _region_vspace = create_vspace(count, sizeof(RegionData));

   if (_region_vspace != 0) {

     _region_data = (RegionData*)_region_vspace->reserved_low_addr();

     _region_count = count;

     return true;

   }

   return false;

 }

 bool ParallelCompactData::initialize_block_data()

 {

   assert(_region_count != 0, "region data must be initialized first");

   const size_t count = _region_count << Log2BlocksPerRegion;

   _block_vspace = create_vspace(count, sizeof(BlockData));

   if (_block_vspace != 0) {

     _block_data = (BlockData*)_block_vspace->reserved_low_addr();

     _block_count = count;

     return true;

   }

   return false;

 }

 void ParallelCompactData::clear()

 {

   memset(_region_data, 0, _region_vspace->committed_size());

   memset(_block_data, 0, _block_vspace->committed_size());

 }

 void ParallelCompactData::clear_range(size_t beg_region, size_t end_region) {

   assert(beg_region <= _region_count, "beg_region out of range");

   assert(end_region <= _region_count, "end_region out of range");

   assert(RegionSize % BlockSize == 0, "RegionSize not a multiple of BlockSize");

   const size_t region_cnt = end_region - beg_region;

   memset(_region_data + beg_region, 0, region_cnt * sizeof(RegionData));

   const size_t beg_block = beg_region * BlocksPerRegion;

   const size_t block_cnt = region_cnt * BlocksPerRegion;

   memset(_block_data + beg_block, 0, block_cnt * sizeof(BlockData));

 }

 HeapWord* ParallelCompactData::partial_obj_end(size_t region_idx) const

 {

   const RegionData* cur_cp = region(region_idx);

   const RegionData* const end_cp = region(region_count() - 1);

   HeapWord* result = region_to_addr(region_idx);

   if (cur_cp < end_cp) {

     do {

       result += cur_cp->partial_obj_size();

     } while (cur_cp->partial_obj_size() == RegionSize && ++cur_cp < end_cp);

   }

   return result;

 }

 void ParallelCompactData::add_obj(HeapWord* addr, size_t len)

 {

   const size_t obj_ofs = pointer_delta(addr, _region_start);

   const size_t beg_region = obj_ofs >> Log2RegionSize;

   const size_t end_region = (obj_ofs + len - 1) >> Log2RegionSize;

   DEBUG_ONLY(Atomic::inc_ptr(&add_obj_count);)

   DEBUG_ONLY(Atomic::add_ptr(len, &add_obj_size);)

   if (beg_region == end_region) {

     // All in one region.

     _region_data[beg_region].add_live_obj(len);

     return;

   }

   // First region.

   const size_t beg_ofs = region_offset(addr);

   _region_data[beg_region].add_live_obj(RegionSize - beg_ofs);

   Klass* klass = ((oop)addr)->klass();

   // Middle regions--completely spanned by this object.

   for (size_t region = beg_region + 1; region < end_region; ++region) {

     _region_data[region].set_partial_obj_size(RegionSize);

     _region_data[region].set_partial_obj_addr(addr);

   }

   // Last region.

   const size_t end_ofs = region_offset(addr + len - 1);

   _region_data[end_region].set_partial_obj_size(end_ofs + 1);

   _region_data[end_region].set_partial_obj_addr(addr);

 }

 void

 ParallelCompactData::summarize_dense_prefix(HeapWord* beg, HeapWord* end)

 {

   assert(region_offset(beg) == 0, "not RegionSize aligned");

   assert(region_offset(end) == 0, "not RegionSize aligned");

   size_t cur_region = addr_to_region_idx(beg);

   const size_t end_region = addr_to_region_idx(end);

   HeapWord* addr = beg;

   while (cur_region < end_region) {

     _region_data[cur_region].set_destination(addr);

     _region_data[cur_region].set_destination_count(0);

     _region_data[cur_region].set_source_region(cur_region);

     _region_data[cur_region].set_data_location(addr);

     // Update live_obj_size so the region appears completely full.

     size_t live_size = RegionSize - _region_data[cur_region].partial_obj_size();

     _region_data[cur_region].set_live_obj_size(live_size);

     ++cur_region;

     addr += RegionSize;

   }

 }

 // Find the point at which a space can be split and, if necessary, record the

 // split point.

 //

 // If the current src region (which overflowed the destination space) doesn't

 // have a partial object, the split point is at the beginning of the current src

 // region (an "easy" split, no extra bookkeeping required).

 //

 // If the current src region has a partial object, the split point is in the

 // region where that partial object starts (call it the split_region).  If

 // split_region has a partial object, then the split point is just after that

 // partial object (a "hard" split where we have to record the split data and

 // zero the partial_obj_size field).  With a "hard" split, we know that the

 // partial_obj ends within split_region because the partial object that caused

 // the overflow starts in split_region.  If split_region doesn't have a partial

 // obj, then the split is at the beginning of split_region (another "easy"

 // split).

 HeapWord*

 ParallelCompactData::summarize_split_space(size_t src_region,

                                            SplitInfo& split_info,

                                            HeapWord* destination,

                                            HeapWord* target_end,

                                            HeapWord** target_next)

 {

   assert(destination <= target_end, "sanity");

   assert(destination + _region_data[src_region].data_size() > target_end,

     "region should not fit into target space");

   assert(is_region_aligned(target_end), "sanity");

   size_t split_region = src_region;

   HeapWord* split_destination = destination;

   size_t partial_obj_size = _region_data[src_region].partial_obj_size();

   if (destination + partial_obj_size > target_end) {

     // The split point is just after the partial object (if any) in the

     // src_region that contains the start of the object that overflowed the

     // destination space.

     //

     // Find the start of the "overflow" object and set split_region to the

     // region containing it.

     HeapWord* const overflow_obj = _region_data[src_region].partial_obj_addr();

     split_region = addr_to_region_idx(overflow_obj);

     // Clear the source_region field of all destination regions whose first word

     // came from data after the split point (a non-null source_region field

     // implies a region must be filled).

     //

     // An alternative to the simple loop below:  clear during post_compact(),

     // which uses memcpy instead of individual stores, and is easy to

     // parallelize.  (The downside is that it clears the entire RegionData

     // object as opposed to just one field.)

     //

     // post_compact() would have to clear the summary data up to the highest

     // address that was written during the summary phase, which would be

     //

     //         max(top, max(new_top, clear_top))

     //

     // where clear_top is a new field in SpaceInfo.  Would have to set clear_top

     // to target_end.

     const RegionData* const sr = region(split_region);

     const size_t beg_idx =

       addr_to_region_idx(region_align_up(sr->destination() +

                                          sr->partial_obj_size()));

     const size_t end_idx = addr_to_region_idx(target_end);

     if (TraceParallelOldGCSummaryPhase) {

         gclog_or_tty->print_cr("split:  clearing source_region field in ["

                                SIZE_FORMAT ", " SIZE_FORMAT ")",

                                beg_idx, end_idx);

     }

     for (size_t idx = beg_idx; idx < end_idx; ++idx) {

       _region_data[idx].set_source_region(0);

     }

     // Set split_destination and partial_obj_size to reflect the split region.

     split_destination = sr->destination();

     partial_obj_size = sr->partial_obj_size();

   }

   // The split is recorded only if a partial object extends onto the region.

   if (partial_obj_size != 0) {

     _region_data[split_region].set_partial_obj_size(0);

     split_info.record(split_region, partial_obj_size, split_destination);

   }

   // Setup the continuation addresses.

   *target_next = split_destination + partial_obj_size;

   HeapWord* const source_next = region_to_addr(split_region) + partial_obj_size;

   if (TraceParallelOldGCSummaryPhase) {

     const char * split_type = partial_obj_size == 0 ? "easy" : "hard";

     gclog_or_tty->print_cr("%s split:  src=" PTR_FORMAT " src_c=" SIZE_FORMAT

                            " pos=" SIZE_FORMAT,

                            split_type, source_next, split_region,

                            partial_obj_size);

     gclog_or_tty->print_cr("%s split:  dst=" PTR_FORMAT " dst_c=" SIZE_FORMAT

                            " tn=" PTR_FORMAT,

                            split_type, split_destination,

                            addr_to_region_idx(split_destination),

                            *target_next);

     if (partial_obj_size != 0) {

       HeapWord* const po_beg = split_info.destination();

       HeapWord* const po_end = po_beg + split_info.partial_obj_size();

       gclog_or_tty->print_cr("%s split:  "

                              "po_beg=" PTR_FORMAT " " SIZE_FORMAT " "

                              "po_end=" PTR_FORMAT " " SIZE_FORMAT,

                              split_type,

                              po_beg, addr_to_region_idx(po_beg),

                              po_end, addr_to_region_idx(po_end));

     }

   }

   return source_next;

 }

 bool ParallelCompactData::summarize(SplitInfo& split_info,

                                     HeapWord* source_beg, HeapWord* source_end,

                                     HeapWord** source_next,

                                     HeapWord* target_beg, HeapWord* target_end,

                                     HeapWord** target_next)

 {

   if (TraceParallelOldGCSummaryPhase) {

     HeapWord* const source_next_val = source_next == NULL ? NULL : *source_next;

     tty->print_cr("sb=" PTR_FORMAT " se=" PTR_FORMAT " sn=" PTR_FORMAT

                   "tb=" PTR_FORMAT " te=" PTR_FORMAT " tn=" PTR_FORMAT,

                   source_beg, source_end, source_next_val,

                   target_beg, target_end, *target_next);

   }

   size_t cur_region = addr_to_region_idx(source_beg);

   const size_t end_region = addr_to_region_idx(region_align_up(source_end));

   HeapWord *dest_addr = target_beg;

   while (cur_region < end_region) {

     // The destination must be set even if the region has no data.

     _region_data[cur_region].set_destination(dest_addr);

     size_t words = _region_data[cur_region].data_size();

     if (words > 0) {

       // If cur_region does not fit entirely into the target space, find a point

       // at which the source space can be 'split' so that part is copied to the

       // target space and the rest is copied elsewhere.

       if (dest_addr + words > target_end) {

         assert(source_next != NULL, "source_next is NULL when splitting");

         *source_next = summarize_split_space(cur_region, split_info, dest_addr,

                                              target_end, target_next);

         return false;

       }

       // Compute the destination_count for cur_region, and if necessary, update

       // source_region for a destination region.  The source_region field is

       // updated if cur_region is the first (left-most) region to be copied to a

       // destination region.

       //

       // The destination_count calculation is a bit subtle.  A region that has

       // data that compacts into itself does not count itself as a destination.

       // This maintains the invariant that a zero count means the region is

       // available and can be claimed and then filled.

       uint destination_count = 0;

       if (split_info.is_split(cur_region)) {

         // The current region has been split:  the partial object will be copied

         // to one destination space and the remaining data will be copied to

         // another destination space.  Adjust the initial destination_count and,

         // if necessary, set the source_region field if the partial object will

         // cross a destination region boundary.

         destination_count = split_info.destination_count();

         if (destination_count == 2) {

           size_t dest_idx = addr_to_region_idx(split_info.dest_region_addr());

           _region_data[dest_idx].set_source_region(cur_region);

         }

       }

       HeapWord* const last_addr = dest_addr + words - 1;

       const size_t dest_region_1 = addr_to_region_idx(dest_addr);

       const size_t dest_region_2 = addr_to_region_idx(last_addr);

       // Initially assume that the destination regions will be the same and

       // adjust the value below if necessary.  Under this assumption, if

       // cur_region == dest_region_2, then cur_region will be compacted

       // completely into itself.

       destination_count += cur_region == dest_region_2 ? 0 : 1;

       if (dest_region_1 != dest_region_2) {

         // Destination regions differ; adjust destination_count.

         destination_count += 1;

         // Data from cur_region will be copied to the start of dest_region_2.

         _region_data[dest_region_2].set_source_region(cur_region);

       } else if (region_offset(dest_addr) == 0) {

         // Data from cur_region will be copied to the start of the destination

         // region.

         _region_data[dest_region_1].set_source_region(cur_region);

       }

       _region_data[cur_region].set_destination_count(destination_count);

       _region_data[cur_region].set_data_location(region_to_addr(cur_region));

       dest_addr += words;

     }

     ++cur_region;

   }

   *target_next = dest_addr;

   return true;

 }

 HeapWord* ParallelCompactData::calc_new_pointer(HeapWord* addr) {

   assert(addr != NULL, "Should detect NULL oop earlier");

   assert(PSParallelCompact::gc_heap()->is_in(addr), "not in heap");

   assert(PSParallelCompact::mark_bitmap()->is_marked(addr), "not marked");

   // Region covering the object.

   RegionData* const region_ptr = addr_to_region_ptr(addr);

   HeapWord* result = region_ptr->destination();

   // If the entire Region is live, the new location is region->destination + the

   // offset of the object within in the Region.

   // Run some performance tests to determine if this special case pays off.  It

   // is worth it for pointers into the dense prefix.  If the optimization to

   // avoid pointer updates in regions that only point to the dense prefix is

   // ever implemented, this should be revisited.

   if (region_ptr->data_size() == RegionSize) {

     result += region_offset(addr);

     return result;

   }

   // Otherwise, the new location is region->destination + block offset + the

   // number of live words in the Block that are (a) to the left of addr and (b)

   // due to objects that start in the Block.

   // Fill in the block table if necessary.  This is unsynchronized, so multiple

   // threads may fill the block table for a region (harmless, since it is

   // idempotent).

   if (!region_ptr->blocks_filled()) {

     PSParallelCompact::fill_blocks(addr_to_region_idx(addr));

     region_ptr->set_blocks_filled();

   }

   HeapWord* const search_start = block_align_down(addr);

   const size_t block_offset = addr_to_block_ptr(addr)->offset();

   const ParMarkBitMap* bitmap = PSParallelCompact::mark_bitmap();

   const size_t live = bitmap->live_words_in_range(search_start, oop(addr));

   result += block_offset + live;

   DEBUG_ONLY(PSParallelCompact::check_new_location(addr, result));

   return result;

 }

 #ifdef ASSERT

 void ParallelCompactData::verify_clear(const PSVirtualSpace* vspace)

 {

   const size_t* const beg = (const size_t*)vspace->committed_low_addr();

   const size_t* const end = (const size_t*)vspace->committed_high_addr();

   for (const size_t* p = beg; p < end; ++p) {

     assert(*p == 0, "not zero");

   }

 }

 void ParallelCompactData::verify_clear()

 {

   verify_clear(_region_vspace);

   verify_clear(_block_vspace);

 }

 #endif  // #ifdef ASSERT

 STWGCTimer          PSParallelCompact::_gc_timer;

 ParallelOldTracer   PSParallelCompact::_gc_tracer;

 elapsedTimer        PSParallelCompact::_accumulated_time;

 unsigned int        PSParallelCompact::_total_invocations = 0;

 unsigned int        PSParallelCompact::_maximum_compaction_gc_num = 0;

 jlong               PSParallelCompact::_time_of_last_gc = 0;

 CollectorCounters*  PSParallelCompact::_counters = NULL;

 ParMarkBitMap       PSParallelCompact::_mark_bitmap;

 ParallelCompactData PSParallelCompact::_summary_data;

 PSParallelCompact::IsAliveClosure PSParallelCompact::_is_alive_closure;

 bool PSParallelCompact::IsAliveClosure::do_object_b(oop p) { return mark_bitmap()->is_marked(p); }

 void PSParallelCompact::KeepAliveClosure::do_oop(oop* p)       { PSParallelCompact::KeepAliveClosure::do_oop_work(p); }

 void PSParallelCompact::KeepAliveClosure::do_oop(narrowOop* p) { PSParallelCompact::KeepAliveClosure::do_oop_work(p); }

 PSParallelCompact::AdjustPointerClosure PSParallelCompact::_adjust_pointer_closure;

 PSParallelCompact::AdjustKlassClosure PSParallelCompact::_adjust_klass_closure;

 void PSParallelCompact::AdjustPointerClosure::do_oop(oop* p)       { adjust_pointer(p); }

 void PSParallelCompact::AdjustPointerClosure::do_oop(narrowOop* p) { adjust_pointer(p); }

 void PSParallelCompact::FollowStackClosure::do_void() { _compaction_manager->follow_marking_stacks(); }

 void PSParallelCompact::MarkAndPushClosure::do_oop(oop* p)       {

   mark_and_push(_compaction_manager, p);

 }

 void PSParallelCompact::MarkAndPushClosure::do_oop(narrowOop* p) { mark_and_push(_compaction_manager, p); }

 void PSParallelCompact::FollowKlassClosure::do_klass(Klass* klass) {

   klass->oops_do(_mark_and_push_closure);

 }

 void PSParallelCompact::AdjustKlassClosure::do_klass(Klass* klass) {

   klass->oops_do(&PSParallelCompact::_adjust_pointer_closure);

 }

 void PSParallelCompact::post_initialize() {

   ParallelScavengeHeap* heap = gc_heap();

   assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity");

   MemRegion mr = heap->reserved_region();

   _ref_processor =

     new ReferenceProcessor(mr,            // span

                            ParallelRefProcEnabled && (ParallelGCThreads > 1), // mt processing

                            (int) ParallelGCThreads, // mt processing degree

                            true,          // mt discovery

                            (int) ParallelGCThreads, // mt discovery degree

                            true,          // atomic_discovery

                            &_is_alive_closure); // non-header is alive closure

   _counters = new CollectorCounters("PSParallelCompact", 1);

   // Initialize static fields in ParCompactionManager.

   ParCompactionManager::initialize(mark_bitmap());

 }

 bool PSParallelCompact::initialize() {

   ParallelScavengeHeap* heap = gc_heap();

   assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity");

   MemRegion mr = heap->reserved_region();

   // Was the old gen get allocated successfully?

   if (!heap->old_gen()->is_allocated()) {

     return false;

   }

   initialize_space_info();

   initialize_dead_wood_limiter();

   if (!_mark_bitmap.initialize(mr)) {

     vm_shutdown_during_initialization(

       err_msg("Unable to allocate " SIZE_FORMAT "KB bitmaps for parallel "

       "garbage collection for the requested " SIZE_FORMAT "KB heap.",

       _mark_bitmap.reserved_byte_size()/K, mr.byte_size()/K));

     return false;

   }

   if (!_summary_data.initialize(mr)) {

     vm_shutdown_during_initialization(

       err_msg("Unable to allocate " SIZE_FORMAT "KB card tables for parallel "

       "garbage collection for the requested " SIZE_FORMAT "KB heap.",

       _summary_data.reserved_byte_size()/K, mr.byte_size()/K));

     return false;

   }

   return true;

 }

 void PSParallelCompact::initialize_space_info()

 {

   memset(&_space_info, 0, sizeof(_space_info));

   ParallelScavengeHeap* heap = gc_heap();

   PSYoungGen* young_gen = heap->young_gen();

   _space_info[old_space_id].set_space(heap->old_gen()->object_space());

   _space_info[eden_space_id].set_space(young_gen->eden_space());

   _space_info[from_space_id].set_space(young_gen->from_space());

   _space_info[to_space_id].set_space(young_gen->to_space());

   _space_info[old_space_id].set_start_array(heap->old_gen()->start_array());

 }

 void PSParallelCompact::initialize_dead_wood_limiter()

 {

   const size_t max = 100;

   _dwl_mean = double(MIN2(ParallelOldDeadWoodLimiterMean, max)) / 100.0;

   _dwl_std_dev = double(MIN2(ParallelOldDeadWoodLimiterStdDev, max)) / 100.0;

   _dwl_first_term = 1.0 / (sqrt(2.0 * M_PI) * _dwl_std_dev);

   DEBUG_ONLY(_dwl_initialized = true;)

   _dwl_adjustment = normal_distribution(1.0);

 }

 // Simple class for storing info about the heap at the start of GC, to be used

 // after GC for comparison/printing.

 class PreGCValues {

 public:

   PreGCValues() { }

   PreGCValues(ParallelScavengeHeap* heap) { fill(heap); }

   void fill(ParallelScavengeHeap* heap) {

     _heap_used      = heap->used();

     _young_gen_used = heap->young_gen()->used_in_bytes();

     _old_gen_used   = heap->old_gen()->used_in_bytes();

     _metadata_used  = MetaspaceAux::used_bytes();

   };

   size_t heap_used() const      { return _heap_used; }

   size_t young_gen_used() const { return _young_gen_used; }

   size_t old_gen_used() const   { return _old_gen_used; }

   size_t metadata_used() const  { return _metadata_used; }

 private:

   size_t _heap_used;

   size_t _young_gen_used;

   size_t _old_gen_used;

   size_t _metadata_used;

 };

 void

 PSParallelCompact::clear_data_covering_space(SpaceId id)

 {

   // At this point, top is the value before GC, new_top() is the value that will

   // be set at the end of GC.  The marking bitmap is cleared to top; nothing

   // should be marked above top.  The summary data is cleared to the larger of

   // top & new_top.

   MutableSpace* const space = _space_info[id].space();

   HeapWord* const bot = space->bottom();

   HeapWord* const top = space->top();

   HeapWord* const max_top = MAX2(top, _space_info[id].new_top());

   const idx_t beg_bit = _mark_bitmap.addr_to_bit(bot);

   const idx_t end_bit = BitMap::word_align_up(_mark_bitmap.addr_to_bit(top));

   _mark_bitmap.clear_range(beg_bit, end_bit);

   const size_t beg_region = _summary_data.addr_to_region_idx(bot);

   const size_t end_region =

     _summary_data.addr_to_region_idx(_summary_data.region_align_up(max_top));

   _summary_data.clear_range(beg_region, end_region);

   // Clear the data used to 'split' regions.

   SplitInfo& split_info = _space_info[id].split_info();

   if (split_info.is_valid()) {

     split_info.clear();

   }

   DEBUG_ONLY(split_info.verify_clear();)

 }

 void PSParallelCompact::pre_compact(PreGCValues* pre_gc_values)

 {

   // Update the from & to space pointers in space_info, since they are swapped

   // at each young gen gc.  Do the update unconditionally (even though a

   // promotion failure does not swap spaces) because an unknown number of minor

   // collections will have swapped the spaces an unknown number of times.

   GCTraceTime tm("pre compact", print_phases(), true, &_gc_timer);

   ParallelScavengeHeap* heap = gc_heap();

   _space_info[from_space_id].set_space(heap->young_gen()->from_space());

   _space_info[to_space_id].set_space(heap->young_gen()->to_space());

   pre_gc_values->fill(heap);

   DEBUG_ONLY(add_obj_count = add_obj_size = 0;)

   DEBUG_ONLY(mark_bitmap_count = mark_bitmap_size = 0;)

   // Increment the invocation count

   heap->increment_total_collections(true);

   // We need to track unique mark sweep invocations as well.

   _total_invocations++;

   heap->print_heap_before_gc();

   heap->trace_heap_before_gc(&_gc_tracer);

   // Fill in TLABs

   heap->accumulate_statistics_all_tlabs();

   heap->ensure_parsability(true);  // retire TLABs

   if (VerifyBeforeGC && heap->total_collections() >= VerifyGCStartAt) {

     HandleMark hm;  // Discard invalid handles created during verification

     Universe::verify(" VerifyBeforeGC:");

   }

   // Verify object start arrays

   if (VerifyObjectStartArray &&

       VerifyBeforeGC) {

     heap->old_gen()->verify_object_start_array();

   }

   DEBUG_ONLY(mark_bitmap()->verify_clear();)

   DEBUG_ONLY(summary_data().verify_clear();)

   // Have worker threads release resources the next time they run a task.

   gc_task_manager()->release_all_resources();

 }

 void PSParallelCompact::post_compact()

 {

   GCTraceTime tm("post compact", print_phases(), true, &_gc_timer);

   for (unsigned int id = old_space_id; id < last_space_id; ++id) {

     // Clear the marking bitmap, summary data and split info.

     clear_data_covering_space(SpaceId(id));

     // Update top().  Must be done after clearing the bitmap and summary data.

     _space_info[id].publish_new_top();

   }

   MutableSpace* const eden_space = _space_info[eden_space_id].space();

   MutableSpace* const from_space = _space_info[from_space_id].space();

   MutableSpace* const to_space   = _space_info[to_space_id].space();

   ParallelScavengeHeap* heap = gc_heap();

   bool eden_empty = eden_space->is_empty();

   if (!eden_empty) {

     eden_empty = absorb_live_data_from_eden(heap->size_policy(),

                                             heap->young_gen(), heap->old_gen());

   }

   // Update heap occupancy information which is used as input to the soft ref

   // clearing policy at the next gc.

   Universe::update_heap_info_at_gc();

   bool young_gen_empty = eden_empty && from_space->is_empty() &&

     to_space->is_empty();

   BarrierSet* bs = heap->barrier_set();

   if (bs->is_a(BarrierSet::ModRef)) {

     ModRefBarrierSet* modBS = (ModRefBarrierSet*)bs;

     MemRegion old_mr = heap->old_gen()->reserved();

     if (young_gen_empty) {

       modBS->clear(MemRegion(old_mr.start(), old_mr.end()));

     } else {

       modBS->invalidate(MemRegion(old_mr.start(), old_mr.end()));

     }

   }

   // Delete metaspaces for unloaded class loaders and clean up loader_data graph

   ClassLoaderDataGraph::purge();

   MetaspaceAux::verify_metrics();

   Threads::gc_epilogue();

   CodeCache::gc_epilogue();

   JvmtiExport::gc_epilogue();

   COMPILER2_PRESENT(DerivedPointerTable::update_pointers());

   ref_processor()->enqueue_discovered_references(NULL);

   if (ZapUnusedHeapArea) {

     heap->gen_mangle_unused_area();

   }

   // Update time of last GC

   reset_millis_since_last_gc();

 }

 HeapWord*

 PSParallelCompact::compute_dense_prefix_via_density(const SpaceId id,

                                                     bool maximum_compaction)

 {

   const size_t region_size = ParallelCompactData::RegionSize;

   const ParallelCompactData& sd = summary_data();

   const MutableSpace* const space = _space_info[id].space();

   HeapWord* const top_aligned_up = sd.region_align_up(space->top());

   const RegionData* const beg_cp = sd.addr_to_region_ptr(space->bottom());

   const RegionData* const end_cp = sd.addr_to_region_ptr(top_aligned_up);

   // Skip full regions at the beginning of the space--they are necessarily part

   // of the dense prefix.

   size_t full_count = 0;

   const RegionData* cp;

   for (cp = beg_cp; cp < end_cp && cp->data_size() == region_size; ++cp) {

     ++full_count;

   }

   assert(total_invocations() >= _maximum_compaction_gc_num, "sanity");

   const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num;

   const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval;

   if (maximum_compaction || cp == end_cp || interval_ended) {

     _maximum_compaction_gc_num = total_invocations();

     return sd.region_to_addr(cp);

   }

   HeapWord* const new_top = _space_info[id].new_top();

   const size_t space_live = pointer_delta(new_top, space->bottom());

   const size_t space_used = space->used_in_words();

   const size_t space_capacity = space->capacity_in_words();

   const double cur_density = double(space_live) / space_capacity;

   const double deadwood_density =

     (1.0 - cur_density) * (1.0 - cur_density) * cur_density * cur_density;

   const size_t deadwood_goal = size_t(space_capacity * deadwood_density);

   if (TraceParallelOldGCDensePrefix) {

     tty->print_cr("cur_dens=%5.3f dw_dens=%5.3f dw_goal=" SIZE_FORMAT,

                   cur_density, deadwood_density, deadwood_goal);

     tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " "

                   "space_cap=" SIZE_FORMAT,

                   space_live, space_used,

                   space_capacity);

   }

   // XXX - Use binary search?

   HeapWord* dense_prefix = sd.region_to_addr(cp);

   const RegionData* full_cp = cp;

   const RegionData* const top_cp = sd.addr_to_region_ptr(space->top() - 1);

   while (cp < end_cp) {

     HeapWord* region_destination = cp->destination();

     const size_t cur_deadwood = pointer_delta(dense_prefix, region_destination);

     if (TraceParallelOldGCDensePrefix && Verbose) {

       tty->print_cr("c#=" SIZE_FORMAT_W(4) " dst=" PTR_FORMAT " "

                     "dp=" SIZE_FORMAT_W(8) " " "cdw=" SIZE_FORMAT_W(8),

                     sd.region(cp), region_destination,

                     dense_prefix, cur_deadwood);

     }

     if (cur_deadwood >= deadwood_goal) {

       // Found the region that has the correct amount of deadwood to the left.

       // This typically occurs after crossing a fairly sparse set of regions, so

       // iterate backwards over those sparse regions, looking for the region

       // that has the lowest density of live objects 'to the right.'

       size_t space_to_left = sd.region(cp) * region_size;

       size_t live_to_left = space_to_left - cur_deadwood;

       size_t space_to_right = space_capacity - space_to_left;

       size_t live_to_right = space_live - live_to_left;

       double density_to_right = double(live_to_right) / space_to_right;

       while (cp > full_cp) {

         --cp;

         const size_t prev_region_live_to_right = live_to_right -

           cp->data_size();

         const size_t prev_region_space_to_right = space_to_right + region_size;

         double prev_region_density_to_right =

           double(prev_region_live_to_right) / prev_region_space_to_right;

         if (density_to_right <= prev_region_density_to_right) {

           return dense_prefix;

         }

         if (TraceParallelOldGCDensePrefix && Verbose) {

           tty->print_cr("backing up from c=" SIZE_FORMAT_W(4) " d2r=%10.8f "

                         "pc_d2r=%10.8f", sd.region(cp), density_to_right,

                         prev_region_density_to_right);

         }

         dense_prefix -= region_size;

         live_to_right = prev_region_live_to_right;

         space_to_right = prev_region_space_to_right;

         density_to_right = prev_region_density_to_right;

       }

       return dense_prefix;

     }

     dense_prefix += region_size;

     ++cp;

   }

   return dense_prefix;

 }

 #ifndef PRODUCT

 void PSParallelCompact::print_dense_prefix_stats(const char* const algorithm,

                                                  const SpaceId id,

                                                  const bool maximum_compaction,

                                                  HeapWord* const addr)

 {

   const size_t region_idx = summary_data().addr_to_region_idx(addr);

   RegionData* const cp = summary_data().region(region_idx);

   const MutableSpace* const space = _space_info[id].space();

   HeapWord* const new_top = _space_info[id].new_top();

   const size_t space_live = pointer_delta(new_top, space->bottom());

   const size_t dead_to_left = pointer_delta(addr, cp->destination());

   const size_t space_cap = space->capacity_in_words();

   const double dead_to_left_pct = double(dead_to_left) / space_cap;

   const size_t live_to_right = new_top - cp->destination();

   const size_t dead_to_right = space->top() - addr - live_to_right;

   tty->print_cr("%s=" PTR_FORMAT " dpc=" SIZE_FORMAT_W(5) " "

                 "spl=" SIZE_FORMAT " "

                 "d2l=" SIZE_FORMAT " d2l%%=%6.4f "

                 "d2r=" SIZE_FORMAT " l2r=" SIZE_FORMAT

                 " ratio=%10.8f",

                 algorithm, addr, region_idx,

                 space_live,

                 dead_to_left, dead_to_left_pct,

                 dead_to_right, live_to_right,

                 double(dead_to_right) / live_to_right);

 }

 #endif  // #ifndef PRODUCT

 // Return a fraction indicating how much of the generation can be treated as

 // "dead wood" (i.e., not reclaimed).  The function uses a normal distribution

 // based on the density of live objects in the generation to determine a limit,

 // which is then adjusted so the return value is min_percent when the density is

 // 1.

 //

 // The following table shows some return values for a different values of the

 // standard deviation (ParallelOldDeadWoodLimiterStdDev); the mean is 0.5 and

 // min_percent is 1.

 //

 //                          fraction allowed as dead wood

 //         -----------------------------------------------------------------

 // density std_dev=70 std_dev=75 std_dev=80 std_dev=85 std_dev=90 std_dev=95

 // ------- ---------- ---------- ---------- ---------- ---------- ----------

 // 0.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000

 // 0.05000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941

 // 0.10000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272

 // 0.15000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066

 // 0.20000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975

 // 0.25000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313

 // 0.30000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132

 // 0.35000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289

 // 0.40000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500

 // 0.45000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386

 // 0.50000 0.13832410 0.11599237 0.09847664 0.08456518 0.07338887 0.06431510

 // 0.55000 0.13687208 0.11481163 0.09750361 0.08375387 0.07270534 0.06373386

 // 0.60000 0.13253818 0.11128511 0.09459590 0.08132834 0.07066107 0.06199500

 // 0.65000 0.12538832 0.10545958 0.08978741 0.07731366 0.06727491 0.05911289

 // 0.70000 0.11553050 0.09741183 0.08313394 0.07175114 0.06257797 0.05511132

 // 0.75000 0.10311208 0.08724696 0.07471205 0.06469760 0.05661313 0.05002313

 // 0.80000 0.08831616 0.07509618 0.06461766 0.05622444 0.04943437 0.04388975

 // 0.85000 0.07135702 0.06111390 0.05296419 0.04641639 0.04110601 0.03676066

 // 0.90000 0.05247504 0.04547452 0.03988045 0.03537016 0.03170171 0.02869272

 // 0.95000 0.03193096 0.02836880 0.02550828 0.02319280 0.02130337 0.01974941

 // 1.00000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000 0.01000000

 double PSParallelCompact::dead_wood_limiter(double density, size_t min_percent)

 {

   assert(_dwl_initialized, "uninitialized");

   // The raw limit is the value of the normal distribution at x = density.

   const double raw_limit = normal_distribution(density);

   // Adjust the raw limit so it becomes the minimum when the density is 1.

   //

   // First subtract the adjustment value (which is simply the precomputed value

   // normal_distribution(1.0)); this yields a value of 0 when the density is 1.

   // Then add the minimum value, so the minimum is returned when the density is

   // 1.  Finally, prevent negative values, which occur when the mean is not 0.5.

   const double min = double(min_percent) / 100.0;

   const double limit = raw_limit - _dwl_adjustment + min;

   return MAX2(limit, 0.0);

 }

 ParallelCompactData::RegionData*

 PSParallelCompact::first_dead_space_region(const RegionData* beg,

                                            const RegionData* end)

 {

   const size_t region_size = ParallelCompactData::RegionSize;

   ParallelCompactData& sd = summary_data();

   size_t left = sd.region(beg);

   size_t right = end > beg ? sd.region(end) - 1 : left;

   // Binary search.

   while (left < right) {

     // Equivalent to (left + right) / 2, but does not overflow.

     const size_t middle = left + (right - left) / 2;

     RegionData* const middle_ptr = sd.region(middle);

     HeapWord* const dest = middle_ptr->destination();

     HeapWord* const addr = sd.region_to_addr(middle);

     assert(dest != NULL, "sanity");

     assert(dest <= addr, "must move left");

     if (middle > left && dest < addr) {

       right = middle - 1;

     } else if (middle < right && middle_ptr->data_size() == region_size) {

       left = middle + 1;

     } else {

       return middle_ptr;

     }

   }

   return sd.region(left);

 }

 ParallelCompactData::RegionData*

 PSParallelCompact::dead_wood_limit_region(const RegionData* beg,

                                           const RegionData* end,

                                           size_t dead_words)

 {

   ParallelCompactData& sd = summary_data();

   size_t left = sd.region(beg);

   size_t right = end > beg ? sd.region(end) - 1 : left;

   // Binary search.

   while (left < right) {

     // Equivalent to (left + right) / 2, but does not overflow.

     const size_t middle = left + (right - left) / 2;

     RegionData* const middle_ptr = sd.region(middle);

     HeapWord* const dest = middle_ptr->destination();

     HeapWord* const addr = sd.region_to_addr(middle);

     assert(dest != NULL, "sanity");

     assert(dest <= addr, "must move left");

     const size_t dead_to_left = pointer_delta(addr, dest);

     if (middle > left && dead_to_left > dead_words) {

       right = middle - 1;

     } else if (middle < right && dead_to_left < dead_words) {

       left = middle + 1;

     } else {

       return middle_ptr;

     }

   }

   return sd.region(left);

 }

 // The result is valid during the summary phase, after the initial summarization

 // of each space into itself, and before final summarization.

 inline double

 PSParallelCompact::reclaimed_ratio(const RegionData* const cp,

                                    HeapWord* const bottom,

                                    HeapWord* const top,

                                    HeapWord* const new_top)

 {

   ParallelCompactData& sd = summary_data();

   assert(cp != NULL, "sanity");

   assert(bottom != NULL, "sanity");

   assert(top != NULL, "sanity");

   assert(new_top != NULL, "sanity");

   assert(top >= new_top, "summary data problem?");

   assert(new_top > bottom, "space is empty; should not be here");

   assert(new_top >= cp->destination(), "sanity");

   assert(top >= sd.region_to_addr(cp), "sanity");

   HeapWord* const destination = cp->destination();

   const size_t dense_prefix_live  = pointer_delta(destination, bottom);

   const size_t compacted_region_live = pointer_delta(new_top, destination);

   const size_t compacted_region_used = pointer_delta(top,

                                                      sd.region_to_addr(cp));

   const size_t reclaimable = compacted_region_used - compacted_region_live;

   const double divisor = dense_prefix_live + 1.25 * compacted_region_live;

   return double(reclaimable) / divisor;

 }

 // Return the address of the end of the dense prefix, a.k.a. the start of the

 // compacted region.  The address is always on a region boundary.

 //

 // Completely full regions at the left are skipped, since no compaction can

 // occur in those regions.  Then the maximum amount of dead wood to allow is

 // computed, based on the density (amount live / capacity) of the generation;

 // the region with approximately that amount of dead space to the left is

 // identified as the limit region.  Regions between the last completely full

 // region and the limit region are scanned and the one that has the best

 // (maximum) reclaimed_ratio() is selected.

 HeapWord*

 PSParallelCompact::compute_dense_prefix(const SpaceId id,

                                         bool maximum_compaction)

 {

   if (ParallelOldGCSplitALot) {

     if (_space_info[id].dense_prefix() != _space_info[id].space()->bottom()) {

       // The value was chosen to provoke splitting a young gen space; use it.

       return _space_info[id].dense_prefix();

     }

   }

   const size_t region_size = ParallelCompactData::RegionSize;

   const ParallelCompactData& sd = summary_data();

   const MutableSpace* const space = _space_info[id].space();

   HeapWord* const top = space->top();

   HeapWord* const top_aligned_up = sd.region_align_up(top);

   HeapWord* const new_top = _space_info[id].new_top();

   HeapWord* const new_top_aligned_up = sd.region_align_up(new_top);

   HeapWord* const bottom = space->bottom();

   const RegionData* const beg_cp = sd.addr_to_region_ptr(bottom);

   const RegionData* const top_cp = sd.addr_to_region_ptr(top_aligned_up);

   const RegionData* const new_top_cp =

     sd.addr_to_region_ptr(new_top_aligned_up);

   // Skip full regions at the beginning of the space--they are necessarily part

   // of the dense prefix.

   const RegionData* const full_cp = first_dead_space_region(beg_cp, new_top_cp);

   assert(full_cp->destination() == sd.region_to_addr(full_cp) ||

          space->is_empty(), "no dead space allowed to the left");

   assert(full_cp->data_size() < region_size || full_cp == new_top_cp - 1,

          "region must have dead space");

   // The gc number is saved whenever a maximum compaction is done, and used to

   // determine when the maximum compaction interval has expired.  This avoids

   // successive max compactions for different reasons.

   assert(total_invocations() >= _maximum_compaction_gc_num, "sanity");

   const size_t gcs_since_max = total_invocations() - _maximum_compaction_gc_num;

   const bool interval_ended = gcs_since_max > HeapMaximumCompactionInterval ||

     total_invocations() == HeapFirstMaximumCompactionCount;

   if (maximum_compaction || full_cp == top_cp || interval_ended) {

     _maximum_compaction_gc_num = total_invocations();

     return sd.region_to_addr(full_cp);

   }

   const size_t space_live = pointer_delta(new_top, bottom);

   const size_t space_used = space->used_in_words();

   const size_t space_capacity = space->capacity_in_words();

   const double density = double(space_live) / double(space_capacity);

   const size_t min_percent_free = MarkSweepDeadRatio;

   const double limiter = dead_wood_limiter(density, min_percent_free);

   const size_t dead_wood_max = space_used - space_live;

   const size_t dead_wood_limit = MIN2(size_t(space_capacity * limiter),

                                       dead_wood_max);

   if (TraceParallelOldGCDensePrefix) {

     tty->print_cr("space_live=" SIZE_FORMAT " " "space_used=" SIZE_FORMAT " "

                   "space_cap=" SIZE_FORMAT,

                   space_live, space_used,

                   space_capacity);

     tty->print_cr("dead_wood_limiter(%6.4f, %d)=%6.4f "

                   "dead_wood_max=" SIZE_FORMAT " dead_wood_limit=" SIZE_FORMAT,

                   density, min_percent_free, limiter,

                   dead_wood_max, dead_wood_limit);

   }

   // Locate the region with the desired amount of dead space to the left.

   const RegionData* const limit_cp =

     dead_wood_limit_region(full_cp, top_cp, dead_wood_limit);

   // Scan from the first region with dead space to the limit region and find the

   // one with the best (largest) reclaimed ratio.

   double best_ratio = 0.0;

   const RegionData* best_cp = full_cp;

   for (const RegionData* cp = full_cp; cp < limit_cp; ++cp) {

     double tmp_ratio = reclaimed_ratio(cp, bottom, top, new_top);

     if (tmp_ratio > best_ratio) {

       best_cp = cp;

       best_ratio = tmp_ratio;

     }

   }

 #if     0

   // Something to consider:  if the region with the best ratio is 'close to' the

   // first region w/free space, choose the first region with free space

   // ("first-free").  The first-free region is usually near the start of the

   // heap, which means we are copying most of the heap already, so copy a bit

   // more to get complete compaction.

   if (pointer_delta(best_cp, full_cp, sizeof(RegionData)) < 4) {

     _maximum_compaction_gc_num = total_invocations();

     best_cp = full_cp;

   }

 #endif  // #if 0

   return sd.region_to_addr(best_cp);

 }

 #ifndef PRODUCT

 void

 PSParallelCompact::fill_with_live_objects(SpaceId id, HeapWord* const start,

                                           size_t words)

 {

   if (TraceParallelOldGCSummaryPhase) {

     tty->print_cr("fill_with_live_objects [" PTR_FORMAT " " PTR_FORMAT ") "

                   SIZE_FORMAT, start, start + words, words);

   }

   ObjectStartArray* const start_array = _space_info[id].start_array();

   CollectedHeap::fill_with_objects(start, words);

   for (HeapWord* p = start; p < start + words; p += oop(p)->size()) {

     _mark_bitmap.mark_obj(p, words);

     _summary_data.add_obj(p, words);

     start_array->allocate_block(p);

   }

 }

 void

 PSParallelCompact::summarize_new_objects(SpaceId id, HeapWord* start)

 {

   ParallelCompactData& sd = summary_data();

   MutableSpace* space = _space_info[id].space();

   // Find the source and destination start addresses.

   HeapWord* const src_addr = sd.region_align_down(start);

   HeapWord* dst_addr;

   if (src_addr < start) {

     dst_addr = sd.addr_to_region_ptr(src_addr)->destination();

   } else if (src_addr > space->bottom()) {

     // The start (the original top() value) is aligned to a region boundary so

     // the associated region does not have a destination.  Compute the

     // destination from the previous region.

     RegionData* const cp = sd.addr_to_region_ptr(src_addr) - 1;

     dst_addr = cp->destination() + cp->data_size();

   } else {

     // Filling the entire space.

     dst_addr = space->bottom();

   }

   assert(dst_addr != NULL, "sanity");

   // Update the summary data.

   bool result = _summary_data.summarize(_space_info[id].split_info(),

                                         src_addr, space->top(), NULL,

                                         dst_addr, space->end(),

                                         _space_info[id].new_top_addr());

   assert(result, "should not fail:  bad filler object size");

 }

 void

 PSParallelCompact::provoke_split_fill_survivor(SpaceId id)

 {

   if (total_invocations() % (ParallelOldGCSplitInterval * 3) != 0) {

     return;

   }

   MutableSpace* const space = _space_info[id].space();

   if (space->is_empty()) {

     HeapWord* b = space->bottom();

     HeapWord* t = b + space->capacity_in_words() / 2;

     space->set_top(t);

     if (ZapUnusedHeapArea) {

       space->set_top_for_allocations();

     }

     size_t min_size = CollectedHeap::min_fill_size();

     size_t obj_len = min_size;

     while (b + obj_len <= t) {

       CollectedHeap::fill_with_object(b, obj_len);

       mark_bitmap()->mark_obj(b, obj_len);

       summary_data().add_obj(b, obj_len);

       b += obj_len;

       obj_len = (obj_len & (min_size*3)) + min_size; // 8 16 24 32 8 16 24 32 ...

     }

     if (b < t) {

       // The loop didn't completely fill to t (top); adjust top downward.

       space->set_top(b);

       if (ZapUnusedHeapArea) {

         space->set_top_for_allocations();

       }

     }

     HeapWord** nta = _space_info[id].new_top_addr();

     bool result = summary_data().summarize(_space_info[id].split_info(),

                                            space->bottom(), space->top(), NULL,

                                            space->bottom(), space->end(), nta);

     assert(result, "space must fit into itself");

   }

 }

 void

 PSParallelCompact::provoke_split(bool & max_compaction)

 {

   if (total_invocations() % ParallelOldGCSplitInterval != 0) {

     return;

   }

   const size_t region_size = ParallelCompactData::RegionSize;

   ParallelCompactData& sd = summary_data();

   MutableSpace* const eden_space = _space_info[eden_space_id].space();

   MutableSpace* const from_space = _space_info[from_space_id].space();

   const size_t eden_live = pointer_delta(eden_space->top(),

                                          _space_info[eden_space_id].new_top());

   const size_t from_live = pointer_delta(from_space->top(),

                                          _space_info[from_space_id].new_top());

   const size_t min_fill_size = CollectedHeap::min_fill_size();

   const size_t eden_free = pointer_delta(eden_space->end(), eden_space->top());

   const size_t eden_fillable = eden_free >= min_fill_size ? eden_free : 0;

   const size_t from_free = pointer_delta(from_space->end(), from_space->top());

   const size_t from_fillable = from_free >= min_fill_size ? from_free : 0;

   // Choose the space to split; need at least 2 regions live (or fillable).

   SpaceId id;

   MutableSpace* space;

   size_t live_words;

   size_t fill_words;

   if (eden_live + eden_fillable >= region_size * 2) {

     id = eden_space_id;

     space = eden_space;

     live_words = eden_live;

     fill_words = eden_fillable;

   } else if (from_live + from_fillable >= region_size * 2) {

     id = from_space_id;

     space = from_space;

     live_words = from_live;

     fill_words = from_fillable;

   } else {

     return; // Give up.

   }

   assert(fill_words == 0 || fill_words >= min_fill_size, "sanity");

   if (live_words < region_size * 2) {

     // Fill from top() to end() w/live objects of mixed sizes.

     HeapWord* const fill_start = space->top();

     live_words += fill_words;

     space->set_top(fill_start + fill_words);

     if (ZapUnusedHeapArea) {

       space->set_top_for_allocations();

     }

     HeapWord* cur_addr = fill_start;

     while (fill_words > 0) {

       const size_t r = (size_t)os::random() % (region_size / 2) + min_fill_size;

       size_t cur_size = MIN2(align_object_size_(r), fill_words);

       if (fill_words - cur_size < min_fill_size) {

         cur_size = fill_words; // Avoid leaving a fragment too small to fill.

       }

       CollectedHeap::fill_with_object(cur_addr, cur_size);

       mark_bitmap()->mark_obj(cur_addr, cur_size);

       sd.add_obj(cur_addr, cur_size);

       cur_addr += cur_size;

       fill_words -= cur_size;

     }

     summarize_new_objects(id, fill_start);

   }

   max_compaction = false;

   // Manipulate the old gen so that it has room for about half of the live data

   // in the target young gen space (live_words / 2).

   id = old_space_id;

   space = _space_info[id].space();

   const size_t free_at_end = space->free_in_words();

   const size_t free_target = align_object_size(live_words / 2);

   const size_t dead = pointer_delta(space->top(), _space_info[id].new_top());

   if (free_at_end >= free_target + min_fill_size) {

     // Fill space above top() and set the dense prefix so everything survives.

     HeapWord* const fill_start = space->top();

     const size_t fill_size = free_at_end - free_target;

     space->set_top(space->top() + fill_size);

     if (ZapUnusedHeapArea) {

       space->set_top_for_allocations();

     }

     fill_with_live_objects(id, fill_start, fill_size);

     summarize_new_objects(id, fill_start);

     _space_info[id].set_dense_prefix(sd.region_align_down(space->top()));

   } else if (dead + free_at_end > free_target) {

     // Find a dense prefix that makes the right amount of space available.

     HeapWord* cur = sd.region_align_down(space->top());

     HeapWord* cur_destination = sd.addr_to_region_ptr(cur)->destination();

     size_t dead_to_right = pointer_delta(space->end(), cur_destination);

     while (dead_to_right < free_target) {

       cur -= region_size;

       cur_destination = sd.addr_to_region_ptr(cur)->destination();

       dead_to_right = pointer_delta(space->end(), cur_destination);

     }

     _space_info[id].set_dense_prefix(cur);

   }

 }

 #endif // #ifndef PRODUCT

 void PSParallelCompact::summarize_spaces_quick()

 {

   for (unsigned int i = 0; i < last_space_id; ++i) {

     const MutableSpace* space = _space_info[i].space();

     HeapWord** nta = _space_info[i].new_top_addr();

     bool result = _summary_data.summarize(_space_info[i].split_info(),

                                           space->bottom(), space->top(), NULL,

                                           space->bottom(), space->end(), nta);

     assert(result, "space must fit into itself");

     _space_info[i].set_dense_prefix(space->bottom());

   }

 #ifndef PRODUCT

   if (ParallelOldGCSplitALot) {

     provoke_split_fill_survivor(to_space_id);

   }

 #endif // #ifndef PRODUCT

 }

 void PSParallelCompact::fill_dense_prefix_end(SpaceId id)

 {

   HeapWord* const dense_prefix_end = dense_prefix(id);

   const RegionData* region = _summary_data.addr_to_region_ptr(dense_prefix_end);

   const idx_t dense_prefix_bit = _mark_bitmap.addr_to_bit(dense_prefix_end);

   if (dead_space_crosses_boundary(region, dense_prefix_bit)) {

     // Only enough dead space is filled so that any remaining dead space to the

     // left is larger than the minimum filler object.  (The remainder is filled

     // during the copy/update phase.)

     //

     // The size of the dead space to the right of the boundary is not a

     // concern, since compaction will be able to use whatever space is

     // available.

     //

     // Here '||' is the boundary, 'x' represents a don't care bit and a box

     // surrounds the space to be filled with an object.

     //

     // In the 32-bit VM, each bit represents two 32-bit words:

     //                              +---+

     // a) beg_bits:  ...  x   x   x | 0 | ||   0   x  x  ...

     //    end_bits:  ...  x   x   x | 0 | ||   0   x  x  ...

     //                              +---+

     //

     // In the 64-bit VM, each bit represents one 64-bit word:

     //                              +------------+

     // b) beg_bits:  ...  x   x   x | 0   ||   0 | x  x  ...

     //    end_bits:  ...  x   x   1 | 0   ||   0 | x  x  ...

     //                              +------------+

     //                          +-------+

     // c) beg_bits:  ...  x   x | 0   0 | ||   0   x  x  ...

     //    end_bits:  ...  x   1 | 0   0 | ||   0   x  x  ...

     //                          +-------+

     //                      +-----------+

     // d) beg_bits:  ...  x | 0   0   0 | ||   0   x  x  ...

     //    end_bits:  ...  1 | 0   0   0 | ||   0   x  x  ...

     //                      +-----------+

     //                          +-------+

     // e) beg_bits:  ...  0   0 | 0   0 | ||   0   x  x  ...

     //    end_bits:  ...  0   0 | 0   0 | ||   0   x  x  ...

     //                          +-------+

     // Initially assume case a, c or e will apply.

     size_t obj_len = CollectedHeap::min_fill_size();

     HeapWord* obj_beg = dense_prefix_end - obj_len;

 #ifdef  _LP64

     if (MinObjAlignment > 1) { // object alignment > heap word size

       // Cases a, c or e.

     } else if (_mark_bitmap.is_obj_end(dense_prefix_bit - 2)) {

       // Case b above.

       obj_beg = dense_prefix_end - 1;

     } else if (!_mark_bitmap.is_obj_end(dense_prefix_bit - 3) &&

                _mark_bitmap.is_obj_end(dense_prefix_bit - 4)) {

       // Case d above.

       obj_beg = dense_prefix_end - 3;

       obj_len = 3;

     }

 #endif  // #ifdef _LP64

     CollectedHeap::fill_with_object(obj_beg, obj_len);

     _mark_bitmap.mark_obj(obj_beg, obj_len);

     _summary_data.add_obj(obj_beg, obj_len);

     assert(start_array(id) != NULL, "sanity");

     start_array(id)->allocate_block(obj_beg);

   }

 }

 void

 PSParallelCompact::clear_source_region(HeapWord* beg_addr, HeapWord* end_addr)

 {

   RegionData* const beg_ptr = _summary_data.addr_to_region_ptr(beg_addr);

   HeapWord* const end_aligned_up = _summary_data.region_align_up(end_addr);

   RegionData* const end_ptr = _summary_data.addr_to_region_ptr(end_aligned_up);

   for (RegionData* cur = beg_ptr; cur < end_ptr; ++cur) {

     cur->set_source_region(0);

   }

 }

 void

 PSParallelCompact::summarize_space(SpaceId id, bool maximum_compaction)

 {

   assert(id < last_space_id, "id out of range");

   assert(_space_info[id].dense_prefix() == _space_info[id].space()->bottom() ||

          ParallelOldGCSplitALot && id == old_space_id,

          "should have been reset in summarize_spaces_quick()");

   const MutableSpace* space = _space_info[id].space();

   if (_space_info[id].new_top() != space->bottom()) {

     HeapWord* dense_prefix_end = compute_dense_prefix(id, maximum_compaction);

     _space_info[id].set_dense_prefix(dense_prefix_end);

 #ifndef PRODUCT

     if (TraceParallelOldGCDensePrefix) {

       print_dense_prefix_stats("ratio", id, maximum_compaction,

                                dense_prefix_end);

       HeapWord* addr = compute_dense_prefix_via_density(id, maximum_compaction);

       print_dense_prefix_stats("density", id, maximum_compaction, addr);

     }

 #endif  // #ifndef PRODUCT

     // Recompute the summary data, taking into account the dense prefix.  If

     // every last byte will be reclaimed, then the existing summary data which

     // compacts everything can be left in place.

     if (!maximum_compaction && dense_prefix_end != space->bottom()) {

       // If dead space crosses the dense prefix boundary, it is (at least

       // partially) filled with a dummy object, marked live and added to the

       // summary data.  This simplifies the copy/update phase and must be done

       // before the final locations of objects are determined, to prevent

       // leaving a fragment of dead space that is too small to fill.

       fill_dense_prefix_end(id);

       // Compute the destination of each Region, and thus each object.

       _summary_data.summarize_dense_prefix(space->bottom(), dense_prefix_end);

       _summary_data.summarize(_space_info[id].split_info(),

                               dense_prefix_end, space->top(), NULL,

                               dense_prefix_end, space->end(),

                               _space_info[id].new_top_addr());

     }

   }

   if (TraceParallelOldGCSummaryPhase) {

     const size_t region_size = ParallelCompactData::RegionSize;

     HeapWord* const dense_prefix_end = _space_info[id].dense_prefix();

     const size_t dp_region = _summary_data.addr_to_region_idx(dense_prefix_end);

     const size_t dp_words = pointer_delta(dense_prefix_end, space->bottom());

     HeapWord* const new_top = _space_info[id].new_top();

     const HeapWord* nt_aligned_up = _summary_data.region_align_up(new_top);

     const size_t cr_words = pointer_delta(nt_aligned_up, dense_prefix_end);

     tty->print_cr("id=%d cap=" SIZE_FORMAT " dp=" PTR_FORMAT " "

                   "dp_region=" SIZE_FORMAT " " "dp_count=" SIZE_FORMAT " "

                   "cr_count=" SIZE_FORMAT " " "nt=" PTR_FORMAT,

                   id, space->capacity_in_words(), dense_prefix_end,

                   dp_region, dp_words / region_size,

                   cr_words / region_size, new_top);

   }

 }

 #ifndef PRODUCT

 void PSParallelCompact::summary_phase_msg(SpaceId dst_space_id,

                                           HeapWord* dst_beg, HeapWord* dst_end,

                                           SpaceId src_space_id,

                                           HeapWord* src_beg, HeapWord* src_end)

 {

   if (TraceParallelOldGCSummaryPhase) {

     tty->print_cr("summarizing %d [%s] into %d [%s]:  "

                   "src=" PTR_FORMAT "-" PTR_FORMAT " "

                   SIZE_FORMAT "-" SIZE_FORMAT " "

                   "dst=" PTR_FORMAT "-" PTR_FORMAT " "

                   SIZE_FORMAT "-" SIZE_FORMAT,

                   src_space_id, space_names[src_space_id],

                   dst_space_id, space_names[dst_space_id],

                   src_beg, src_end,

                   _summary_data.addr_to_region_idx(src_beg),

                   _summary_data.addr_to_region_idx(src_end),

                   dst_beg, dst_end,

                   _summary_data.addr_to_region_idx(dst_beg),

                   _summary_data.addr_to_region_idx(dst_end));

   }

 }

 #endif  // #ifndef PRODUCT

 void PSParallelCompact::summary_phase(ParCompactionManager* cm,

                                       bool maximum_compaction)

 {

   GCTraceTime tm("summary phase", print_phases(), true, &_gc_timer);

   // trace("2");

 #ifdef  ASSERT

   if (TraceParallelOldGCMarkingPhase) {

     tty->print_cr("add_obj_count=" SIZE_FORMAT " "

                   "add_obj_bytes=" SIZE_FORMAT,

                   add_obj_count, add_obj_size * HeapWordSize);

     tty->print_cr("mark_bitmap_count=" SIZE_FORMAT " "

                   "mark_bitmap_bytes=" SIZE_FORMAT,

                   mark_bitmap_count, mark_bitmap_size * HeapWordSize);

   }

 #endif  // #ifdef ASSERT

   // Quick summarization of each space into itself, to see how much is live.

   summarize_spaces_quick();

   if (TraceParallelOldGCSummaryPhase) {

     tty->print_cr("summary_phase:  after summarizing each space to self");

     Universe::print();

     NOT_PRODUCT(print_region_ranges());

     if (Verbose) {

       NOT_PRODUCT(print_initial_summary_data(_summary_data, _space_info));

     }

   }

   // The amount of live data that will end up in old space (assuming it fits).

   size_t old_space_total_live = 0;

   for (unsigned int id = old_space_id; id < last_space_id; ++id) {

     old_space_total_live += pointer_delta(_space_info[id].new_top(),

                                           _space_info[id].space()->bottom());

   }

   MutableSpace* const old_space = _space_info[old_space_id].space();

   const size_t old_capacity = old_space->capacity_in_words();

   if (old_space_total_live > old_capacity) {

     // XXX - should also try to expand

     maximum_compaction = true;

   }

 #ifndef PRODUCT

   if (ParallelOldGCSplitALot && old_space_total_live < old_capacity) {

     provoke_split(maximum_compaction);

   }

 #endif // #ifndef PRODUCT

   // Old generations.

   summarize_space(old_space_id, maximum_compaction);

   // Summarize the remaining spaces in the young gen.  The initial target space

   // is the old gen.  If a space does not fit entirely into the target, then the

   // remainder is compacted into the space itself and that space becomes the new

   // target.

   SpaceId dst_space_id = old_space_id;

   HeapWord* dst_space_end = old_space->end();

   HeapWord** new_top_addr = _space_info[dst_space_id].new_top_addr();

   for (unsigned int id = eden_space_id; id < last_space_id; ++id) {

     const MutableSpace* space = _space_info[id].space();

     const size_t live = pointer_delta(_space_info[id].new_top(),

                                       space->bottom());

     const size_t available = pointer_delta(dst_space_end, *new_top_addr);

     NOT_PRODUCT(summary_phase_msg(dst_space_id, *new_top_addr, dst_space_end,

                                   SpaceId(id), space->bottom(), space->top());)

     if (live > 0 && live <= available) {

       // All the live data will fit.

       bool done = _summary_data.summarize(_space_info[id].split_info(),

                                           space->bottom(), space->top(),

                                           NULL,

                                           *new_top_addr, dst_space_end,

                                           new_top_addr);

       assert(done, "space must fit into old gen");

       // Reset the new_top value for the space.

       _space_info[id].set_new_top(space->bottom());

     } else if (live > 0) {

       // Attempt to fit part of the source space into the target space.

       HeapWord* next_src_addr = NULL;

       bool done = _summary_data.summarize(_space_info[id].split_info(),

                                           space->bottom(), space->top(),

                                           &next_src_addr,

                                           *new_top_addr, dst_space_end,

                                           new_top_addr);

       assert(!done, "space should not fit into old gen");

       assert(next_src_addr != NULL, "sanity");

       // The source space becomes the new target, so the remainder is compacted

       // within the space itself.

       dst_space_id = SpaceId(id);

       dst_space_end = space->end();

       new_top_addr = _space_info[id].new_top_addr();

       NOT_PRODUCT(summary_phase_msg(dst_space_id,

                                     space->bottom(), dst_space_end,

                                     SpaceId(id), next_src_addr, space->top());)

       done = _summary_data.summarize(_space_info[id].split_info(),

                                      next_src_addr, space->top(),

                                      NULL,

                                      space->bottom(), dst_space_end,

                                      new_top_addr);

       assert(done, "space must fit when compacted into itself");

       assert(*new_top_addr <= space->top(), "usage should not grow");

     }

   }

   if (TraceParallelOldGCSummaryPhase) {

     tty->print_cr("summary_phase:  after final summarization");

     Universe::print();

     NOT_PRODUCT(print_region_ranges());

     if (Verbose) {

       NOT_PRODUCT(print_generic_summary_data(_summary_data, _space_info));

     }

   }

 }

 // This method should contain all heap-specific policy for invoking a full

 // collection.  invoke_no_policy() will only attempt to compact the heap; it

 // will do nothing further.  If we need to bail out for policy reasons, scavenge

 // before full gc, or any other specialized behavior, it needs to be added here.

 //

 // Note that this method should only be called from the vm_thread while at a

 // safepoint.

 //

 // Note that the all_soft_refs_clear flag in the collector policy

 // may be true because this method can be called without intervening

 // activity.  For example when the heap space is tight and full measure

 // are being taken to free space.

 void PSParallelCompact::invoke(bool maximum_heap_compaction) {

   assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");

   assert(Thread::current() == (Thread*)VMThread::vm_thread(),

          "should be in vm thread");

   ParallelScavengeHeap* heap = gc_heap();

   GCCause::Cause gc_cause = heap->gc_cause();

   assert(!heap->is_gc_active(), "not reentrant");

   PSAdaptiveSizePolicy* policy = heap->size_policy();

   IsGCActiveMark mark;

   if (ScavengeBeforeFullGC) {

     PSScavenge::invoke_no_policy();

   }

   const bool clear_all_soft_refs =

     heap->collector_policy()->should_clear_all_soft_refs();

   PSParallelCompact::invoke_no_policy(clear_all_soft_refs ||

                                       maximum_heap_compaction);

 }

 // This method contains no policy. You should probably

 // be calling invoke() instead.

 bool PSParallelCompact::invoke_no_policy(bool maximum_heap_compaction) {

   assert(SafepointSynchronize::is_at_safepoint(), "must be at a safepoint");

   assert(ref_processor() != NULL, "Sanity");

   if (GC_locker::check_active_before_gc()) {

     return false;

   }

   ParallelScavengeHeap* heap = gc_heap();

   _gc_timer.register_gc_start();

   _gc_tracer.report_gc_start(heap->gc_cause(), _gc_timer.gc_start());

   TimeStamp marking_start;

   TimeStamp compaction_start;

   TimeStamp collection_exit;

   GCCause::Cause gc_cause = heap->gc_cause();

   PSYoungGen* young_gen = heap->young_gen();

   PSOldGen* old_gen = heap->old_gen();

   PSAdaptiveSizePolicy* size_policy = heap->size_policy();

   // The scope of casr should end after code that can change

   // CollectorPolicy::_should_clear_all_soft_refs.

   ClearedAllSoftRefs casr(maximum_heap_compaction,

                           heap->collector_policy());

   if (ZapUnusedHeapArea) {

     // Save information needed to minimize mangling

     heap->record_gen_tops_before_GC();

   }

   heap->pre_full_gc_dump(&_gc_timer);

   _print_phases = PrintGCDetails && PrintParallelOldGCPhaseTimes;

   // Make sure data structures are sane, make the heap parsable, and do other

   // miscellaneous bookkeeping.

   PreGCValues pre_gc_values;

   pre_compact(&pre_gc_values);

   // Get the compaction manager reserved for the VM thread.

   ParCompactionManager* const vmthread_cm =

     ParCompactionManager::manager_array(gc_task_manager()->workers());

   // Place after pre_compact() where the number of invocations is incremented.

   AdaptiveSizePolicyOutput(size_policy, heap->total_collections());

   {

     ResourceMark rm;

     HandleMark hm;

     // Set the number of GC threads to be used in this collection

     gc_task_manager()->set_active_gang();

     gc_task_manager()->task_idle_workers();

     heap->set_par_threads(gc_task_manager()->active_workers());

     gclog_or_tty->date_stamp(PrintGC && PrintGCDateStamps);

     TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);

     GCTraceTime t1(GCCauseString("Full GC", gc_cause), PrintGC, !PrintGCDetails, NULL);

     TraceCollectorStats tcs(counters());

     TraceMemoryManagerStats tms(true /* Full GC */,gc_cause);

     if (TraceGen1Time) accumulated_time()->start();

     // Let the size policy know we're starting

     size_policy->major_collection_begin();

     CodeCache::gc_prologue();

     Threads::gc_prologue();

     COMPILER2_PRESENT(DerivedPointerTable::clear());

     ref_processor()->enable_discovery(true /*verify_disabled*/, true /*verify_no_refs*/);

     ref_processor()->setup_policy(maximum_heap_compaction);

     bool marked_for_unloading = false;

     marking_start.update();

     marking_phase(vmthread_cm, maximum_heap_compaction, &_gc_tracer);

     bool max_on_system_gc = UseMaximumCompactionOnSystemGC

       && gc_cause == GCCause::_java_lang_system_gc;

     summary_phase(vmthread_cm, maximum_heap_compaction || max_on_system_gc);

     COMPILER2_PRESENT(assert(DerivedPointerTable::is_active(), "Sanity"));

     COMPILER2_PRESENT(DerivedPointerTable::set_active(false));

     // adjust_roots() updates Universe::_intArrayKlassObj which is

     // needed by the compaction for filling holes in the dense prefix.

     adjust_roots();

     compaction_start.update();

     compact();

     // Reset the mark bitmap, summary data, and do other bookkeeping.  Must be

     // done before resizing.

     post_compact();

     // Let the size policy know we're done

     size_policy->major_collection_end(old_gen->used_in_bytes(), gc_cause);

     /* 2014/2/12/ Liao: In UseOldNUMA, the size of old-gen should not be changed as the young-gen,

      * when minorGC happens, the eden size will be emptied, so we can change the size of the eden

      * size and then bind to each NUMA group, but in fullGC, the old-gen should not be emptied, so

      * we let the old-gen should not be changed here. */

     if (UseAdaptiveSizePolicy && !UseOldNUMA) {

       if (PrintAdaptiveSizePolicy) {

         gclog_or_tty->print("AdaptiveSizeStart: ");

         gclog_or_tty->stamp();

         gclog_or_tty->print_cr(" collection: %d ",

                        heap->total_collections());

         if (Verbose) {

           gclog_or_tty->print("old_gen_capacity: %d young_gen_capacity: %d",

             old_gen->capacity_in_bytes(), young_gen->capacity_in_bytes());

         }

       }

       // Don't check if the size_policy is ready here.  Let

       // the size_policy check that internally.

       if (UseAdaptiveGenerationSizePolicyAtMajorCollection &&

           ((gc_cause != GCCause::_java_lang_system_gc) ||

             UseAdaptiveSizePolicyWithSystemGC)) {

         // Calculate optimal free space amounts

         assert(young_gen->max_size() >

           young_gen->from_space()->capacity_in_bytes() +

           young_gen->to_space()->capacity_in_bytes(),

           "Sizes of space in young gen are out-of-bounds");

         size_t young_live = young_gen->used_in_bytes();

         size_t eden_live = young_gen->eden_space()->used_in_bytes();

         size_t old_live = old_gen->used_in_bytes();

         size_t cur_eden = young_gen->eden_space()->capacity_in_bytes();

         size_t max_old_gen_size = old_gen->max_gen_size();

         size_t max_eden_size = young_gen->max_size() -

           young_gen->from_space()->capacity_in_bytes() -

           young_gen->to_space()->capacity_in_bytes();

         // Used for diagnostics

         size_policy->clear_generation_free_space_flags();

         size_policy->compute_generations_free_space(young_live,

                                                     eden_live,

                                                     old_live,

                                                     cur_eden,

                                                     max_old_gen_size,

                                                     max_eden_size,

                                                     true /* full gc*/);

         size_policy->check_gc_overhead_limit(young_live,

                                              eden_live,

                                              max_old_gen_size,

                                              max_eden_size,

                                              true /* full gc*/,

                                              gc_cause,

                                              heap->collector_policy());

         size_policy->decay_supplemental_growth(true /* full gc*/);

         heap->resize_old_gen(

           size_policy->calculated_old_free_size_in_bytes());

         // Don't resize the young generation at an major collection.  A

         // desired young generation size may have been calculated but

         // resizing the young generation complicates the code because the

         // resizing of the old generation may have moved the boundary

         // between the young generation and the old generation.  Let the

         // young generation resizing happen at the minor collections.

       }

       if (PrintAdaptiveSizePolicy) {

         gclog_or_tty->print_cr("AdaptiveSizeStop: collection: %d ",

                        heap->total_collections());

       }

     }

     if (UsePerfData) {

       PSGCAdaptivePolicyCounters* const counters = heap->gc_policy_counters();

       counters->update_counters();

       counters->update_old_capacity(old_gen->capacity_in_bytes());

       counters->update_young_capacity(young_gen->capacity_in_bytes());

     }

     heap->resize_all_tlabs();

     // Resize the metaspace capactiy after a collection

     MetaspaceGC::compute_new_size();

     if (TraceGen1Time) accumulated_time()->stop();

     if (PrintGC) {

       if (PrintGCDetails) {

         // No GC timestamp here.  This is after GC so it would be confusing.

         young_gen->print_used_change(pre_gc_values.young_gen_used());

         old_gen->print_used_change(pre_gc_values.old_gen_used());

         heap->print_heap_change(pre_gc_values.heap_used());

         MetaspaceAux::print_metaspace_change(pre_gc_values.metadata_used());

       } else {

         heap->print_heap_change(pre_gc_values.heap_used());

       }

     }

     // Track memory usage and detect low memory

     MemoryService::track_memory_usage();

     heap->update_counters();

     gc_task_manager()->release_idle_workers();

   }

 #ifdef ASSERT

   for (size_t i = 0; i < ParallelGCThreads + 1; ++i) {

     ParCompactionManager* const cm =

       ParCompactionManager::manager_array(int(i));

     assert(cm->marking_stack()->is_empty(),       "should be empty");

     assert(ParCompactionManager::region_list(int(i))->is_empty(), "should be empty");

   }

 #endif // ASSERT

   if (VerifyAfterGC && heap->total_collections() >= VerifyGCStartAt) {

     HandleMark hm;  // Discard invalid handles created during verification

     Universe::verify(" VerifyAfterGC:");

   }

   // Re-verify object start arrays

   if (VerifyObjectStartArray &&

       VerifyAfterGC) {

     old_gen->verify_object_start_array();

   }

   if (ZapUnusedHeapArea) {

     old_gen->object_space()->check_mangled_unused_area_complete();

   }

   NOT_PRODUCT(ref_processor()->verify_no_references_recorded());

   collection_exit.update();

   heap->print_heap_after_gc();

   heap->trace_heap_after_gc(&_gc_tracer);

   if (PrintGCTaskTimeStamps) {

     gclog_or_tty->print_cr("VM-Thread " INT64_FORMAT " " INT64_FORMAT " "

                            INT64_FORMAT,

                            marking_start.ticks(), compaction_start.ticks(),

                            collection_exit.ticks());

     gc_task_manager()->print_task_time_stamps();

   }

   heap->post_full_gc_dump(&_gc_timer);

 #ifdef TRACESPINNING

   ParallelTaskTerminator::print_termination_counts();

 #endif

   _gc_timer.register_gc_end();

   _gc_tracer.report_dense_prefix(dense_prefix(old_space_id));

   _gc_tracer.report_gc_end(_gc_timer.gc_end(), _gc_timer.time_partitions());

   return true;

 }

 bool PSParallelCompact::absorb_live_data_from_eden(PSAdaptiveSizePolicy* size_policy,

                                              PSYoungGen* young_gen,

                                              PSOldGen* old_gen) {

   MutableSpace* const eden_space = young_gen->eden_space();

   assert(!eden_space->is_empty(), "eden must be non-empty");

   assert(young_gen->virtual_space()->alignment() ==

          old_gen->virtual_space()->alignment(), "alignments do not match");

   if (!(UseAdaptiveSizePolicy && UseAdaptiveGCBoundary)) {

     return false;

   }

   // Both generations must be completely committed.

   if (young_gen->virtual_space()->uncommitted_size() != 0) {

     return false;

   }

   if (old_gen->virtual_space()->uncommitted_size() != 0) {

     return false;

   }

   // Figure out how much to take from eden.  Include the average amount promoted

   // in the total; otherwise the next young gen GC will simply bail out to a

   // full GC.

   const size_t alignment = old_gen->virtual_space()->alignment();

   const size_t eden_used = eden_space->used_in_bytes();

   const size_t promoted = (size_t)size_policy->avg_promoted()->padded_average();

   const size_t absorb_size = align_size_up(eden_used + promoted, alignment);

   const size_t eden_capacity = eden_space->capacity_in_bytes();

   if (absorb_size >= eden_capacity) {

     return false; // Must leave some space in eden.

   }

   const size_t new_young_size = young_gen->capacity_in_bytes() - absorb_size;

   if (new_young_size < young_gen->min_gen_size()) {

     return false; // Respect young gen minimum size.

   }

   if (TraceAdaptiveGCBoundary && Verbose) {

     gclog_or_tty->print(" absorbing " SIZE_FORMAT "K:  "

                         "eden " SIZE_FORMAT "K->" SIZE_FORMAT "K "

                         "from " SIZE_FORMAT "K, to " SIZE_FORMAT "K "

                         "young_gen " SIZE_FORMAT "K->" SIZE_FORMAT "K ",

                         absorb_size / K,

                         eden_capacity / K, (eden_capacity - absorb_size) / K,

                         young_gen->from_space()->used_in_bytes() / K,

                         young_gen->to_space()->used_in_bytes() / K,

                         young_gen->capacity_in_bytes() / K, new_young_size / K);

   }

   // Fill the unused part of the old gen.

   MutableSpace* const old_space = old_gen->object_space();

   HeapWord* const unused_start = old_space->top();

   size_t const unused_words = pointer_delta(old_space->end(), unused_start);

   if (unused_words > 0) {

     if (unused_words < CollectedHeap::min_fill_size()) {

       return false;  // If the old gen cannot be filled, must give up.

     }

     CollectedHeap::fill_with_objects(unused_start, unused_words);

   }

   // Take the live data from eden and set both top and end in the old gen to

   // eden top.  (Need to set end because reset_after_change() mangles the region

   // from end to virtual_space->high() in debug builds).

   HeapWord* const new_top = eden_space->top();

   old_gen->virtual_space()->expand_into(young_gen->virtual_space(),

                                         absorb_size);

   young_gen->reset_after_change();

   old_space->set_top(new_top);

   old_space->set_end(new_top);

   old_gen->reset_after_change();

   // Update the object start array for the filler object and the data from eden.

   ObjectStartArray* const start_array = old_gen->start_array();

   for (HeapWord* p = unused_start; p < new_top; p += oop(p)->size()) {

     start_array->allocate_block(p);

   }

   // Could update the promoted average here, but it is not typically updated at

   // full GCs and the value to use is unclear.  Something like

   //

   // cur_promoted_avg + absorb_size / number_of_scavenges_since_last_full_gc.

   size_policy->set_bytes_absorbed_from_eden(absorb_size);

   return true;

 }

 GCTaskManager* const PSParallelCompact::gc_task_manager() {

   assert(ParallelScavengeHeap::gc_task_manager() != NULL,

     "shouldn't return NULL");

   return ParallelScavengeHeap::gc_task_manager();

 }

 void PSParallelCompact::marking_phase(ParCompactionManager* cm,

                                       bool maximum_heap_compaction,

                                       ParallelOldTracer *gc_tracer) {

   // Recursively traverse all live objects and mark them

   GCTraceTime tm("marking phase", print_phases(), true, &_gc_timer);

   ParallelScavengeHeap* heap = gc_heap();

   uint parallel_gc_threads = heap->gc_task_manager()->workers();

   uint active_gc_threads = heap->gc_task_manager()->active_workers();

   TaskQueueSetSuper* qset = ParCompactionManager::region_array();

   ParallelTaskTerminator terminator(active_gc_threads, qset);

   PSParallelCompact::MarkAndPushClosure mark_and_push_closure(cm);

   PSParallelCompact::FollowStackClosure follow_stack_closure(cm);

   // Need new claim bits before marking starts.

   ClassLoaderDataGraph::clear_claimed_marks();

   {

     GCTraceTime tm_m("par mark", print_phases(), true, &_gc_timer);

     ParallelScavengeHeap::ParStrongRootsScope psrs;

     GCTaskQueue* q = GCTaskQueue::create();

     q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::universe));

     q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jni_handles));

     // We scan the thread roots in parallel

     Threads::create_thread_roots_marking_tasks(q);

     q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::object_synchronizer));

     q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::flat_profiler));

     q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::management));

     q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::system_dictionary));

     q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::class_loader_data));

     q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::jvmti));

     q->enqueue(new MarkFromRootsTask(MarkFromRootsTask::code_cache));

     if (active_gc_threads > 1) {

       for (uint j = 0; j < active_gc_threads; j++) {

         q->enqueue(new StealMarkingTask(&terminator));

       }

     }

     gc_task_manager()->execute_and_wait(q);

   }

   // Process reference objects found during marking

   {

     GCTraceTime tm_r("reference processing", print_phases(), true, &_gc_timer);

     ReferenceProcessorStats stats;

     if (ref_processor()->processing_is_mt()) {

       RefProcTaskExecutor task_executor;

       stats = ref_processor()->process_discovered_references(

         is_alive_closure(), &mark_and_push_closure, &follow_stack_closure,

         &task_executor, &_gc_timer);

     } else {

       stats = ref_processor()->process_discovered_references(

         is_alive_closure(), &mark_and_push_closure, &follow_stack_closure, NULL,

         &_gc_timer);

     }

     gc_tracer->report_gc_reference_stats(stats);

   }

   GCTraceTime tm_c("class unloading", print_phases(), true, &_gc_timer);

   // This is the point where the entire marking should have completed.

   assert(cm->marking_stacks_empty(), "Marking should have completed");

   // Follow system dictionary roots and unload classes.

   bool purged_class = SystemDictionary::do_unloading(is_alive_closure());

   // Unload nmethods.

   CodeCache::do_unloading(is_alive_closure(), purged_class);

   // Prune dead klasses from subklass/sibling/implementor lists.

   Klass::clean_weak_klass_links(is_alive_closure());

   // Delete entries for dead interned strings.

   StringTable::unlink(is_alive_closure());

   // Clean up unreferenced symbols in symbol table.

   SymbolTable::unlink();

   _gc_tracer.report_object_count_after_gc(is_alive_closure());

 }

 void PSParallelCompact::follow_class_loader(ParCompactionManager* cm,

                                             ClassLoaderData* cld) {

   PSParallelCompact::MarkAndPushClosure mark_and_push_closure(cm);

   PSParallelCompact::FollowKlassClosure follow_klass_closure(&mark_and_push_closure);

   cld->oops_do(&mark_and_push_closure, &follow_klass_closure, true);

 }

 // This should be moved to the shared markSweep code!

 class PSAlwaysTrueClosure: public BoolObjectClosure {

 public:

   bool do_object_b(oop p) { return true; }

 };

 static PSAlwaysTrueClosure always_true;

 void PSParallelCompact::adjust_roots() {

   // Adjust the pointers to reflect the new locations

   GCTraceTime tm("adjust roots", print_phases(), true, &_gc_timer);

   // Need new claim bits when tracing through and adjusting pointers.

   ClassLoaderDataGraph::clear_claimed_marks();

   // General strong roots.

   Universe::oops_do(adjust_pointer_closure());

   JNIHandles::oops_do(adjust_pointer_closure());   // Global (strong) JNI handles

   CLDToOopClosure adjust_from_cld(adjust_pointer_closure());

   Threads::oops_do(adjust_pointer_closure(), &adjust_from_cld, NULL);

   ObjectSynchronizer::oops_do(adjust_pointer_closure());

   FlatProfiler::oops_do(adjust_pointer_closure());

   Management::oops_do(adjust_pointer_closure());

   JvmtiExport::oops_do(adjust_pointer_closure());

   // SO_AllClasses

   SystemDictionary::oops_do(adjust_pointer_closure());

   ClassLoaderDataGraph::oops_do(adjust_pointer_closure(), adjust_klass_closure(), true);

   // Now adjust pointers in remaining weak roots.  (All of which should

   // have been cleared if they pointed to non-surviving objects.)

   // Global (weak) JNI handles

   JNIHandles::weak_oops_do(&always_true, adjust_pointer_closure());

   CodeCache::oops_do(adjust_pointer_closure());

   StringTable::oops_do(adjust_pointer_closure());

   ref_processor()->weak_oops_do(adjust_pointer_closure());

   // Roots were visited so references into the young gen in roots

   // may have been scanned.  Process them also.

   // Should the reference processor have a span that excludes

   // young gen objects?

   PSScavenge::reference_processor()->weak_oops_do(adjust_pointer_closure());

 }

 void PSParallelCompact::enqueue_region_draining_tasks(GCTaskQueue* q,

                                                       uint parallel_gc_threads)

 {

   GCTraceTime tm("drain task setup", print_phases(), true, &_gc_timer);

   // Find the threads that are active

   unsigned int which = 0;

   const uint task_count = MAX2(parallel_gc_threads, 1U);

   for (uint j = 0; j < task_count; j++) {

     q->enqueue(new DrainStacksCompactionTask(j));

     ParCompactionManager::verify_region_list_empty(j);

     // Set the region stacks variables to "no" region stack values

     // so that they will be recognized and needing a region stack

     // in the stealing tasks if they do not get one by executing

     // a draining stack.

     ParCompactionManager* cm = ParCompactionManager::manager_array(j);

     cm->set_region_stack(NULL);

     cm->set_region_stack_index((uint)max_uintx);

   }

   ParCompactionManager::reset_recycled_stack_index();

   // Find all regions that are available (can be filled immediately) and

   // distribute them to the thread stacks.  The iteration is done in reverse

   // order (high to low) so the regions will be removed in ascending order.

   const ParallelCompactData& sd = PSParallelCompact::summary_data();

   size_t fillable_regions = 0;   // A count for diagnostic purposes.

   // A region index which corresponds to the tasks created above.

   // "which" must be 0 <= which < task_count

   which = 0;

   // id + 1 is used to test termination so unsigned  can

   // be used with an old_space_id == 0.

   for (unsigned int id = to_space_id; id + 1 > old_space_id; --id) {

     SpaceInfo* const space_info = _space_info + id;

     MutableSpace* const space = space_info->space();

     HeapWord* const new_top = space_info->new_top();

     const size_t beg_region = sd.addr_to_region_idx(space_info->dense_prefix());

     const size_t end_region =

       sd.addr_to_region_idx(sd.region_align_up(new_top));

     for (size_t cur = end_region - 1; cur + 1 > beg_region; --cur) {

       if (sd.region(cur)->claim_unsafe()) {

         ParCompactionManager::region_list_push(which, cur);

         if (TraceParallelOldGCCompactionPhase && Verbose) {

           const size_t count_mod_8 = fillable_regions & 7;

           if (count_mod_8 == 0) gclog_or_tty->print("fillable: ");

           gclog_or_tty->print(" " SIZE_FORMAT_W(7), cur);

           if (count_mod_8 == 7) gclog_or_tty->cr();

         }

         NOT_PRODUCT(++fillable_regions;)

         // Assign regions to tasks in round-robin fashion.

         if (++which == task_count) {

           assert(which <= parallel_gc_threads,

             "Inconsistent number of workers");

           which = 0;

         }

       }

     }

   }

   if (TraceParallelOldGCCompactionPhase) {

     if (Verbose && (fillable_regions & 7) != 0) gclog_or_tty->cr();

     gclog_or_tty->print_cr("%u initially fillable regions", fillable_regions);

   }

 }

 #define PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING 4

 void PSParallelCompact::enqueue_dense_prefix_tasks(GCTaskQueue* q,

                                                     uint parallel_gc_threads) {

   GCTraceTime tm("dense prefix task setup", print_phases(), true, &_gc_timer);

   ParallelCompactData& sd = PSParallelCompact::summary_data();

   // Iterate over all the spaces adding tasks for updating

   // regions in the dense prefix.  Assume that 1 gc thread

   // will work on opening the gaps and the remaining gc threads

   // will work on the dense prefix.

   unsigned int space_id;

   for (space_id = old_space_id; space_id < last_space_id; ++ space_id) {

     HeapWord* const dense_prefix_end = _space_info[space_id].dense_prefix();

     const MutableSpace* const space = _space_info[space_id].space();

     if (dense_prefix_end == space->bottom()) {

       // There is no dense prefix for this space.

       continue;

     }

     // The dense prefix is before this region.

     size_t region_index_end_dense_prefix =

         sd.addr_to_region_idx(dense_prefix_end);

     RegionData* const dense_prefix_cp =

       sd.region(region_index_end_dense_prefix);

     assert(dense_prefix_end == space->end() ||

            dense_prefix_cp->available() ||

            dense_prefix_cp->claimed(),

            "The region after the dense prefix should always be ready to fill");

     size_t region_index_start = sd.addr_to_region_idx(space->bottom());

     // Is there dense prefix work?

     size_t total_dense_prefix_regions =

       region_index_end_dense_prefix - region_index_start;

     // How many regions of the dense prefix should be given to

     // each thread?

     if (total_dense_prefix_regions > 0) {

       uint tasks_for_dense_prefix = 1;

       if (total_dense_prefix_regions <=

           (parallel_gc_threads * PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING)) {

         // Don't over partition.  This assumes that

         // PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING is a small integer value

         // so there are not many regions to process.

         tasks_for_dense_prefix = parallel_gc_threads;

       } else {

         // Over partition

         tasks_for_dense_prefix = parallel_gc_threads *

           PAR_OLD_DENSE_PREFIX_OVER_PARTITIONING;

       }

       size_t regions_per_thread = total_dense_prefix_regions /

         tasks_for_dense_prefix;

       // Give each thread at least 1 region.

       if (regions_per_thread == 0) {

         regions_per_thread = 1;

       }

       for (uint k = 0; k < tasks_for_dense_prefix; k++) {

         if (region_index_start >= region_index_end_dense_prefix) {

           break;

         }

         // region_index_end is not processed

         size_t region_index_end = MIN2(region_index_start + regions_per_thread,

                                        region_index_end_dense_prefix);

         q->enqueue(new UpdateDensePrefixTask(SpaceId(space_id),

                                              region_index_start,

                                              region_index_end));

         region_index_start = region_index_end;

       }

     }

     // This gets any part of the dense prefix that did not

     // fit evenly.

     if (region_index_start < region_index_end_dense_prefix) {

       q->enqueue(new UpdateDensePrefixTask(SpaceId(space_id),