--- a/src/share/vm/gc_implementation/parallelScavenge/parallelScavengeHeap.hpp	Mon May 27 12:58:42 2013 +0200
+++ b/src/share/vm/gc_implementation/parallelScavenge/parallelScavengeHeap.hpp	Thu May 30 13:04:51 2013 -0700
@@ -116,7 +116,7 @@
 
   // The alignment used for eden and survivors within the young gen
   // and for boundary between young gen and old gen.
-  size_t intra_heap_alignment() const { return 64 * K; }
+  size_t intra_heap_alignment() const { return 64 * K * HeapWordSize; }
 
   size_t capacity() const;
   size_t used() const;

--- a/src/share/vm/gc_implementation/parallelScavenge/psParallelCompact.cpp	Mon May 27 12:58:42 2013 +0200
+++ b/src/share/vm/gc_implementation/parallelScavenge/psParallelCompact.cpp	Thu May 30 13:04:51 2013 -0700
@@ -59,13 +59,25 @@
 #include <math.h>
 
 // All sizes are in HeapWords.
-const size_t ParallelCompactData::Log2RegionSize  = 9; // 512 words
+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::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;
@@ -359,6 +371,10 @@
   _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)
@@ -372,8 +388,7 @@
   assert((region_size & RegionSizeOffsetMask) == 0,
          "region size not a multiple of RegionSize");
 
-  bool result = initialize_region_data(region_size);
-
+  bool result = initialize_region_data(region_size) && initialize_block_data();
   return result;
 }
 
@@ -418,17 +433,36 @@
   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
@@ -707,49 +741,48 @@
 
 HeapWord* ParallelCompactData::calc_new_pointer(HeapWord* addr) {
   assert(addr != NULL, "Should detect NULL oop earlier");
-  assert(PSParallelCompact::gc_heap()->is_in(addr), "addr not in heap");
-#ifdef ASSERT
-  if (PSParallelCompact::mark_bitmap()->is_unmarked(addr)) {
-    gclog_or_tty->print_cr("calc_new_pointer:: addr " PTR_FORMAT, addr);
-  }
-#endif
-  assert(PSParallelCompact::mark_bitmap()->is_marked(addr), "obj not marked");
+  assert(PSParallelCompact::gc_heap()->is_in(addr), "not in heap");
+  assert(PSParallelCompact::mark_bitmap()->is_marked(addr), "not marked");
 
   // Region covering the object.
-  size_t region_index = addr_to_region_idx(addr);
-  const RegionData* const region_ptr = region(region_index);
-  HeapWord* const region_addr = region_align_down(addr);
-
-  assert(addr < region_addr + RegionSize, "Region does not cover object");
-  assert(addr_to_region_ptr(region_addr) == region_ptr, "sanity check");
-
+  RegionData* const region_ptr = addr_to_region_ptr(addr);
   HeapWord* result = region_ptr->destination();
 
-  // If all the data in the region is live, then the new location of the object
-  // can be calculated from the destination of the region plus the offset of the
-  // object in the region.
+  // 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 += pointer_delta(addr, region_addr);
-    DEBUG_ONLY(PSParallelCompact::check_new_location(addr, result);)
+    result += region_offset(addr);
     return result;
   }
 
-  // The new location of the object is
-  //    region destination +
-  //    size of the partial object extending onto the region +
-  //    sizes of the live objects in the Region that are to the left of addr
-  const size_t partial_obj_size = region_ptr->partial_obj_size();
-  HeapWord* const search_start = region_addr + partial_obj_size;
+  // 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();
-  size_t live_to_left = bitmap->live_words_in_range(search_start, oop(addr));
-
-  result += partial_obj_size + live_to_left;
-  DEBUG_ONLY(PSParallelCompact::check_new_location(addr, result);)
+  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
+#ifdef ASSERT
 void ParallelCompactData::verify_clear(const PSVirtualSpace* vspace)
 {
   const size_t* const beg = (const size_t*)vspace->committed_low_addr();
@@ -762,16 +795,10 @@
 void ParallelCompactData::verify_clear()
 {
   verify_clear(_region_vspace);
+  verify_clear(_block_vspace);
 }
 #endif  // #ifdef ASSERT
 
-#ifdef NOT_PRODUCT
-ParallelCompactData::RegionData* debug_region(size_t region_index) {
-  ParallelCompactData& sd = PSParallelCompact::summary_data();
-  return sd.region(region_index);
-}
-#endif
-
 elapsedTimer        PSParallelCompact::_accumulated_time;
 unsigned int        PSParallelCompact::_total_invocations = 0;
 unsigned int        PSParallelCompact::_maximum_compaction_gc_num = 0;
@@ -1961,11 +1988,6 @@
                                       maximum_heap_compaction);
 }
 
-bool ParallelCompactData::region_contains(size_t region_index, HeapWord* addr) {
-  size_t addr_region_index = addr_to_region_idx(addr);
-  return region_index == addr_region_index;
-}
-
 // This method contains no policy. You should probably
 // be calling invoke() instead.
 bool PSParallelCompact::invoke_no_policy(bool maximum_heap_compaction) {
@@ -2627,6 +2649,41 @@
   }
 }
 
+#ifdef ASSERT
+// Write a histogram of the number of times the block table was filled for a
+// region.
+void PSParallelCompact::write_block_fill_histogram(outputStream* const out)
+{
+  if (!TraceParallelOldGCCompactionPhase) return;
+
+  typedef ParallelCompactData::RegionData rd_t;
+  ParallelCompactData& sd = summary_data();
+
+  for (unsigned int id = old_space_id; id < last_space_id; ++id) {
+    MutableSpace* const spc = _space_info[id].space();
+    if (spc->bottom() != spc->top()) {
+      const rd_t* const beg = sd.addr_to_region_ptr(spc->bottom());
+      HeapWord* const top_aligned_up = sd.region_align_up(spc->top());
+      const rd_t* const end = sd.addr_to_region_ptr(top_aligned_up);
+
+      size_t histo[5] = { 0, 0, 0, 0, 0 };
+      const size_t histo_len = sizeof(histo) / sizeof(size_t);
+      const size_t region_cnt = pointer_delta(end, beg, sizeof(rd_t));
+
+      for (const rd_t* cur = beg; cur < end; ++cur) {
+        ++histo[MIN2(cur->blocks_filled_count(), histo_len - 1)];
+      }
+      out->print("%u %-4s" SIZE_FORMAT_W(5), id, space_names[id], region_cnt);
+      for (size_t i = 0; i < histo_len; ++i) {
+        out->print(" " SIZE_FORMAT_W(5) " %5.1f%%",
+                   histo[i], 100.0 * histo[i] / region_cnt);
+      }
+      out->cr();
+    }
+  }
+}
+#endif // #ifdef ASSERT
+
 void PSParallelCompact::compact() {
   // trace("5");
   TraceTime tm("compaction phase", print_phases(), true, gclog_or_tty);
@@ -2666,6 +2723,8 @@
       update_deferred_objects(cm, SpaceId(id));
     }
   }
+
+  DEBUG_ONLY(write_block_fill_histogram(gclog_or_tty));
 }
 
 #ifdef  ASSERT
@@ -3130,6 +3189,57 @@
   } while (true);
 }
 
+void PSParallelCompact::fill_blocks(size_t region_idx)
+{
+  // Fill in the block table elements for the specified region.  Each block
+  // table element holds the number of live words in the region that are to the
+  // left of the first object that starts in the block.  Thus only blocks in
+  // which an object starts need to be filled.
+  //
+  // The algorithm scans the section of the bitmap that corresponds to the
+  // region, keeping a running total of the live words.  When an object start is
+  // found, if it's the first to start in the block that contains it, the
+  // current total is written to the block table element.
+  const size_t Log2BlockSize = ParallelCompactData::Log2BlockSize;
+  const size_t Log2RegionSize = ParallelCompactData::Log2RegionSize;
+  const size_t RegionSize = ParallelCompactData::RegionSize;
+
+  ParallelCompactData& sd = summary_data();
+  const size_t partial_obj_size = sd.region(region_idx)->partial_obj_size();
+  if (partial_obj_size >= RegionSize) {
+    return; // No objects start in this region.
+  }
+
+  // Ensure the first loop iteration decides that the block has changed.
+  size_t cur_block = sd.block_count();
+
+  const ParMarkBitMap* const bitmap = mark_bitmap();
+
+  const size_t Log2BitsPerBlock = Log2BlockSize - LogMinObjAlignment;
+  assert((size_t)1 << Log2BitsPerBlock ==
+         bitmap->words_to_bits(ParallelCompactData::BlockSize), "sanity");
+
+  size_t beg_bit = bitmap->words_to_bits(region_idx << Log2RegionSize);
+  const size_t range_end = beg_bit + bitmap->words_to_bits(RegionSize);
+  size_t live_bits = bitmap->words_to_bits(partial_obj_size);
+  beg_bit = bitmap->find_obj_beg(beg_bit + live_bits, range_end);
+  while (beg_bit < range_end) {
+    const size_t new_block = beg_bit >> Log2BitsPerBlock;
+    if (new_block != cur_block) {
+      cur_block = new_block;
+      sd.block(cur_block)->set_offset(bitmap->bits_to_words(live_bits));
+    }
+
+    const size_t end_bit = bitmap->find_obj_end(beg_bit, range_end);
+    if (end_bit < range_end - 1) {
+      live_bits += end_bit - beg_bit + 1;
+      beg_bit = bitmap->find_obj_beg(end_bit + 1, range_end);
+    } else {
+      return;
+    }
+  }
+}
+
 void
 PSParallelCompact::move_and_update(ParCompactionManager* cm, SpaceId space_id) {
   const MutableSpace* sp = space(space_id);

--- a/src/share/vm/gc_implementation/parallelScavenge/psParallelCompact.hpp	Mon May 27 12:58:42 2013 +0200
+++ b/src/share/vm/gc_implementation/parallelScavenge/psParallelCompact.hpp	Thu May 30 13:04:51 2013 -0700
@@ -220,6 +220,17 @@
   // Mask for the bits in a pointer to get the address of the start of a region.
   static const size_t RegionAddrMask;
 
+  static const size_t Log2BlockSize;
+  static const size_t BlockSize;
+  static const size_t BlockSizeBytes;
+
+  static const size_t BlockSizeOffsetMask;
+  static const size_t BlockAddrOffsetMask;
+  static const size_t BlockAddrMask;
+
+  static const size_t BlocksPerRegion;
+  static const size_t Log2BlocksPerRegion;
+
   class RegionData
   {
   public:
@@ -272,6 +283,12 @@
     inline uint destination_count() const;
     inline uint destination_count_raw() const;
 
+    // Whether the block table for this region has been filled.
+    inline bool blocks_filled() const;
+
+    // Number of times the block table was filled.
+    DEBUG_ONLY(inline size_t blocks_filled_count() const;)
+
     // The location of the java heap data that corresponds to this region.
     inline HeapWord* data_location() const;
 
@@ -296,6 +313,7 @@
     void set_partial_obj_size(size_t words)    {
       _partial_obj_size = (region_sz_t) words;
     }
+    inline void set_blocks_filled();
 
     inline void set_destination_count(uint count);
     inline void set_live_obj_size(size_t words);
@@ -328,7 +346,11 @@
     HeapWord*            _partial_obj_addr;
     region_sz_t          _partial_obj_size;
     region_sz_t volatile _dc_and_los;
+    bool                 _blocks_filled;
+
 #ifdef ASSERT
+    size_t               _blocks_filled_count;   // Number of block table fills.
+
     // These enable optimizations that are only partially implemented.  Use
     // debug builds to prevent the code fragments from breaking.
     HeapWord*            _data_location;
@@ -337,11 +359,26 @@
 
 #ifdef ASSERT
    public:
-    uint            _pushed;   // 0 until region is pushed onto a worker's stack
+    uint                 _pushed;   // 0 until region is pushed onto a stack
    private:
 #endif
   };
 
+  // "Blocks" allow shorter sections of the bitmap to be searched.  Each Block
+  // holds an offset, which is the amount of live data in the Region to the left
+  // of the first live object that starts in the Block.
+  class BlockData
+  {
+  public:
+    typedef unsigned short int blk_ofs_t;
+
+    blk_ofs_t offset() const    { return _offset; }
+    void set_offset(size_t val) { _offset = (blk_ofs_t)val; }
+
+  private:
+    blk_ofs_t _offset;
+  };
+
 public:
   ParallelCompactData();
   bool initialize(MemRegion covered_region);
@@ -353,8 +390,9 @@
   inline RegionData* region(size_t region_idx) const;
   inline size_t     region(const RegionData* const region_ptr) const;
 
-  // Returns true if the given address is contained within the region
-  bool region_contains(size_t region_index, HeapWord* addr);
+  size_t block_count() const { return _block_count; }
+  inline BlockData* block(size_t block_idx) const;
+  inline size_t     block(const BlockData* block_ptr) const;
 
   void add_obj(HeapWord* addr, size_t len);
   void add_obj(oop p, size_t len) { add_obj((HeapWord*)p, len); }
@@ -394,11 +432,24 @@
   inline HeapWord*  region_align_up(HeapWord* addr) const;
   inline bool       is_region_aligned(HeapWord* addr) const;
 
+  // Analogous to region_offset() for blocks.
+  size_t     block_offset(const HeapWord* addr) const;
+  size_t     addr_to_block_idx(const HeapWord* addr) const;
+  size_t     addr_to_block_idx(const oop obj) const {
+    return addr_to_block_idx((HeapWord*) obj);
+  }
+  inline BlockData* addr_to_block_ptr(const HeapWord* addr) const;
+  inline HeapWord*  block_to_addr(size_t block) const;
+  inline size_t     region_to_block_idx(size_t region) const;
+
+  inline HeapWord*  block_align_down(HeapWord* addr) const;
+  inline HeapWord*  block_align_up(HeapWord* addr) const;
+  inline bool       is_block_aligned(HeapWord* addr) const;
+
   // Return the address one past the end of the partial object.
   HeapWord* partial_obj_end(size_t region_idx) const;
 
-  // Return the new location of the object p after the
-  // the compaction.
+  // Return the location of the object after compaction.
   HeapWord* calc_new_pointer(HeapWord* addr);
 
   HeapWord* calc_new_pointer(oop p) {
@@ -411,6 +462,7 @@
 #endif  // #ifdef ASSERT
 
 private:
+  bool initialize_block_data();
   bool initialize_region_data(size_t region_size);
   PSVirtualSpace* create_vspace(size_t count, size_t element_size);
 
@@ -424,6 +476,10 @@
   size_t          _reserved_byte_size;
   RegionData*     _region_data;
   size_t          _region_count;
+
+  PSVirtualSpace* _block_vspace;
+  BlockData*      _block_data;
+  size_t          _block_count;
 };
 
 inline uint
@@ -438,6 +494,28 @@
   return destination_count_raw() >> dc_shift;
 }
 
+inline bool
+ParallelCompactData::RegionData::blocks_filled() const
+{
+  return _blocks_filled;
+}
+
+#ifdef ASSERT
+inline size_t
+ParallelCompactData::RegionData::blocks_filled_count() const
+{
+  return _blocks_filled_count;
+}
+#endif // #ifdef ASSERT
+
+inline void
+ParallelCompactData::RegionData::set_blocks_filled()
+{
+  _blocks_filled = true;
+  // Debug builds count the number of times the table was filled.
+  DEBUG_ONLY(Atomic::inc_ptr(&_blocks_filled_count));
+}
+
 inline void
 ParallelCompactData::RegionData::set_destination_count(uint count)
 {
@@ -532,6 +610,12 @@
   return pointer_delta(region_ptr, _region_data, sizeof(RegionData));
 }
 
+inline ParallelCompactData::BlockData*
+ParallelCompactData::block(size_t n) const {
+  assert(n < block_count(), "bad arg");
+  return _block_data + n;
+}
+
 inline size_t
 ParallelCompactData::region_offset(const HeapWord* addr) const
 {
@@ -598,6 +682,63 @@
   return region_offset(addr) == 0;
 }
 
+inline size_t
+ParallelCompactData::block_offset(const HeapWord* addr) const
+{
+  assert(addr >= _region_start, "bad addr");
+  assert(addr <= _region_end, "bad addr");
+  return (size_t(addr) & BlockAddrOffsetMask) >> LogHeapWordSize;
+}
+
+inline size_t
+ParallelCompactData::addr_to_block_idx(const HeapWord* addr) const
+{
+  assert(addr >= _region_start, "bad addr");
+  assert(addr <= _region_end, "bad addr");
+  return pointer_delta(addr, _region_start) >> Log2BlockSize;
+}
+
+inline ParallelCompactData::BlockData*
+ParallelCompactData::addr_to_block_ptr(const HeapWord* addr) const
+{
+  return block(addr_to_block_idx(addr));
+}
+
+inline HeapWord*
+ParallelCompactData::block_to_addr(size_t block) const
+{
+  assert(block < _block_count, "block out of range");
+  return _region_start + (block << Log2BlockSize);
+}
+
+inline size_t
+ParallelCompactData::region_to_block_idx(size_t region) const
+{
+  return region << Log2BlocksPerRegion;
+}
+
+inline HeapWord*
+ParallelCompactData::block_align_down(HeapWord* addr) const
+{
+  assert(addr >= _region_start, "bad addr");
+  assert(addr < _region_end + RegionSize, "bad addr");
+  return (HeapWord*)(size_t(addr) & BlockAddrMask);
+}
+
+inline HeapWord*
+ParallelCompactData::block_align_up(HeapWord* addr) const
+{
+  assert(addr >= _region_start, "bad addr");
+  assert(addr <= _region_end, "bad addr");
+  return block_align_down(addr + BlockSizeOffsetMask);
+}
+
+inline bool
+ParallelCompactData::is_block_aligned(HeapWord* addr) const
+{
+  return block_offset(addr) == 0;
+}
+
 // Abstract closure for use with ParMarkBitMap::iterate(), which will invoke the
 // do_addr() method.
 //
@@ -775,6 +916,7 @@
   // Convenient access to type names.
   typedef ParMarkBitMap::idx_t idx_t;
   typedef ParallelCompactData::RegionData RegionData;
+  typedef ParallelCompactData::BlockData BlockData;
 
   typedef enum {
     old_space_id, eden_space_id,
@@ -962,6 +1104,8 @@
   // Adjust addresses in roots.  Does not adjust addresses in heap.
   static void adjust_roots();
 
+  DEBUG_ONLY(static void write_block_fill_histogram(outputStream* const out);)
+
   // Move objects to new locations.
   static void compact_perm(ParCompactionManager* cm);
   static void compact();
@@ -1128,6 +1272,9 @@
     fill_region(cm, region);
   }
 
+  // Fill in the block table for the specified region.
+  static void fill_blocks(size_t region_idx);
+
   // Update the deferred objects in the space.
   static void update_deferred_objects(ParCompactionManager* cm, SpaceId id);