--- a/src/share/vm/gc_implementation/concurrentMarkSweep/compactibleFreeListSpace.cpp	Tue Apr 07 11:20:51 2015 +0200
+++ b/src/share/vm/gc_implementation/concurrentMarkSweep/compactibleFreeListSpace.cpp	Tue Apr 07 14:58:49 2015 +0200
@@ -33,12 +33,14 @@
 #include "memory/allocation.inline.hpp"
 #include "memory/blockOffsetTable.inline.hpp"
 #include "memory/resourceArea.hpp"
+#include "memory/space.inline.hpp"
 #include "memory/universe.inline.hpp"
 #include "oops/oop.inline.hpp"
 #include "runtime/globals.hpp"
 #include "runtime/handles.inline.hpp"
 #include "runtime/init.hpp"
 #include "runtime/java.hpp"
+#include "runtime/orderAccess.inline.hpp"
 #include "runtime/vmThread.hpp"
 #include "utilities/copy.hpp"
 
@@ -793,53 +795,6 @@
   }
 }
 
-// Apply the given closure to each oop in the space \intersect memory region.
-void CompactibleFreeListSpace::oop_iterate(MemRegion mr, ExtendedOopClosure* cl) {
-  assert_lock_strong(freelistLock());
-  if (is_empty()) {
-    return;
-  }
-  MemRegion cur = MemRegion(bottom(), end());
-  mr = mr.intersection(cur);
-  if (mr.is_empty()) {
-    return;
-  }
-  if (mr.equals(cur)) {
-    oop_iterate(cl);
-    return;
-  }
-  assert(mr.end() <= end(), "just took an intersection above");
-  HeapWord* obj_addr = block_start(mr.start());
-  HeapWord* t = mr.end();
-
-  SpaceMemRegionOopsIterClosure smr_blk(cl, mr);
-  if (block_is_obj(obj_addr)) {
-    // Handle first object specially.
-    oop obj = oop(obj_addr);
-    obj_addr += adjustObjectSize(obj->oop_iterate(&smr_blk));
-  } else {
-    FreeChunk* fc = (FreeChunk*)obj_addr;
-    obj_addr += fc->size();
-  }
-  while (obj_addr < t) {
-    HeapWord* obj = obj_addr;
-    obj_addr += block_size(obj_addr);
-    // If "obj_addr" is not greater than top, then the
-    // entire object "obj" is within the region.
-    if (obj_addr <= t) {
-      if (block_is_obj(obj)) {
-        oop(obj)->oop_iterate(cl);
-      }
-    } else {
-      // "obj" extends beyond end of region
-      if (block_is_obj(obj)) {
-        oop(obj)->oop_iterate(&smr_blk);
-      }
-      break;
-    }
-  }
-}
-
 // NOTE: In the following methods, in order to safely be able to
 // apply the closure to an object, we need to be sure that the
 // object has been initialized. We are guaranteed that an object
@@ -898,42 +853,60 @@
                                                   UpwardsObjectClosure* cl) {
   assert_locked(freelistLock());
   NOT_PRODUCT(verify_objects_initialized());
-  Space::object_iterate_mem(mr, cl);
+  assert(!mr.is_empty(), "Should be non-empty");
+  // We use MemRegion(bottom(), end()) rather than used_region() below
+  // because the two are not necessarily equal for some kinds of
+  // spaces, in particular, certain kinds of free list spaces.
+  // We could use the more complicated but more precise:
+  // MemRegion(used_region().start(), round_to(used_region().end(), CardSize))
+  // but the slight imprecision seems acceptable in the assertion check.
+  assert(MemRegion(bottom(), end()).contains(mr),
+         "Should be within used space");
+  HeapWord* prev = cl->previous();   // max address from last time
+  if (prev >= mr.end()) { // nothing to do
+    return;
+  }
+  // This assert will not work when we go from cms space to perm
+  // space, and use same closure. Easy fix deferred for later. XXX YSR
+  // assert(prev == NULL || contains(prev), "Should be within space");
+
+  bool last_was_obj_array = false;
+  HeapWord *blk_start_addr, *region_start_addr;
+  if (prev > mr.start()) {
+    region_start_addr = prev;
+    blk_start_addr    = prev;
+    // The previous invocation may have pushed "prev" beyond the
+    // last allocated block yet there may be still be blocks
+    // in this region due to a particular coalescing policy.
+    // Relax the assertion so that the case where the unallocated
+    // block is maintained and "prev" is beyond the unallocated
+    // block does not cause the assertion to fire.
+    assert((BlockOffsetArrayUseUnallocatedBlock &&
+            (!is_in(prev))) ||
+           (blk_start_addr == block_start(region_start_addr)), "invariant");
+  } else {
+    region_start_addr = mr.start();
+    blk_start_addr    = block_start(region_start_addr);
+  }
+  HeapWord* region_end_addr = mr.end();
+  MemRegion derived_mr(region_start_addr, region_end_addr);
+  while (blk_start_addr < region_end_addr) {
+    const size_t size = block_size(blk_start_addr);
+    if (block_is_obj(blk_start_addr)) {
+      last_was_obj_array = cl->do_object_bm(oop(blk_start_addr), derived_mr);
+    } else {
+      last_was_obj_array = false;
+    }
+    blk_start_addr += size;
+  }
+  if (!last_was_obj_array) {
+    assert((bottom() <= blk_start_addr) && (blk_start_addr <= end()),
+           "Should be within (closed) used space");
+    assert(blk_start_addr > prev, "Invariant");
+    cl->set_previous(blk_start_addr); // min address for next time
+  }
 }
 
-// Callers of this iterator beware: The closure application should
-// be robust in the face of uninitialized objects and should (always)
-// return a correct size so that the next addr + size below gives us a
-// valid block boundary. [See for instance,
-// ScanMarkedObjectsAgainCarefullyClosure::do_object_careful()
-// in ConcurrentMarkSweepGeneration.cpp.]
-HeapWord*
-CompactibleFreeListSpace::object_iterate_careful(ObjectClosureCareful* cl) {
-  assert_lock_strong(freelistLock());
-  HeapWord *addr, *last;
-  size_t size;
-  for (addr = bottom(), last  = end();
-       addr < last; addr += size) {
-    FreeChunk* fc = (FreeChunk*)addr;
-    if (fc->is_free()) {
-      // Since we hold the free list lock, which protects direct
-      // allocation in this generation by mutators, a free object
-      // will remain free throughout this iteration code.
-      size = fc->size();
-    } else {
-      // Note that the object need not necessarily be initialized,
-      // because (for instance) the free list lock does NOT protect
-      // object initialization. The closure application below must
-      // therefore be correct in the face of uninitialized objects.
-      size = cl->do_object_careful(oop(addr));
-      if (size == 0) {
-        // An unparsable object found. Signal early termination.
-        return addr;
-      }
-    }
-  }
-  return NULL;
-}
 
 // Callers of this iterator beware: The closure application should
 // be robust in the face of uninitialized objects and should (always)
@@ -2668,7 +2641,7 @@
   // Get the #blocks we want to claim
   size_t n_blks = (size_t)_blocks_to_claim[word_sz].average();
   assert(n_blks > 0, "Error");
-  assert(ResizePLAB || n_blks == OldPLABSize, "Error");
+  assert(ResizeOldPLAB || n_blks == OldPLABSize, "Error");
   // In some cases, when the application has a phase change,
   // there may be a sudden and sharp shift in the object survival
   // profile, and updating the counts at the end of a scavenge
@@ -2760,10 +2733,12 @@
   }
 }
 
-void CompactibleFreeListSpace:: par_get_chunk_of_blocks(size_t word_sz, size_t n, AdaptiveFreeList<FreeChunk>* fl) {
-  assert(fl->count() == 0, "Precondition.");
-  assert(word_sz < CompactibleFreeListSpace::IndexSetSize,
-         "Precondition");
+// Used by par_get_chunk_of_blocks() for the chunks from the
+// indexed_free_lists.  Looks for a chunk with size that is a multiple
+// of "word_sz" and if found, splits it into "word_sz" chunks and add
+// to the free list "fl".  "n" is the maximum number of chunks to
+// be added to "fl".
+bool CompactibleFreeListSpace:: par_get_chunk_of_blocks_IFL(size_t word_sz, size_t n, AdaptiveFreeList<FreeChunk>* fl) {
 
   // We'll try all multiples of word_sz in the indexed set, starting with
   // word_sz itself and, if CMSSplitIndexedFreeListBlocks, try larger multiples,
@@ -2844,11 +2819,15 @@
                         Mutex::_no_safepoint_check_flag);
         ssize_t births = _indexedFreeList[word_sz].split_births() + num;
         _indexedFreeList[word_sz].set_split_births(births);
-        return;
+        return true;
       }
     }
+    return found;
   }
-  // Otherwise, we'll split a block from the dictionary.
+}
+
+FreeChunk* CompactibleFreeListSpace::get_n_way_chunk_to_split(size_t word_sz, size_t n) {
+
   FreeChunk* fc = NULL;
   FreeChunk* rem_fc = NULL;
   size_t rem;
@@ -2859,16 +2838,12 @@
       fc = dictionary()->get_chunk(MAX2(n * word_sz, _dictionary->min_size()),
                                   FreeBlockDictionary<FreeChunk>::atLeast);
       if (fc != NULL) {
-        _bt.allocated((HeapWord*)fc, fc->size(), true /* reducing */);  // update _unallocated_blk
-        dictionary()->dict_census_update(fc->size(),
-                                       true /*split*/,
-                                       false /*birth*/);
         break;
       } else {
         n--;
       }
     }
-    if (fc == NULL) return;
+    if (fc == NULL) return NULL;
     // Otherwise, split up that block.
     assert((ssize_t)n >= 1, "Control point invariant");
     assert(fc->is_free(), "Error: should be a free block");
@@ -2890,10 +2865,14 @@
     // dictionary and return, leaving "fl" empty.
     if (n == 0) {
       returnChunkToDictionary(fc);
-      assert(fl->count() == 0, "We never allocated any blocks");
-      return;
+      return NULL;
     }
 
+    _bt.allocated((HeapWord*)fc, fc->size(), true /* reducing */);  // update _unallocated_blk
+    dictionary()->dict_census_update(fc->size(),
+                                     true /*split*/,
+                                     false /*birth*/);
+
     // First return the remainder, if any.
     // Note that we hold the lock until we decide if we're going to give
     // back the remainder to the dictionary, since a concurrent allocation
@@ -2926,7 +2905,24 @@
     _indexedFreeList[rem].return_chunk_at_head(rem_fc);
     smallSplitBirth(rem);
   }
-  assert((ssize_t)n > 0 && fc != NULL, "Consistency");
+  assert(n * word_sz == fc->size(),
+    err_msg("Chunk size " SIZE_FORMAT " is not exactly splittable by "
+    SIZE_FORMAT " sized chunks of size " SIZE_FORMAT,
+    fc->size(), n, word_sz));
+  return fc;
+}
+
+void CompactibleFreeListSpace:: par_get_chunk_of_blocks_dictionary(size_t word_sz, size_t targetted_number_of_chunks, AdaptiveFreeList<FreeChunk>* fl) {
+
+  FreeChunk* fc = get_n_way_chunk_to_split(word_sz, targetted_number_of_chunks);
+
+  if (fc == NULL) {
+    return;
+  }
+
+  size_t n = fc->size() / word_sz;
+
+  assert((ssize_t)n > 0, "Consistency");
   // Now do the splitting up.
   // Must do this in reverse order, so that anybody attempting to
   // access the main chunk sees it as a single free block until we
@@ -2974,6 +2970,20 @@
   assert(fl->tail()->next() == NULL, "List invariant.");
 }
 
+void CompactibleFreeListSpace:: par_get_chunk_of_blocks(size_t word_sz, size_t n, AdaptiveFreeList<FreeChunk>* fl) {
+  assert(fl->count() == 0, "Precondition.");
+  assert(word_sz < CompactibleFreeListSpace::IndexSetSize,
+         "Precondition");
+
+  if (par_get_chunk_of_blocks_IFL(word_sz, n, fl)) {
+    // Got it
+    return;
+  }
+
+  // Otherwise, we'll split a block from the dictionary.
+  par_get_chunk_of_blocks_dictionary(word_sz, n, fl);
+}
+
 // Set up the space's par_seq_tasks structure for work claiming
 // for parallel rescan. See CMSParRemarkTask where this is currently used.
 // XXX Need to suitably abstract and generalize this and the next