4396719: Mark Sweep stack overflow on deeply nested Object arrays

Summary: Use an explicit stack for object arrays and process them in chunks.
Reviewed-by: iveresov, apetrusenko

26 files changed, 418 insertions, 181 deletions

diff --git a/src/share/vm/gc_implementation/g1/concurrentMark.hpp b/src/share/vm/gc_implementation/g1/concurrentMark.hpp
index f6902c863..f2b9d56ec 100644
--- a/src/share/vm/gc_implementation/g1/concurrentMark.hpp
+++ b/src/share/vm/gc_implementation/g1/concurrentMark.hpp
@@ -24,8 +24,8 @@
 
 class G1CollectedHeap;
 class CMTask;
-typedef GenericTaskQueue<oop> CMTaskQueue;
-typedef GenericTaskQueueSet<oop> CMTaskQueueSet;
+typedef GenericTaskQueue<oop>            CMTaskQueue;
+typedef GenericTaskQueueSet<CMTaskQueue> CMTaskQueueSet;
 
 // A generic CM bit map.  This is essentially a wrapper around the BitMap
 // class, with one bit per (1<<_shifter) HeapWords.
diff --git a/src/share/vm/gc_implementation/g1/g1CollectedHeap.hpp b/src/share/vm/gc_implementation/g1/g1CollectedHeap.hpp
index 692b3e7f3..c9c586750 100644
--- a/src/share/vm/gc_implementation/g1/g1CollectedHeap.hpp
+++ b/src/share/vm/gc_implementation/g1/g1CollectedHeap.hpp
@@ -56,8 +56,8 @@ class ConcurrentZFThread;
 #  define IF_G1_DETAILED_STATS(code)
 #endif
 
-typedef GenericTaskQueue<StarTask>    RefToScanQueue;
-typedef GenericTaskQueueSet<StarTask> RefToScanQueueSet;
+typedef GenericTaskQueue<StarTask>          RefToScanQueue;
+typedef GenericTaskQueueSet<RefToScanQueue> RefToScanQueueSet;
 
 typedef int RegionIdx_t;   // needs to hold [ 0..max_regions() )
 typedef int CardIdx_t;     // needs to hold [ 0..CardsPerRegion )
diff --git a/src/share/vm/gc_implementation/g1/g1MarkSweep.cpp b/src/share/vm/gc_implementation/g1/g1MarkSweep.cpp
index 53ea2f33a..a874bcac5 100644
--- a/src/share/vm/gc_implementation/g1/g1MarkSweep.cpp
+++ b/src/share/vm/gc_implementation/g1/g1MarkSweep.cpp
@@ -101,6 +101,8 @@ void G1MarkSweep::allocate_stacks() {
 
   GenMarkSweep::_marking_stack =
     new (ResourceObj::C_HEAP) GrowableArray<oop>(4000, true);
+  GenMarkSweep::_objarray_stack =
+    new (ResourceObj::C_HEAP) GrowableArray<ObjArrayTask>(50, true);
 
   int size = SystemDictionary::number_of_classes() * 2;
   GenMarkSweep::_revisit_klass_stack =
diff --git a/src/share/vm/gc_implementation/includeDB_gc_parallelScavenge b/src/share/vm/gc_implementation/includeDB_gc_parallelScavenge
index bdf672163..e787c2a5a 100644
--- a/src/share/vm/gc_implementation/includeDB_gc_parallelScavenge
+++ b/src/share/vm/gc_implementation/includeDB_gc_parallelScavenge
@@ -175,6 +175,7 @@ psAdaptiveSizePolicy.hpp                gcUtil.hpp
 psAdaptiveSizePolicy.hpp		adaptiveSizePolicy.hpp
 
 psCompactionManager.cpp                 gcTaskManager.hpp
+psCompactionManager.cpp                 objArrayKlass.inline.hpp
 psCompactionManager.cpp                 objectStartArray.hpp
 psCompactionManager.cpp                 oop.hpp
 psCompactionManager.cpp                 oop.inline.hpp
@@ -189,6 +190,9 @@ psCompactionManager.cpp                 systemDictionary.hpp
 psCompactionManager.hpp                 allocation.hpp
 psCompactionManager.hpp                 taskqueue.hpp
 
+psCompactionManager.inline.hpp		psCompactionManager.hpp
+psCompactionManager.inline.hpp		psParallelCompact.hpp
+
 psGCAdaptivePolicyCounters.hpp		gcAdaptivePolicyCounters.hpp
 psGCAdaptivePolicyCounters.hpp          gcPolicyCounters.hpp
 psGCAdaptivePolicyCounters.hpp          psAdaptiveSizePolicy.hpp
@@ -379,12 +383,12 @@ pcTasks.cpp                             fprofiler.hpp
 pcTasks.cpp                             jniHandles.hpp
 pcTasks.cpp                             jvmtiExport.hpp
 pcTasks.cpp                             management.hpp
+pcTasks.cpp                             objArrayKlass.inline.hpp
 pcTasks.cpp                             psParallelCompact.hpp
 pcTasks.cpp                             pcTasks.hpp
 pcTasks.cpp                             oop.inline.hpp
 pcTasks.cpp                             oop.pcgc.inline.hpp
 pcTasks.cpp                             systemDictionary.hpp
-pcTasks.cpp                             taskqueue.hpp
 pcTasks.cpp                             thread.hpp
 pcTasks.cpp                             universe.hpp
 pcTasks.cpp                             vmThread.hpp
diff --git a/src/share/vm/gc_implementation/parallelScavenge/pcTasks.cpp b/src/share/vm/gc_implementation/parallelScavenge/pcTasks.cpp
index 979bdaae7..a58fcbfb5 100644
--- a/src/share/vm/gc_implementation/parallelScavenge/pcTasks.cpp
+++ b/src/share/vm/gc_implementation/parallelScavenge/pcTasks.cpp
@@ -48,7 +48,7 @@ void ThreadRootsMarkingTask::do_it(GCTaskManager* manager, uint which) {
     _vm_thread->oops_do(&mark_and_push_closure, &mark_and_push_in_blobs);
 
   // Do the real work
-  cm->drain_marking_stacks(&mark_and_push_closure);
+  cm->follow_marking_stacks();
 }
 
 
@@ -118,7 +118,7 @@ void MarkFromRootsTask::do_it(GCTaskManager* manager, uint which) {
   }
 
   // Do the real work
-  cm->drain_marking_stacks(&mark_and_push_closure);
+  cm->follow_marking_stacks();
   // cm->deallocate_stacks();
 }
 
@@ -196,17 +196,19 @@ void StealMarkingTask::do_it(GCTaskManager* manager, uint which) {
   PSParallelCompact::MarkAndPushClosure mark_and_push_closure(cm);
 
   oop obj = NULL;
+  ObjArrayTask task;
   int random_seed = 17;
-  while(true) {
-    if (ParCompactionManager::steal(which, &random_seed, obj)) {
+  do {
+    while (ParCompactionManager::steal_objarray(which, &random_seed, task)) {
+      objArrayKlass* const k = (objArrayKlass*)task.obj()->blueprint();
+      k->oop_follow_contents(cm, task.obj(), task.index());
+      cm->follow_marking_stacks();
+    }
+    while (ParCompactionManager::steal(which, &random_seed, obj)) {
       obj->follow_contents(cm);
-      cm->drain_marking_stacks(&mark_and_push_closure);
-    } else {
-      if (terminator()->offer_termination()) {
-        break;
-      }
+      cm->follow_marking_stacks();
     }
-  }
+  } while (!terminator()->offer_termination());
 }
 
 //
diff --git a/src/share/vm/gc_implementation/parallelScavenge/psCompactionManager.cpp b/src/share/vm/gc_implementation/parallelScavenge/psCompactionManager.cpp
index 5d8284b43..a9723f3cd 100644
--- a/src/share/vm/gc_implementation/parallelScavenge/psCompactionManager.cpp
+++ b/src/share/vm/gc_implementation/parallelScavenge/psCompactionManager.cpp
@@ -28,6 +28,8 @@
 PSOldGen*            ParCompactionManager::_old_gen = NULL;
 ParCompactionManager**  ParCompactionManager::_manager_array = NULL;
 OopTaskQueueSet*     ParCompactionManager::_stack_array = NULL;
+ParCompactionManager::ObjArrayTaskQueueSet*
+  ParCompactionManager::_objarray_queues = NULL;
 ObjectStartArray*    ParCompactionManager::_start_array = NULL;
 ParMarkBitMap*       ParCompactionManager::_mark_bitmap = NULL;
 RegionTaskQueueSet*   ParCompactionManager::_region_array = NULL;
@@ -46,6 +48,11 @@ ParCompactionManager::ParCompactionManager() :
 
   // We want the overflow stack to be permanent
   _overflow_stack = new (ResourceObj::C_HEAP) GrowableArray<oop>(10, true);
+
+  _objarray_queue.initialize();
+  _objarray_overflow_stack =
+    new (ResourceObj::C_HEAP) ObjArrayOverflowStack(10, true);
+
 #ifdef USE_RegionTaskQueueWithOverflow
   region_stack()->initialize();
 #else
@@ -69,6 +76,7 @@ ParCompactionManager::ParCompactionManager() :
 
 ParCompactionManager::~ParCompactionManager() {
   delete _overflow_stack;
+  delete _objarray_overflow_stack;
   delete _revisit_klass_stack;
   delete _revisit_mdo_stack;
   // _manager_array and _stack_array are statics
@@ -86,18 +94,21 @@ void ParCompactionManager::initialize(ParMarkBitMap* mbm) {
 
   assert(_manager_array == NULL, "Attempt to initialize twice");
   _manager_array = NEW_C_HEAP_ARRAY(ParCompactionManager*, parallel_gc_threads+1 );
-  guarantee(_manager_array != NULL, "Could not initialize promotion manager");
+  guarantee(_manager_array != NULL, "Could not allocate manager_array");
 
   _stack_array = new OopTaskQueueSet(parallel_gc_threads);
-  guarantee(_stack_array != NULL, "Count not initialize promotion manager");
+  guarantee(_stack_array != NULL, "Could not allocate stack_array");
+  _objarray_queues = new ObjArrayTaskQueueSet(parallel_gc_threads);
+  guarantee(_objarray_queues != NULL, "Could not allocate objarray_queues");
   _region_array = new RegionTaskQueueSet(parallel_gc_threads);
-  guarantee(_region_array != NULL, "Count not initialize promotion manager");
+  guarantee(_region_array != NULL, "Could not allocate region_array");
 
   // Create and register the ParCompactionManager(s) for the worker threads.
   for(uint i=0; i<parallel_gc_threads; i++) {
     _manager_array[i] = new ParCompactionManager();
     guarantee(_manager_array[i] != NULL, "Could not create ParCompactionManager");
     stack_array()->register_queue(i, _manager_array[i]->marking_stack());
+    _objarray_queues->register_queue(i, &_manager_array[i]->_objarray_queue);
 #ifdef USE_RegionTaskQueueWithOverflow
     region_array()->register_queue(i, _manager_array[i]->region_stack()->task_queue());
 #else
@@ -203,36 +214,30 @@ void ParCompactionManager::reset() {
   }
 }
 
-void ParCompactionManager::drain_marking_stacks(OopClosure* blk) {
-#ifdef ASSERT
-  ParallelScavengeHeap* heap = (ParallelScavengeHeap*)Universe::heap();
-  assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity");
-  MutableSpace* to_space = heap->young_gen()->to_space();
-  MutableSpace* old_space = heap->old_gen()->object_space();
-  MutableSpace* perm_space = heap->perm_gen()->object_space();
-#endif /* ASSERT */
-
-
+void ParCompactionManager::follow_marking_stacks() {
   do {
-
-    // Drain overflow stack first, so other threads can steal from
-    // claimed stack while we work.
-    while(!overflow_stack()->is_empty()) {
-      oop obj = overflow_stack()->pop();
-      obj->follow_contents(this);
+    // Drain the overflow stack first, to allow stealing from the marking stack.
+    while (!overflow_stack()->is_empty()) {
+      overflow_stack()->pop()->follow_contents(this);
     }
-
     oop obj;
-    // obj is a reference!!!
     while (marking_stack()->pop_local(obj)) {
-      // It would be nice to assert about the type of objects we might
-      // pop, but they can come from anywhere, unfortunately.
       obj->follow_contents(this);
     }
-  } while((marking_stack()->size() != 0) || (overflow_stack()->length() != 0));
 
-  assert(marking_stack()->size() == 0, "Sanity");
-  assert(overflow_stack()->length() == 0, "Sanity");
+    ObjArrayTask task;
+    while (!_objarray_overflow_stack->is_empty()) {
+      task = _objarray_overflow_stack->pop();
+      objArrayKlass* const k = (objArrayKlass*)task.obj()->blueprint();
+      k->oop_follow_contents(this, task.obj(), task.index());
+    }
+    while (_objarray_queue.pop_local(task)) {
+      objArrayKlass* const k = (objArrayKlass*)task.obj()->blueprint();
+      k->oop_follow_contents(this, task.obj(), task.index());
+    }
+  } while (!marking_stacks_empty());
+
+  assert(marking_stacks_empty(), "Sanity");
 }
 
 void ParCompactionManager::drain_region_overflow_stack() {
diff --git a/src/share/vm/gc_implementation/parallelScavenge/psCompactionManager.hpp b/src/share/vm/gc_implementation/parallelScavenge/psCompactionManager.hpp
index 596c427a1..5a5eab370 100644
--- a/src/share/vm/gc_implementation/parallelScavenge/psCompactionManager.hpp
+++ b/src/share/vm/gc_implementation/parallelScavenge/psCompactionManager.hpp
@@ -22,18 +22,6 @@
  *
  */
 
-//
-// psPromotionManager is used by a single thread to manage object survival
-// during a scavenge. The promotion manager contains thread local data only.
-//
-// NOTE! Be carefull when allocating the stacks on cheap. If you are going
-// to use a promotion manager in more than one thread, the stacks MUST be
-// on cheap. This can lead to memory leaks, though, as they are not auto
-// deallocated.
-//
-// FIX ME FIX ME Add a destructor, and don't rely on the user to drain/flush/deallocate!
-//
-
 // Move to some global location
 #define HAS_BEEN_MOVED 0x1501d01d
 // End move to some global location
@@ -46,8 +34,6 @@ class ObjectStartArray;
 class ParallelCompactData;
 class ParMarkBitMap;
 
-// Move to it's own file if this works out.
-
 class ParCompactionManager : public CHeapObj {
   friend class ParallelTaskTerminator;
   friend class ParMarkBitMap;
@@ -72,14 +58,27 @@ class ParCompactionManager : public CHeapObj {
 // ------------------------  End don't putback if not needed
 
  private:
+  // 32-bit:  4K * 8 = 32KiB; 64-bit:  8K * 16 = 128KiB
+  #define OBJARRAY_QUEUE_SIZE (1 << NOT_LP64(12) LP64_ONLY(13))
+  typedef GenericTaskQueue<ObjArrayTask, OBJARRAY_QUEUE_SIZE> ObjArrayTaskQueue;
+  typedef GenericTaskQueueSet<ObjArrayTaskQueue> ObjArrayTaskQueueSet;
+  #undef OBJARRAY_QUEUE_SIZE
+
   static ParCompactionManager** _manager_array;
   static OopTaskQueueSet*       _stack_array;
+  static ObjArrayTaskQueueSet*  _objarray_queues;
   static ObjectStartArray*      _start_array;
   static RegionTaskQueueSet*    _region_array;
   static PSOldGen*              _old_gen;
 
+private:
   OopTaskQueue                  _marking_stack;
   GrowableArray<oop>*           _overflow_stack;
+
+  typedef GrowableArray<ObjArrayTask> ObjArrayOverflowStack;
+  ObjArrayTaskQueue             _objarray_queue;
+  ObjArrayOverflowStack*        _objarray_overflow_stack;
+
   // Is there a way to reuse the _marking_stack for the
   // saving empty regions?  For now just create a different
   // type of TaskQueue.
@@ -128,8 +127,8 @@ class ParCompactionManager : public CHeapObj {
   // Pushes onto the region stack.  If the region stack is full,
   // pushes onto the region overflow stack.
   void region_stack_push(size_t region_index);
- public:
 
+public:
   Action action() { return _action; }
   void set_action(Action v) { _action = v; }
 
@@ -163,6 +162,8 @@ class ParCompactionManager : public CHeapObj {
   // Get a oop for scanning.  If returns null, no oop were found.
   oop retrieve_for_scanning();
 
+  inline void push_objarray(oop obj, size_t index);
+
   // Save region for later processing.  Must not fail.
   void save_for_processing(size_t region_index);
   // Get a region for processing.  If returns null, no region were found.
@@ -175,12 +176,17 @@ class ParCompactionManager : public CHeapObj {
     return stack_array()->steal(queue_num, seed, t);
   }
 
+  static bool steal_objarray(int queue_num, int* seed, ObjArrayTask& t) {
+    return _objarray_queues->steal(queue_num, seed, t);
+  }
+
   static bool steal(int queue_num, int* seed, RegionTask& t) {
     return region_array()->steal(queue_num, seed, t);
   }
 
-  // Process tasks remaining on any stack
-  void drain_marking_stacks(OopClosure *blk);
+  // Process tasks remaining on any marking stack
+  void follow_marking_stacks();
+  inline bool marking_stacks_empty() const;
 
   // Process tasks remaining on any stack
   void drain_region_stacks();
@@ -200,3 +206,8 @@ inline ParCompactionManager* ParCompactionManager::manager_array(int index) {
     "out of range manager_array access");
   return _manager_array[index];
 }
+
+bool ParCompactionManager::marking_stacks_empty() const {
+  return _marking_stack.size() == 0 && _overflow_stack->is_empty() &&
+    _objarray_queue.size() == 0 && _objarray_overflow_stack->is_empty();
+}
diff --git a/src/share/vm/gc_implementation/parallelScavenge/psCompactionManager.inline.hpp b/src/share/vm/gc_implementation/parallelScavenge/psCompactionManager.inline.hpp
new file mode 100644
index 000000000..7a78dbcda
--- /dev/null
+++ b/src/share/vm/gc_implementation/parallelScavenge/psCompactionManager.inline.hpp
@@ -0,0 +1,32 @@
+/*
+ * Copyright 2010 Sun Microsystems, Inc.  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 Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
+ * CA 95054 USA or visit www.sun.com if you need additional information or
+ * have any questions.
+ *
+ */
+
+void ParCompactionManager::push_objarray(oop obj, size_t index)
+{
+  ObjArrayTask task(obj, index);
+  assert(task.is_valid(), "bad ObjArrayTask");
+  if (!_objarray_queue.push(task)) {
+    _objarray_overflow_stack->push(task);
+  }
+}
diff --git a/src/share/vm/gc_implementation/parallelScavenge/psMarkSweep.cpp b/src/share/vm/gc_implementation/parallelScavenge/psMarkSweep.cpp
index 1ea38984e..fcf6eb6ed 100644
--- a/src/share/vm/gc_implementation/parallelScavenge/psMarkSweep.cpp
+++ b/src/share/vm/gc_implementation/parallelScavenge/psMarkSweep.cpp
@@ -479,6 +479,7 @@ void PSMarkSweep::allocate_stacks() {
   _preserved_oop_stack = NULL;
 
   _marking_stack = new (ResourceObj::C_HEAP) GrowableArray<oop>(4000, true);
+  _objarray_stack = new (ResourceObj::C_HEAP) GrowableArray<ObjArrayTask>(50, true);
 
   int size = SystemDictionary::number_of_classes() * 2;
   _revisit_klass_stack = new (ResourceObj::C_HEAP) GrowableArray<Klass*>(size, true);
@@ -497,6 +498,7 @@ void PSMarkSweep::deallocate_stacks() {
   }
 
   delete _marking_stack;
+  delete _objarray_stack;
   delete _revisit_klass_stack;
   delete _revisit_mdo_stack;
 }
diff --git a/src/share/vm/gc_implementation/parallelScavenge/psParallelCompact.cpp b/src/share/vm/gc_implementation/parallelScavenge/psParallelCompact.cpp
index 37e8159a0..e7a973deb 100644
--- a/src/share/vm/gc_implementation/parallelScavenge/psParallelCompact.cpp
+++ b/src/share/vm/gc_implementation/parallelScavenge/psParallelCompact.cpp
@@ -785,7 +785,7 @@ PSParallelCompact::AdjustPointerClosure PSParallelCompact::_adjust_pointer_closu
 void PSParallelCompact::AdjustPointerClosure::do_oop(oop* p)       { adjust_pointer(p, _is_root); }
 void PSParallelCompact::AdjustPointerClosure::do_oop(narrowOop* p) { adjust_pointer(p, _is_root); }
 
-void PSParallelCompact::FollowStackClosure::do_void() { follow_stack(_compaction_manager); }
+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); }
@@ -2376,7 +2376,7 @@ void PSParallelCompact::marking_phase(ParCompactionManager* cm,
   // Follow code cache roots.
   CodeCache::do_unloading(is_alive_closure(), &mark_and_push_closure,
                           purged_class);
-  follow_stack(cm); // Flush marking stack.
+  cm->follow_marking_stacks(); // Flush marking stack.
 
   // Update subklass/sibling/implementor links of live klasses
   // revisit_klass_stack is used in follow_weak_klass_links().
@@ -2389,8 +2389,7 @@ void PSParallelCompact::marking_phase(ParCompactionManager* cm,
   SymbolTable::unlink(is_alive_closure());
   StringTable::unlink(is_alive_closure());
 
-  assert(cm->marking_stack()->size() == 0, "stack should be empty by now");
-  assert(cm->overflow_stack()->is_empty(), "stack should be empty by now");
+  assert(cm->marking_stacks_empty(), "marking stacks should be empty");
 }
 
 // This should be moved to the shared markSweep code!
@@ -2709,22 +2708,6 @@ void PSParallelCompact::compact_serial(ParCompactionManager* cm) {
   young_gen->move_and_update(cm);
 }
 
-
-void PSParallelCompact::follow_stack(ParCompactionManager* cm) {
-  while(!cm->overflow_stack()->is_empty()) {
-    oop obj = cm->overflow_stack()->pop();
-    obj->follow_contents(cm);
-  }
-
-  oop obj;
-  // obj is a reference!!!
-  while (cm->marking_stack()->pop_local(obj)) {
-    // It would be nice to assert about the type of objects we might
-    // pop, but they can come from anywhere, unfortunately.
-    obj->follow_contents(cm);
-  }
-}
-
 void
 PSParallelCompact::follow_weak_klass_links() {
   // All klasses on the revisit stack are marked at this point.
@@ -2745,7 +2728,7 @@ PSParallelCompact::follow_weak_klass_links() {
         &keep_alive_closure);
     }
     // revisit_klass_stack is cleared in reset()
-    follow_stack(cm);
+    cm->follow_marking_stacks();
   }
 }
 
@@ -2776,7 +2759,7 @@ void PSParallelCompact::follow_mdo_weak_refs() {
       rms->at(j)->follow_weak_refs(is_alive_closure());
     }
     // revisit_mdo_stack is cleared in reset()
-    follow_stack(cm);
+    cm->follow_marking_stacks();
   }
 }
 
diff --git a/src/share/vm/gc_implementation/parallelScavenge/psParallelCompact.hpp b/src/share/vm/gc_implementation/parallelScavenge/psParallelCompact.hpp
index d8c361f33..a445fc29e 100644
--- a/src/share/vm/gc_implementation/parallelScavenge/psParallelCompact.hpp
+++ b/src/share/vm/gc_implementation/parallelScavenge/psParallelCompact.hpp
@@ -901,7 +901,6 @@ class PSParallelCompact : AllStatic {
   // Mark live objects
   static void marking_phase(ParCompactionManager* cm,
                             bool maximum_heap_compaction);
-  static void follow_stack(ParCompactionManager* cm);
   static void follow_weak_klass_links();
   static void follow_mdo_weak_refs();
 
@@ -1276,7 +1275,7 @@ inline void PSParallelCompact::follow_root(ParCompactionManager* cm, T* p) {
       }
     }
   }
-  follow_stack(cm);
+  cm->follow_marking_stacks();
 }
 
 template <class T>
diff --git a/src/share/vm/gc_implementation/shared/markSweep.cpp b/src/share/vm/gc_implementation/shared/markSweep.cpp
index 3f1468358..776730da7 100644
--- a/src/share/vm/gc_implementation/shared/markSweep.cpp
+++ b/src/share/vm/gc_implementation/shared/markSweep.cpp
@@ -25,8 +25,9 @@
 #include "incls/_precompiled.incl"
 #include "incls/_markSweep.cpp.incl"
 
-GrowableArray<oop>*     MarkSweep::_marking_stack       = NULL;
-GrowableArray<Klass*>*  MarkSweep::_revisit_klass_stack = NULL;
+GrowableArray<oop>*          MarkSweep::_marking_stack = NULL;
+GrowableArray<ObjArrayTask>* MarkSweep::_objarray_stack = NULL;
+GrowableArray<Klass*>*       MarkSweep::_revisit_klass_stack = NULL;
 GrowableArray<DataLayout*>*  MarkSweep::_revisit_mdo_stack = NULL;
 
 GrowableArray<oop>*     MarkSweep::_preserved_oop_stack = NULL;
@@ -104,11 +105,18 @@ void MarkSweep::MarkAndPushClosure::do_oop(oop* p)       { mark_and_push(p); }
 void MarkSweep::MarkAndPushClosure::do_oop(narrowOop* p) { mark_and_push(p); }
 
 void MarkSweep::follow_stack() {
-  while (!_marking_stack->is_empty()) {
-    oop obj = _marking_stack->pop();
-    assert (obj->is_gc_marked(), "p must be marked");
-    obj->follow_contents();
-  }
+  do {
+    while (!_marking_stack->is_empty()) {
+      oop obj = _marking_stack->pop();
+      assert (obj->is_gc_marked(), "p must be marked");
+      obj->follow_contents();
+    }
+    while (!_objarray_stack->is_empty()) {
+      ObjArrayTask task = _objarray_stack->pop();
+      objArrayKlass* const k = (objArrayKlass*)task.obj()->blueprint();
+      k->oop_follow_contents(task.obj(), task.index());
+    }
+  } while (!_marking_stack->is_empty() || !_objarray_stack->is_empty());
 }
 
 MarkSweep::FollowStackClosure MarkSweep::follow_stack_closure;
diff --git a/src/share/vm/gc_implementation/shared/markSweep.hpp b/src/share/vm/gc_implementation/shared/markSweep.hpp
index 8b5825edc..d92c4a181 100644
--- a/src/share/vm/gc_implementation/shared/markSweep.hpp
+++ b/src/share/vm/gc_implementation/shared/markSweep.hpp
@@ -110,8 +110,9 @@ class MarkSweep : AllStatic {
   // Vars
   //
  protected:
-  // Traversal stack used during phase1
+  // Traversal stacks used during phase1
   static GrowableArray<oop>*             _marking_stack;
+  static GrowableArray<ObjArrayTask>*    _objarray_stack;
   // Stack for live klasses to revisit at end of marking phase
   static GrowableArray<Klass*>*          _revisit_klass_stack;
   // Set (stack) of MDO's to revisit at end of marking phase
@@ -188,6 +189,7 @@ class MarkSweep : AllStatic {
   template <class T> static inline void mark_and_follow(T* p);
   // Check mark and maybe push on marking stack
   template <class T> static inline void mark_and_push(T* p);
+  static inline void push_objarray(oop obj, size_t index);
 
   static void follow_stack();   // Empty marking stack.
 
diff --git a/src/share/vm/gc_implementation/shared/markSweep.inline.hpp b/src/share/vm/gc_implementation/shared/markSweep.inline.hpp
index 7596433cc..1caed8f35 100644
--- a/src/share/vm/gc_implementation/shared/markSweep.inline.hpp
+++ b/src/share/vm/gc_implementation/shared/markSweep.inline.hpp
@@ -77,6 +77,12 @@ template <class T> inline void MarkSweep::mark_and_push(T* p) {
   }
 }
 
+void MarkSweep::push_objarray(oop obj, size_t index) {
+  ObjArrayTask task(obj, index);
+  assert(task.is_valid(), "bad ObjArrayTask");
+  _objarray_stack->push(task);
+}
+
 template <class T> inline void MarkSweep::adjust_pointer(T* p, bool isroot) {
   T heap_oop = oopDesc::load_heap_oop(p);
   if (!oopDesc::is_null(heap_oop)) {
diff --git a/src/share/vm/includeDB_core b/src/share/vm/includeDB_core
index 74b5f71a5..a9451f135 100644
--- a/src/share/vm/includeDB_core
+++ b/src/share/vm/includeDB_core
@@ -2724,8 +2724,10 @@ markOop.inline.hpp                      markOop.hpp
 
 markSweep.cpp                           compileBroker.hpp
 markSweep.cpp                           methodDataOop.hpp
+markSweep.cpp				objArrayKlass.inline.hpp
 
 markSweep.hpp                           collectedHeap.hpp
+markSweep.hpp				taskqueue.hpp
 
 memRegion.cpp                           globals.hpp
 memRegion.cpp                           memRegion.hpp
@@ -3054,8 +3056,10 @@ objArrayKlass.cpp                       copy.hpp
 objArrayKlass.cpp                       genOopClosures.inline.hpp
 objArrayKlass.cpp                       handles.inline.hpp
 objArrayKlass.cpp                       instanceKlass.hpp
+objArrayKlass.cpp                       markSweep.inline.hpp
 objArrayKlass.cpp                       mutexLocker.hpp
 objArrayKlass.cpp                       objArrayKlass.hpp
+objArrayKlass.cpp                       objArrayKlass.inline.hpp
 objArrayKlass.cpp                       objArrayKlassKlass.hpp
 objArrayKlass.cpp                       objArrayOop.hpp
 objArrayKlass.cpp                       oop.inline.hpp
@@ -3066,11 +3070,12 @@ objArrayKlass.cpp                       systemDictionary.hpp
 objArrayKlass.cpp                       universe.inline.hpp
 objArrayKlass.cpp                       vmSymbols.hpp
 
-
 objArrayKlass.hpp                       arrayKlass.hpp
 objArrayKlass.hpp                       instanceKlass.hpp
 objArrayKlass.hpp                       specialized_oop_closures.hpp
 
+objArrayKlass.inline.hpp		objArrayKlass.hpp
+
 objArrayKlassKlass.cpp                  collectedHeap.inline.hpp
 objArrayKlassKlass.cpp                  instanceKlass.hpp
 objArrayKlassKlass.cpp                  javaClasses.hpp
@@ -4096,6 +4101,7 @@ task.cpp                                timer.hpp
 task.hpp                                top.hpp
 
 taskqueue.cpp                           debug.hpp
+taskqueue.cpp				oop.inline.hpp
 taskqueue.cpp                           os.hpp
 taskqueue.cpp                           taskqueue.hpp
 taskqueue.cpp                           thread_<os_family>.inline.hpp
diff --git a/src/share/vm/includeDB_gc_parallel b/src/share/vm/includeDB_gc_parallel
index 2d1c45a0c..54301b1d1 100644
--- a/src/share/vm/includeDB_gc_parallel
+++ b/src/share/vm/includeDB_gc_parallel
@@ -115,10 +115,14 @@ objArrayKlass.cpp                       heapRegionSeq.inline.hpp
 objArrayKlass.cpp                       g1CollectedHeap.inline.hpp
 objArrayKlass.cpp                       g1OopClosures.inline.hpp
 objArrayKlass.cpp                       oop.pcgc.inline.hpp
+objArrayKlass.cpp                       psCompactionManager.hpp
 objArrayKlass.cpp                       psPromotionManager.inline.hpp
 objArrayKlass.cpp                       psScavenge.inline.hpp
 objArrayKlass.cpp                       parOopClosures.inline.hpp
 
+objArrayKlass.inline.hpp		psCompactionManager.inline.hpp
+objArrayKlass.inline.hpp		psParallelCompact.hpp
+
 oop.pcgc.inline.hpp                     parNewGeneration.hpp
 oop.pcgc.inline.hpp                     parallelScavengeHeap.hpp
 oop.pcgc.inline.hpp                     psCompactionManager.hpp
diff --git a/src/share/vm/memory/genMarkSweep.cpp b/src/share/vm/memory/genMarkSweep.cpp
index 8f3223778..290141c4f 100644
--- a/src/share/vm/memory/genMarkSweep.cpp
+++ b/src/share/vm/memory/genMarkSweep.cpp
@@ -159,6 +159,7 @@ void GenMarkSweep::allocate_stacks() {
   _preserved_oop_stack = NULL;
 
   _marking_stack       = new (ResourceObj::C_HEAP) GrowableArray<oop>(4000, true);
+  _objarray_stack      = new (ResourceObj::C_HEAP) GrowableArray<ObjArrayTask>(50, true);
 
   int size = SystemDictionary::number_of_classes() * 2;
   _revisit_klass_stack = new (ResourceObj::C_HEAP) GrowableArray<Klass*>(size, true);
@@ -194,7 +195,6 @@ void GenMarkSweep::allocate_stacks() {
 
 
 void GenMarkSweep::deallocate_stacks() {
-
   if (!UseG1GC) {
     GenCollectedHeap* gch = GenCollectedHeap::heap();
     gch->release_scratch();
@@ -208,6 +208,7 @@ void GenMarkSweep::deallocate_stacks() {
   }
 
   delete _marking_stack;
+  delete _objarray_stack;
   delete _revisit_klass_stack;
   delete _revisit_mdo_stack;
 
diff --git a/src/share/vm/memory/genOopClosures.hpp b/src/share/vm/memory/genOopClosures.hpp
index 8f8fe5caa..c614811b2 100644
--- a/src/share/vm/memory/genOopClosures.hpp
+++ b/src/share/vm/memory/genOopClosures.hpp
@@ -28,10 +28,10 @@ class CardTableRS;
 class CardTableModRefBS;
 class DefNewGeneration;
 
-template<class E> class GenericTaskQueue;
-typedef GenericTaskQueue<oop> OopTaskQueue;
-template<class E> class GenericTaskQueueSet;
-typedef GenericTaskQueueSet<oop> OopTaskQueueSet;
+template<class E, unsigned int N> class GenericTaskQueue;
+typedef GenericTaskQueue<oop, TASKQUEUE_SIZE> OopTaskQueue;
+template<class T> class GenericTaskQueueSet;
+typedef GenericTaskQueueSet<OopTaskQueue> OopTaskQueueSet;
 
 // Closure for iterating roots from a particular generation
 // Note: all classes deriving from this MUST call this do_barrier
diff --git a/src/share/vm/oops/objArrayKlass.cpp b/src/share/vm/oops/objArrayKlass.cpp
index cb6884a9c..4224f1706 100644
--- a/src/share/vm/oops/objArrayKlass.cpp
+++ b/src/share/vm/oops/objArrayKlass.cpp
@@ -314,24 +314,24 @@ void objArrayKlass::initialize(TRAPS) {
 
 void objArrayKlass::oop_follow_contents(oop obj) {
   assert (obj->is_array(), "obj must be array");
-  objArrayOop a = objArrayOop(obj);
-  a->follow_header();
-  ObjArrayKlass_OOP_ITERATE( \
-    a, p, \
-    /* we call mark_and_follow here to avoid excessive marking stack usage */ \
-    MarkSweep::mark_and_follow(p))
+  objArrayOop(obj)->follow_header();
+  if (UseCompressedOops) {
+    objarray_follow_contents<narrowOop>(obj, 0);
+  } else {
+    objarray_follow_contents<oop>(obj, 0);
+  }
 }
 
 #ifndef SERIALGC
 void objArrayKlass::oop_follow_contents(ParCompactionManager* cm,
                                         oop obj) {
-  assert (obj->is_array(), "obj must be array");
-  objArrayOop a = objArrayOop(obj);
-  a->follow_header(cm);
-  ObjArrayKlass_OOP_ITERATE( \
-    a, p, \
-    /* we call mark_and_follow here to avoid excessive marking stack usage */ \
-    PSParallelCompact::mark_and_follow(cm, p))
+  assert(obj->is_array(), "obj must be array");
+  objArrayOop(obj)->follow_header(cm);
+  if (UseCompressedOops) {
+    objarray_follow_contents<narrowOop>(cm, obj, 0);
+  } else {
+    objarray_follow_contents<oop>(cm, obj, 0);
+  }
 }
 #endif // SERIALGC
 
diff --git a/src/share/vm/oops/objArrayKlass.hpp b/src/share/vm/oops/objArrayKlass.hpp
index fba1069b3..289e3916f 100644
--- a/src/share/vm/oops/objArrayKlass.hpp
+++ b/src/share/vm/oops/objArrayKlass.hpp
@@ -91,10 +91,18 @@ class objArrayKlass : public arrayKlass {
 
   // Garbage collection
   void oop_follow_contents(oop obj);
+  inline void oop_follow_contents(oop obj, int index);
+  template <class T> inline void objarray_follow_contents(oop obj, int index);
+
   int  oop_adjust_pointers(oop obj);
 
   // Parallel Scavenge and Parallel Old
   PARALLEL_GC_DECLS
+#ifndef SERIALGC
+  inline void oop_follow_contents(ParCompactionManager* cm, oop obj, int index);
+  template <class T> inline void
+    objarray_follow_contents(ParCompactionManager* cm, oop obj, int index);
+#endif // !SERIALGC
 
   // Iterators
   int oop_oop_iterate(oop obj, OopClosure* blk) {
@@ -131,5 +139,4 @@ class objArrayKlass : public arrayKlass {
   void oop_verify_on(oop obj, outputStream* st);
   void oop_verify_old_oop(oop obj, oop* p, bool allow_dirty);
   void oop_verify_old_oop(oop obj, narrowOop* p, bool allow_dirty);
-
 };
diff --git a/src/share/vm/oops/objArrayKlass.inline.hpp b/src/share/vm/oops/objArrayKlass.inline.hpp
new file mode 100644
index 000000000..97ea32bff
--- /dev/null
+++ b/src/share/vm/oops/objArrayKlass.inline.hpp
@@ -0,0 +1,89 @@
+/*
+ * Copyright 2010 Sun Microsystems, Inc.  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 Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
+ * CA 95054 USA or visit www.sun.com if you need additional information or
+ * have any questions.
+ *
+ */
+
+void objArrayKlass::oop_follow_contents(oop obj, int index) {
+  if (UseCompressedOops) {
+    objarray_follow_contents<narrowOop>(obj, index);
+  } else {
+    objarray_follow_contents<oop>(obj, index);
+  }
+}
+
+template <class T>
+void objArrayKlass::objarray_follow_contents(oop obj, int index) {
+  objArrayOop a = objArrayOop(obj);
+  const size_t len = size_t(a->length());
+  const size_t beg_index = size_t(index);
+  assert(beg_index < len || len == 0, "index too large");
+
+  const size_t stride = MIN2(len - beg_index, ObjArrayMarkingStride);
+  const size_t end_index = beg_index + stride;
+  T* const base = (T*)a->base();
+  T* const beg = base + beg_index;
+  T* const end = base + end_index;
+
+  // Push the non-NULL elements of the next stride on the marking stack.
+  for (T* e = beg; e < end; e++) {
+    MarkSweep::mark_and_push<T>(e);
+  }
+
+  if (end_index < len) {
+    MarkSweep::push_objarray(a, end_index); // Push the continuation.
+  }
+}
+
+#ifndef SERIALGC
+void objArrayKlass::oop_follow_contents(ParCompactionManager* cm, oop obj,
+                                        int index) {
+  if (UseCompressedOops) {
+    objarray_follow_contents<narrowOop>(cm, obj, index);
+  } else {
+    objarray_follow_contents<oop>(cm, obj, index);
+  }
+}
+
+template <class T>
+void objArrayKlass::objarray_follow_contents(ParCompactionManager* cm, oop obj,
+                                             int index) {
+  objArrayOop a = objArrayOop(obj);
+  const size_t len = size_t(a->length());
+  const size_t beg_index = size_t(index);
+  assert(beg_index < len || len == 0, "index too large");
+
+  const size_t stride = MIN2(len - beg_index, ObjArrayMarkingStride);
+  const size_t end_index = beg_index + stride;
+  T* const base = (T*)a->base();
+  T* const beg = base + beg_index;
+  T* const end = base + end_index;
+
+  // Push the non-NULL elements of the next stride on the marking stack.
+  for (T* e = beg; e < end; e++) {
+    PSParallelCompact::mark_and_push<T>(cm, e);
+  }
+
+  if (end_index < len) {
+    cm->push_objarray(a, end_index); // Push the continuation.
+  }
+}
+#endif // #ifndef SERIALGC
diff --git a/src/share/vm/runtime/arguments.cpp b/src/share/vm/runtime/arguments.cpp
index 11a632046..70fafe644 100644
--- a/src/share/vm/runtime/arguments.cpp
+++ b/src/share/vm/runtime/arguments.cpp
@@ -1346,9 +1346,7 @@ void Arguments::set_g1_gc_flags() {
   }
 
   if (FLAG_IS_DEFAULT(MarkStackSize)) {
-    // Size as a multiple of TaskQueueSuper::N which is larger
-    // for 64-bit.
-    FLAG_SET_DEFAULT(MarkStackSize, 128 * TaskQueueSuper::total_size());
+    FLAG_SET_DEFAULT(MarkStackSize, 128 * TASKQUEUE_SIZE);
   }
   if (PrintGCDetails && Verbose) {
     tty->print_cr("MarkStackSize: %uk  MarkStackSizeMax: %uk",
diff --git a/src/share/vm/runtime/globals.hpp b/src/share/vm/runtime/globals.hpp
index 4baf31700..d097c321b 100644
--- a/src/share/vm/runtime/globals.hpp
+++ b/src/share/vm/runtime/globals.hpp
@@ -1795,6 +1795,10 @@ class CommandLineFlags {
   product(uintx, PreserveMarkStackSize, 1024,                               \
           "Size for stack used in promotion failure handling")              \
                                                                             \
+  develop(uintx, ObjArrayMarkingStride, 512,                                \
+          "Number of ObjArray elements to push onto the marking stack"      \
+          "before pushing a continuation entry")                            \
+                                                                            \
   product_pd(bool, UseTLAB, "Use thread-local object allocation")           \
                                                                             \
   product_pd(bool, ResizeTLAB,                                              \
diff --git a/src/share/vm/utilities/globalDefinitions.hpp b/src/share/vm/utilities/globalDefinitions.hpp
index a83f4ba23..08898cc78 100644
--- a/src/share/vm/utilities/globalDefinitions.hpp
+++ b/src/share/vm/utilities/globalDefinitions.hpp
@@ -827,6 +827,8 @@ const int      badCodeHeapFreeVal = 0xDD;                   // value used to zap
 #define       badHeapWord       (::badHeapWordVal)
 #define       badJNIHandle      ((oop)::badJNIHandleVal)
 
+// Default TaskQueue size is 16K (32-bit) or 128K (64-bit)
+#define TASKQUEUE_SIZE (NOT_LP64(1<<14) LP64_ONLY(1<<17))
 
 //----------------------------------------------------------------------------------------------------
 // Utility functions for bitfield manipulations
diff --git a/src/share/vm/utilities/taskqueue.cpp b/src/share/vm/utilities/taskqueue.cpp
index 3a41b5754..35895d95c 100644
--- a/src/share/vm/utilities/taskqueue.cpp
+++ b/src/share/vm/utilities/taskqueue.cpp
@@ -31,10 +31,6 @@ uint ParallelTaskTerminator::_total_spins = 0;
 uint ParallelTaskTerminator::_total_peeks = 0;
 #endif
 
-bool TaskQueueSuper::peek() {
-  return _bottom != _age.top();
-}
-
 int TaskQueueSetSuper::randomParkAndMiller(int *seed0) {
   const int a =      16807;
   const int m = 2147483647;
@@ -180,6 +176,13 @@ void ParallelTaskTerminator::reset_for_reuse() {
   }
 }
 
+#ifdef ASSERT
+bool ObjArrayTask::is_valid() const {
+  return _obj != NULL && _obj->is_objArray() && _index > 0 &&
+    _index < objArrayOop(_obj)->length();
+}
+#endif // ASSERT
+
 bool RegionTaskQueueWithOverflow::is_empty() {
   return (_region_queue.size() == 0) &&
          (_overflow_stack->length() == 0);
diff --git a/src/share/vm/utilities/taskqueue.hpp b/src/share/vm/utilities/taskqueue.hpp
index 5376bc09f..45cba3e1d 100644
--- a/src/share/vm/utilities/taskqueue.hpp
+++ b/src/share/vm/utilities/taskqueue.hpp
@@ -22,6 +22,7 @@
  *
  */
 
+template <unsigned int N>
 class TaskQueueSuper: public CHeapObj {
 protected:
   // Internal type for indexing the queue; also used for the tag.
@@ -30,10 +31,7 @@ protected:
   // The first free element after the last one pushed (mod N).
   volatile uint _bottom;
 
-  enum {
-    N = 1 << NOT_LP64(14) LP64_ONLY(17), // Queue size: 16K or 128K
-    MOD_N_MASK = N - 1                   // To compute x mod N efficiently.
-  };
+  enum { MOD_N_MASK = N - 1 };
 
   class Age {
   public:
@@ -84,12 +82,12 @@ protected:
 
   // Returns a number in the range [0..N).  If the result is "N-1", it should be
   // interpreted as 0.
-  uint dirty_size(uint bot, uint top) {
+  uint dirty_size(uint bot, uint top) const {
     return (bot - top) & MOD_N_MASK;
   }
 
   // Returns the size corresponding to the given "bot" and "top".
-  uint size(uint bot, uint top) {
+  uint size(uint bot, uint top) const {
     uint sz = dirty_size(bot, top);
     // Has the queue "wrapped", so that bottom is less than top?  There's a
     // complicated special case here.  A pair of threads could perform pop_local
@@ -111,17 +109,17 @@ protected:
 public:
   TaskQueueSuper() : _bottom(0), _age() {}
 
-  // Return "true" if the TaskQueue contains any tasks.
-  bool peek();
+  // Return true if the TaskQueue contains any tasks.
+  bool peek() { return _bottom != _age.top(); }
 
   // Return an estimate of the number of elements in the queue.
   // The "careful" version admits the possibility of pop_local/pop_global
   // races.
-  uint size() {
+  uint size() const {
     return size(_bottom, _age.top());
   }
 
-  uint dirty_size() {
+  uint dirty_size() const {
     return dirty_size(_bottom, _age.top());
   }
 
@@ -132,19 +130,36 @@ public:
 
   // Maximum number of elements allowed in the queue.  This is two less
   // than the actual queue size, for somewhat complicated reasons.
-  uint max_elems() { return N - 2; }
+  uint max_elems() const { return N - 2; }
 
   // Total size of queue.
   static const uint total_size() { return N; }
 };
 
-template<class E> class GenericTaskQueue: public TaskQueueSuper {
+template<class E, unsigned int N = TASKQUEUE_SIZE>
+class GenericTaskQueue: public TaskQueueSuper<N> {
+protected:
+  typedef typename TaskQueueSuper<N>::Age Age;
+  typedef typename TaskQueueSuper<N>::idx_t idx_t;
+
+  using TaskQueueSuper<N>::_bottom;
+  using TaskQueueSuper<N>::_age;
+  using TaskQueueSuper<N>::increment_index;
+  using TaskQueueSuper<N>::decrement_index;
+  using TaskQueueSuper<N>::dirty_size;
+
+public:
+  using TaskQueueSuper<N>::max_elems;
+  using TaskQueueSuper<N>::size;
+
 private:
   // Slow paths for push, pop_local.  (pop_global has no fast path.)
   bool push_slow(E t, uint dirty_n_elems);
   bool pop_local_slow(uint localBot, Age oldAge);
 
 public:
+  typedef E element_type;
+
   // Initializes the queue to empty.
   GenericTaskQueue();
 
@@ -175,19 +190,19 @@ private:
   volatile E* _elems;
 };
 
-template<class E>
-GenericTaskQueue<E>::GenericTaskQueue():TaskQueueSuper() {
+template<class E, unsigned int N>
+GenericTaskQueue<E, N>::GenericTaskQueue() {
   assert(sizeof(Age) == sizeof(size_t), "Depends on this.");
 }
 
-template<class E>
-void GenericTaskQueue<E>::initialize() {
+template<class E, unsigned int N>
+void GenericTaskQueue<E, N>::initialize() {
   _elems = NEW_C_HEAP_ARRAY(E, N);
   guarantee(_elems != NULL, "Allocation failed.");
 }
 
-template<class E>
-void GenericTaskQueue<E>::oops_do(OopClosure* f) {
+template<class E, unsigned int N>
+void GenericTaskQueue<E, N>::oops_do(OopClosure* f) {
   // tty->print_cr("START OopTaskQueue::oops_do");
   uint iters = size();
   uint index = _bottom;
@@ -203,21 +218,21 @@ void GenericTaskQueue<E>::oops_do(OopClosure* f) {
   // tty->print_cr("END OopTaskQueue::oops_do");
 }
 
-
-template<class E>
-bool GenericTaskQueue<E>::push_slow(E t, uint dirty_n_elems) {
+template<class E, unsigned int N>
+bool GenericTaskQueue<E, N>::push_slow(E t, uint dirty_n_elems) {
   if (dirty_n_elems == N - 1) {
     // Actually means 0, so do the push.
     uint localBot = _bottom;
-    _elems[localBot] = t;
+    // g++ complains if the volatile result of the assignment is unused.
+    const_cast<E&>(_elems[localBot] = t);
     OrderAccess::release_store(&_bottom, increment_index(localBot));
     return true;
   }
   return false;
 }
 
-template<class E>
-bool GenericTaskQueue<E>::
+template<class E, unsigned int N>
+bool GenericTaskQueue<E, N>::
 pop_local_slow(uint localBot, Age oldAge) {
   // This queue was observed to contain exactly one element; either this
   // thread will claim it, or a competing "pop_global".  In either case,
@@ -249,8 +264,8 @@ pop_local_slow(uint localBot, Age oldAge) {
   return false;
 }
 
-template<class E>
-bool GenericTaskQueue<E>::pop_global(E& t) {
+template<class E, unsigned int N>
+bool GenericTaskQueue<E, N>::pop_global(E& t) {
   Age oldAge = _age.get();
   uint localBot = _bottom;
   uint n_elems = size(localBot, oldAge.top());
@@ -258,7 +273,7 @@ bool GenericTaskQueue<E>::pop_global(E& t) {
     return false;
   }
 
-  t = _elems[oldAge.top()];
+  const_cast<E&>(t = _elems[oldAge.top()]);
   Age newAge(oldAge);
   newAge.increment();
   Age resAge = _age.cmpxchg(newAge, oldAge);
@@ -269,8 +284,8 @@ bool GenericTaskQueue<E>::pop_global(E& t) {
   return resAge == oldAge;
 }
 
-template<class E>
-GenericTaskQueue<E>::~GenericTaskQueue() {
+template<class E, unsigned int N>
+GenericTaskQueue<E, N>::~GenericTaskQueue() {
   FREE_C_HEAP_ARRAY(E, _elems);
 }
 
@@ -283,16 +298,18 @@ public:
   virtual bool peek() = 0;
 };
 
-template<class E> class GenericTaskQueueSet: public TaskQueueSetSuper {
+template<class T>
+class GenericTaskQueueSet: public TaskQueueSetSuper {
 private:
   uint _n;
-  GenericTaskQueue<E>** _queues;
+  T** _queues;
 
 public:
+  typedef typename T::element_type E;
+
   GenericTaskQueueSet(int n) : _n(n) {
-    typedef GenericTaskQueue<E>* GenericTaskQueuePtr;
+    typedef T* GenericTaskQueuePtr;
     _queues = NEW_C_HEAP_ARRAY(GenericTaskQueuePtr, n);
-    guarantee(_queues != NULL, "Allocation failure.");
     for (int i = 0; i < n; i++) {
       _queues[i] = NULL;
     }
@@ -302,9 +319,9 @@ public:
   bool steal_best_of_2(uint queue_num, int* seed, E& t);
   bool steal_best_of_all(uint queue_num, int* seed, E& t);
 
-  void register_queue(uint i, GenericTaskQueue<E>* q);
+  void register_queue(uint i, T* q);
 
-  GenericTaskQueue<E>* queue(uint n);
+  T* queue(uint n);
 
   // The thread with queue number "queue_num" (and whose random number seed
   // is at "seed") is trying to steal a task from some other queue.  (It
@@ -316,27 +333,27 @@ public:
   bool peek();
 };
 
-template<class E>
-void GenericTaskQueueSet<E>::register_queue(uint i, GenericTaskQueue<E>* q) {
+template<class T> void
+GenericTaskQueueSet<T>::register_queue(uint i, T* q) {
   assert(i < _n, "index out of range.");
   _queues[i] = q;
 }
 
-template<class E>
-GenericTaskQueue<E>* GenericTaskQueueSet<E>::queue(uint i) {
+template<class T> T*
+GenericTaskQueueSet<T>::queue(uint i) {
   return _queues[i];
 }
 
-template<class E>
-bool GenericTaskQueueSet<E>::steal(uint queue_num, int* seed, E& t) {
+template<class T> bool
+GenericTaskQueueSet<T>::steal(uint queue_num, int* seed, E& t) {
   for (uint i = 0; i < 2 * _n; i++)
     if (steal_best_of_2(queue_num, seed, t))
       return true;
   return false;
 }
 
-template<class E>
-bool GenericTaskQueueSet<E>::steal_best_of_all(uint queue_num, int* seed, E& t) {
+template<class T> bool
+GenericTaskQueueSet<T>::steal_best_of_all(uint queue_num, int* seed, E& t) {
   if (_n > 2) {
     int best_k;
     uint best_sz = 0;
@@ -359,8 +376,8 @@ bool GenericTaskQueueSet<E>::steal_best_of_all(uint queue_num, int* seed, E& t)
   }
 }
 
-template<class E>
-bool GenericTaskQueueSet<E>::steal_1_random(uint queue_num, int* seed, E& t) {
+template<class T> bool
+GenericTaskQueueSet<T>::steal_1_random(uint queue_num, int* seed, E& t) {
   if (_n > 2) {
     uint k = queue_num;
     while (k == queue_num) k = randomParkAndMiller(seed) % _n;
@@ -375,8 +392,8 @@ bool GenericTaskQueueSet<E>::steal_1_random(uint queue_num, int* seed, E& t) {
   }
 }
 
-template<class E>
-bool GenericTaskQueueSet<E>::steal_best_of_2(uint queue_num, int* seed, E& t) {
+template<class T> bool
+GenericTaskQueueSet<T>::steal_best_of_2(uint queue_num, int* seed, E& t) {
   if (_n > 2) {
     uint k1 = queue_num;
     while (k1 == queue_num) k1 = randomParkAndMiller(seed) % _n;
@@ -397,8 +414,8 @@ bool GenericTaskQueueSet<E>::steal_best_of_2(uint queue_num, int* seed, E& t) {
   }
 }
 
-template<class E>
-bool GenericTaskQueueSet<E>::peek() {
+template<class T>
+bool GenericTaskQueueSet<T>::peek() {
   // Try all the queues.
   for (uint j = 0; j < _n; j++) {
     if (_queues[j]->peek())
@@ -468,14 +485,16 @@ public:
 #endif
 };
 
-template<class E> inline bool GenericTaskQueue<E>::push(E t) {
+template<class E, unsigned int N> inline bool
+GenericTaskQueue<E, N>::push(E t) {
   uint localBot = _bottom;
   assert((localBot >= 0) && (localBot < N), "_bottom out of range.");
   idx_t top = _age.top();
   uint dirty_n_elems = dirty_size(localBot, top);
-  assert((dirty_n_elems >= 0) && (dirty_n_elems < N), "n_elems out of range.");
+  assert(dirty_n_elems < N, "n_elems out of range.");
   if (dirty_n_elems < max_elems()) {
-    _elems[localBot] = t;
+    // g++ complains if the volatile result of the assignment is unused.
+    const_cast<E&>(_elems[localBot] = t);
     OrderAccess::release_store(&_bottom, increment_index(localBot));
     return true;
   } else {
@@ -483,7 +502,8 @@ template<class E> inline bool GenericTaskQueue<E>::push(E t) {
   }
 }
 
-template<class E> inline bool GenericTaskQueue<E>::pop_local(E& t) {
+template<class E, unsigned int N> inline bool
+GenericTaskQueue<E, N>::pop_local(E& t) {
   uint localBot = _bottom;
   // This value cannot be N-1.  That can only occur as a result of
   // the assignment to bottom in this method.  If it does, this method
@@ -497,7 +517,7 @@ template<class E> inline bool GenericTaskQueue<E>::pop_local(E& t) {
   // This is necessary to prevent any read below from being reordered
   // before the store just above.
   OrderAccess::fence();
-  t = _elems[localBot];
+  const_cast<E&>(t = _elems[localBot]);
   // This is a second read of "age"; the "size()" above is the first.
   // If there's still at least one element in the queue, based on the
   // "_bottom" and "age" we've read, then there can be no interference with
@@ -514,17 +534,23 @@ template<class E> inline bool GenericTaskQueue<E>::pop_local(E& t) {
 }
 
 typedef oop Task;
-typedef GenericTaskQueue<Task>         OopTaskQueue;
-typedef GenericTaskQueueSet<Task>      OopTaskQueueSet;
+typedef GenericTaskQueue<Task>            OopTaskQueue;
+typedef GenericTaskQueueSet<OopTaskQueue> OopTaskQueueSet;
 
-
-#define COMPRESSED_OOP_MASK  1
+#ifdef _MSC_VER
+#pragma warning(push)
+// warning C4522: multiple assignment operators specified
+#pragma warning(disable:4522)
+#endif
 
 // This is a container class for either an oop* or a narrowOop*.
 // Both are pushed onto a task queue and the consumer will test is_narrow()
 // to determine which should be processed.
 class StarTask {
   void*  _holder;        // either union oop* or narrowOop*
+
+  enum { COMPRESSED_OOP_MASK = 1 };
+
  public:
   StarTask(narrowOop* p) {
     assert(((uintptr_t)p & COMPRESSED_OOP_MASK) == 0, "Information loss!");
@@ -540,20 +566,61 @@ class StarTask {
     return (narrowOop*)((uintptr_t)_holder & ~COMPRESSED_OOP_MASK);
   }
 
-  // Operators to preserve const/volatile in assignments required by gcc
-  void operator=(const volatile StarTask& t) volatile { _holder = t._holder; }
+  StarTask& operator=(const StarTask& t) {
+    _holder = t._holder;
+    return *this;
+  }
+  volatile StarTask& operator=(const volatile StarTask& t) volatile {
+    _holder = t._holder;
+    return *this;
+  }
 
   bool is_narrow() const {
     return (((uintptr_t)_holder & COMPRESSED_OOP_MASK) != 0);
   }
 };
 
-typedef GenericTaskQueue<StarTask>     OopStarTaskQueue;
-typedef GenericTaskQueueSet<StarTask>  OopStarTaskQueueSet;
+class ObjArrayTask
+{
+public:
+  ObjArrayTask(oop o = NULL, int idx = 0): _obj(o), _index(idx) { }
+  ObjArrayTask(oop o, size_t idx): _obj(o), _index(int(idx)) {
+    assert(idx <= size_t(max_jint), "too big");
+  }
+  ObjArrayTask(const ObjArrayTask& t): _obj(t._obj), _index(t._index) { }
+
+  ObjArrayTask& operator =(const ObjArrayTask& t) {
+    _obj = t._obj;
+    _index = t._index;
+    return *this;
+  }
+  volatile ObjArrayTask&
+  operator =(const volatile ObjArrayTask& t) volatile {
+    _obj = t._obj;
+    _index = t._index;
+    return *this;
+  }
+
+  inline oop obj()   const { return _obj; }
+  inline int index() const { return _index; }
+
+  DEBUG_ONLY(bool is_valid() const); // Tasks to be pushed/popped must be valid.
+
+private:
+  oop _obj;
+  int _index;
+};
+
+#ifdef _MSC_VER
+#pragma warning(pop)
+#endif
+
+typedef GenericTaskQueue<StarTask>            OopStarTaskQueue;
+typedef GenericTaskQueueSet<OopStarTaskQueue> OopStarTaskQueueSet;
 
 typedef size_t RegionTask;  // index for region
-typedef GenericTaskQueue<RegionTask>    RegionTaskQueue;
-typedef GenericTaskQueueSet<RegionTask> RegionTaskQueueSet;
+typedef GenericTaskQueue<RegionTask>         RegionTaskQueue;
+typedef GenericTaskQueueSet<RegionTaskQueue> RegionTaskQueueSet;
 
 class RegionTaskQueueWithOverflow: public CHeapObj {
  protected: