8264788: Make SequentialSubTasksDone use-once

Reviewed-by: ayang, sjohanss

src/hotspot/share/gc/parallel/psParallelCompact.cpp

@@ -2056,18 +2056,15 @@ public:
   MarkFromRootsTask(uint active_workers) :
       AbstractGangTask("MarkFromRootsTask"),
       _strong_roots_scope(active_workers),
-      _subtasks(),
+      _subtasks(ParallelRootType::sentinel),
       _terminator(active_workers, ParCompactionManager::oop_task_queues()),
       _active_workers(active_workers) {
-    _subtasks.set_n_threads(active_workers);
-    _subtasks.set_n_tasks(ParallelRootType::sentinel);
   }
 
   virtual void work(uint worker_id) {
     for (uint task = 0; _subtasks.try_claim_task(task); /*empty*/ ) {
       mark_from_roots_work(static_cast<ParallelRootType::Value>(task), worker_id);
     }
-    _subtasks.all_tasks_completed();
 
     PCAddThreadRootsMarkingTaskClosure closure(worker_id);
     Threads::possibly_parallel_threads_do(true /*parallel */, &closure);

src/hotspot/share/gc/parallel/psScavenge.cpp

@@ -304,14 +304,12 @@ public:
                     bool is_empty) :
       AbstractGangTask("ScavengeRootsTask"),
       _strong_roots_scope(active_workers),
-      _subtasks(),
+      _subtasks(ParallelRootType::sentinel),
       _old_gen(old_gen),
       _gen_top(gen_top),
       _active_workers(active_workers),
       _is_empty(is_empty),
       _terminator(active_workers, PSPromotionManager::vm_thread_promotion_manager()->stack_array_depth()) {
-    _subtasks.set_n_threads(active_workers);
-    _subtasks.set_n_tasks(ParallelRootType::sentinel);
   }
 
   virtual void work(uint worker_id) {
@@ -346,7 +344,6 @@ public:
     for (uint root_type = 0; _subtasks.try_claim_task(root_type); /* empty */ ) {
       scavenge_roots_work(static_cast<ParallelRootType::Value>(root_type), worker_id);
     }
-    _subtasks.all_tasks_completed();
 
     PSThreadRootsTaskClosure closure(worker_id);
     Threads::possibly_parallel_threads_do(true /*parallel */, &closure);

src/hotspot/share/gc/shared/preservedMarks.cpp

@@ -104,18 +104,15 @@ public:
     while (_sub_tasks.try_claim_task(/* reference */ task_id)) {
       _preserved_marks_set->get(task_id)->restore_and_increment(_total_size_addr);
     }
-    _sub_tasks.all_tasks_completed();
   }
 
-  ParRestoreTask(uint worker_num,
+  ParRestoreTask(PreservedMarksSet* preserved_marks_set,
-                 PreservedMarksSet* preserved_marks_set,
                  volatile size_t* total_size_addr)
       : AbstractGangTask("Parallel Preserved Mark Restoration"),
         _preserved_marks_set(preserved_marks_set),
+        _sub_tasks(preserved_marks_set->num()),
         _total_size_addr(total_size_addr) {
-    _sub_tasks.set_n_threads(worker_num);
+    }
-    _sub_tasks.set_n_tasks(preserved_marks_set->num());
-  }
 };
 
 void PreservedMarksSet::restore(WorkGang* workers) {
@@ -127,7 +124,7 @@ void PreservedMarksSet::restore(WorkGang* workers) {
   for (uint i = 0; i < _num; i += 1) {
     total_size_before += get(i)->size();
   }
-#endif // def ASSERT
+#endif // ASSERT
 
   if (workers == NULL) {
     for (uint i = 0; i < num(); i += 1) {
@@ -135,7 +132,7 @@ void PreservedMarksSet::restore(WorkGang* workers) {
       get(i)->restore();
     }
   } else {
-    ParRestoreTask task(workers->active_workers(), this, &total_size);
+    ParRestoreTask task(this, &total_size);
     workers->run_task(&task);
   }
 

src/hotspot/share/gc/shared/workgroup.cpp

@@ -399,39 +399,10 @@ SubTasksDone::~SubTasksDone() {
 
 // *** SequentialSubTasksDone
 
-void SequentialSubTasksDone::clear() {
-  _n_tasks   = _n_claimed   = 0;
-  _n_threads = _n_completed = 0;
-}
-
-bool SequentialSubTasksDone::valid() {
-  return _n_threads > 0;
-}
-
 bool SequentialSubTasksDone::try_claim_task(uint& t) {
-  t = _n_claimed;
+  t = _num_claimed;
-  while (t < _n_tasks) {
+  if (t < _num_tasks) {
-    uint res = Atomic::cmpxchg(&_n_claimed, t, t+1);
+    t = Atomic::add(&_num_claimed, 1u) - 1;
-    if (res == t) {
-      return true;
-    }
-    t = res;
   }
-  return false;
+  return t < _num_tasks;
-}
-
-bool SequentialSubTasksDone::all_tasks_completed() {
-  uint complete = _n_completed;
-  while (true) {
-    uint res = Atomic::cmpxchg(&_n_completed, complete, complete+1);
-    if (res == complete) {
-      break;
-    }
-    complete = res;
-  }
-  if (complete+1 == _n_threads) {
-    clear();
-    return true;
-  }
-  return false;
 }

src/hotspot/share/gc/shared/workgroup.hpp

@@ -346,57 +346,27 @@ public:
 // sub-tasks from a set (possibly an enumeration), claim sub-tasks
 // in sequential order. This is ideal for claiming dynamically
 // partitioned tasks (like striding in the parallel remembered
-// set scanning). Note that unlike the above class this is
+// set scanning).
-// a stack object - is there any reason for it not to be?
 
-class SequentialSubTasksDone : public StackObj {
+class SequentialSubTasksDone : public CHeapObj<mtInternal> {
-protected:
-  uint _n_tasks;     // Total number of tasks available.
-  volatile uint _n_claimed;   // Number of tasks claimed.
-  // _n_threads is used to determine when a sub task is done.
-  // See comments on SubTasksDone::_n_threads
-  uint _n_threads;   // Total number of parallel threads.
-  volatile uint _n_completed; // Number of completed threads.
 
-  void clear();
+  uint _num_tasks;     // Total number of tasks available.
+  volatile uint _num_claimed;   // Number of tasks claimed.
+
+  NONCOPYABLE(SequentialSubTasksDone);
 
 public:
-  SequentialSubTasksDone() {
+  SequentialSubTasksDone(uint num_tasks) : _num_tasks(num_tasks), _num_claimed(0) { }
-    clear();
+  ~SequentialSubTasksDone() {
+    // Claiming may try to claim more tasks than there are.
+    assert(_num_claimed >= _num_tasks, "Claimed %u tasks of %u", _num_claimed, _num_tasks);
   }
-  ~SequentialSubTasksDone() {}
-
-  // True iff the object is in a valid state.
-  bool valid();
-
-  // number of tasks
-  uint n_tasks() const { return _n_tasks; }
-
-  // Get/set the number of parallel threads doing the tasks to t.
-  // Should be called before the task starts but it is safe
-  // to call this once a task is running provided that all
-  // threads agree on the number of threads.
-  uint n_threads() { return _n_threads; }
-  void set_n_threads(uint t) { _n_threads = t; }
-
-  // Set the number of tasks to be claimed to t. As above,
-  // should be called before the tasks start but it is safe
-  // to call this once a task is running provided all threads
-  // agree on the number of tasks.
-  void set_n_tasks(uint t) { _n_tasks = t; }
 
   // Attempt to claim the next unclaimed task in the sequence,
   // returning true if successful, with t set to the index of the
-  // claimed task.  Returns false if there are no more unclaimed tasks
+  // claimed task. Returns false if there are no more unclaimed tasks
-  // in the sequence.
+  // in the sequence. In this case t is undefined.
   bool try_claim_task(uint& t);
-
-  // The calling thread asserts that it has attempted to claim
-  // all the tasks it possibly can in the sequence. Every thread
-  // claiming tasks must promise call this. Returns true if this
-  // is the last thread to complete so that the thread can perform
-  // cleanup if necessary.
-  bool all_tasks_completed();
 };
 
 #endif // SHARE_GC_SHARED_WORKGROUP_HPP