1 /*
   2  * Copyright (c) 2001, 2016, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #ifndef SHARE_VM_UTILITIES_TASKQUEUE_HPP
  26 #define SHARE_VM_UTILITIES_TASKQUEUE_HPP
  27 
  28 #include "memory/allocation.hpp"
  29 #include "memory/allocation.inline.hpp"
  30 #include "runtime/mutex.hpp"
  31 #include "runtime/orderAccess.inline.hpp"
  32 #include "utilities/globalDefinitions.hpp"
  33 #include "utilities/stack.hpp"
  34 
  35 // Simple TaskQueue stats that are collected by default in debug builds.
  36 
  37 #if !defined(TASKQUEUE_STATS) && defined(ASSERT)
  38 #define TASKQUEUE_STATS 1
  39 #elif !defined(TASKQUEUE_STATS)
  40 #define TASKQUEUE_STATS 0
  41 #endif
  42 
  43 #if TASKQUEUE_STATS
  44 #define TASKQUEUE_STATS_ONLY(code) code
  45 #else
  46 #define TASKQUEUE_STATS_ONLY(code)
  47 #endif // TASKQUEUE_STATS
  48 
  49 #if TASKQUEUE_STATS
  50 class TaskQueueStats {
  51 public:
  52   enum StatId {
  53     push,             // number of taskqueue pushes
  54     pop,              // number of taskqueue pops
  55     pop_slow,         // subset of taskqueue pops that were done slow-path
  56     steal_attempt,    // number of taskqueue steal attempts
  57     steal,            // number of taskqueue steals
  58     overflow,         // number of overflow pushes
  59     overflow_max_len, // max length of overflow stack
  60     last_stat_id
  61   };
  62 
  63 public:
  64   inline TaskQueueStats()       { reset(); }
  65 
  66   inline void record_push()     { ++_stats[push]; }
  67   inline void record_pop()      { ++_stats[pop]; }
  68   inline void record_pop_slow() { record_pop(); ++_stats[pop_slow]; }
  69   inline void record_steal(bool success);
  70   inline void record_overflow(size_t new_length);
  71 
  72   TaskQueueStats & operator +=(const TaskQueueStats & addend);
  73 
  74   inline size_t get(StatId id) const { return _stats[id]; }
  75   inline const size_t* get() const   { return _stats; }
  76 
  77   inline void reset();
  78 
  79   // Print the specified line of the header (does not include a line separator).
  80   static void print_header(unsigned int line, outputStream* const stream = tty,
  81                            unsigned int width = 10);
  82   // Print the statistics (does not include a line separator).
  83   void print(outputStream* const stream = tty, unsigned int width = 10) const;
  84 
  85   DEBUG_ONLY(void verify() const;)
  86 
  87 private:
  88   size_t                    _stats[last_stat_id];
  89   static const char * const _names[last_stat_id];
  90 };
  91 
  92 void TaskQueueStats::record_steal(bool success) {
  93   ++_stats[steal_attempt];
  94   if (success) ++_stats[steal];
  95 }
  96 
  97 void TaskQueueStats::record_overflow(size_t new_len) {
  98   ++_stats[overflow];
  99   if (new_len > _stats[overflow_max_len]) _stats[overflow_max_len] = new_len;
 100 }
 101 
 102 void TaskQueueStats::reset() {
 103   memset(_stats, 0, sizeof(_stats));
 104 }
 105 #endif // TASKQUEUE_STATS
 106 
 107 // TaskQueueSuper collects functionality common to all GenericTaskQueue instances.
 108 
 109 template <unsigned int N, MEMFLAGS F>
 110 class TaskQueueSuper: public CHeapObj<F> {
 111 protected:
 112   // Internal type for indexing the queue; also used for the tag.
 113   typedef NOT_LP64(uint16_t) LP64_ONLY(uint32_t) idx_t;
 114 
 115   // The first free element after the last one pushed (mod N).
 116   volatile uint _bottom;
 117 
 118   enum { MOD_N_MASK = N - 1 };
 119 
 120   class Age {
 121   public:
 122     Age(size_t data = 0)         { _data = data; }
 123     Age(const Age& age)          { _data = age._data; }
 124     Age(idx_t top, idx_t tag)    { _fields._top = top; _fields._tag = tag; }
 125 
 126     Age   get()        const volatile { return _data; }
 127     void  set(Age age) volatile       { _data = age._data; }
 128 
 129     idx_t top()        const volatile { return _fields._top; }
 130     idx_t tag()        const volatile { return _fields._tag; }
 131 
 132     // Increment top; if it wraps, increment tag also.
 133     void increment() {
 134       _fields._top = increment_index(_fields._top);
 135       if (_fields._top == 0) ++_fields._tag;
 136     }
 137 
 138     Age cmpxchg(const Age new_age, const Age old_age) volatile {
 139       return (size_t) Atomic::cmpxchg_ptr((intptr_t)new_age._data,
 140                                           (volatile intptr_t *)&_data,
 141                                           (intptr_t)old_age._data);
 142     }
 143 
 144     bool operator ==(const Age& other) const { return _data == other._data; }
 145 
 146   private:
 147     struct fields {
 148       idx_t _top;
 149       idx_t _tag;
 150     };
 151     union {
 152       size_t _data;
 153       fields _fields;
 154     };
 155   };
 156 
 157   volatile Age _age;
 158 
 159   // These both operate mod N.
 160   static uint increment_index(uint ind) {
 161     return (ind + 1) & MOD_N_MASK;
 162   }
 163   static uint decrement_index(uint ind) {
 164     return (ind - 1) & MOD_N_MASK;
 165   }
 166 
 167   // Returns a number in the range [0..N).  If the result is "N-1", it should be
 168   // interpreted as 0.
 169   uint dirty_size(uint bot, uint top) const {
 170     return (bot - top) & MOD_N_MASK;
 171   }
 172 
 173   // Returns the size corresponding to the given "bot" and "top".
 174   uint size(uint bot, uint top) const {
 175     uint sz = dirty_size(bot, top);
 176     // Has the queue "wrapped", so that bottom is less than top?  There's a
 177     // complicated special case here.  A pair of threads could perform pop_local
 178     // and pop_global operations concurrently, starting from a state in which
 179     // _bottom == _top+1.  The pop_local could succeed in decrementing _bottom,
 180     // and the pop_global in incrementing _top (in which case the pop_global
 181     // will be awarded the contested queue element.)  The resulting state must
 182     // be interpreted as an empty queue.  (We only need to worry about one such
 183     // event: only the queue owner performs pop_local's, and several concurrent
 184     // threads attempting to perform the pop_global will all perform the same
 185     // CAS, and only one can succeed.)  Any stealing thread that reads after
 186     // either the increment or decrement will see an empty queue, and will not
 187     // join the competitors.  The "sz == -1 || sz == N-1" state will not be
 188     // modified by concurrent queues, so the owner thread can reset the state to
 189     // _bottom == top so subsequent pushes will be performed normally.
 190     return (sz == N - 1) ? 0 : sz;
 191   }
 192 
 193 public:
 194   TaskQueueSuper() : _bottom(0), _age() {}
 195 
 196   // Return true if the TaskQueue contains/does not contain any tasks.
 197   bool peek()     const { return _bottom != _age.top(); }
 198   bool is_empty() const { return size() == 0; }
 199 
 200   // Return an estimate of the number of elements in the queue.
 201   // The "careful" version admits the possibility of pop_local/pop_global
 202   // races.
 203   uint size() const {
 204     return size(_bottom, _age.top());
 205   }
 206 
 207   uint dirty_size() const {
 208     return dirty_size(_bottom, _age.top());
 209   }
 210 
 211   void set_empty() {
 212     _bottom = 0;
 213     _age.set(0);
 214   }
 215 
 216   // Maximum number of elements allowed in the queue.  This is two less
 217   // than the actual queue size, for somewhat complicated reasons.
 218   uint max_elems() const { return N - 2; }
 219 
 220   // Total size of queue.
 221   static const uint total_size() { return N; }
 222 
 223   TASKQUEUE_STATS_ONLY(TaskQueueStats stats;)
 224 };
 225 
 226 //
 227 // GenericTaskQueue implements an ABP, Aurora-Blumofe-Plaxton, double-
 228 // ended-queue (deque), intended for use in work stealing. Queue operations
 229 // are non-blocking.
 230 //
 231 // A queue owner thread performs push() and pop_local() operations on one end
 232 // of the queue, while other threads may steal work using the pop_global()
 233 // method.
 234 //
 235 // The main difference to the original algorithm is that this
 236 // implementation allows wrap-around at the end of its allocated
 237 // storage, which is an array.
 238 //
 239 // The original paper is:
 240 //
 241 // Arora, N. S., Blumofe, R. D., and Plaxton, C. G.
 242 // Thread scheduling for multiprogrammed multiprocessors.
 243 // Theory of Computing Systems 34, 2 (2001), 115-144.
 244 //
 245 // The following paper provides an correctness proof and an
 246 // implementation for weakly ordered memory models including (pseudo-)
 247 // code containing memory barriers for a Chase-Lev deque. Chase-Lev is
 248 // similar to ABP, with the main difference that it allows resizing of the
 249 // underlying storage:
 250 //
 251 // Le, N. M., Pop, A., Cohen A., and Nardell, F. Z.
 252 // Correct and efficient work-stealing for weak memory models
 253 // Proceedings of the 18th ACM SIGPLAN symposium on Principles and
 254 // practice of parallel programming (PPoPP 2013), 69-80
 255 //
 256 
 257 template <class E, MEMFLAGS F, unsigned int N = TASKQUEUE_SIZE>
 258 class GenericTaskQueue: public TaskQueueSuper<N, F> {
 259   ArrayAllocator<E, F> _array_allocator;
 260 protected:
 261   typedef typename TaskQueueSuper<N, F>::Age Age;
 262   typedef typename TaskQueueSuper<N, F>::idx_t idx_t;
 263 
 264   using TaskQueueSuper<N, F>::_bottom;
 265   using TaskQueueSuper<N, F>::_age;
 266   using TaskQueueSuper<N, F>::increment_index;
 267   using TaskQueueSuper<N, F>::decrement_index;
 268   using TaskQueueSuper<N, F>::dirty_size;
 269 
 270 public:
 271   using TaskQueueSuper<N, F>::max_elems;
 272   using TaskQueueSuper<N, F>::size;
 273 
 274 #if  TASKQUEUE_STATS
 275   using TaskQueueSuper<N, F>::stats;
 276 #endif
 277 
 278 private:
 279   // Slow paths for push, pop_local.  (pop_global has no fast path.)
 280   bool push_slow(E t, uint dirty_n_elems);
 281   bool pop_local_slow(uint localBot, Age oldAge);
 282 
 283 public:
 284   typedef E element_type;
 285 
 286   // Initializes the queue to empty.
 287   GenericTaskQueue();
 288 
 289   void initialize();
 290 
 291   // Push the task "t" on the queue.  Returns "false" iff the queue is full.
 292   inline bool push(E t);
 293 
 294   // Attempts to claim a task from the "local" end of the queue (the most
 295   // recently pushed).  If successful, returns true and sets t to the task;
 296   // otherwise, returns false (the queue is empty).
 297   inline bool pop_local(volatile E& t);
 298 
 299   // Like pop_local(), but uses the "global" end of the queue (the least
 300   // recently pushed).
 301   bool pop_global(volatile E& t);
 302 
 303   // Delete any resource associated with the queue.
 304   ~GenericTaskQueue();
 305 
 306   // apply the closure to all elements in the task queue
 307   void oops_do(OopClosure* f);
 308 
 309 private:
 310   // Element array.
 311   volatile E* _elems;
 312 };
 313 
 314 template<class E, MEMFLAGS F, unsigned int N>
 315 GenericTaskQueue<E, F, N>::GenericTaskQueue() {
 316   assert(sizeof(Age) == sizeof(size_t), "Depends on this.");
 317 }
 318 
 319 template<class E, MEMFLAGS F, unsigned int N>
 320 void GenericTaskQueue<E, F, N>::initialize() {
 321   _elems = _array_allocator.allocate(N);
 322 }
 323 
 324 template<class E, MEMFLAGS F, unsigned int N>
 325 void GenericTaskQueue<E, F, N>::oops_do(OopClosure* f) {
 326   // tty->print_cr("START OopTaskQueue::oops_do");
 327   uint iters = size();
 328   uint index = _bottom;
 329   for (uint i = 0; i < iters; ++i) {
 330     index = decrement_index(index);
 331     // tty->print_cr("  doing entry %d," INTPTR_T " -> " INTPTR_T,
 332     //            index, &_elems[index], _elems[index]);
 333     E* t = (E*)&_elems[index];      // cast away volatility
 334     oop* p = (oop*)t;
 335     // G1 does its own checking
 336     assert(UseG1GC || (*t)->is_oop_or_null(), "Not an oop or null");
 337     f->do_oop(p);
 338   }
 339   // tty->print_cr("END OopTaskQueue::oops_do");
 340 }
 341 
 342 template<class E, MEMFLAGS F, unsigned int N>
 343 bool GenericTaskQueue<E, F, N>::push_slow(E t, uint dirty_n_elems) {
 344   if (dirty_n_elems == N - 1) {
 345     // Actually means 0, so do the push.
 346     uint localBot = _bottom;
 347     // g++ complains if the volatile result of the assignment is
 348     // unused, so we cast the volatile away.  We cannot cast directly
 349     // to void, because gcc treats that as not using the result of the
 350     // assignment.  However, casting to E& means that we trigger an
 351     // unused-value warning.  So, we cast the E& to void.
 352     (void)const_cast<E&>(_elems[localBot] = t);
 353     OrderAccess::release_store(&_bottom, increment_index(localBot));
 354     TASKQUEUE_STATS_ONLY(stats.record_push());
 355     return true;
 356   }
 357   return false;
 358 }
 359 
 360 // pop_local_slow() is done by the owning thread and is trying to
 361 // get the last task in the queue.  It will compete with pop_global()
 362 // that will be used by other threads.  The tag age is incremented
 363 // whenever the queue goes empty which it will do here if this thread
 364 // gets the last task or in pop_global() if the queue wraps (top == 0
 365 // and pop_global() succeeds, see pop_global()).
 366 template<class E, MEMFLAGS F, unsigned int N>
 367 bool GenericTaskQueue<E, F, N>::pop_local_slow(uint localBot, Age oldAge) {
 368   // This queue was observed to contain exactly one element; either this
 369   // thread will claim it, or a competing "pop_global".  In either case,
 370   // the queue will be logically empty afterwards.  Create a new Age value
 371   // that represents the empty queue for the given value of "_bottom".  (We
 372   // must also increment "tag" because of the case where "bottom == 1",
 373   // "top == 0".  A pop_global could read the queue element in that case,
 374   // then have the owner thread do a pop followed by another push.  Without
 375   // the incrementing of "tag", the pop_global's CAS could succeed,
 376   // allowing it to believe it has claimed the stale element.)
 377   Age newAge((idx_t)localBot, oldAge.tag() + 1);
 378   // Perhaps a competing pop_global has already incremented "top", in which
 379   // case it wins the element.
 380   if (localBot == oldAge.top()) {
 381     // No competing pop_global has yet incremented "top"; we'll try to
 382     // install new_age, thus claiming the element.
 383     Age tempAge = _age.cmpxchg(newAge, oldAge);
 384     if (tempAge == oldAge) {
 385       // We win.
 386       assert(dirty_size(localBot, _age.top()) != N - 1, "sanity");
 387       TASKQUEUE_STATS_ONLY(stats.record_pop_slow());
 388       return true;
 389     }
 390   }
 391   // We lose; a completing pop_global gets the element.  But the queue is empty
 392   // and top is greater than bottom.  Fix this representation of the empty queue
 393   // to become the canonical one.
 394   _age.set(newAge);
 395   assert(dirty_size(localBot, _age.top()) != N - 1, "sanity");
 396   return false;
 397 }
 398 
 399 template<class E, MEMFLAGS F, unsigned int N>
 400 bool GenericTaskQueue<E, F, N>::pop_global(volatile E& t) {
 401   Age oldAge = _age.get();
 402   // Architectures with weak memory model require a barrier here
 403   // to guarantee that bottom is not older than age,
 404   // which is crucial for the correctness of the algorithm.
 405 #if !(defined SPARC || defined IA32 || defined AMD64)
 406   OrderAccess::fence();
 407 #endif
 408   uint localBot = OrderAccess::load_acquire((volatile juint*)&_bottom);
 409   uint n_elems = size(localBot, oldAge.top());
 410   if (n_elems == 0) {
 411     return false;
 412   }
 413 
 414   // g++ complains if the volatile result of the assignment is
 415   // unused, so we cast the volatile away.  We cannot cast directly
 416   // to void, because gcc treats that as not using the result of the
 417   // assignment.  However, casting to E& means that we trigger an
 418   // unused-value warning.  So, we cast the E& to void.
 419   (void) const_cast<E&>(t = _elems[oldAge.top()]);
 420   Age newAge(oldAge);
 421   newAge.increment();
 422   Age resAge = _age.cmpxchg(newAge, oldAge);
 423 
 424   // Note that using "_bottom" here might fail, since a pop_local might
 425   // have decremented it.
 426   assert(dirty_size(localBot, newAge.top()) != N - 1, "sanity");
 427   return resAge == oldAge;
 428 }
 429 
 430 template<class E, MEMFLAGS F, unsigned int N>
 431 GenericTaskQueue<E, F, N>::~GenericTaskQueue() {
 432   FREE_C_HEAP_ARRAY(E, _elems, F);
 433 }
 434 
 435 // OverflowTaskQueue is a TaskQueue that also includes an overflow stack for
 436 // elements that do not fit in the TaskQueue.
 437 //
 438 // This class hides two methods from super classes:
 439 //
 440 // push() - push onto the task queue or, if that fails, onto the overflow stack
 441 // is_empty() - return true if both the TaskQueue and overflow stack are empty
 442 //
 443 // Note that size() is not hidden--it returns the number of elements in the
 444 // TaskQueue, and does not include the size of the overflow stack.  This
 445 // simplifies replacement of GenericTaskQueues with OverflowTaskQueues.
 446 template<class E, MEMFLAGS F, unsigned int N = TASKQUEUE_SIZE>
 447 class OverflowTaskQueue: public GenericTaskQueue<E, F, N>
 448 {
 449 public:
 450   typedef Stack<E, F>               overflow_t;
 451   typedef GenericTaskQueue<E, F, N> taskqueue_t;
 452 
 453   TASKQUEUE_STATS_ONLY(using taskqueue_t::stats;)
 454 
 455   // Push task t onto the queue or onto the overflow stack.  Return true.
 456   inline bool push(E t);
 457 
 458   // Try to push task t onto the queue only. Returns true if successful, false otherwise.
 459   inline bool try_push_to_taskqueue(E t);
 460 
 461   // Attempt to pop from the overflow stack; return true if anything was popped.
 462   inline bool pop_overflow(E& t);
 463 
 464   inline overflow_t* overflow_stack() { return &_overflow_stack; }
 465 
 466   inline bool taskqueue_empty() const { return taskqueue_t::is_empty(); }
 467   inline bool overflow_empty()  const { return _overflow_stack.is_empty(); }
 468   inline bool is_empty()        const {
 469     return taskqueue_empty() && overflow_empty();
 470   }
 471 
 472 private:
 473   overflow_t _overflow_stack;
 474 };
 475 
 476 template <class E, MEMFLAGS F, unsigned int N>
 477 bool OverflowTaskQueue<E, F, N>::push(E t)
 478 {
 479   if (!taskqueue_t::push(t)) {
 480     overflow_stack()->push(t);
 481     TASKQUEUE_STATS_ONLY(stats.record_overflow(overflow_stack()->size()));
 482   }
 483   return true;
 484 }
 485 
 486 template <class E, MEMFLAGS F, unsigned int N>
 487 bool OverflowTaskQueue<E, F, N>::pop_overflow(E& t)
 488 {
 489   if (overflow_empty()) return false;
 490   t = overflow_stack()->pop();
 491   return true;
 492 }
 493 
 494 template <class E, MEMFLAGS F, unsigned int N>
 495 bool OverflowTaskQueue<E, F, N>::try_push_to_taskqueue(E t) {
 496   return taskqueue_t::push(t);
 497 }
 498 class TaskQueueSetSuper {
 499 protected:
 500   static int randomParkAndMiller(int* seed0);
 501 public:
 502   // Returns "true" if some TaskQueue in the set contains a task.
 503   virtual bool peek() = 0;
 504 };
 505 
 506 template <MEMFLAGS F> class TaskQueueSetSuperImpl: public CHeapObj<F>, public TaskQueueSetSuper {
 507 };
 508 
 509 template<class T, MEMFLAGS F>
 510 class GenericTaskQueueSet: public TaskQueueSetSuperImpl<F> {
 511 private:
 512   uint _n;
 513   T** _queues;
 514 
 515 public:
 516   typedef typename T::element_type E;
 517 
 518   GenericTaskQueueSet(int n) : _n(n) {
 519     typedef T* GenericTaskQueuePtr;
 520     _queues = NEW_C_HEAP_ARRAY(GenericTaskQueuePtr, n, F);
 521     for (int i = 0; i < n; i++) {
 522       _queues[i] = NULL;
 523     }
 524   }
 525 
 526   bool steal_best_of_2(uint queue_num, int* seed, E& t);
 527 
 528   void register_queue(uint i, T* q);
 529 
 530   T* queue(uint n);
 531 
 532   // The thread with queue number "queue_num" (and whose random number seed is
 533   // at "seed") is trying to steal a task from some other queue.  (It may try
 534   // several queues, according to some configuration parameter.)  If some steal
 535   // succeeds, returns "true" and sets "t" to the stolen task, otherwise returns
 536   // false.
 537   bool steal(uint queue_num, int* seed, E& t);
 538 
 539   bool peek();
 540 };
 541 
 542 template<class T, MEMFLAGS F> void
 543 GenericTaskQueueSet<T, F>::register_queue(uint i, T* q) {
 544   assert(i < _n, "index out of range.");
 545   _queues[i] = q;
 546 }
 547 
 548 template<class T, MEMFLAGS F> T*
 549 GenericTaskQueueSet<T, F>::queue(uint i) {
 550   return _queues[i];
 551 }
 552 
 553 template<class T, MEMFLAGS F> bool
 554 GenericTaskQueueSet<T, F>::steal(uint queue_num, int* seed, E& t) {
 555   for (uint i = 0; i < 2 * _n; i++) {
 556     if (steal_best_of_2(queue_num, seed, t)) {
 557       TASKQUEUE_STATS_ONLY(queue(queue_num)->stats.record_steal(true));
 558       return true;
 559     }
 560   }
 561   TASKQUEUE_STATS_ONLY(queue(queue_num)->stats.record_steal(false));
 562   return false;
 563 }
 564 
 565 template<class T, MEMFLAGS F> bool
 566 GenericTaskQueueSet<T, F>::steal_best_of_2(uint queue_num, int* seed, E& t) {
 567   if (_n > 2) {
 568     uint k1 = queue_num;
 569     while (k1 == queue_num) k1 = TaskQueueSetSuper::randomParkAndMiller(seed) % _n;
 570     uint k2 = queue_num;
 571     while (k2 == queue_num || k2 == k1) k2 = TaskQueueSetSuper::randomParkAndMiller(seed) % _n;
 572     // Sample both and try the larger.
 573     uint sz1 = _queues[k1]->size();
 574     uint sz2 = _queues[k2]->size();
 575     if (sz2 > sz1) return _queues[k2]->pop_global(t);
 576     else return _queues[k1]->pop_global(t);
 577   } else if (_n == 2) {
 578     // Just try the other one.
 579     uint k = (queue_num + 1) % 2;
 580     return _queues[k]->pop_global(t);
 581   } else {
 582     assert(_n == 1, "can't be zero.");
 583     return false;
 584   }
 585 }
 586 
 587 template<class T, MEMFLAGS F>
 588 bool GenericTaskQueueSet<T, F>::peek() {
 589   // Try all the queues.
 590   for (uint j = 0; j < _n; j++) {
 591     if (_queues[j]->peek())
 592       return true;
 593   }
 594   return false;
 595 }
 596 
 597 // When to terminate from the termination protocol.
 598 class TerminatorTerminator: public CHeapObj<mtInternal> {
 599 public:
 600   virtual bool should_exit_termination() = 0;
 601 };
 602 
 603 // A class to aid in the termination of a set of parallel tasks using
 604 // TaskQueueSet's for work stealing.
 605 
 606 #undef TRACESPINNING
 607 
 608 class ParallelTaskTerminator: public StackObj {
 609 private:
 610   int _n_threads;
 611   TaskQueueSetSuper* _queue_set;
 612   char _pad_before[DEFAULT_CACHE_LINE_SIZE];
 613   int _offered_termination;
 614   char _pad_after[DEFAULT_CACHE_LINE_SIZE];
 615 
 616 #ifdef TRACESPINNING
 617   static uint _total_yields;
 618   static uint _total_spins;
 619   static uint _total_peeks;
 620 #endif
 621 
 622   bool peek_in_queue_set();
 623 protected:
 624   virtual void yield();
 625   void sleep(uint millis);
 626 
 627 public:
 628 
 629   // "n_threads" is the number of threads to be terminated.  "queue_set" is a
 630   // queue sets of work queues of other threads.
 631   ParallelTaskTerminator(int n_threads, TaskQueueSetSuper* queue_set);
 632 
 633   // The current thread has no work, and is ready to terminate if everyone
 634   // else is.  If returns "true", all threads are terminated.  If returns
 635   // "false", available work has been observed in one of the task queues,
 636   // so the global task is not complete.
 637   bool offer_termination() {
 638     return offer_termination(NULL);
 639   }
 640 
 641   // As above, but it also terminates if the should_exit_termination()
 642   // method of the terminator parameter returns true. If terminator is
 643   // NULL, then it is ignored.
 644   bool offer_termination(TerminatorTerminator* terminator);
 645 
 646   // Reset the terminator, so that it may be reused again.
 647   // The caller is responsible for ensuring that this is done
 648   // in an MT-safe manner, once the previous round of use of
 649   // the terminator is finished.
 650   void reset_for_reuse();
 651   // Same as above but the number of parallel threads is set to the
 652   // given number.
 653   void reset_for_reuse(int n_threads);
 654 
 655 #ifdef TRACESPINNING
 656   static uint total_yields() { return _total_yields; }
 657   static uint total_spins() { return _total_spins; }
 658   static uint total_peeks() { return _total_peeks; }
 659   static void print_termination_counts();
 660 #endif
 661 };
 662 
 663 template<class E, MEMFLAGS F, unsigned int N> inline bool
 664 GenericTaskQueue<E, F, N>::push(E t) {
 665   uint localBot = _bottom;
 666   assert(localBot < N, "_bottom out of range.");
 667   idx_t top = _age.top();
 668   uint dirty_n_elems = dirty_size(localBot, top);
 669   assert(dirty_n_elems < N, "n_elems out of range.");
 670   if (dirty_n_elems < max_elems()) {
 671     // g++ complains if the volatile result of the assignment is
 672     // unused, so we cast the volatile away.  We cannot cast directly
 673     // to void, because gcc treats that as not using the result of the
 674     // assignment.  However, casting to E& means that we trigger an
 675     // unused-value warning.  So, we cast the E& to void.
 676     (void) const_cast<E&>(_elems[localBot] = t);
 677     OrderAccess::release_store(&_bottom, increment_index(localBot));
 678     TASKQUEUE_STATS_ONLY(stats.record_push());
 679     return true;
 680   } else {
 681     return push_slow(t, dirty_n_elems);
 682   }
 683 }
 684 
 685 template<class E, MEMFLAGS F, unsigned int N> inline bool
 686 GenericTaskQueue<E, F, N>::pop_local(volatile E& t) {
 687   uint localBot = _bottom;
 688   // This value cannot be N-1.  That can only occur as a result of
 689   // the assignment to bottom in this method.  If it does, this method
 690   // resets the size to 0 before the next call (which is sequential,
 691   // since this is pop_local.)
 692   uint dirty_n_elems = dirty_size(localBot, _age.top());
 693   assert(dirty_n_elems != N - 1, "Shouldn't be possible...");
 694   if (dirty_n_elems == 0) return false;
 695   localBot = decrement_index(localBot);
 696   _bottom = localBot;
 697   // This is necessary to prevent any read below from being reordered
 698   // before the store just above.
 699   OrderAccess::fence();
 700   // g++ complains if the volatile result of the assignment is
 701   // unused, so we cast the volatile away.  We cannot cast directly
 702   // to void, because gcc treats that as not using the result of the
 703   // assignment.  However, casting to E& means that we trigger an
 704   // unused-value warning.  So, we cast the E& to void.
 705   (void) const_cast<E&>(t = _elems[localBot]);
 706   // This is a second read of "age"; the "size()" above is the first.
 707   // If there's still at least one element in the queue, based on the
 708   // "_bottom" and "age" we've read, then there can be no interference with
 709   // a "pop_global" operation, and we're done.
 710   idx_t tp = _age.top();    // XXX
 711   if (size(localBot, tp) > 0) {
 712     assert(dirty_size(localBot, tp) != N - 1, "sanity");
 713     TASKQUEUE_STATS_ONLY(stats.record_pop());
 714     return true;
 715   } else {
 716     // Otherwise, the queue contained exactly one element; we take the slow
 717     // path.
 718 
 719     // The barrier is required to prevent reordering the two reads of _age:
 720     // one is the _age.get() below, and the other is _age.top() above the if-stmt.
 721     // The algorithm may fail if _age.get() reads an older value than _age.top().
 722     OrderAccess::loadload();
 723     return pop_local_slow(localBot, _age.get());
 724   }
 725 }
 726 
 727 typedef GenericTaskQueue<oop, mtGC>             OopTaskQueue;
 728 typedef GenericTaskQueueSet<OopTaskQueue, mtGC> OopTaskQueueSet;
 729 
 730 #ifdef _MSC_VER
 731 #pragma warning(push)
 732 // warning C4522: multiple assignment operators specified
 733 #pragma warning(disable:4522)
 734 #endif
 735 
 736 // This is a container class for either an oop* or a narrowOop*.
 737 // Both are pushed onto a task queue and the consumer will test is_narrow()
 738 // to determine which should be processed.
 739 class StarTask {
 740   void*  _holder;        // either union oop* or narrowOop*
 741 
 742   enum { COMPRESSED_OOP_MASK = 1 };
 743 
 744  public:
 745   StarTask(narrowOop* p) {
 746     assert(((uintptr_t)p & COMPRESSED_OOP_MASK) == 0, "Information loss!");
 747     _holder = (void *)((uintptr_t)p | COMPRESSED_OOP_MASK);
 748   }
 749   StarTask(oop* p)       {
 750     assert(((uintptr_t)p & COMPRESSED_OOP_MASK) == 0, "Information loss!");
 751     _holder = (void*)p;
 752   }
 753   StarTask()             { _holder = NULL; }
 754   operator oop*()        { return (oop*)_holder; }
 755   operator narrowOop*()  {
 756     return (narrowOop*)((uintptr_t)_holder & ~COMPRESSED_OOP_MASK);
 757   }
 758 
 759   StarTask& operator=(const StarTask& t) {
 760     _holder = t._holder;
 761     return *this;
 762   }
 763   volatile StarTask& operator=(const volatile StarTask& t) volatile {
 764     _holder = t._holder;
 765     return *this;
 766   }
 767 
 768   bool is_narrow() const {
 769     return (((uintptr_t)_holder & COMPRESSED_OOP_MASK) != 0);
 770   }
 771 };
 772 
 773 class ObjArrayTask
 774 {
 775 public:
 776   ObjArrayTask(oop o = NULL, int idx = 0): _obj(o), _index(idx) { }
 777   ObjArrayTask(oop o, size_t idx): _obj(o), _index(int(idx)) {
 778     assert(idx <= size_t(max_jint), "too big");
 779   }
 780   ObjArrayTask(const ObjArrayTask& t): _obj(t._obj), _index(t._index) { }
 781 
 782   ObjArrayTask& operator =(const ObjArrayTask& t) {
 783     _obj = t._obj;
 784     _index = t._index;
 785     return *this;
 786   }
 787   volatile ObjArrayTask&
 788   operator =(const volatile ObjArrayTask& t) volatile {
 789     (void)const_cast<oop&>(_obj = t._obj);
 790     _index = t._index;
 791     return *this;
 792   }
 793 
 794   inline oop obj()   const { return _obj; }
 795   inline int index() const { return _index; }
 796 
 797   DEBUG_ONLY(bool is_valid() const); // Tasks to be pushed/popped must be valid.
 798 
 799 private:
 800   oop _obj;
 801   int _index;
 802 };
 803 
 804 #ifdef _MSC_VER
 805 #pragma warning(pop)
 806 #endif
 807 
 808 typedef OverflowTaskQueue<StarTask, mtClass>           OopStarTaskQueue;
 809 typedef GenericTaskQueueSet<OopStarTaskQueue, mtClass> OopStarTaskQueueSet;
 810 
 811 typedef OverflowTaskQueue<size_t, mtInternal>             RegionTaskQueue;
 812 typedef GenericTaskQueueSet<RegionTaskQueue, mtClass>     RegionTaskQueueSet;
 813 
 814 
 815 #endif // SHARE_VM_UTILITIES_TASKQUEUE_HPP