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