1 /*
   2  * Copyright (c) 2001, 2011, 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 "utilities/stack.hpp"
  32 #ifdef TARGET_OS_ARCH_linux_x86
  33 # include "orderAccess_linux_x86.inline.hpp"
  34 #endif
  35 #ifdef TARGET_OS_ARCH_linux_sparc
  36 # include "orderAccess_linux_sparc.inline.hpp"
  37 #endif
  38 #ifdef TARGET_OS_ARCH_linux_zero
  39 # include "orderAccess_linux_zero.inline.hpp"
  40 #endif
  41 #ifdef TARGET_OS_ARCH_solaris_x86
  42 # include "orderAccess_solaris_x86.inline.hpp"
  43 #endif
  44 #ifdef TARGET_OS_ARCH_solaris_sparc
  45 # include "orderAccess_solaris_sparc.inline.hpp"
  46 #endif
  47 #ifdef TARGET_OS_ARCH_windows_x86
  48 # include "orderAccess_windows_x86.inline.hpp"
  49 #endif
  50 #ifdef TARGET_OS_ARCH_linux_arm
  51 # include "orderAccess_linux_arm.inline.hpp"
  52 #endif
  53 #ifdef TARGET_OS_ARCH_linux_ppc
  54 # include "orderAccess_linux_ppc.inline.hpp"
  55 #endif
  56 #ifdef TARGET_OS_ARCH_bsd_x86
  57 # include "orderAccess_bsd_x86.inline.hpp"
  58 #endif
  59 #ifdef TARGET_OS_ARCH_bsd_zero
  60 # include "orderAccess_bsd_zero.inline.hpp"
  61 #endif
  62 
  63 // Simple TaskQueue stats that are collected by default in debug builds.
  64 
  65 #if !defined(TASKQUEUE_STATS) && defined(ASSERT)
  66 #define TASKQUEUE_STATS 1
  67 #elif !defined(TASKQUEUE_STATS)
  68 #define TASKQUEUE_STATS 0
  69 #endif
  70 
  71 #if TASKQUEUE_STATS
  72 #define TASKQUEUE_STATS_ONLY(code) code
  73 #else
  74 #define TASKQUEUE_STATS_ONLY(code)
  75 #endif // TASKQUEUE_STATS
  76 
  77 #if TASKQUEUE_STATS
  78 class TaskQueueStats {
  79 public:
  80   enum StatId {
  81     push,             // number of taskqueue pushes
  82     pop,              // number of taskqueue pops
  83     pop_slow,         // subset of taskqueue pops that were done slow-path
  84     steal_attempt,    // number of taskqueue steal attempts
  85     steal,            // number of taskqueue steals
  86     overflow,         // number of overflow pushes
  87     overflow_max_len, // max length of overflow stack
  88     last_stat_id
  89   };
  90 
  91 public:
  92   inline TaskQueueStats()       { reset(); }
  93 
  94   inline void record_push()     { ++_stats[push]; }
  95   inline void record_pop()      { ++_stats[pop]; }
  96   inline void record_pop_slow() { record_pop(); ++_stats[pop_slow]; }
  97   inline void record_steal(bool success);
  98   inline void record_overflow(size_t new_length);
  99 
 100   TaskQueueStats & operator +=(const TaskQueueStats & addend);
 101 
 102   inline size_t get(StatId id) const { return _stats[id]; }
 103   inline const size_t* get() const   { return _stats; }
 104 
 105   inline void reset();
 106 
 107   // Print the specified line of the header (does not include a line separator).
 108   static void print_header(unsigned int line, outputStream* const stream = tty,
 109                            unsigned int width = 10);
 110   // Print the statistics (does not include a line separator).
 111   void print(outputStream* const stream = tty, unsigned int width = 10) const;
 112 
 113   DEBUG_ONLY(void verify() const;)
 114 
 115 private:
 116   size_t                    _stats[last_stat_id];
 117   static const char * const _names[last_stat_id];
 118 };
 119 
 120 void TaskQueueStats::record_steal(bool success) {
 121   ++_stats[steal_attempt];
 122   if (success) ++_stats[steal];
 123 }
 124 
 125 void TaskQueueStats::record_overflow(size_t new_len) {
 126   ++_stats[overflow];
 127   if (new_len > _stats[overflow_max_len]) _stats[overflow_max_len] = new_len;
 128 }
 129 
 130 void TaskQueueStats::reset() {
 131   memset(_stats, 0, sizeof(_stats));
 132 }
 133 #endif // TASKQUEUE_STATS
 134 
 135 template <unsigned int N, MEMFLAGS F>
 136 class TaskQueueSuper: public CHeapObj<F> {
 137 protected:
 138   // Internal type for indexing the queue; also used for the tag.
 139   typedef NOT_LP64(uint16_t) LP64_ONLY(uint32_t) idx_t;
 140 
 141   // The first free element after the last one pushed (mod N).
 142   volatile uint _bottom;
 143 
 144   enum { MOD_N_MASK = N - 1 };
 145 
 146   class Age {
 147   public:
 148     Age(size_t data = 0)         { _data = data; }
 149     Age(const Age& age)          { _data = age._data; }
 150     Age(idx_t top, idx_t tag)    { _fields._top = top; _fields._tag = tag; }
 151 
 152     Age   get()        const volatile { return _data; }
 153     void  set(Age age) volatile       { _data = age._data; }
 154 
 155     idx_t top()        const volatile { return _fields._top; }
 156     idx_t tag()        const volatile { return _fields._tag; }
 157 
 158     // Increment top; if it wraps, increment tag also.
 159     void increment() {
 160       _fields._top = increment_index(_fields._top);
 161       if (_fields._top == 0) ++_fields._tag;
 162     }
 163 
 164     Age cmpxchg(const Age new_age, const Age old_age) volatile {
 165       return (size_t) Atomic::cmpxchg_ptr((intptr_t)new_age._data,
 166                                           (volatile intptr_t *)&_data,
 167                                           (intptr_t)old_age._data);
 168     }
 169 
 170     bool operator ==(const Age& other) const { return _data == other._data; }
 171 
 172   private:
 173     struct fields {
 174       idx_t _top;
 175       idx_t _tag;
 176     };
 177     union {
 178       size_t _data;
 179       fields _fields;
 180     };
 181   };
 182 
 183   volatile Age _age;
 184 
 185   // These both operate mod N.
 186   static uint increment_index(uint ind) {
 187     return (ind + 1) & MOD_N_MASK;
 188   }
 189   static uint decrement_index(uint ind) {
 190     return (ind - 1) & MOD_N_MASK;
 191   }
 192 
 193   // Returns a number in the range [0..N).  If the result is "N-1", it should be
 194   // interpreted as 0.
 195   uint dirty_size(uint bot, uint top) const {
 196     return (bot - top) & MOD_N_MASK;
 197   }
 198 
 199   // Returns the size corresponding to the given "bot" and "top".
 200   uint size(uint bot, uint top) const {
 201     uint sz = dirty_size(bot, top);
 202     // Has the queue "wrapped", so that bottom is less than top?  There's a
 203     // complicated special case here.  A pair of threads could perform pop_local
 204     // and pop_global operations concurrently, starting from a state in which
 205     // _bottom == _top+1.  The pop_local could succeed in decrementing _bottom,
 206     // and the pop_global in incrementing _top (in which case the pop_global
 207     // will be awarded the contested queue element.)  The resulting state must
 208     // be interpreted as an empty queue.  (We only need to worry about one such
 209     // event: only the queue owner performs pop_local's, and several concurrent
 210     // threads attempting to perform the pop_global will all perform the same
 211     // CAS, and only one can succeed.)  Any stealing thread that reads after
 212     // either the increment or decrement will see an empty queue, and will not
 213     // join the competitors.  The "sz == -1 || sz == N-1" state will not be
 214     // modified by concurrent queues, so the owner thread can reset the state to
 215     // _bottom == top so subsequent pushes will be performed normally.
 216     return (sz == N - 1) ? 0 : sz;
 217   }
 218 
 219 public:
 220   TaskQueueSuper() : _bottom(0), _age() {}
 221 
 222   // Return true if the TaskQueue contains/does not contain any tasks.
 223   bool peek()     const { return _bottom != _age.top(); }
 224   bool is_empty() const { return size() == 0; }
 225 
 226   // Return an estimate of the number of elements in the queue.
 227   // The "careful" version admits the possibility of pop_local/pop_global
 228   // races.
 229   uint size() const {
 230     return size(_bottom, _age.top());
 231   }
 232 
 233   uint dirty_size() const {
 234     return dirty_size(_bottom, _age.top());
 235   }
 236 
 237   void set_empty() {
 238     _bottom = 0;
 239     _age.set(0);
 240   }
 241 
 242   // Maximum number of elements allowed in the queue.  This is two less
 243   // than the actual queue size, for somewhat complicated reasons.
 244   uint max_elems() const { return N - 2; }
 245 
 246   // Total size of queue.
 247   static const uint total_size() { return N; }
 248 
 249   TASKQUEUE_STATS_ONLY(TaskQueueStats stats;)
 250 };
 251 
 252 
 253 
 254 template <class E, MEMFLAGS F, unsigned int N = TASKQUEUE_SIZE>
 255 class GenericTaskQueue: public TaskQueueSuper<N, F> {
 256   ArrayAllocator<E, F> _array_allocator;
 257 protected:
 258   typedef typename TaskQueueSuper<N, F>::Age Age;
 259   typedef typename TaskQueueSuper<N, F>::idx_t idx_t;
 260 
 261   using TaskQueueSuper<N, F>::_bottom;
 262   using TaskQueueSuper<N, F>::_age;
 263   using TaskQueueSuper<N, F>::increment_index;
 264   using TaskQueueSuper<N, F>::decrement_index;
 265   using TaskQueueSuper<N, F>::dirty_size;
 266 
 267 public:
 268   using TaskQueueSuper<N, F>::max_elems;
 269   using TaskQueueSuper<N, F>::size;
 270 
 271 #if  TASKQUEUE_STATS
 272   using TaskQueueSuper<N, F>::stats;
 273 #endif
 274 
 275 private:
 276   // Slow paths for push, pop_local.  (pop_global has no fast path.)
 277   bool push_slow(E t, uint dirty_n_elems);
 278   bool pop_local_slow(uint localBot, Age oldAge);
 279 
 280 public:
 281   typedef E element_type;
 282 
 283   // Initializes the queue to empty.
 284   GenericTaskQueue();
 285 
 286   void initialize();
 287 
 288   // Push the task "t" on the queue.  Returns "false" iff the queue is full.
 289   inline bool push(E t);
 290 
 291   // Attempts to claim a task from the "local" end of the queue (the most
 292   // recently pushed).  If successful, returns true and sets t to the task;
 293   // otherwise, returns false (the queue is empty).
 294   inline bool pop_local(E& t);
 295 
 296   // Like pop_local(), but uses the "global" end of the queue (the least
 297   // recently pushed).
 298   bool pop_global(E& t);
 299 
 300   // Delete any resource associated with the queue.
 301   ~GenericTaskQueue();
 302 
 303   // apply the closure to all elements in the task queue
 304   void oops_do(OopClosure* f);
 305 
 306 private:
 307   // Element array.
 308   volatile E* _elems;
 309 };
 310 
 311 template<class E, MEMFLAGS F, unsigned int N>
 312 GenericTaskQueue<E, F, N>::GenericTaskQueue() {
 313   assert(sizeof(Age) == sizeof(size_t), "Depends on this.");
 314 }
 315 
 316 template<class E, MEMFLAGS F, unsigned int N>
 317 void GenericTaskQueue<E, F, N>::initialize() {
 318   _elems = _array_allocator.allocate(N);
 319 }
 320 
 321 template<class E, MEMFLAGS F, unsigned int N>
 322 void GenericTaskQueue<E, F, N>::oops_do(OopClosure* f) {
 323   // tty->print_cr("START OopTaskQueue::oops_do");
 324   uint iters = size();
 325   uint index = _bottom;
 326   for (uint i = 0; i < iters; ++i) {
 327     index = decrement_index(index);
 328     // tty->print_cr("  doing entry %d," INTPTR_T " -> " INTPTR_T,
 329     //            index, &_elems[index], _elems[index]);
 330     E* t = (E*)&_elems[index];      // cast away volatility
 331     oop* p = (oop*)t;
 332     assert((*t)->is_oop_or_null(), "Not an oop or null");
 333     f->do_oop(p);
 334   }
 335   // tty->print_cr("END OopTaskQueue::oops_do");
 336 }
 337 
 338 template<class E, MEMFLAGS F, unsigned int N>
 339 bool GenericTaskQueue<E, F, N>::push_slow(E t, uint dirty_n_elems) {
 340   if (dirty_n_elems == N - 1) {
 341     // Actually means 0, so do the push.
 342     uint localBot = _bottom;
 343     // g++ complains if the volatile result of the assignment is unused.
 344     const_cast<E&>(_elems[localBot] = t);
 345     OrderAccess::release_store(&_bottom, increment_index(localBot));
 346     TASKQUEUE_STATS_ONLY(stats.record_push());
 347     return true;
 348   }
 349   return false;
 350 }
 351 
 352 // pop_local_slow() is done by the owning thread and is trying to
 353 // get the last task in the queue.  It will compete with pop_global()
 354 // that will be used by other threads.  The tag age is incremented
 355 // whenever the queue goes empty which it will do here if this thread
 356 // gets the last task or in pop_global() if the queue wraps (top == 0
 357 // and pop_global() succeeds, see pop_global()).
 358 template<class E, MEMFLAGS F, unsigned int N>
 359 bool GenericTaskQueue<E, F, N>::pop_local_slow(uint localBot, Age oldAge) {
 360   // This queue was observed to contain exactly one element; either this
 361   // thread will claim it, or a competing "pop_global".  In either case,
 362   // the queue will be logically empty afterwards.  Create a new Age value
 363   // that represents the empty queue for the given value of "_bottom".  (We
 364   // must also increment "tag" because of the case where "bottom == 1",
 365   // "top == 0".  A pop_global could read the queue element in that case,
 366   // then have the owner thread do a pop followed by another push.  Without
 367   // the incrementing of "tag", the pop_global's CAS could succeed,
 368   // allowing it to believe it has claimed the stale element.)
 369   Age newAge((idx_t)localBot, oldAge.tag() + 1);
 370   // Perhaps a competing pop_global has already incremented "top", in which
 371   // case it wins the element.
 372   if (localBot == oldAge.top()) {
 373     // No competing pop_global has yet incremented "top"; we'll try to
 374     // install new_age, thus claiming the element.
 375     Age tempAge = _age.cmpxchg(newAge, oldAge);
 376     if (tempAge == oldAge) {
 377       // We win.
 378       assert(dirty_size(localBot, _age.top()) != N - 1, "sanity");
 379       TASKQUEUE_STATS_ONLY(stats.record_pop_slow());
 380       return true;
 381     }
 382   }
 383   // We lose; a completing pop_global gets the element.  But the queue is empty
 384   // and top is greater than bottom.  Fix this representation of the empty queue
 385   // to become the canonical one.
 386   _age.set(newAge);
 387   assert(dirty_size(localBot, _age.top()) != N - 1, "sanity");
 388   return false;
 389 }
 390 
 391 template<class E, MEMFLAGS F, unsigned int N>
 392 bool GenericTaskQueue<E, F, N>::pop_global(E& t) {
 393   Age oldAge = _age.get();
 394   // Architectures with weak memory model require a fence here. The
 395   // fence has a cumulative effect on getting age and getting bottom.
 396   // This way it is guaranteed that bottom is not older than age,
 397   // which is crucial for the correctness of the algorithm.
 398 #if !(defined SPARC || defined IA32 || defined AMD64)
 399   OrderAccess::fence();
 400 #endif
 401   uint localBot = OrderAccess::load_acquire((volatile juint*)&_bottom);
 402   uint n_elems = size(localBot, oldAge.top());
 403   if (n_elems == 0) {
 404     return false;
 405   }
 406 
 407   const_cast<E&>(t = _elems[oldAge.top()]);
 408   Age newAge(oldAge);
 409   newAge.increment();
 410   Age resAge = _age.cmpxchg(newAge, oldAge);
 411 
 412   // Note that using "_bottom" here might fail, since a pop_local might
 413   // have decremented it.
 414   assert(dirty_size(localBot, newAge.top()) != N - 1, "sanity");
 415   return resAge == oldAge;
 416 }
 417 
 418 template<class E, MEMFLAGS F, unsigned int N>
 419 GenericTaskQueue<E, F, N>::~GenericTaskQueue() {
 420   FREE_C_HEAP_ARRAY(E, _elems, F);
 421 }
 422 
 423 // OverflowTaskQueue is a TaskQueue that also includes an overflow stack for
 424 // elements that do not fit in the TaskQueue.
 425 //
 426 // This class hides two methods from super classes:
 427 //
 428 // push() - push onto the task queue or, if that fails, onto the overflow stack
 429 // is_empty() - return true if both the TaskQueue and overflow stack are empty
 430 //
 431 // Note that size() is not hidden--it returns the number of elements in the
 432 // TaskQueue, and does not include the size of the overflow stack.  This
 433 // simplifies replacement of GenericTaskQueues with OverflowTaskQueues.
 434 template<class E, MEMFLAGS F, unsigned int N = TASKQUEUE_SIZE>
 435 class OverflowTaskQueue: public GenericTaskQueue<E, F, N>
 436 {
 437 public:
 438   typedef Stack<E, F>               overflow_t;
 439   typedef GenericTaskQueue<E, F, N> taskqueue_t;
 440 
 441   TASKQUEUE_STATS_ONLY(using taskqueue_t::stats;)
 442 
 443   // Push task t onto the queue or onto the overflow stack.  Return true.
 444   inline bool push(E t);
 445 
 446   // Attempt to pop from the overflow stack; return true if anything was popped.
 447   inline bool pop_overflow(E& t);
 448 
 449   inline overflow_t* overflow_stack() { return &_overflow_stack; }
 450 
 451   inline bool taskqueue_empty() const { return taskqueue_t::is_empty(); }
 452   inline bool overflow_empty()  const { return _overflow_stack.is_empty(); }
 453   inline bool is_empty()        const {
 454     return taskqueue_empty() && overflow_empty();
 455   }
 456 
 457 private:
 458   overflow_t _overflow_stack;
 459 };
 460 
 461 template <class E, MEMFLAGS F, unsigned int N>
 462 bool OverflowTaskQueue<E, F, N>::push(E t)
 463 {
 464   if (!taskqueue_t::push(t)) {
 465     overflow_stack()->push(t);
 466     TASKQUEUE_STATS_ONLY(stats.record_overflow(overflow_stack()->size()));
 467   }
 468   return true;
 469 }
 470 
 471 template <class E, MEMFLAGS F, unsigned int N>
 472 bool OverflowTaskQueue<E, F, N>::pop_overflow(E& t)
 473 {
 474   if (overflow_empty()) return false;
 475   t = overflow_stack()->pop();
 476   return true;
 477 }
 478 
 479 class TaskQueueSetSuper {
 480 protected:
 481   static int randomParkAndMiller(int* seed0);
 482 public:
 483   // Returns "true" if some TaskQueue in the set contains a task.
 484   virtual bool peek() = 0;
 485 };
 486 
 487 template <MEMFLAGS F> class TaskQueueSetSuperImpl: public CHeapObj<F>, public TaskQueueSetSuper {
 488 };
 489 
 490 template<class T, MEMFLAGS F>
 491 class GenericTaskQueueSet: public TaskQueueSetSuperImpl<F> {
 492 private:
 493   uint _n;
 494   T** _queues;
 495 
 496 public:
 497   typedef typename T::element_type E;
 498 
 499   GenericTaskQueueSet(int n) : _n(n) {
 500     typedef T* GenericTaskQueuePtr;
 501     _queues = NEW_C_HEAP_ARRAY(GenericTaskQueuePtr, n, F);
 502     for (int i = 0; i < n; i++) {
 503       _queues[i] = NULL;
 504     }
 505   }
 506 
 507   bool steal_1_random(uint queue_num, int* seed, E& t);
 508   bool steal_best_of_2(uint queue_num, int* seed, E& t);
 509   bool steal_best_of_all(uint queue_num, int* seed, E& t);
 510 
 511   void register_queue(uint i, T* q);
 512 
 513   T* queue(uint n);
 514 
 515   // The thread with queue number "queue_num" (and whose random number seed is
 516   // at "seed") is trying to steal a task from some other queue.  (It may try
 517   // several queues, according to some configuration parameter.)  If some steal
 518   // succeeds, returns "true" and sets "t" to the stolen task, otherwise returns
 519   // false.
 520   bool steal(uint queue_num, int* seed, E& t);
 521 
 522   bool peek();
 523 };
 524 
 525 template<class T, MEMFLAGS F> void
 526 GenericTaskQueueSet<T, F>::register_queue(uint i, T* q) {
 527   assert(i < _n, "index out of range.");
 528   _queues[i] = q;
 529 }
 530 
 531 template<class T, MEMFLAGS F> T*
 532 GenericTaskQueueSet<T, F>::queue(uint i) {
 533   return _queues[i];
 534 }
 535 
 536 template<class T, MEMFLAGS F> bool
 537 GenericTaskQueueSet<T, F>::steal(uint queue_num, int* seed, E& t) {
 538   for (uint i = 0; i < 2 * _n; i++) {
 539     if (steal_best_of_2(queue_num, seed, t)) {
 540       TASKQUEUE_STATS_ONLY(queue(queue_num)->stats.record_steal(true));
 541       return true;
 542     }
 543   }
 544   TASKQUEUE_STATS_ONLY(queue(queue_num)->stats.record_steal(false));
 545   return false;
 546 }
 547 
 548 template<class T, MEMFLAGS F> bool
 549 GenericTaskQueueSet<T, F>::steal_best_of_all(uint queue_num, int* seed, E& t) {
 550   if (_n > 2) {
 551     int best_k;
 552     uint best_sz = 0;
 553     for (uint k = 0; k < _n; k++) {
 554       if (k == queue_num) continue;
 555       uint sz = _queues[k]->size();
 556       if (sz > best_sz) {
 557         best_sz = sz;
 558         best_k = k;
 559       }
 560     }
 561     return best_sz > 0 && _queues[best_k]->pop_global(t);
 562   } else if (_n == 2) {
 563     // Just try the other one.
 564     int k = (queue_num + 1) % 2;
 565     return _queues[k]->pop_global(t);
 566   } else {
 567     assert(_n == 1, "can't be zero.");
 568     return false;
 569   }
 570 }
 571 
 572 template<class T, MEMFLAGS F> bool
 573 GenericTaskQueueSet<T, F>::steal_1_random(uint queue_num, int* seed, E& t) {
 574   if (_n > 2) {
 575     uint k = queue_num;
 576     while (k == queue_num) k = TaskQueueSetSuper::randomParkAndMiller(seed) % _n;
 577     return _queues[2]->pop_global(t);
 578   } else if (_n == 2) {
 579     // Just try the other one.
 580     int k = (queue_num + 1) % 2;
 581     return _queues[k]->pop_global(t);
 582   } else {
 583     assert(_n == 1, "can't be zero.");
 584     return false;
 585   }
 586 }
 587 
 588 template<class T, MEMFLAGS F> bool
 589 GenericTaskQueueSet<T, F>::steal_best_of_2(uint queue_num, int* seed, E& t) {
 590   if (_n > 2) {
 591     uint k1 = queue_num;
 592     while (k1 == queue_num) k1 = TaskQueueSetSuper::randomParkAndMiller(seed) % _n;
 593     uint k2 = queue_num;
 594     while (k2 == queue_num || k2 == k1) k2 = TaskQueueSetSuper::randomParkAndMiller(seed) % _n;
 595     // Sample both and try the larger.
 596     uint sz1 = _queues[k1]->size();
 597     uint sz2 = _queues[k2]->size();
 598     if (sz2 > sz1) return _queues[k2]->pop_global(t);
 599     else return _queues[k1]->pop_global(t);
 600   } else if (_n == 2) {
 601     // Just try the other one.
 602     uint k = (queue_num + 1) % 2;
 603     return _queues[k]->pop_global(t);
 604   } else {
 605     assert(_n == 1, "can't be zero.");
 606     return false;
 607   }
 608 }
 609 
 610 template<class T, MEMFLAGS F>
 611 bool GenericTaskQueueSet<T, F>::peek() {
 612   // Try all the queues.
 613   for (uint j = 0; j < _n; j++) {
 614     if (_queues[j]->peek())
 615       return true;
 616   }
 617   return false;
 618 }
 619 
 620 // When to terminate from the termination protocol.
 621 class TerminatorTerminator: public CHeapObj<mtInternal> {
 622 public:
 623   virtual bool should_exit_termination() = 0;
 624 };
 625 
 626 // A class to aid in the termination of a set of parallel tasks using
 627 // TaskQueueSet's for work stealing.
 628 
 629 #undef TRACESPINNING
 630 
 631 class ParallelTaskTerminator: public StackObj {
 632 private:
 633   int _n_threads;
 634   TaskQueueSetSuper* _queue_set;
 635   int _offered_termination;
 636 
 637 #ifdef TRACESPINNING
 638   static uint _total_yields;
 639   static uint _total_spins;
 640   static uint _total_peeks;
 641 #endif
 642 
 643   bool peek_in_queue_set();
 644 protected:
 645   virtual void yield();
 646   void sleep(uint millis);
 647 
 648 public:
 649 
 650   // "n_threads" is the number of threads to be terminated.  "queue_set" is a
 651   // queue sets of work queues of other threads.
 652   ParallelTaskTerminator(int n_threads, TaskQueueSetSuper* queue_set);
 653 
 654   // The current thread has no work, and is ready to terminate if everyone
 655   // else is.  If returns "true", all threads are terminated.  If returns
 656   // "false", available work has been observed in one of the task queues,
 657   // so the global task is not complete.
 658   bool offer_termination() {
 659     return offer_termination(NULL);
 660   }
 661 
 662   // As above, but it also terminates if the should_exit_termination()
 663   // method of the terminator parameter returns true. If terminator is
 664   // NULL, then it is ignored.
 665   bool offer_termination(TerminatorTerminator* terminator);
 666 
 667   // Reset the terminator, so that it may be reused again.
 668   // The caller is responsible for ensuring that this is done
 669   // in an MT-safe manner, once the previous round of use of
 670   // the terminator is finished.
 671   void reset_for_reuse();
 672   // Same as above but the number of parallel threads is set to the
 673   // given number.
 674   void reset_for_reuse(int n_threads);
 675 
 676 #ifdef TRACESPINNING
 677   static uint total_yields() { return _total_yields; }
 678   static uint total_spins() { return _total_spins; }
 679   static uint total_peeks() { return _total_peeks; }
 680   static void print_termination_counts();
 681 #endif
 682 };
 683 
 684 template<class E, MEMFLAGS F, unsigned int N> inline bool
 685 GenericTaskQueue<E, F, N>::push(E t) {
 686   uint localBot = _bottom;
 687   assert(localBot < N, "_bottom out of range.");
 688   idx_t top = _age.top();
 689   uint dirty_n_elems = dirty_size(localBot, top);
 690   assert(dirty_n_elems < N, "n_elems out of range.");
 691   if (dirty_n_elems < max_elems()) {
 692     // g++ complains if the volatile result of the assignment is unused.
 693     const_cast<E&>(_elems[localBot] = t);
 694     OrderAccess::release_store(&_bottom, increment_index(localBot));
 695     TASKQUEUE_STATS_ONLY(stats.record_push());
 696     return true;
 697   } else {
 698     return push_slow(t, dirty_n_elems);
 699   }
 700 }
 701 
 702 template<class E, MEMFLAGS F, unsigned int N> inline bool
 703 GenericTaskQueue<E, F, N>::pop_local(E& t) {
 704   uint localBot = _bottom;
 705   // This value cannot be N-1.  That can only occur as a result of
 706   // the assignment to bottom in this method.  If it does, this method
 707   // resets the size to 0 before the next call (which is sequential,
 708   // since this is pop_local.)
 709   uint dirty_n_elems = dirty_size(localBot, _age.top());
 710   assert(dirty_n_elems != N - 1, "Shouldn't be possible...");
 711   if (dirty_n_elems == 0) return false;
 712   localBot = decrement_index(localBot);
 713   _bottom = localBot;
 714   // This is necessary to prevent any read below from being reordered
 715   // before the store just above.
 716   OrderAccess::fence();
 717   const_cast<E&>(t = _elems[localBot]);
 718   // This is a second read of "age"; the "size()" above is the first.
 719   // If there's still at least one element in the queue, based on the
 720   // "_bottom" and "age" we've read, then there can be no interference with
 721   // a "pop_global" operation, and we're done.
 722   idx_t tp = _age.top();    // XXX
 723   if (size(localBot, tp) > 0) {
 724     assert(dirty_size(localBot, tp) != N - 1, "sanity");
 725     TASKQUEUE_STATS_ONLY(stats.record_pop());
 726     return true;
 727   } else {
 728     // Otherwise, the queue contained exactly one element; we take the slow
 729     // path.
 730     return pop_local_slow(localBot, _age.get());
 731   }
 732 }
 733 
 734 typedef GenericTaskQueue<oop, mtGC>             OopTaskQueue;
 735 typedef GenericTaskQueueSet<OopTaskQueue, mtGC> OopTaskQueueSet;
 736 
 737 #ifdef _MSC_VER
 738 #pragma warning(push)
 739 // warning C4522: multiple assignment operators specified
 740 #pragma warning(disable:4522)
 741 #endif
 742 
 743 // This is a container class for either an oop* or a narrowOop*.
 744 // Both are pushed onto a task queue and the consumer will test is_narrow()
 745 // to determine which should be processed.
 746 class StarTask {
 747   void*  _holder;        // either union oop* or narrowOop*
 748 
 749   enum { COMPRESSED_OOP_MASK = 1 };
 750 
 751  public:
 752   StarTask(narrowOop* p) {
 753     assert(((uintptr_t)p & COMPRESSED_OOP_MASK) == 0, "Information loss!");
 754     _holder = (void *)((uintptr_t)p | COMPRESSED_OOP_MASK);
 755   }
 756   StarTask(oop* p)       {
 757     assert(((uintptr_t)p & COMPRESSED_OOP_MASK) == 0, "Information loss!");
 758     _holder = (void*)p;
 759   }
 760   StarTask()             { _holder = NULL; }
 761   operator oop*()        { return (oop*)_holder; }
 762   operator narrowOop*()  {
 763     return (narrowOop*)((uintptr_t)_holder & ~COMPRESSED_OOP_MASK);
 764   }
 765 
 766   StarTask& operator=(const StarTask& t) {
 767     _holder = t._holder;
 768     return *this;
 769   }
 770   volatile StarTask& operator=(const volatile StarTask& t) volatile {
 771     _holder = t._holder;
 772     return *this;
 773   }
 774 
 775   bool is_narrow() const {
 776     return (((uintptr_t)_holder & COMPRESSED_OOP_MASK) != 0);
 777   }
 778 };
 779 
 780 class ObjArrayTask
 781 {
 782 public:
 783   ObjArrayTask(oop o = NULL, int idx = 0): _obj(o), _index(idx) { }
 784   ObjArrayTask(oop o, size_t idx): _obj(o), _index(int(idx)) {
 785     assert(idx <= size_t(max_jint), "too big");
 786   }
 787   ObjArrayTask(const ObjArrayTask& t): _obj(t._obj), _index(t._index) { }
 788 
 789   ObjArrayTask& operator =(const ObjArrayTask& t) {
 790     _obj = t._obj;
 791     _index = t._index;
 792     return *this;
 793   }
 794   volatile ObjArrayTask&
 795   operator =(const volatile ObjArrayTask& t) volatile {
 796     _obj = t._obj;
 797     _index = t._index;
 798     return *this;
 799   }
 800 
 801   inline oop obj()   const { return _obj; }
 802   inline int index() const { return _index; }
 803 
 804   DEBUG_ONLY(bool is_valid() const); // Tasks to be pushed/popped must be valid.
 805 
 806 private:
 807   oop _obj;
 808   int _index;
 809 };
 810 
 811 #ifdef _MSC_VER
 812 #pragma warning(pop)
 813 #endif
 814 
 815 typedef OverflowTaskQueue<StarTask, mtClass>           OopStarTaskQueue;
 816 typedef GenericTaskQueueSet<OopStarTaskQueue, mtClass> OopStarTaskQueueSet;
 817 
 818 typedef OverflowTaskQueue<size_t, mtInternal>             RegionTaskQueue;
 819 typedef GenericTaskQueueSet<RegionTaskQueue, mtClass>     RegionTaskQueueSet;
 820 
 821 
 822 #endif // SHARE_VM_UTILITIES_TASKQUEUE_HPP