src/share/vm/utilities/taskqueue.hpp

Tue, 23 Nov 2010 13:22:55 -0800

author
stefank
date
Tue, 23 Nov 2010 13:22:55 -0800
changeset 2314
f95d63e2154a
parent 2191
894b1d7c7e01
child 2508
b92c45f2bc75
permissions
-rw-r--r--

6989984: Use standard include model for Hospot
Summary: Replaced MakeDeps and the includeDB files with more standardized solutions.
Reviewed-by: coleenp, kvn, kamg

     1 /*
     2  * Copyright (c) 2001, 2010, 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  */
    25 #ifndef SHARE_VM_UTILITIES_TASKQUEUE_HPP
    26 #define SHARE_VM_UTILITIES_TASKQUEUE_HPP
    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
    51 // Simple TaskQueue stats that are collected by default in debug builds.
    53 #if !defined(TASKQUEUE_STATS) && defined(ASSERT)
    54 #define TASKQUEUE_STATS 1
    55 #elif !defined(TASKQUEUE_STATS)
    56 #define TASKQUEUE_STATS 0
    57 #endif
    59 #if TASKQUEUE_STATS
    60 #define TASKQUEUE_STATS_ONLY(code) code
    61 #else
    62 #define TASKQUEUE_STATS_ONLY(code)
    63 #endif // TASKQUEUE_STATS
    65 #if TASKQUEUE_STATS
    66 class TaskQueueStats {
    67 public:
    68   enum StatId {
    69     push,             // number of taskqueue pushes
    70     pop,              // number of taskqueue pops
    71     pop_slow,         // subset of taskqueue pops that were done slow-path
    72     steal_attempt,    // number of taskqueue steal attempts
    73     steal,            // number of taskqueue steals
    74     overflow,         // number of overflow pushes
    75     overflow_max_len, // max length of overflow stack
    76     last_stat_id
    77   };
    79 public:
    80   inline TaskQueueStats()       { reset(); }
    82   inline void record_push()     { ++_stats[push]; }
    83   inline void record_pop()      { ++_stats[pop]; }
    84   inline void record_pop_slow() { record_pop(); ++_stats[pop_slow]; }
    85   inline void record_steal(bool success);
    86   inline void record_overflow(size_t new_length);
    88   TaskQueueStats & operator +=(const TaskQueueStats & addend);
    90   inline size_t get(StatId id) const { return _stats[id]; }
    91   inline const size_t* get() const   { return _stats; }
    93   inline void reset();
    95   // Print the specified line of the header (does not include a line separator).
    96   static void print_header(unsigned int line, outputStream* const stream = tty,
    97                            unsigned int width = 10);
    98   // Print the statistics (does not include a line separator).
    99   void print(outputStream* const stream = tty, unsigned int width = 10) const;
   101   DEBUG_ONLY(void verify() const;)
   103 private:
   104   size_t                    _stats[last_stat_id];
   105   static const char * const _names[last_stat_id];
   106 };
   108 void TaskQueueStats::record_steal(bool success) {
   109   ++_stats[steal_attempt];
   110   if (success) ++_stats[steal];
   111 }
   113 void TaskQueueStats::record_overflow(size_t new_len) {
   114   ++_stats[overflow];
   115   if (new_len > _stats[overflow_max_len]) _stats[overflow_max_len] = new_len;
   116 }
   118 void TaskQueueStats::reset() {
   119   memset(_stats, 0, sizeof(_stats));
   120 }
   121 #endif // TASKQUEUE_STATS
   123 template <unsigned int N>
   124 class TaskQueueSuper: public CHeapObj {
   125 protected:
   126   // Internal type for indexing the queue; also used for the tag.
   127   typedef NOT_LP64(uint16_t) LP64_ONLY(uint32_t) idx_t;
   129   // The first free element after the last one pushed (mod N).
   130   volatile uint _bottom;
   132   enum { MOD_N_MASK = N - 1 };
   134   class Age {
   135   public:
   136     Age(size_t data = 0)         { _data = data; }
   137     Age(const Age& age)          { _data = age._data; }
   138     Age(idx_t top, idx_t tag)    { _fields._top = top; _fields._tag = tag; }
   140     Age   get()        const volatile { return _data; }
   141     void  set(Age age) volatile       { _data = age._data; }
   143     idx_t top()        const volatile { return _fields._top; }
   144     idx_t tag()        const volatile { return _fields._tag; }
   146     // Increment top; if it wraps, increment tag also.
   147     void increment() {
   148       _fields._top = increment_index(_fields._top);
   149       if (_fields._top == 0) ++_fields._tag;
   150     }
   152     Age cmpxchg(const Age new_age, const Age old_age) volatile {
   153       return (size_t) Atomic::cmpxchg_ptr((intptr_t)new_age._data,
   154                                           (volatile intptr_t *)&_data,
   155                                           (intptr_t)old_age._data);
   156     }
   158     bool operator ==(const Age& other) const { return _data == other._data; }
   160   private:
   161     struct fields {
   162       idx_t _top;
   163       idx_t _tag;
   164     };
   165     union {
   166       size_t _data;
   167       fields _fields;
   168     };
   169   };
   171   volatile Age _age;
   173   // These both operate mod N.
   174   static uint increment_index(uint ind) {
   175     return (ind + 1) & MOD_N_MASK;
   176   }
   177   static uint decrement_index(uint ind) {
   178     return (ind - 1) & MOD_N_MASK;
   179   }
   181   // Returns a number in the range [0..N).  If the result is "N-1", it should be
   182   // interpreted as 0.
   183   uint dirty_size(uint bot, uint top) const {
   184     return (bot - top) & MOD_N_MASK;
   185   }
   187   // Returns the size corresponding to the given "bot" and "top".
   188   uint size(uint bot, uint top) const {
   189     uint sz = dirty_size(bot, top);
   190     // Has the queue "wrapped", so that bottom is less than top?  There's a
   191     // complicated special case here.  A pair of threads could perform pop_local
   192     // and pop_global operations concurrently, starting from a state in which
   193     // _bottom == _top+1.  The pop_local could succeed in decrementing _bottom,
   194     // and the pop_global in incrementing _top (in which case the pop_global
   195     // will be awarded the contested queue element.)  The resulting state must
   196     // be interpreted as an empty queue.  (We only need to worry about one such
   197     // event: only the queue owner performs pop_local's, and several concurrent
   198     // threads attempting to perform the pop_global will all perform the same
   199     // CAS, and only one can succeed.)  Any stealing thread that reads after
   200     // either the increment or decrement will see an empty queue, and will not
   201     // join the competitors.  The "sz == -1 || sz == N-1" state will not be
   202     // modified by concurrent queues, so the owner thread can reset the state to
   203     // _bottom == top so subsequent pushes will be performed normally.
   204     return (sz == N - 1) ? 0 : sz;
   205   }
   207 public:
   208   TaskQueueSuper() : _bottom(0), _age() {}
   210   // Return true if the TaskQueue contains/does not contain any tasks.
   211   bool peek()     const { return _bottom != _age.top(); }
   212   bool is_empty() const { return size() == 0; }
   214   // Return an estimate of the number of elements in the queue.
   215   // The "careful" version admits the possibility of pop_local/pop_global
   216   // races.
   217   uint size() const {
   218     return size(_bottom, _age.top());
   219   }
   221   uint dirty_size() const {
   222     return dirty_size(_bottom, _age.top());
   223   }
   225   void set_empty() {
   226     _bottom = 0;
   227     _age.set(0);
   228   }
   230   // Maximum number of elements allowed in the queue.  This is two less
   231   // than the actual queue size, for somewhat complicated reasons.
   232   uint max_elems() const { return N - 2; }
   234   // Total size of queue.
   235   static const uint total_size() { return N; }
   237   TASKQUEUE_STATS_ONLY(TaskQueueStats stats;)
   238 };
   240 template<class E, unsigned int N = TASKQUEUE_SIZE>
   241 class GenericTaskQueue: public TaskQueueSuper<N> {
   242 protected:
   243   typedef typename TaskQueueSuper<N>::Age Age;
   244   typedef typename TaskQueueSuper<N>::idx_t idx_t;
   246   using TaskQueueSuper<N>::_bottom;
   247   using TaskQueueSuper<N>::_age;
   248   using TaskQueueSuper<N>::increment_index;
   249   using TaskQueueSuper<N>::decrement_index;
   250   using TaskQueueSuper<N>::dirty_size;
   252 public:
   253   using TaskQueueSuper<N>::max_elems;
   254   using TaskQueueSuper<N>::size;
   255   TASKQUEUE_STATS_ONLY(using TaskQueueSuper<N>::stats;)
   257 private:
   258   // Slow paths for push, pop_local.  (pop_global has no fast path.)
   259   bool push_slow(E t, uint dirty_n_elems);
   260   bool pop_local_slow(uint localBot, Age oldAge);
   262 public:
   263   typedef E element_type;
   265   // Initializes the queue to empty.
   266   GenericTaskQueue();
   268   void initialize();
   270   // Push the task "t" on the queue.  Returns "false" iff the queue is full.
   271   inline bool push(E t);
   273   // Attempts to claim a task from the "local" end of the queue (the most
   274   // recently pushed).  If successful, returns true and sets t to the task;
   275   // otherwise, returns false (the queue is empty).
   276   inline bool pop_local(E& t);
   278   // Like pop_local(), but uses the "global" end of the queue (the least
   279   // recently pushed).
   280   bool pop_global(E& t);
   282   // Delete any resource associated with the queue.
   283   ~GenericTaskQueue();
   285   // apply the closure to all elements in the task queue
   286   void oops_do(OopClosure* f);
   288 private:
   289   // Element array.
   290   volatile E* _elems;
   291 };
   293 template<class E, unsigned int N>
   294 GenericTaskQueue<E, N>::GenericTaskQueue() {
   295   assert(sizeof(Age) == sizeof(size_t), "Depends on this.");
   296 }
   298 template<class E, unsigned int N>
   299 void GenericTaskQueue<E, N>::initialize() {
   300   _elems = NEW_C_HEAP_ARRAY(E, N);
   301 }
   303 template<class E, unsigned int N>
   304 void GenericTaskQueue<E, N>::oops_do(OopClosure* f) {
   305   // tty->print_cr("START OopTaskQueue::oops_do");
   306   uint iters = size();
   307   uint index = _bottom;
   308   for (uint i = 0; i < iters; ++i) {
   309     index = decrement_index(index);
   310     // tty->print_cr("  doing entry %d," INTPTR_T " -> " INTPTR_T,
   311     //            index, &_elems[index], _elems[index]);
   312     E* t = (E*)&_elems[index];      // cast away volatility
   313     oop* p = (oop*)t;
   314     assert((*t)->is_oop_or_null(), "Not an oop or null");
   315     f->do_oop(p);
   316   }
   317   // tty->print_cr("END OopTaskQueue::oops_do");
   318 }
   320 template<class E, unsigned int N>
   321 bool GenericTaskQueue<E, N>::push_slow(E t, uint dirty_n_elems) {
   322   if (dirty_n_elems == N - 1) {
   323     // Actually means 0, so do the push.
   324     uint localBot = _bottom;
   325     // g++ complains if the volatile result of the assignment is unused.
   326     const_cast<E&>(_elems[localBot] = t);
   327     OrderAccess::release_store(&_bottom, increment_index(localBot));
   328     TASKQUEUE_STATS_ONLY(stats.record_push());
   329     return true;
   330   }
   331   return false;
   332 }
   334 // pop_local_slow() is done by the owning thread and is trying to
   335 // get the last task in the queue.  It will compete with pop_global()
   336 // that will be used by other threads.  The tag age is incremented
   337 // whenever the queue goes empty which it will do here if this thread
   338 // gets the last task or in pop_global() if the queue wraps (top == 0
   339 // and pop_global() succeeds, see pop_global()).
   340 template<class E, unsigned int N>
   341 bool GenericTaskQueue<E, N>::pop_local_slow(uint localBot, Age oldAge) {
   342   // This queue was observed to contain exactly one element; either this
   343   // thread will claim it, or a competing "pop_global".  In either case,
   344   // the queue will be logically empty afterwards.  Create a new Age value
   345   // that represents the empty queue for the given value of "_bottom".  (We
   346   // must also increment "tag" because of the case where "bottom == 1",
   347   // "top == 0".  A pop_global could read the queue element in that case,
   348   // then have the owner thread do a pop followed by another push.  Without
   349   // the incrementing of "tag", the pop_global's CAS could succeed,
   350   // allowing it to believe it has claimed the stale element.)
   351   Age newAge((idx_t)localBot, oldAge.tag() + 1);
   352   // Perhaps a competing pop_global has already incremented "top", in which
   353   // case it wins the element.
   354   if (localBot == oldAge.top()) {
   355     // No competing pop_global has yet incremented "top"; we'll try to
   356     // install new_age, thus claiming the element.
   357     Age tempAge = _age.cmpxchg(newAge, oldAge);
   358     if (tempAge == oldAge) {
   359       // We win.
   360       assert(dirty_size(localBot, _age.top()) != N - 1, "sanity");
   361       TASKQUEUE_STATS_ONLY(stats.record_pop_slow());
   362       return true;
   363     }
   364   }
   365   // We lose; a completing pop_global gets the element.  But the queue is empty
   366   // and top is greater than bottom.  Fix this representation of the empty queue
   367   // to become the canonical one.
   368   _age.set(newAge);
   369   assert(dirty_size(localBot, _age.top()) != N - 1, "sanity");
   370   return false;
   371 }
   373 template<class E, unsigned int N>
   374 bool GenericTaskQueue<E, N>::pop_global(E& t) {
   375   Age oldAge = _age.get();
   376   uint localBot = _bottom;
   377   uint n_elems = size(localBot, oldAge.top());
   378   if (n_elems == 0) {
   379     return false;
   380   }
   382   const_cast<E&>(t = _elems[oldAge.top()]);
   383   Age newAge(oldAge);
   384   newAge.increment();
   385   Age resAge = _age.cmpxchg(newAge, oldAge);
   387   // Note that using "_bottom" here might fail, since a pop_local might
   388   // have decremented it.
   389   assert(dirty_size(localBot, newAge.top()) != N - 1, "sanity");
   390   return resAge == oldAge;
   391 }
   393 template<class E, unsigned int N>
   394 GenericTaskQueue<E, N>::~GenericTaskQueue() {
   395   FREE_C_HEAP_ARRAY(E, _elems);
   396 }
   398 // OverflowTaskQueue is a TaskQueue that also includes an overflow stack for
   399 // elements that do not fit in the TaskQueue.
   400 //
   401 // This class hides two methods from super classes:
   402 //
   403 // push() - push onto the task queue or, if that fails, onto the overflow stack
   404 // is_empty() - return true if both the TaskQueue and overflow stack are empty
   405 //
   406 // Note that size() is not hidden--it returns the number of elements in the
   407 // TaskQueue, and does not include the size of the overflow stack.  This
   408 // simplifies replacement of GenericTaskQueues with OverflowTaskQueues.
   409 template<class E, unsigned int N = TASKQUEUE_SIZE>
   410 class OverflowTaskQueue: public GenericTaskQueue<E, N>
   411 {
   412 public:
   413   typedef Stack<E>               overflow_t;
   414   typedef GenericTaskQueue<E, N> taskqueue_t;
   416   TASKQUEUE_STATS_ONLY(using taskqueue_t::stats;)
   418   // Push task t onto the queue or onto the overflow stack.  Return true.
   419   inline bool push(E t);
   421   // Attempt to pop from the overflow stack; return true if anything was popped.
   422   inline bool pop_overflow(E& t);
   424   inline overflow_t* overflow_stack() { return &_overflow_stack; }
   426   inline bool taskqueue_empty() const { return taskqueue_t::is_empty(); }
   427   inline bool overflow_empty()  const { return _overflow_stack.is_empty(); }
   428   inline bool is_empty()        const {
   429     return taskqueue_empty() && overflow_empty();
   430   }
   432 private:
   433   overflow_t _overflow_stack;
   434 };
   436 template <class E, unsigned int N>
   437 bool OverflowTaskQueue<E, N>::push(E t)
   438 {
   439   if (!taskqueue_t::push(t)) {
   440     overflow_stack()->push(t);
   441     TASKQUEUE_STATS_ONLY(stats.record_overflow(overflow_stack()->size()));
   442   }
   443   return true;
   444 }
   446 template <class E, unsigned int N>
   447 bool OverflowTaskQueue<E, N>::pop_overflow(E& t)
   448 {
   449   if (overflow_empty()) return false;
   450   t = overflow_stack()->pop();
   451   return true;
   452 }
   454 class TaskQueueSetSuper: public CHeapObj {
   455 protected:
   456   static int randomParkAndMiller(int* seed0);
   457 public:
   458   // Returns "true" if some TaskQueue in the set contains a task.
   459   virtual bool peek() = 0;
   460 };
   462 template<class T>
   463 class GenericTaskQueueSet: public TaskQueueSetSuper {
   464 private:
   465   uint _n;
   466   T** _queues;
   468 public:
   469   typedef typename T::element_type E;
   471   GenericTaskQueueSet(int n) : _n(n) {
   472     typedef T* GenericTaskQueuePtr;
   473     _queues = NEW_C_HEAP_ARRAY(GenericTaskQueuePtr, n);
   474     for (int i = 0; i < n; i++) {
   475       _queues[i] = NULL;
   476     }
   477   }
   479   bool steal_1_random(uint queue_num, int* seed, E& t);
   480   bool steal_best_of_2(uint queue_num, int* seed, E& t);
   481   bool steal_best_of_all(uint queue_num, int* seed, E& t);
   483   void register_queue(uint i, T* q);
   485   T* queue(uint n);
   487   // The thread with queue number "queue_num" (and whose random number seed is
   488   // at "seed") is trying to steal a task from some other queue.  (It may try
   489   // several queues, according to some configuration parameter.)  If some steal
   490   // succeeds, returns "true" and sets "t" to the stolen task, otherwise returns
   491   // false.
   492   bool steal(uint queue_num, int* seed, E& t);
   494   bool peek();
   495 };
   497 template<class T> void
   498 GenericTaskQueueSet<T>::register_queue(uint i, T* q) {
   499   assert(i < _n, "index out of range.");
   500   _queues[i] = q;
   501 }
   503 template<class T> T*
   504 GenericTaskQueueSet<T>::queue(uint i) {
   505   return _queues[i];
   506 }
   508 template<class T> bool
   509 GenericTaskQueueSet<T>::steal(uint queue_num, int* seed, E& t) {
   510   for (uint i = 0; i < 2 * _n; i++) {
   511     if (steal_best_of_2(queue_num, seed, t)) {
   512       TASKQUEUE_STATS_ONLY(queue(queue_num)->stats.record_steal(true));
   513       return true;
   514     }
   515   }
   516   TASKQUEUE_STATS_ONLY(queue(queue_num)->stats.record_steal(false));
   517   return false;
   518 }
   520 template<class T> bool
   521 GenericTaskQueueSet<T>::steal_best_of_all(uint queue_num, int* seed, E& t) {
   522   if (_n > 2) {
   523     int best_k;
   524     uint best_sz = 0;
   525     for (uint k = 0; k < _n; k++) {
   526       if (k == queue_num) continue;
   527       uint sz = _queues[k]->size();
   528       if (sz > best_sz) {
   529         best_sz = sz;
   530         best_k = k;
   531       }
   532     }
   533     return best_sz > 0 && _queues[best_k]->pop_global(t);
   534   } else if (_n == 2) {
   535     // Just try the other one.
   536     int k = (queue_num + 1) % 2;
   537     return _queues[k]->pop_global(t);
   538   } else {
   539     assert(_n == 1, "can't be zero.");
   540     return false;
   541   }
   542 }
   544 template<class T> bool
   545 GenericTaskQueueSet<T>::steal_1_random(uint queue_num, int* seed, E& t) {
   546   if (_n > 2) {
   547     uint k = queue_num;
   548     while (k == queue_num) k = randomParkAndMiller(seed) % _n;
   549     return _queues[2]->pop_global(t);
   550   } else if (_n == 2) {
   551     // Just try the other one.
   552     int k = (queue_num + 1) % 2;
   553     return _queues[k]->pop_global(t);
   554   } else {
   555     assert(_n == 1, "can't be zero.");
   556     return false;
   557   }
   558 }
   560 template<class T> bool
   561 GenericTaskQueueSet<T>::steal_best_of_2(uint queue_num, int* seed, E& t) {
   562   if (_n > 2) {
   563     uint k1 = queue_num;
   564     while (k1 == queue_num) k1 = randomParkAndMiller(seed) % _n;
   565     uint k2 = queue_num;
   566     while (k2 == queue_num || k2 == k1) k2 = randomParkAndMiller(seed) % _n;
   567     // Sample both and try the larger.
   568     uint sz1 = _queues[k1]->size();
   569     uint sz2 = _queues[k2]->size();
   570     if (sz2 > sz1) return _queues[k2]->pop_global(t);
   571     else return _queues[k1]->pop_global(t);
   572   } else if (_n == 2) {
   573     // Just try the other one.
   574     uint k = (queue_num + 1) % 2;
   575     return _queues[k]->pop_global(t);
   576   } else {
   577     assert(_n == 1, "can't be zero.");
   578     return false;
   579   }
   580 }
   582 template<class T>
   583 bool GenericTaskQueueSet<T>::peek() {
   584   // Try all the queues.
   585   for (uint j = 0; j < _n; j++) {
   586     if (_queues[j]->peek())
   587       return true;
   588   }
   589   return false;
   590 }
   592 // When to terminate from the termination protocol.
   593 class TerminatorTerminator: public CHeapObj {
   594 public:
   595   virtual bool should_exit_termination() = 0;
   596 };
   598 // A class to aid in the termination of a set of parallel tasks using
   599 // TaskQueueSet's for work stealing.
   601 #undef TRACESPINNING
   603 class ParallelTaskTerminator: public StackObj {
   604 private:
   605   int _n_threads;
   606   TaskQueueSetSuper* _queue_set;
   607   int _offered_termination;
   609 #ifdef TRACESPINNING
   610   static uint _total_yields;
   611   static uint _total_spins;
   612   static uint _total_peeks;
   613 #endif
   615   bool peek_in_queue_set();
   616 protected:
   617   virtual void yield();
   618   void sleep(uint millis);
   620 public:
   622   // "n_threads" is the number of threads to be terminated.  "queue_set" is a
   623   // queue sets of work queues of other threads.
   624   ParallelTaskTerminator(int n_threads, TaskQueueSetSuper* queue_set);
   626   // The current thread has no work, and is ready to terminate if everyone
   627   // else is.  If returns "true", all threads are terminated.  If returns
   628   // "false", available work has been observed in one of the task queues,
   629   // so the global task is not complete.
   630   bool offer_termination() {
   631     return offer_termination(NULL);
   632   }
   634   // As above, but it also terminates if the should_exit_termination()
   635   // method of the terminator parameter returns true. If terminator is
   636   // NULL, then it is ignored.
   637   bool offer_termination(TerminatorTerminator* terminator);
   639   // Reset the terminator, so that it may be reused again.
   640   // The caller is responsible for ensuring that this is done
   641   // in an MT-safe manner, once the previous round of use of
   642   // the terminator is finished.
   643   void reset_for_reuse();
   644   // Same as above but the number of parallel threads is set to the
   645   // given number.
   646   void reset_for_reuse(int n_threads);
   648 #ifdef TRACESPINNING
   649   static uint total_yields() { return _total_yields; }
   650   static uint total_spins() { return _total_spins; }
   651   static uint total_peeks() { return _total_peeks; }
   652   static void print_termination_counts();
   653 #endif
   654 };
   656 template<class E, unsigned int N> inline bool
   657 GenericTaskQueue<E, N>::push(E t) {
   658   uint localBot = _bottom;
   659   assert((localBot >= 0) && (localBot < N), "_bottom out of range.");
   660   idx_t top = _age.top();
   661   uint dirty_n_elems = dirty_size(localBot, top);
   662   assert(dirty_n_elems < N, "n_elems out of range.");
   663   if (dirty_n_elems < max_elems()) {
   664     // g++ complains if the volatile result of the assignment is unused.
   665     const_cast<E&>(_elems[localBot] = t);
   666     OrderAccess::release_store(&_bottom, increment_index(localBot));
   667     TASKQUEUE_STATS_ONLY(stats.record_push());
   668     return true;
   669   } else {
   670     return push_slow(t, dirty_n_elems);
   671   }
   672 }
   674 template<class E, unsigned int N> inline bool
   675 GenericTaskQueue<E, N>::pop_local(E& t) {
   676   uint localBot = _bottom;
   677   // This value cannot be N-1.  That can only occur as a result of
   678   // the assignment to bottom in this method.  If it does, this method
   679   // resets the size to 0 before the next call (which is sequential,
   680   // since this is pop_local.)
   681   uint dirty_n_elems = dirty_size(localBot, _age.top());
   682   assert(dirty_n_elems != N - 1, "Shouldn't be possible...");
   683   if (dirty_n_elems == 0) return false;
   684   localBot = decrement_index(localBot);
   685   _bottom = localBot;
   686   // This is necessary to prevent any read below from being reordered
   687   // before the store just above.
   688   OrderAccess::fence();
   689   const_cast<E&>(t = _elems[localBot]);
   690   // This is a second read of "age"; the "size()" above is the first.
   691   // If there's still at least one element in the queue, based on the
   692   // "_bottom" and "age" we've read, then there can be no interference with
   693   // a "pop_global" operation, and we're done.
   694   idx_t tp = _age.top();    // XXX
   695   if (size(localBot, tp) > 0) {
   696     assert(dirty_size(localBot, tp) != N - 1, "sanity");
   697     TASKQUEUE_STATS_ONLY(stats.record_pop());
   698     return true;
   699   } else {
   700     // Otherwise, the queue contained exactly one element; we take the slow
   701     // path.
   702     return pop_local_slow(localBot, _age.get());
   703   }
   704 }
   706 typedef GenericTaskQueue<oop>             OopTaskQueue;
   707 typedef GenericTaskQueueSet<OopTaskQueue> OopTaskQueueSet;
   709 #ifdef _MSC_VER
   710 #pragma warning(push)
   711 // warning C4522: multiple assignment operators specified
   712 #pragma warning(disable:4522)
   713 #endif
   715 // This is a container class for either an oop* or a narrowOop*.
   716 // Both are pushed onto a task queue and the consumer will test is_narrow()
   717 // to determine which should be processed.
   718 class StarTask {
   719   void*  _holder;        // either union oop* or narrowOop*
   721   enum { COMPRESSED_OOP_MASK = 1 };
   723  public:
   724   StarTask(narrowOop* p) {
   725     assert(((uintptr_t)p & COMPRESSED_OOP_MASK) == 0, "Information loss!");
   726     _holder = (void *)((uintptr_t)p | COMPRESSED_OOP_MASK);
   727   }
   728   StarTask(oop* p)       {
   729     assert(((uintptr_t)p & COMPRESSED_OOP_MASK) == 0, "Information loss!");
   730     _holder = (void*)p;
   731   }
   732   StarTask()             { _holder = NULL; }
   733   operator oop*()        { return (oop*)_holder; }
   734   operator narrowOop*()  {
   735     return (narrowOop*)((uintptr_t)_holder & ~COMPRESSED_OOP_MASK);
   736   }
   738   StarTask& operator=(const StarTask& t) {
   739     _holder = t._holder;
   740     return *this;
   741   }
   742   volatile StarTask& operator=(const volatile StarTask& t) volatile {
   743     _holder = t._holder;
   744     return *this;
   745   }
   747   bool is_narrow() const {
   748     return (((uintptr_t)_holder & COMPRESSED_OOP_MASK) != 0);
   749   }
   750 };
   752 class ObjArrayTask
   753 {
   754 public:
   755   ObjArrayTask(oop o = NULL, int idx = 0): _obj(o), _index(idx) { }
   756   ObjArrayTask(oop o, size_t idx): _obj(o), _index(int(idx)) {
   757     assert(idx <= size_t(max_jint), "too big");
   758   }
   759   ObjArrayTask(const ObjArrayTask& t): _obj(t._obj), _index(t._index) { }
   761   ObjArrayTask& operator =(const ObjArrayTask& t) {
   762     _obj = t._obj;
   763     _index = t._index;
   764     return *this;
   765   }
   766   volatile ObjArrayTask&
   767   operator =(const volatile ObjArrayTask& t) volatile {
   768     _obj = t._obj;
   769     _index = t._index;
   770     return *this;
   771   }
   773   inline oop obj()   const { return _obj; }
   774   inline int index() const { return _index; }
   776   DEBUG_ONLY(bool is_valid() const); // Tasks to be pushed/popped must be valid.
   778 private:
   779   oop _obj;
   780   int _index;
   781 };
   783 #ifdef _MSC_VER
   784 #pragma warning(pop)
   785 #endif
   787 typedef OverflowTaskQueue<StarTask>           OopStarTaskQueue;
   788 typedef GenericTaskQueueSet<OopStarTaskQueue> OopStarTaskQueueSet;
   790 typedef OverflowTaskQueue<size_t>             RegionTaskQueue;
   791 typedef GenericTaskQueueSet<RegionTaskQueue>  RegionTaskQueueSet;
   794 #endif // SHARE_VM_UTILITIES_TASKQUEUE_HPP

mercurial