1 /* 2 * Copyright (c) 2001, 2012, 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 344 // unused, so we cast the volatile away. We cannot cast directly 345 // to void, because gcc treats that as not using the result of the 346 // assignment. However, casting to E& means that we trigger an 347 // unused-value warning. So, we cast the E& to void. 348 (void)const_cast<E&>(_elems[localBot] = t); 349 OrderAccess::release_store(&_bottom, increment_index(localBot)); 350 TASKQUEUE_STATS_ONLY(stats.record_push()); 351 return true; 352 } 353 return false; 354 } 355 356 // pop_local_slow() is done by the owning thread and is trying to 357 // get the last task in the queue. It will compete with pop_global() 358 // that will be used by other threads. The tag age is incremented 359 // whenever the queue goes empty which it will do here if this thread 360 // gets the last task or in pop_global() if the queue wraps (top == 0 361 // and pop_global() succeeds, see pop_global()). 362 template<class E, MEMFLAGS F, unsigned int N> 363 bool GenericTaskQueue<E, F, N>::pop_local_slow(uint localBot, Age oldAge) { 364 // This queue was observed to contain exactly one element; either this 365 // thread will claim it, or a competing "pop_global". In either case, 366 // the queue will be logically empty afterwards. Create a new Age value 367 // that represents the empty queue for the given value of "_bottom". (We 368 // must also increment "tag" because of the case where "bottom == 1", 369 // "top == 0". A pop_global could read the queue element in that case, 370 // then have the owner thread do a pop followed by another push. Without 371 // the incrementing of "tag", the pop_global's CAS could succeed, 372 // allowing it to believe it has claimed the stale element.) 373 Age newAge((idx_t)localBot, oldAge.tag() + 1); 374 // Perhaps a competing pop_global has already incremented "top", in which 375 // case it wins the element. 376 if (localBot == oldAge.top()) { 377 // No competing pop_global has yet incremented "top"; we'll try to 378 // install new_age, thus claiming the element. 379 Age tempAge = _age.cmpxchg(newAge, oldAge); 380 if (tempAge == oldAge) { 381 // We win. 382 assert(dirty_size(localBot, _age.top()) != N - 1, "sanity"); 383 TASKQUEUE_STATS_ONLY(stats.record_pop_slow()); 384 return true; 385 } 386 } 387 // We lose; a completing pop_global gets the element. But the queue is empty 388 // and top is greater than bottom. Fix this representation of the empty queue 389 // to become the canonical one. 390 _age.set(newAge); 391 assert(dirty_size(localBot, _age.top()) != N - 1, "sanity"); 392 return false; 393 } 394 395 template<class E, MEMFLAGS F, unsigned int N> 396 bool GenericTaskQueue<E, F, N>::pop_global(E& t) { 397 Age oldAge = _age.get(); 398 uint localBot = _bottom; 399 uint n_elems = size(localBot, oldAge.top()); 400 if (n_elems == 0) { 401 return false; 402 } 403 404 // g++ complains if the volatile result of the assignment is 405 // unused, so we cast the volatile away. We cannot cast directly 406 // to void, because gcc treats that as not using the result of the 407 // assignment. However, casting to E& means that we trigger an 408 // unused-value warning. So, we cast the E& to void. 409 (void) const_cast<E&>(t = _elems[oldAge.top()]); 410 Age newAge(oldAge); 411 newAge.increment(); 412 Age resAge = _age.cmpxchg(newAge, oldAge); 413 414 // Note that using "_bottom" here might fail, since a pop_local might 415 // have decremented it. 416 assert(dirty_size(localBot, newAge.top()) != N - 1, "sanity"); 417 return resAge == oldAge; 418 } 419 420 template<class E, MEMFLAGS F, unsigned int N> 421 GenericTaskQueue<E, F, N>::~GenericTaskQueue() { 422 FREE_C_HEAP_ARRAY(E, _elems, F); 423 } 424 425 // OverflowTaskQueue is a TaskQueue that also includes an overflow stack for 426 // elements that do not fit in the TaskQueue. 427 // 428 // This class hides two methods from super classes: 429 // 430 // push() - push onto the task queue or, if that fails, onto the overflow stack 431 // is_empty() - return true if both the TaskQueue and overflow stack are empty 432 // 433 // Note that size() is not hidden--it returns the number of elements in the 434 // TaskQueue, and does not include the size of the overflow stack. This 435 // simplifies replacement of GenericTaskQueues with OverflowTaskQueues. 436 template<class E, MEMFLAGS F, unsigned int N = TASKQUEUE_SIZE> 437 class OverflowTaskQueue: public GenericTaskQueue<E, F, N> 438 { 439 public: 440 typedef Stack<E, F> overflow_t; 441 typedef GenericTaskQueue<E, F, N> taskqueue_t; 442 443 TASKQUEUE_STATS_ONLY(using taskqueue_t::stats;) 444 445 // Push task t onto the queue or onto the overflow stack. Return true. 446 inline bool push(E t); 447 448 // Attempt to pop from the overflow stack; return true if anything was popped. 449 inline bool pop_overflow(E& t); 450 451 inline overflow_t* overflow_stack() { return &_overflow_stack; } 452 453 inline bool taskqueue_empty() const { return taskqueue_t::is_empty(); } 454 inline bool overflow_empty() const { return _overflow_stack.is_empty(); } 455 inline bool is_empty() const { 456 return taskqueue_empty() && overflow_empty(); 457 } 458 459 private: 460 overflow_t _overflow_stack; 461 }; 462 463 template <class E, MEMFLAGS F, unsigned int N> 464 bool OverflowTaskQueue<E, F, N>::push(E t) 465 { 466 if (!taskqueue_t::push(t)) { 467 overflow_stack()->push(t); 468 TASKQUEUE_STATS_ONLY(stats.record_overflow(overflow_stack()->size())); 469 } 470 return true; 471 } 472 473 template <class E, MEMFLAGS F, unsigned int N> 474 bool OverflowTaskQueue<E, F, N>::pop_overflow(E& t) 475 { 476 if (overflow_empty()) return false; 477 t = overflow_stack()->pop(); 478 return true; 479 } 480 481 class TaskQueueSetSuper { 482 protected: 483 static int randomParkAndMiller(int* seed0); 484 public: 485 // Returns "true" if some TaskQueue in the set contains a task. 486 virtual bool peek() = 0; 487 }; 488 489 template <MEMFLAGS F> class TaskQueueSetSuperImpl: public CHeapObj<F>, public TaskQueueSetSuper { 490 }; 491 492 template<class T, MEMFLAGS F> 493 class GenericTaskQueueSet: public TaskQueueSetSuperImpl<F> { 494 private: 495 uint _n; 496 T** _queues; 497 498 public: 499 typedef typename T::element_type E; 500 501 GenericTaskQueueSet(int n) : _n(n) { 502 typedef T* GenericTaskQueuePtr; 503 _queues = NEW_C_HEAP_ARRAY(GenericTaskQueuePtr, n, F); 504 for (int i = 0; i < n; i++) { 505 _queues[i] = NULL; 506 } 507 } 508 509 bool steal_best_of_2(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_2(uint queue_num, int* seed, E& t) { 550 if (_n > 2) { 551 uint k1 = queue_num; 552 while (k1 == queue_num) k1 = TaskQueueSetSuper::randomParkAndMiller(seed) % _n; 553 uint k2 = queue_num; 554 while (k2 == queue_num || k2 == k1) k2 = TaskQueueSetSuper::randomParkAndMiller(seed) % _n; 555 // Sample both and try the larger. 556 uint sz1 = _queues[k1]->size(); 557 uint sz2 = _queues[k2]->size(); 558 if (sz2 > sz1) return _queues[k2]->pop_global(t); 559 else return _queues[k1]->pop_global(t); 560 } else if (_n == 2) { 561 // Just try the other one. 562 uint k = (queue_num + 1) % 2; 563 return _queues[k]->pop_global(t); 564 } else { 565 assert(_n == 1, "can't be zero."); 566 return false; 567 } 568 } 569 570 template<class T, MEMFLAGS F> 571 bool GenericTaskQueueSet<T, F>::peek() { 572 // Try all the queues. 573 for (uint j = 0; j < _n; j++) { 574 if (_queues[j]->peek()) 575 return true; 576 } 577 return false; 578 } 579 580 // When to terminate from the termination protocol. 581 class TerminatorTerminator: public CHeapObj<mtInternal> { 582 public: 583 virtual bool should_exit_termination() = 0; 584 }; 585 586 // A class to aid in the termination of a set of parallel tasks using 587 // TaskQueueSet's for work stealing. 588 589 #undef TRACESPINNING 590 591 class ParallelTaskTerminator: public StackObj { 592 private: 593 int _n_threads; 594 TaskQueueSetSuper* _queue_set; 595 int _offered_termination; 596 597 #ifdef TRACESPINNING 598 static uint _total_yields; 599 static uint _total_spins; 600 static uint _total_peeks; 601 #endif 602 603 bool peek_in_queue_set(); 604 protected: 605 virtual void yield(); 606 void sleep(uint millis); 607 608 public: 609 610 // "n_threads" is the number of threads to be terminated. "queue_set" is a 611 // queue sets of work queues of other threads. 612 ParallelTaskTerminator(int n_threads, TaskQueueSetSuper* queue_set); 613 614 // The current thread has no work, and is ready to terminate if everyone 615 // else is. If returns "true", all threads are terminated. If returns 616 // "false", available work has been observed in one of the task queues, 617 // so the global task is not complete. 618 bool offer_termination() { 619 return offer_termination(NULL); 620 } 621 622 // As above, but it also terminates if the should_exit_termination() 623 // method of the terminator parameter returns true. If terminator is 624 // NULL, then it is ignored. 625 bool offer_termination(TerminatorTerminator* terminator); 626 627 // Reset the terminator, so that it may be reused again. 628 // The caller is responsible for ensuring that this is done 629 // in an MT-safe manner, once the previous round of use of 630 // the terminator is finished. 631 void reset_for_reuse(); 632 // Same as above but the number of parallel threads is set to the 633 // given number. 634 void reset_for_reuse(int n_threads); 635 636 #ifdef TRACESPINNING 637 static uint total_yields() { return _total_yields; } 638 static uint total_spins() { return _total_spins; } 639 static uint total_peeks() { return _total_peeks; } 640 static void print_termination_counts(); 641 #endif 642 }; 643 644 template<class E, MEMFLAGS F, unsigned int N> inline bool 645 GenericTaskQueue<E, F, N>::push(E t) { 646 uint localBot = _bottom; 647 assert((localBot >= 0) && (localBot < N), "_bottom out of range."); 648 idx_t top = _age.top(); 649 uint dirty_n_elems = dirty_size(localBot, top); 650 assert(dirty_n_elems < N, "n_elems out of range."); 651 if (dirty_n_elems < max_elems()) { 652 // g++ complains if the volatile result of the assignment is 653 // unused, so we cast the volatile away. We cannot cast directly 654 // to void, because gcc treats that as not using the result of the 655 // assignment. However, casting to E& means that we trigger an 656 // unused-value warning. So, we cast the E& to void. 657 (void) const_cast<E&>(_elems[localBot] = t); 658 OrderAccess::release_store(&_bottom, increment_index(localBot)); 659 TASKQUEUE_STATS_ONLY(stats.record_push()); 660 return true; 661 } else { 662 return push_slow(t, dirty_n_elems); 663 } 664 } 665 666 template<class E, MEMFLAGS F, unsigned int N> inline bool 667 GenericTaskQueue<E, F, N>::pop_local(E& t) { 668 uint localBot = _bottom; 669 // This value cannot be N-1. That can only occur as a result of 670 // the assignment to bottom in this method. If it does, this method 671 // resets the size to 0 before the next call (which is sequential, 672 // since this is pop_local.) 673 uint dirty_n_elems = dirty_size(localBot, _age.top()); 674 assert(dirty_n_elems != N - 1, "Shouldn't be possible..."); 675 if (dirty_n_elems == 0) return false; 676 localBot = decrement_index(localBot); 677 _bottom = localBot; 678 // This is necessary to prevent any read below from being reordered 679 // before the store just above. 680 OrderAccess::fence(); 681 // g++ complains if the volatile result of the assignment is 682 // unused, so we cast the volatile away. We cannot cast directly 683 // to void, because gcc treats that as not using the result of the 684 // assignment. However, casting to E& means that we trigger an 685 // unused-value warning. So, we cast the E& to void. 686 (void) const_cast<E&>(t = _elems[localBot]); 687 // This is a second read of "age"; the "size()" above is the first. 688 // If there's still at least one element in the queue, based on the 689 // "_bottom" and "age" we've read, then there can be no interference with 690 // a "pop_global" operation, and we're done. 691 idx_t tp = _age.top(); // XXX 692 if (size(localBot, tp) > 0) { 693 assert(dirty_size(localBot, tp) != N - 1, "sanity"); 694 TASKQUEUE_STATS_ONLY(stats.record_pop()); 695 return true; 696 } else { 697 // Otherwise, the queue contained exactly one element; we take the slow 698 // path. 699 return pop_local_slow(localBot, _age.get()); 700 } 701 } 702 703 typedef GenericTaskQueue<oop, mtGC> OopTaskQueue; 704 typedef GenericTaskQueueSet<OopTaskQueue, mtGC> OopTaskQueueSet; 705 706 #ifdef _MSC_VER 707 #pragma warning(push) 708 // warning C4522: multiple assignment operators specified 709 #pragma warning(disable:4522) 710 #endif 711 712 // This is a container class for either an oop* or a narrowOop*. 713 // Both are pushed onto a task queue and the consumer will test is_narrow() 714 // to determine which should be processed. 715 class StarTask { 716 void* _holder; // either union oop* or narrowOop* 717 718 enum { COMPRESSED_OOP_MASK = 1 }; 719 720 public: 721 StarTask(narrowOop* p) { 722 assert(((uintptr_t)p & COMPRESSED_OOP_MASK) == 0, "Information loss!"); 723 _holder = (void *)((uintptr_t)p | COMPRESSED_OOP_MASK); 724 } 725 StarTask(oop* p) { 726 assert(((uintptr_t)p & COMPRESSED_OOP_MASK) == 0, "Information loss!"); 727 _holder = (void*)p; 728 } 729 StarTask() { _holder = NULL; } 730 operator oop*() { return (oop*)_holder; } 731 operator narrowOop*() { 732 return (narrowOop*)((uintptr_t)_holder & ~COMPRESSED_OOP_MASK); 733 } 734 735 StarTask& operator=(const StarTask& t) { 736 _holder = t._holder; 737 return *this; 738 } 739 volatile StarTask& operator=(const volatile StarTask& t) volatile { 740 _holder = t._holder; 741 return *this; 742 } 743 744 bool is_narrow() const { 745 return (((uintptr_t)_holder & COMPRESSED_OOP_MASK) != 0); 746 } 747 }; 748 749 class ObjArrayTask 750 { 751 public: 752 ObjArrayTask(oop o = NULL, int idx = 0): _obj(o), _index(idx) { } 753 ObjArrayTask(oop o, size_t idx): _obj(o), _index(int(idx)) { 754 assert(idx <= size_t(max_jint), "too big"); 755 } 756 ObjArrayTask(const ObjArrayTask& t): _obj(t._obj), _index(t._index) { } 757 758 ObjArrayTask& operator =(const ObjArrayTask& t) { 759 _obj = t._obj; 760 _index = t._index; 761 return *this; 762 } 763 volatile ObjArrayTask& 764 operator =(const volatile ObjArrayTask& t) volatile { 765 _obj = t._obj; 766 _index = t._index; 767 return *this; 768 } 769 770 inline oop obj() const { return _obj; } 771 inline int index() const { return _index; } 772 773 DEBUG_ONLY(bool is_valid() const); // Tasks to be pushed/popped must be valid. 774 775 private: 776 oop _obj; 777 int _index; 778 }; 779 780 #ifdef _MSC_VER 781 #pragma warning(pop) 782 #endif 783 784 typedef OverflowTaskQueue<StarTask, mtClass> OopStarTaskQueue; 785 typedef GenericTaskQueueSet<OopStarTaskQueue, mtClass> OopStarTaskQueueSet; 786 787 typedef OverflowTaskQueue<size_t, mtInternal> RegionTaskQueue; 788 typedef GenericTaskQueueSet<RegionTaskQueue, mtClass> RegionTaskQueueSet; 789 790 791 #endif // SHARE_VM_UTILITIES_TASKQUEUE_HPP