1 /*
2 * Copyright (c) 2001, 2018, 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_GC_SHARED_TASKQUEUE_HPP
26 #define SHARE_VM_GC_SHARED_TASKQUEUE_HPP
27
28 #include "memory/allocation.hpp"
29 #include "oops/oopsHierarchy.hpp"
30 #include "utilities/ostream.hpp"
31 #include "utilities/stack.hpp"
32
33 // Simple TaskQueue stats that are collected by default in debug builds.
34
35 #if !defined(TASKQUEUE_STATS) && defined(ASSERT)
36 #define TASKQUEUE_STATS 1
37 #elif !defined(TASKQUEUE_STATS)
38 #define TASKQUEUE_STATS 0
39 #endif
40
41 #if TASKQUEUE_STATS
42 #define TASKQUEUE_STATS_ONLY(code) code
43 #else
44 #define TASKQUEUE_STATS_ONLY(code)
45 #endif // TASKQUEUE_STATS
46
47 #if TASKQUEUE_STATS
48 class TaskQueueStats {
49 public:
50 enum StatId {
51 push, // number of taskqueue pushes
52 pop, // number of taskqueue pops
53 pop_slow, // subset of taskqueue pops that were done slow-path
54 steal_attempt, // number of taskqueue steal attempts
55 steal, // number of taskqueue steals
56 overflow, // number of overflow pushes
57 overflow_max_len, // max length of overflow stack
58 last_stat_id
59 };
60
61 public:
62 inline TaskQueueStats() { reset(); }
63
64 inline void record_push() { ++_stats[push]; }
65 inline void record_pop() { ++_stats[pop]; }
66 inline void record_pop_slow() { record_pop(); ++_stats[pop_slow]; }
67 inline void record_steal(bool success);
68 inline void record_overflow(size_t new_length);
69
70 TaskQueueStats & operator +=(const TaskQueueStats & addend);
71
72 inline size_t get(StatId id) const { return _stats[id]; }
73 inline const size_t* get() const { return _stats; }
74
75 inline void reset();
76
77 // Print the specified line of the header (does not include a line separator).
78 static void print_header(unsigned int line, outputStream* const stream = tty,
79 unsigned int width = 10);
80 // Print the statistics (does not include a line separator).
81 void print(outputStream* const stream = tty, unsigned int width = 10) const;
82
83 DEBUG_ONLY(void verify() const;)
84
85 private:
86 size_t _stats[last_stat_id];
87 static const char * const _names[last_stat_id];
88 };
89
90 void TaskQueueStats::record_steal(bool success) {
91 ++_stats[steal_attempt];
92 if (success) ++_stats[steal];
93 }
94
95 void TaskQueueStats::record_overflow(size_t new_len) {
96 ++_stats[overflow];
97 if (new_len > _stats[overflow_max_len]) _stats[overflow_max_len] = new_len;
98 }
99
100 void TaskQueueStats::reset() {
101 memset(_stats, 0, sizeof(_stats));
102 }
103 #endif // TASKQUEUE_STATS
104
105 // TaskQueueSuper collects functionality common to all GenericTaskQueue instances.
106
107 template <unsigned int N, MEMFLAGS F>
108 class TaskQueueSuper: public CHeapObj<F> {
109 protected:
110 // Internal type for indexing the queue; also used for the tag.
111 typedef NOT_LP64(uint16_t) LP64_ONLY(uint32_t) idx_t;
112
113 // The first free element after the last one pushed (mod N).
114 volatile uint _bottom;
115
116 enum { MOD_N_MASK = N - 1 };
117
118 class Age {
119 public:
120 Age(size_t data = 0) { _data = data; }
121 Age(const Age& age) { _data = age._data; }
122 Age(idx_t top, idx_t tag) { _fields._top = top; _fields._tag = tag; }
123
124 Age get() const volatile { return _data; }
125 void set(Age age) volatile { _data = age._data; }
126
127 idx_t top() const volatile { return _fields._top; }
128 idx_t tag() const volatile { return _fields._tag; }
129
130 // Increment top; if it wraps, increment tag also.
131 void increment() {
132 _fields._top = increment_index(_fields._top);
133 if (_fields._top == 0) ++_fields._tag;
134 }
135
136 Age cmpxchg(const Age new_age, const Age old_age) volatile;
137
138 bool operator ==(const Age& other) const { return _data == other._data; }
139
140 private:
141 struct fields {
142 idx_t _top;
143 idx_t _tag;
144 };
145 union {
146 size_t _data;
147 fields _fields;
148 };
149 };
150
151 volatile Age _age;
152
153 // These both operate mod N.
154 static uint increment_index(uint ind) {
155 return (ind + 1) & MOD_N_MASK;
156 }
157 static uint decrement_index(uint ind) {
158 return (ind - 1) & MOD_N_MASK;
159 }
160
161 // Returns a number in the range [0..N). If the result is "N-1", it should be
162 // interpreted as 0.
163 uint dirty_size(uint bot, uint top) const {
164 return (bot - top) & MOD_N_MASK;
165 }
166
167 // Returns the size corresponding to the given "bot" and "top".
168 uint size(uint bot, uint top) const {
169 uint sz = dirty_size(bot, top);
170 // Has the queue "wrapped", so that bottom is less than top? There's a
171 // complicated special case here. A pair of threads could perform pop_local
172 // and pop_global operations concurrently, starting from a state in which
173 // _bottom == _top+1. The pop_local could succeed in decrementing _bottom,
174 // and the pop_global in incrementing _top (in which case the pop_global
175 // will be awarded the contested queue element.) The resulting state must
176 // be interpreted as an empty queue. (We only need to worry about one such
177 // event: only the queue owner performs pop_local's, and several concurrent
178 // threads attempting to perform the pop_global will all perform the same
179 // CAS, and only one can succeed.) Any stealing thread that reads after
180 // either the increment or decrement will see an empty queue, and will not
181 // join the competitors. The "sz == -1 || sz == N-1" state will not be
182 // modified by concurrent queues, so the owner thread can reset the state to
183 // _bottom == top so subsequent pushes will be performed normally.
184 return (sz == N - 1) ? 0 : sz;
185 }
186
187 public:
188 TaskQueueSuper() : _bottom(0), _age() {}
189
190 // Return true if the TaskQueue contains/does not contain any tasks.
191 bool peek() const { return _bottom != _age.top(); }
192 bool is_empty() const { return size() == 0; }
193
194 // Return an estimate of the number of elements in the queue.
195 // The "careful" version admits the possibility of pop_local/pop_global
196 // races.
197 uint size() const {
198 return size(_bottom, _age.top());
199 }
200
201 uint dirty_size() const {
202 return dirty_size(_bottom, _age.top());
203 }
204
205 void set_empty() {
206 _bottom = 0;
207 _age.set(0);
208 }
209
210 // Maximum number of elements allowed in the queue. This is two less
211 // than the actual queue size, for somewhat complicated reasons.
212 uint max_elems() const { return N - 2; }
213
214 // Total size of queue.
215 static const uint total_size() { return N; }
216
217 TASKQUEUE_STATS_ONLY(TaskQueueStats stats;)
218 };
219
220 //
221 // GenericTaskQueue implements an ABP, Aurora-Blumofe-Plaxton, double-
222 // ended-queue (deque), intended for use in work stealing. Queue operations
223 // are non-blocking.
224 //
225 // A queue owner thread performs push() and pop_local() operations on one end
226 // of the queue, while other threads may steal work using the pop_global()
227 // method.
228 //
229 // The main difference to the original algorithm is that this
230 // implementation allows wrap-around at the end of its allocated
231 // storage, which is an array.
232 //
233 // The original paper is:
234 //
235 // Arora, N. S., Blumofe, R. D., and Plaxton, C. G.
236 // Thread scheduling for multiprogrammed multiprocessors.
237 // Theory of Computing Systems 34, 2 (2001), 115-144.
238 //
239 // The following paper provides an correctness proof and an
240 // implementation for weakly ordered memory models including (pseudo-)
241 // code containing memory barriers for a Chase-Lev deque. Chase-Lev is
242 // similar to ABP, with the main difference that it allows resizing of the
243 // underlying storage:
244 //
245 // Le, N. M., Pop, A., Cohen A., and Nardell, F. Z.
246 // Correct and efficient work-stealing for weak memory models
247 // Proceedings of the 18th ACM SIGPLAN symposium on Principles and
248 // practice of parallel programming (PPoPP 2013), 69-80
249 //
250
251 template <class E, MEMFLAGS F, unsigned int N = TASKQUEUE_SIZE>
252 class GenericTaskQueue: public TaskQueueSuper<N, F> {
253 protected:
254 typedef typename TaskQueueSuper<N, F>::Age Age;
255 typedef typename TaskQueueSuper<N, F>::idx_t idx_t;
256
257 using TaskQueueSuper<N, F>::_bottom;
258 using TaskQueueSuper<N, F>::_age;
259 using TaskQueueSuper<N, F>::increment_index;
260 using TaskQueueSuper<N, F>::decrement_index;
261 using TaskQueueSuper<N, F>::dirty_size;
262
263 public:
264 using TaskQueueSuper<N, F>::max_elems;
265 using TaskQueueSuper<N, F>::size;
266
267 #if TASKQUEUE_STATS
268 using TaskQueueSuper<N, F>::stats;
269 #endif
270
271 private:
272 // Slow paths for push, pop_local. (pop_global has no fast path.)
273 bool push_slow(E t, uint dirty_n_elems);
274 bool pop_local_slow(uint localBot, Age oldAge);
275
276 public:
277 typedef E element_type;
278
279 // Initializes the queue to empty.
280 GenericTaskQueue();
281
282 void initialize();
283
284 // Push the task "t" on the queue. Returns "false" iff the queue is full.
285 inline bool push(E t);
286
287 // Attempts to claim a task from the "local" end of the queue (the most
288 // recently pushed) as long as the number of entries exceeds the threshold.
289 // If successful, returns true and sets t to the task; otherwise, returns false
290 // (the queue is empty or the number of elements below the threshold).
291 inline bool pop_local(volatile E& t, uint threshold = 0);
292
293 // Like pop_local(), but uses the "global" end of the queue (the least
294 // recently pushed).
295 bool pop_global(volatile E& t);
296
297 // Delete any resource associated with the queue.
298 ~GenericTaskQueue();
299
300 // Apply fn to each element in the task queue. The queue must not
301 // be modified while iterating.
302 template<typename Fn> void iterate(Fn fn);
303
304 private:
305 // Element array.
306 volatile E* _elems;
307 };
308
309 template<class E, MEMFLAGS F, unsigned int N>
310 GenericTaskQueue<E, F, N>::GenericTaskQueue() {
311 assert(sizeof(Age) == sizeof(size_t), "Depends on this.");
312 }
313
314 // OverflowTaskQueue is a TaskQueue that also includes an overflow stack for
315 // elements that do not fit in the TaskQueue.
316 //
317 // This class hides two methods from super classes:
318 //
319 // push() - push onto the task queue or, if that fails, onto the overflow stack
320 // is_empty() - return true if both the TaskQueue and overflow stack are empty
321 //
322 // Note that size() is not hidden--it returns the number of elements in the
323 // TaskQueue, and does not include the size of the overflow stack. This
324 // simplifies replacement of GenericTaskQueues with OverflowTaskQueues.
325 template<class E, MEMFLAGS F, unsigned int N = TASKQUEUE_SIZE>
326 class OverflowTaskQueue: public GenericTaskQueue<E, F, N>
327 {
328 public:
329 typedef Stack<E, F> overflow_t;
330 typedef GenericTaskQueue<E, F, N> taskqueue_t;
331
332 TASKQUEUE_STATS_ONLY(using taskqueue_t::stats;)
333
334 // Push task t onto the queue or onto the overflow stack. Return true.
335 inline bool push(E t);
336 // Try to push task t onto the queue only. Returns true if successful, false otherwise.
337 inline bool try_push_to_taskqueue(E t);
338
339 // Attempt to pop from the overflow stack; return true if anything was popped.
340 inline bool pop_overflow(E& t);
341
342 inline overflow_t* overflow_stack() { return &_overflow_stack; }
343
344 inline bool taskqueue_empty() const { return taskqueue_t::is_empty(); }
345 inline bool overflow_empty() const { return _overflow_stack.is_empty(); }
346 inline bool is_empty() const {
347 return taskqueue_empty() && overflow_empty();
348 }
349
350 private:
351 overflow_t _overflow_stack;
352 };
353
354 template<class E, MEMFLAGS F, unsigned int N = TASKQUEUE_SIZE>
355 class BufferedOverflowTaskQueue: public OverflowTaskQueue<E, F, N>
356 {
357 public:
358 typedef OverflowTaskQueue<E, F, N> taskqueue_t;
359
360 BufferedOverflowTaskQueue() : _buf_empty(true) {};
361
362 TASKQUEUE_STATS_ONLY(using taskqueue_t::stats;)
363
364 // Push task t onto:
365 // - first, try buffer;
366 // - then, try the queue;
367 // - then, overflow stack.
368 // Return true.
369 inline bool push(E t);
370
371 // Attempt to pop from the buffer; return true if anything was popped.
372 inline bool pop_buffer(E &t);
373
374 inline void clear_buffer() { _buf_empty = true; }
375 inline bool buffer_empty() const { return _buf_empty; }
376 inline bool is_empty() const {
377 return taskqueue_t::is_empty() && buffer_empty();
378 }
379
380 private:
381 bool _buf_empty;
382 E _elem;
383 };
384
385 class TaskQueueSetSuper {
386 protected:
387 static int randomParkAndMiller(int* seed0);
388 public:
389 // Returns "true" if some TaskQueue in the set contains a task.
390 virtual bool peek() = 0;
391 virtual size_t tasks() = 0;
392 };
393
394 template <MEMFLAGS F> class TaskQueueSetSuperImpl: public CHeapObj<F>, public TaskQueueSetSuper {
395 };
396
397 template<class T, MEMFLAGS F>
398 class GenericTaskQueueSet: public TaskQueueSetSuperImpl<F> {
399 private:
400 uint _n;
401 T** _queues;
402
403 public:
404 typedef typename T::element_type E;
405
406 GenericTaskQueueSet(int n);
407 ~GenericTaskQueueSet();
408
409 bool steal_best_of_2(uint queue_num, int* seed, E& t);
410
411 void register_queue(uint i, T* q);
412
413 T* queue(uint n);
414
415 // The thread with queue number "queue_num" (and whose random number seed is
416 // at "seed") is trying to steal a task from some other queue. (It may try
417 // several queues, according to some configuration parameter.) If some steal
418 // succeeds, returns "true" and sets "t" to the stolen task, otherwise returns
419 // false.
420 bool steal(uint queue_num, int* seed, E& t);
421
422 bool peek();
423 size_t tasks();
424
425 uint size() const { return _n; }
426 };
427
428 template<class T, MEMFLAGS F> void
429 GenericTaskQueueSet<T, F>::register_queue(uint i, T* q) {
430 assert(i < _n, "index out of range.");
431 _queues[i] = q;
432 }
433
434 template<class T, MEMFLAGS F> T*
435 GenericTaskQueueSet<T, F>::queue(uint i) {
436 return _queues[i];
437 }
438
439 template<class T, MEMFLAGS F>
440 bool GenericTaskQueueSet<T, F>::peek() {
441 // Try all the queues.
442 for (uint j = 0; j < _n; j++) {
443 if (_queues[j]->peek())
444 return true;
445 }
446 return false;
447 }
448
449 template<class T, MEMFLAGS F>
450 size_t GenericTaskQueueSet<T, F>::tasks() {
451 size_t n = 0;
452 for (uint j = 0; j < _n; j++) {
453 n += _queues[j]->size();
454 }
455 return n;
456 }
457
458
459 // When to terminate from the termination protocol.
460 class TerminatorTerminator: public CHeapObj<mtInternal> {
461 public:
462 virtual bool should_exit_termination() = 0;
463 virtual bool should_force_termination() { return false; }
464 };
465
466 // A class to aid in the termination of a set of parallel tasks using
467 // TaskQueueSet's for work stealing.
468
469 #undef TRACESPINNING
470
471 class ParallelTaskTerminator: public StackObj {
472 protected:
473 uint _n_threads;
474 TaskQueueSetSuper* _queue_set;
475 volatile uint _offered_termination;
476
477 #ifdef TRACESPINNING
478 static uint _total_yields;
479 static uint _total_spins;
480 static uint _total_peeks;
481 #endif
482
483 bool peek_in_queue_set();
484 protected:
485 virtual void yield();
486 void sleep(uint millis);
487
488 public:
489
490 // "n_threads" is the number of threads to be terminated. "queue_set" is a
491 // queue sets of work queues of other threads.
492 ParallelTaskTerminator(uint n_threads, TaskQueueSetSuper* queue_set);
493
494 // The current thread has no work, and is ready to terminate if everyone
495 // else is. If returns "true", all threads are terminated. If returns
496 // "false", available work has been observed in one of the task queues,
497 // so the global task is not complete.
498 // If force is set to true, it terminates even if there's remaining work left
499 bool offer_termination() {
500 return offer_termination(NULL);
501 }
502
503 // As above, but it also terminates if the should_exit_termination()
504 // method of the terminator parameter returns true. If terminator is
505 // NULL, then it is ignored.
506 // If force is set to true, it terminates even if there's remaining work left
507 virtual bool offer_termination(TerminatorTerminator* terminator);
508
509 // Reset the terminator, so that it may be reused again.
510 // The caller is responsible for ensuring that this is done
511 // in an MT-safe manner, once the previous round of use of
512 // the terminator is finished.
513 void reset_for_reuse();
514 // Same as above but the number of parallel threads is set to the
515 // given number.
516 void reset_for_reuse(uint n_threads);
517
518 #ifdef TRACESPINNING
519 static uint total_yields() { return _total_yields; }
520 static uint total_spins() { return _total_spins; }
521 static uint total_peeks() { return _total_peeks; }
522 static void print_termination_counts();
523 #endif
524 };
525
526 typedef GenericTaskQueue<oop, mtGC> OopTaskQueue;
527 typedef GenericTaskQueueSet<OopTaskQueue, mtGC> OopTaskQueueSet;
528
529 #ifdef _MSC_VER
530 #pragma warning(push)
531 // warning C4522: multiple assignment operators specified
532 #pragma warning(disable:4522)
533 #endif
534
535 // This is a container class for either an oop* or a narrowOop*.
536 // Both are pushed onto a task queue and the consumer will test is_narrow()
537 // to determine which should be processed.
538 class StarTask {
539 void* _holder; // either union oop* or narrowOop*
540
541 enum { COMPRESSED_OOP_MASK = 1 };
542
543 public:
544 StarTask(narrowOop* p) {
545 assert(((uintptr_t)p & COMPRESSED_OOP_MASK) == 0, "Information loss!");
546 _holder = (void *)((uintptr_t)p | COMPRESSED_OOP_MASK);
547 }
548 StarTask(oop* p) {
549 assert(((uintptr_t)p & COMPRESSED_OOP_MASK) == 0, "Information loss!");
550 _holder = (void*)p;
551 }
552 StarTask() { _holder = NULL; }
553 operator oop*() { return (oop*)_holder; }
554 operator narrowOop*() {
555 return (narrowOop*)((uintptr_t)_holder & ~COMPRESSED_OOP_MASK);
556 }
557
558 StarTask& operator=(const StarTask& t) {
559 _holder = t._holder;
560 return *this;
561 }
562 volatile StarTask& operator=(const volatile StarTask& t) volatile {
563 _holder = t._holder;
564 return *this;
565 }
566
567 bool is_narrow() const {
568 return (((uintptr_t)_holder & COMPRESSED_OOP_MASK) != 0);
569 }
570 };
571
572 class ObjArrayTask
573 {
574 public:
575 ObjArrayTask(oop o = NULL, int idx = 0): _obj(o), _index(idx) { }
576 ObjArrayTask(oop o, size_t idx): _obj(o), _index(int(idx)) {
577 assert(idx <= size_t(max_jint), "too big");
578 }
579 ObjArrayTask(const ObjArrayTask& t): _obj(t._obj), _index(t._index) { }
580
581 ObjArrayTask& operator =(const ObjArrayTask& t) {
582 _obj = t._obj;
583 _index = t._index;
584 return *this;
585 }
586 volatile ObjArrayTask&
587 operator =(const volatile ObjArrayTask& t) volatile {
588 (void)const_cast<oop&>(_obj = t._obj);
589 _index = t._index;
590 return *this;
591 }
592
593 inline oop obj() const { return _obj; }
594 inline int index() const { return _index; }
595
596 DEBUG_ONLY(bool is_valid() const); // Tasks to be pushed/popped must be valid.
597
598 private:
599 oop _obj;
600 int _index;
601 };
602
603 // ObjArrayChunkedTask
604 //
605 // Encodes both regular oops, and the array oops plus chunking data for parallel array processing.
606 // The design goal is to make the regular oop ops very fast, because that would be the prevailing
607 // case. On the other hand, it should not block parallel array processing from efficiently dividing
608 // the array work.
609 //
610 // The idea is to steal the bits from the 64-bit oop to encode array data, if needed. For the
611 // proper divide-and-conquer strategies, we want to encode the "blocking" data. It turns out, the
612 // most efficient way to do this is to encode the array block as (chunk * 2^pow), where it is assumed
613 // that the block has the size of 2^pow. This requires for pow to have only 5 bits (2^32) to encode
614 // all possible arrays.
615 //
616 // |---------oop---------|-pow-|--chunk---|
617 // 0 49 54 64
618 //
619 // By definition, chunk == 0 means "no chunk", i.e. chunking starts from 1.
620 //
621 // This encoding gives a few interesting benefits:
622 //
623 // a) Encoding/decoding regular oops is very simple, because the upper bits are zero in that task:
624 //
625 // |---------oop---------|00000|0000000000| // no chunk data
626 //
627 // This helps the most ubiquitous path. The initialization amounts to putting the oop into the word
628 // with zero padding. Testing for "chunkedness" is testing for zero with chunk mask.
629 //
630 // b) Splitting tasks for divide-and-conquer is possible. Suppose we have chunk <C, P> that covers
631 // interval [ (C-1)*2^P; C*2^P ). We can then split it into two chunks:
632 // <2*C - 1, P-1>, that covers interval [ (2*C - 2)*2^(P-1); (2*C - 1)*2^(P-1) )
633 // <2*C, P-1>, that covers interval [ (2*C - 1)*2^(P-1); 2*C*2^(P-1) )
634 //
635 // Observe that the union of these two intervals is:
636 // [ (2*C - 2)*2^(P-1); 2*C*2^(P-1) )
637 //
638 // ...which is the original interval:
639 // [ (C-1)*2^P; C*2^P )
640 //
641 // c) The divide-and-conquer strategy could even start with chunk <1, round-log2-len(arr)>, and split
642 // down in the parallel threads, which alleviates the upfront (serial) splitting costs.
643 //
644 // Encoding limitations caused by current bitscales mean:
645 // 10 bits for chunk: max 1024 blocks per array
646 // 5 bits for power: max 2^32 array
647 // 49 bits for oop: max 512 TB of addressable space
648 //
649 // Stealing bits from oop trims down the addressable space. Stealing too few bits for chunk ID limits
650 // potential parallelism. Stealing too few bits for pow limits the maximum array size that can be handled.
651 // In future, these might be rebalanced to favor one degree of freedom against another. For example,
652 // if/when Arrays 2.0 bring 2^64-sized arrays, we might need to steal another bit for power. We could regain
653 // some bits back if chunks are counted in ObjArrayMarkingStride units.
654 //
655 // There is also a fallback version that uses plain fields, when we don't have enough space to steal the
656 // bits from the native pointer. It is useful to debug the _LP64 version.
657 //
658 #ifdef _LP64
659 class ObjArrayChunkedTask
660 {
661 public:
662 enum {
663 chunk_bits = 10,
664 pow_bits = 5,
665 oop_bits = sizeof(uintptr_t)*8 - chunk_bits - pow_bits,
666 };
667 enum {
668 oop_shift = 0,
669 pow_shift = oop_shift + oop_bits,
670 chunk_shift = pow_shift + pow_bits,
671 };
672
673 public:
674 ObjArrayChunkedTask(oop o = NULL) {
675 _obj = ((uintptr_t)(void*) o) << oop_shift;
676 }
677 ObjArrayChunkedTask(oop o, int chunk, int mult) {
678 assert(0 <= chunk && chunk < nth_bit(chunk_bits), "chunk is sane: %d", chunk);
679 assert(0 <= mult && mult < nth_bit(pow_bits), "pow is sane: %d", mult);
680 uintptr_t t_b = ((uintptr_t) chunk) << chunk_shift;
681 uintptr_t t_m = ((uintptr_t) mult) << pow_shift;
682 uintptr_t obj = (uintptr_t)(void*)o;
683 assert(obj < nth_bit(oop_bits), "obj ref is sane: " PTR_FORMAT, obj);
684 intptr_t t_o = obj << oop_shift;
685 _obj = t_o | t_m | t_b;
686 }
687 ObjArrayChunkedTask(const ObjArrayChunkedTask& t): _obj(t._obj) { }
688
689 ObjArrayChunkedTask& operator =(const ObjArrayChunkedTask& t) {
690 _obj = t._obj;
691 return *this;
692 }
693 volatile ObjArrayChunkedTask&
694 operator =(const volatile ObjArrayChunkedTask& t) volatile {
695 (void)const_cast<uintptr_t&>(_obj = t._obj);
696 return *this;
697 }
698
699 inline oop obj() const { return (oop) reinterpret_cast<void*>((_obj >> oop_shift) & right_n_bits(oop_bits)); }
700 inline int chunk() const { return (int) (_obj >> chunk_shift) & right_n_bits(chunk_bits); }
701 inline int pow() const { return (int) ((_obj >> pow_shift) & right_n_bits(pow_bits)); }
702 inline bool is_not_chunked() const { return (_obj & ~right_n_bits(oop_bits + pow_bits)) == 0; }
703
704 DEBUG_ONLY(bool is_valid() const); // Tasks to be pushed/popped must be valid.
705
706 static size_t max_addressable() {
707 return nth_bit(oop_bits);
708 }
709
710 static int chunk_size() {
711 return nth_bit(chunk_bits);
712 }
713
714 private:
715 uintptr_t _obj;
716 };
717 #else
718 class ObjArrayChunkedTask
719 {
720 public:
721 enum {
722 chunk_bits = 10,
723 pow_bits = 5,
724 };
725 public:
726 ObjArrayChunkedTask(oop o = NULL, int chunk = 0, int pow = 0): _obj(o) {
727 assert(0 <= chunk && chunk < nth_bit(chunk_bits), "chunk is sane: %d", chunk);
728 assert(0 <= pow && pow < nth_bit(pow_bits), "pow is sane: %d", pow);
729 _chunk = chunk;
730 _pow = pow;
731 }
732 ObjArrayChunkedTask(const ObjArrayChunkedTask& t): _obj(t._obj), _chunk(t._chunk), _pow(t._pow) { }
733
734 ObjArrayChunkedTask& operator =(const ObjArrayChunkedTask& t) {
735 _obj = t._obj;
736 _chunk = t._chunk;
737 _pow = t._pow;
738 return *this;
739 }
740 volatile ObjArrayChunkedTask&
741 operator =(const volatile ObjArrayChunkedTask& t) volatile {
742 (void)const_cast<oop&>(_obj = t._obj);
743 _chunk = t._chunk;
744 _pow = t._pow;
745 return *this;
746 }
747
748 inline oop obj() const { return _obj; }
749 inline int chunk() const { return _chunk; }
750 inline int pow() const { return _pow; }
751
752 inline bool is_not_chunked() const { return _chunk == 0; }
753
754 DEBUG_ONLY(bool is_valid() const); // Tasks to be pushed/popped must be valid.
755
756 static size_t max_addressable() {
757 return sizeof(oop);
758 }
759
760 static int chunk_size() {
761 return nth_bit(chunk_bits);
762 }
763
764 private:
765 oop _obj;
766 int _chunk;
767 int _pow;
768 };
769 #endif
770
771 #ifdef _MSC_VER
772 #pragma warning(pop)
773 #endif
774
775 typedef OverflowTaskQueue<StarTask, mtGC> OopStarTaskQueue;
776 typedef GenericTaskQueueSet<OopStarTaskQueue, mtGC> OopStarTaskQueueSet;
777
778 typedef OverflowTaskQueue<size_t, mtGC> RegionTaskQueue;
779 typedef GenericTaskQueueSet<RegionTaskQueue, mtGC> RegionTaskQueueSet;
780
781 #endif // SHARE_VM_GC_SHARED_TASKQUEUE_HPP