1 |
/* |
2 |
* Written by Doug Lea with assistance from members of JCP JSR-166 |
3 |
* Expert Group and released to the public domain, as explained at |
4 |
* http://creativecommons.org/publicdomain/zero/1.0/ |
5 |
*/ |
6 |
|
7 |
package jsr166e; |
8 |
|
9 |
/** |
10 |
* A {@link ForkJoinTask} with a completion action performed when |
11 |
* triggered and there are no remaining pending |
12 |
* actions. CountedCompleters are in general more robust in the |
13 |
* presence of subtask stalls and blockage than are other forms of |
14 |
* ForkJoinTasks, but are less intuitive to program. Uses of |
15 |
* CountedCompleter are similar to those of other completion based |
16 |
* components (such as {@link java.nio.channels.CompletionHandler}) |
17 |
* except that multiple <em>pending</em> completions may be necessary |
18 |
* to trigger the {@link #onCompletion} action, not just one. Unless |
19 |
* initialized otherwise, the {@link #getPendingCount pending count} |
20 |
* starts at zero, but may be (atomically) changed using methods |
21 |
* {@link #setPendingCount}, {@link #addToPendingCount}, and {@link |
22 |
* #compareAndSetPendingCount}. Upon invocation of {@link |
23 |
* #tryComplete}, if the pending action count is nonzero, it is |
24 |
* decremented; otherwise, the completion action is performed, and if |
25 |
* this completer itself has a completer, the process is continued |
26 |
* with its completer. As is the case with related synchronization |
27 |
* components such as {@link java.util.concurrent.Phaser Phaser} and |
28 |
* {@link java.util.concurrent.Semaphore Semaphore}, these methods |
29 |
* affect only internal counts; they do not establish any further |
30 |
* internal bookkeeping. In particular, the identities of pending |
31 |
* tasks are not maintained. As illustrated below, you can create |
32 |
* subclasses that do record some or all pending tasks or their |
33 |
* results when needed. As illustrated below, utility methods |
34 |
* supporting customization of completion traversals are also |
35 |
* provided. However, because CountedCompleters provide only basic |
36 |
* synchronization mechanisms, it may be useful to create further |
37 |
* abstract subclasses that maintain linkages, fields, and additional |
38 |
* support methods appropriate for a set of related usages. |
39 |
* |
40 |
* <p>A concrete CountedCompleter class must define method {@link |
41 |
* #compute}, that should in most cases (as illustrated below), invoke |
42 |
* {@code tryComplete()} once before returning. The class may also |
43 |
* optionally override method {@link #onCompletion} to perform an |
44 |
* action upon normal completion, and method {@link |
45 |
* #onExceptionalCompletion} to perform an action upon any exception. |
46 |
* |
47 |
* <p>CountedCompleters most often do not bear results, in which case |
48 |
* they are normally declared as {@code CountedCompleter<Void>}, and |
49 |
* will always return {@code null} as a result value. In other cases, |
50 |
* you should override method {@link #getRawResult} to provide a |
51 |
* result from {@code join(), invoke()}, and related methods. In |
52 |
* general, this method should return the value of a field (or a |
53 |
* function of one or more fields) of the CountedCompleter object that |
54 |
* holds the result upon completion. Method {@link #setRawResult} by |
55 |
* default plays no role in CountedCompleters. It is possible, but |
56 |
* rarely applicable, to override this method to maintain other |
57 |
* objects or fields holding result data. |
58 |
* |
59 |
* <p>A CountedCompleter that does not itself have a completer (i.e., |
60 |
* one for which {@link #getCompleter} returns {@code null}) can be |
61 |
* used as a regular ForkJoinTask with this added functionality. |
62 |
* However, any completer that in turn has another completer serves |
63 |
* only as an internal helper for other computations, so its own task |
64 |
* status (as reported in methods such as {@link ForkJoinTask#isDone}) |
65 |
* is arbitrary; this status changes only upon explicit invocations of |
66 |
* {@link #complete}, {@link ForkJoinTask#cancel}, {@link |
67 |
* ForkJoinTask#completeExceptionally} or upon exceptional completion |
68 |
* of method {@code compute}. Upon any exceptional completion, the |
69 |
* exception may be relayed to a task's completer (and its completer, |
70 |
* and so on), if one exists and it has not otherwise already |
71 |
* completed. Similarly, cancelling an internal CountedCompleter has |
72 |
* only a local effect on that completer, so is not often useful. |
73 |
* |
74 |
* <p><b>Sample Usages.</b> |
75 |
* |
76 |
* <p><b>Parallel recursive decomposition.</b> CountedCompleters may |
77 |
* be arranged in trees similar to those often used with {@link |
78 |
* RecursiveAction}s, although the constructions involved in setting |
79 |
* them up typically vary. Here, the completer of each task is its |
80 |
* parent in the computation tree. Even though they entail a bit more |
81 |
* bookkeeping, CountedCompleters may be better choices when applying |
82 |
* a possibly time-consuming operation (that cannot be further |
83 |
* subdivided) to each element of an array or collection; especially |
84 |
* when the operation takes a significantly different amount of time |
85 |
* to complete for some elements than others, either because of |
86 |
* intrinsic variation (for example I/O) or auxiliary effects such as |
87 |
* garbage collection. Because CountedCompleters provide their own |
88 |
* continuations, other threads need not block waiting to perform |
89 |
* them. |
90 |
* |
91 |
* <p>For example, here is an initial version of a class that uses |
92 |
* divide-by-two recursive decomposition to divide work into single |
93 |
* pieces (leaf tasks). Even when work is split into individual calls, |
94 |
* tree-based techniques are usually preferable to directly forking |
95 |
* leaf tasks, because they reduce inter-thread communication and |
96 |
* improve load balancing. In the recursive case, the second of each |
97 |
* pair of subtasks to finish triggers completion of its parent |
98 |
* (because no result combination is performed, the default no-op |
99 |
* implementation of method {@code onCompletion} is not overridden). A |
100 |
* static utility method sets up the base task and invokes it |
101 |
* (here, implicitly using the {@link ForkJoinPool#commonPool()}). |
102 |
* |
103 |
* <pre> {@code |
104 |
* class MyOperation<E> { void apply(E e) { ... } } |
105 |
* |
106 |
* class ForEach<E> extends CountedCompleter<Void> { |
107 |
* |
108 |
* public static <E> void forEach(E[] array, MyOperation<E> op) { |
109 |
* new ForEach<E>(null, array, op, 0, array.length).invoke(); |
110 |
* } |
111 |
* |
112 |
* final E[] array; final MyOperation<E> op; final int lo, hi; |
113 |
* ForEach(CountedCompleter<?> p, E[] array, MyOperation<E> op, int lo, int hi) { |
114 |
* super(p); |
115 |
* this.array = array; this.op = op; this.lo = lo; this.hi = hi; |
116 |
* } |
117 |
* |
118 |
* public void compute() { // version 1 |
119 |
* if (hi - lo >= 2) { |
120 |
* int mid = (lo + hi) >>> 1; |
121 |
* setPendingCount(2); // must set pending count before fork |
122 |
* new ForEach(this, array, op, mid, hi).fork(); // right child |
123 |
* new ForEach(this, array, op, lo, mid).fork(); // left child |
124 |
* } |
125 |
* else if (hi > lo) |
126 |
* op.apply(array[lo]); |
127 |
* tryComplete(); |
128 |
* } |
129 |
* }}</pre> |
130 |
* |
131 |
* This design can be improved by noticing that in the recursive case, |
132 |
* the task has nothing to do after forking its right task, so can |
133 |
* directly invoke its left task before returning. (This is an analog |
134 |
* of tail recursion removal.) Also, because the task returns upon |
135 |
* executing its left task (rather than falling through to invoke |
136 |
* {@code tryComplete}) the pending count is set to one: |
137 |
* |
138 |
* <pre> {@code |
139 |
* class ForEach<E> ... |
140 |
* public void compute() { // version 2 |
141 |
* if (hi - lo >= 2) { |
142 |
* int mid = (lo + hi) >>> 1; |
143 |
* setPendingCount(1); // only one pending |
144 |
* new ForEach(this, array, op, mid, hi).fork(); // right child |
145 |
* new ForEach(this, array, op, lo, mid).compute(); // direct invoke |
146 |
* } |
147 |
* else { |
148 |
* if (hi > lo) |
149 |
* op.apply(array[lo]); |
150 |
* tryComplete(); |
151 |
* } |
152 |
* } |
153 |
* }</pre> |
154 |
* |
155 |
* As a further improvement, notice that the left task need not even |
156 |
* exist. Instead of creating a new one, we can iterate using the |
157 |
* original task, and add a pending count for each fork. Additionally, |
158 |
* because no task in this tree implements an {@link #onCompletion} |
159 |
* method, {@code tryComplete()} can be replaced with {@link |
160 |
* #propagateCompletion}. |
161 |
* |
162 |
* <pre> {@code |
163 |
* class ForEach<E> ... |
164 |
* public void compute() { // version 3 |
165 |
* int l = lo, h = hi; |
166 |
* while (h - l >= 2) { |
167 |
* int mid = (l + h) >>> 1; |
168 |
* addToPendingCount(1); |
169 |
* new ForEach(this, array, op, mid, h).fork(); // right child |
170 |
* h = mid; |
171 |
* } |
172 |
* if (h > l) |
173 |
* op.apply(array[l]); |
174 |
* propagateCompletion(); |
175 |
* } |
176 |
* }</pre> |
177 |
* |
178 |
* Additional improvements of such classes might entail precomputing |
179 |
* pending counts so that they can be established in constructors, |
180 |
* specializing classes for leaf steps, subdividing by say, four, |
181 |
* instead of two per iteration, and using an adaptive threshold |
182 |
* instead of always subdividing down to single elements. |
183 |
* |
184 |
* <p><b>Searching.</b> A tree of CountedCompleters can search for a |
185 |
* value or property in different parts of a data structure, and |
186 |
* report a result in an {@link |
187 |
* java.util.concurrent.atomic.AtomicReference AtomicReference} as |
188 |
* soon as one is found. The others can poll the result to avoid |
189 |
* unnecessary work. (You could additionally {@linkplain #cancel |
190 |
* cancel} other tasks, but it is usually simpler and more efficient |
191 |
* to just let them notice that the result is set and if so skip |
192 |
* further processing.) Illustrating again with an array using full |
193 |
* partitioning (again, in practice, leaf tasks will almost always |
194 |
* process more than one element): |
195 |
* |
196 |
* <pre> {@code |
197 |
* class Searcher<E> extends CountedCompleter<E> { |
198 |
* final E[] array; final AtomicReference<E> result; final int lo, hi; |
199 |
* Searcher(CountedCompleter<?> p, E[] array, AtomicReference<E> result, int lo, int hi) { |
200 |
* super(p); |
201 |
* this.array = array; this.result = result; this.lo = lo; this.hi = hi; |
202 |
* } |
203 |
* public E getRawResult() { return result.get(); } |
204 |
* public void compute() { // similar to ForEach version 3 |
205 |
* int l = lo, h = hi; |
206 |
* while (result.get() == null && h >= l) { |
207 |
* if (h - l >= 2) { |
208 |
* int mid = (l + h) >>> 1; |
209 |
* addToPendingCount(1); |
210 |
* new Searcher(this, array, result, mid, h).fork(); |
211 |
* h = mid; |
212 |
* } |
213 |
* else { |
214 |
* E x = array[l]; |
215 |
* if (matches(x) && result.compareAndSet(null, x)) |
216 |
* quietlyCompleteRoot(); // root task is now joinable |
217 |
* break; |
218 |
* } |
219 |
* } |
220 |
* tryComplete(); // normally complete whether or not found |
221 |
* } |
222 |
* boolean matches(E e) { ... } // return true if found |
223 |
* |
224 |
* public static <E> E search(E[] array) { |
225 |
* return new Searcher<E>(null, array, new AtomicReference<E>(), 0, array.length).invoke(); |
226 |
* } |
227 |
* }}</pre> |
228 |
* |
229 |
* In this example, as well as others in which tasks have no other |
230 |
* effects except to compareAndSet a common result, the trailing |
231 |
* unconditional invocation of {@code tryComplete} could be made |
232 |
* conditional ({@code if (result.get() == null) tryComplete();}) |
233 |
* because no further bookkeeping is required to manage completions |
234 |
* once the root task completes. |
235 |
* |
236 |
* <p><b>Recording subtasks.</b> CountedCompleter tasks that combine |
237 |
* results of multiple subtasks usually need to access these results |
238 |
* in method {@link #onCompletion}. As illustrated in the following |
239 |
* class (that performs a simplified form of map-reduce where mappings |
240 |
* and reductions are all of type {@code E}), one way to do this in |
241 |
* divide and conquer designs is to have each subtask record its |
242 |
* sibling, so that it can be accessed in method {@code onCompletion}. |
243 |
* This technique applies to reductions in which the order of |
244 |
* combining left and right results does not matter; ordered |
245 |
* reductions require explicit left/right designations. Variants of |
246 |
* other streamlinings seen in the above examples may also apply. |
247 |
* |
248 |
* <pre> {@code |
249 |
* class MyMapper<E> { E apply(E v) { ... } } |
250 |
* class MyReducer<E> { E apply(E x, E y) { ... } } |
251 |
* class MapReducer<E> extends CountedCompleter<E> { |
252 |
* final E[] array; final MyMapper<E> mapper; |
253 |
* final MyReducer<E> reducer; final int lo, hi; |
254 |
* MapReducer<E> sibling; |
255 |
* E result; |
256 |
* MapReducer(CountedCompleter<?> p, E[] array, MyMapper<E> mapper, |
257 |
* MyReducer<E> reducer, int lo, int hi) { |
258 |
* super(p); |
259 |
* this.array = array; this.mapper = mapper; |
260 |
* this.reducer = reducer; this.lo = lo; this.hi = hi; |
261 |
* } |
262 |
* public void compute() { |
263 |
* if (hi - lo >= 2) { |
264 |
* int mid = (lo + hi) >>> 1; |
265 |
* MapReducer<E> left = new MapReducer(this, array, mapper, reducer, lo, mid); |
266 |
* MapReducer<E> right = new MapReducer(this, array, mapper, reducer, mid, hi); |
267 |
* left.sibling = right; |
268 |
* right.sibling = left; |
269 |
* setPendingCount(1); // only right is pending |
270 |
* right.fork(); |
271 |
* left.compute(); // directly execute left |
272 |
* } |
273 |
* else { |
274 |
* if (hi > lo) |
275 |
* result = mapper.apply(array[lo]); |
276 |
* tryComplete(); |
277 |
* } |
278 |
* } |
279 |
* public void onCompletion(CountedCompleter<?> caller) { |
280 |
* if (caller != this) { |
281 |
* MapReducer<E> child = (MapReducer<E>)caller; |
282 |
* MapReducer<E> sib = child.sibling; |
283 |
* if (sib == null || sib.result == null) |
284 |
* result = child.result; |
285 |
* else |
286 |
* result = reducer.apply(child.result, sib.result); |
287 |
* } |
288 |
* } |
289 |
* public E getRawResult() { return result; } |
290 |
* |
291 |
* public static <E> E mapReduce(E[] array, MyMapper<E> mapper, MyReducer<E> reducer) { |
292 |
* return new MapReducer<E>(null, array, mapper, reducer, |
293 |
* 0, array.length).invoke(); |
294 |
* } |
295 |
* }}</pre> |
296 |
* |
297 |
* Here, method {@code onCompletion} takes a form common to many |
298 |
* completion designs that combine results. This callback-style method |
299 |
* is triggered once per task, in either of the two different contexts |
300 |
* in which the pending count is, or becomes, zero: (1) by a task |
301 |
* itself, if its pending count is zero upon invocation of {@code |
302 |
* tryComplete}, or (2) by any of its subtasks when they complete and |
303 |
* decrement the pending count to zero. The {@code caller} argument |
304 |
* distinguishes cases. Most often, when the caller is {@code this}, |
305 |
* no action is necessary. Otherwise the caller argument can be used |
306 |
* (usually via a cast) to supply a value (and/or links to other |
307 |
* values) to be combined. Assuming proper use of pending counts, the |
308 |
* actions inside {@code onCompletion} occur (once) upon completion of |
309 |
* a task and its subtasks. No additional synchronization is required |
310 |
* within this method to ensure thread safety of accesses to fields of |
311 |
* this task or other completed tasks. |
312 |
* |
313 |
* <p><b>Completion Traversals</b>. If using {@code onCompletion} to |
314 |
* process completions is inapplicable or inconvenient, you can use |
315 |
* methods {@link #firstComplete} and {@link #nextComplete} to create |
316 |
* custom traversals. For example, to define a MapReducer that only |
317 |
* splits out right-hand tasks in the form of the third ForEach |
318 |
* example, the completions must cooperatively reduce along |
319 |
* unexhausted subtask links, which can be done as follows: |
320 |
* |
321 |
* <pre> {@code |
322 |
* class MapReducer<E> extends CountedCompleter<E> { // version 2 |
323 |
* final E[] array; final MyMapper<E> mapper; |
324 |
* final MyReducer<E> reducer; final int lo, hi; |
325 |
* MapReducer<E> forks, next; // record subtask forks in list |
326 |
* E result; |
327 |
* MapReducer(CountedCompleter<?> p, E[] array, MyMapper<E> mapper, |
328 |
* MyReducer<E> reducer, int lo, int hi, MapReducer<E> next) { |
329 |
* super(p); |
330 |
* this.array = array; this.mapper = mapper; |
331 |
* this.reducer = reducer; this.lo = lo; this.hi = hi; |
332 |
* this.next = next; |
333 |
* } |
334 |
* public void compute() { |
335 |
* int l = lo, h = hi; |
336 |
* while (h - l >= 2) { |
337 |
* int mid = (l + h) >>> 1; |
338 |
* addToPendingCount(1); |
339 |
* (forks = new MapReducer(this, array, mapper, reducer, mid, h, forks)).fork; |
340 |
* h = mid; |
341 |
* } |
342 |
* if (h > l) |
343 |
* result = mapper.apply(array[l]); |
344 |
* // process completions by reducing along and advancing subtask links |
345 |
* for (CountedCompleter<?> c = firstComplete(); c != null; c = c.nextComplete()) { |
346 |
* for (MapReducer t = (MapReducer)c, s = t.forks; s != null; s = t.forks = s.next) |
347 |
* t.result = reducer.apply(t.result, s.result); |
348 |
* } |
349 |
* } |
350 |
* public E getRawResult() { return result; } |
351 |
* |
352 |
* public static <E> E mapReduce(E[] array, MyMapper<E> mapper, MyReducer<E> reducer) { |
353 |
* return new MapReducer<E>(null, array, mapper, reducer, |
354 |
* 0, array.length, null).invoke(); |
355 |
* } |
356 |
* }}</pre> |
357 |
* |
358 |
* <p><b>Triggers.</b> Some CountedCompleters are themselves never |
359 |
* forked, but instead serve as bits of plumbing in other designs; |
360 |
* including those in which the completion of one of more async tasks |
361 |
* triggers another async task. For example: |
362 |
* |
363 |
* <pre> {@code |
364 |
* class HeaderBuilder extends CountedCompleter<...> { ... } |
365 |
* class BodyBuilder extends CountedCompleter<...> { ... } |
366 |
* class PacketSender extends CountedCompleter<...> { |
367 |
* PacketSender(...) { super(null, 1); ... } // trigger on second completion |
368 |
* public void compute() { } // never called |
369 |
* public void onCompletion(CountedCompleter<?> caller) { sendPacket(); } |
370 |
* } |
371 |
* // sample use: |
372 |
* PacketSender p = new PacketSender(); |
373 |
* new HeaderBuilder(p, ...).fork(); |
374 |
* new BodyBuilder(p, ...).fork(); |
375 |
* }</pre> |
376 |
* |
377 |
* @since 1.8 |
378 |
* @author Doug Lea |
379 |
*/ |
380 |
public abstract class CountedCompleter<T> extends ForkJoinTask<T> { |
381 |
private static final long serialVersionUID = 5232453752276485070L; |
382 |
|
383 |
/** This task's completer, or null if none */ |
384 |
final CountedCompleter<?> completer; |
385 |
/** The number of pending tasks until completion */ |
386 |
volatile int pending; |
387 |
|
388 |
/** |
389 |
* Creates a new CountedCompleter with the given completer |
390 |
* and initial pending count. |
391 |
* |
392 |
* @param completer this task's completer, or {@code null} if none |
393 |
* @param initialPendingCount the initial pending count |
394 |
*/ |
395 |
protected CountedCompleter(CountedCompleter<?> completer, |
396 |
int initialPendingCount) { |
397 |
this.completer = completer; |
398 |
this.pending = initialPendingCount; |
399 |
} |
400 |
|
401 |
/** |
402 |
* Creates a new CountedCompleter with the given completer |
403 |
* and an initial pending count of zero. |
404 |
* |
405 |
* @param completer this task's completer, or {@code null} if none |
406 |
*/ |
407 |
protected CountedCompleter(CountedCompleter<?> completer) { |
408 |
this.completer = completer; |
409 |
} |
410 |
|
411 |
/** |
412 |
* Creates a new CountedCompleter with no completer |
413 |
* and an initial pending count of zero. |
414 |
*/ |
415 |
protected CountedCompleter() { |
416 |
this.completer = null; |
417 |
} |
418 |
|
419 |
/** |
420 |
* The main computation performed by this task. |
421 |
*/ |
422 |
public abstract void compute(); |
423 |
|
424 |
/** |
425 |
* Performs an action when method {@link #tryComplete} is invoked |
426 |
* and the pending count is zero, or when the unconditional |
427 |
* method {@link #complete} is invoked. By default, this method |
428 |
* does nothing. You can distinguish cases by checking the |
429 |
* identity of the given caller argument. If not equal to {@code |
430 |
* this}, then it is typically a subtask that may contain results |
431 |
* (and/or links to other results) to combine. |
432 |
* |
433 |
* @param caller the task invoking this method (which may |
434 |
* be this task itself). |
435 |
*/ |
436 |
public void onCompletion(CountedCompleter<?> caller) { |
437 |
} |
438 |
|
439 |
/** |
440 |
* Performs an action when method {@link #completeExceptionally} |
441 |
* is invoked or method {@link #compute} throws an exception, and |
442 |
* this task has not otherwise already completed normally. On |
443 |
* entry to this method, this task {@link |
444 |
* ForkJoinTask#isCompletedAbnormally}. The return value of this |
445 |
* method controls further propagation: If {@code true} and this |
446 |
* task has a completer, then this completer is also completed |
447 |
* exceptionally. The default implementation of this method does |
448 |
* nothing except return {@code true}. |
449 |
* |
450 |
* @param ex the exception |
451 |
* @param caller the task invoking this method (which may |
452 |
* be this task itself). |
453 |
* @return true if this exception should be propagated to this |
454 |
* task's completer, if one exists. |
455 |
*/ |
456 |
public boolean onExceptionalCompletion(Throwable ex, CountedCompleter<?> caller) { |
457 |
return true; |
458 |
} |
459 |
|
460 |
/** |
461 |
* Returns the completer established in this task's constructor, |
462 |
* or {@code null} if none. |
463 |
* |
464 |
* @return the completer |
465 |
*/ |
466 |
public final CountedCompleter<?> getCompleter() { |
467 |
return completer; |
468 |
} |
469 |
|
470 |
/** |
471 |
* Returns the current pending count. |
472 |
* |
473 |
* @return the current pending count |
474 |
*/ |
475 |
public final int getPendingCount() { |
476 |
return pending; |
477 |
} |
478 |
|
479 |
/** |
480 |
* Sets the pending count to the given value. |
481 |
* |
482 |
* @param count the count |
483 |
*/ |
484 |
public final void setPendingCount(int count) { |
485 |
pending = count; |
486 |
} |
487 |
|
488 |
/** |
489 |
* Adds (atomically) the given value to the pending count. |
490 |
* |
491 |
* @param delta the value to add |
492 |
*/ |
493 |
public final void addToPendingCount(int delta) { |
494 |
int c; // note: can replace with intrinsic in jdk8 |
495 |
do {} while (!U.compareAndSwapInt(this, PENDING, c = pending, c+delta)); |
496 |
} |
497 |
|
498 |
/** |
499 |
* Sets (atomically) the pending count to the given count only if |
500 |
* it currently holds the given expected value. |
501 |
* |
502 |
* @param expected the expected value |
503 |
* @param count the new value |
504 |
* @return true if successful |
505 |
*/ |
506 |
public final boolean compareAndSetPendingCount(int expected, int count) { |
507 |
return U.compareAndSwapInt(this, PENDING, expected, count); |
508 |
} |
509 |
|
510 |
/** |
511 |
* If the pending count is nonzero, (atomically) decrements it. |
512 |
* |
513 |
* @return the initial (undecremented) pending count holding on entry |
514 |
* to this method |
515 |
*/ |
516 |
public final int decrementPendingCountUnlessZero() { |
517 |
int c; |
518 |
do {} while ((c = pending) != 0 && |
519 |
!U.compareAndSwapInt(this, PENDING, c, c - 1)); |
520 |
return c; |
521 |
} |
522 |
|
523 |
/** |
524 |
* Returns the root of the current computation; i.e., this |
525 |
* task if it has no completer, else its completer's root. |
526 |
* |
527 |
* @return the root of the current computation |
528 |
*/ |
529 |
public final CountedCompleter<?> getRoot() { |
530 |
CountedCompleter<?> a = this, p; |
531 |
while ((p = a.completer) != null) |
532 |
a = p; |
533 |
return a; |
534 |
} |
535 |
|
536 |
/** |
537 |
* If the pending count is nonzero, decrements the count; |
538 |
* otherwise invokes {@link #onCompletion} and then similarly |
539 |
* tries to complete this task's completer, if one exists, |
540 |
* else marks this task as complete. |
541 |
*/ |
542 |
public final void tryComplete() { |
543 |
CountedCompleter<?> a = this, s = a; |
544 |
for (int c;;) { |
545 |
if ((c = a.pending) == 0) { |
546 |
a.onCompletion(s); |
547 |
if ((a = (s = a).completer) == null) { |
548 |
s.quietlyComplete(); |
549 |
return; |
550 |
} |
551 |
} |
552 |
else if (U.compareAndSwapInt(a, PENDING, c, c - 1)) |
553 |
return; |
554 |
} |
555 |
} |
556 |
|
557 |
/** |
558 |
* Equivalent to {@link #tryComplete} but does not invoke {@link |
559 |
* #onCompletion} along the completion path: If the pending count |
560 |
* is nonzero, decrements the count; otherwise, similarly tries to |
561 |
* complete this task's completer, if one exists, else marks this |
562 |
* task as complete. This method may be useful in cases where |
563 |
* {@code onCompletion} should not, or need not, be invoked for |
564 |
* each completer in a computation. |
565 |
*/ |
566 |
public final void propagateCompletion() { |
567 |
CountedCompleter<?> a = this, s = a; |
568 |
for (int c;;) { |
569 |
if ((c = a.pending) == 0) { |
570 |
if ((a = (s = a).completer) == null) { |
571 |
s.quietlyComplete(); |
572 |
return; |
573 |
} |
574 |
} |
575 |
else if (U.compareAndSwapInt(a, PENDING, c, c - 1)) |
576 |
return; |
577 |
} |
578 |
} |
579 |
|
580 |
/** |
581 |
* Regardless of pending count, invokes {@link #onCompletion}, |
582 |
* marks this task as complete and further triggers {@link |
583 |
* #tryComplete} on this task's completer, if one exists. The |
584 |
* given rawResult is used as an argument to {@link #setRawResult} |
585 |
* before invoking {@link #onCompletion} or marking this task as |
586 |
* complete; its value is meaningful only for classes overriding |
587 |
* {@code setRawResult}. |
588 |
* |
589 |
* <p>This method may be useful when forcing completion as soon as |
590 |
* any one (versus all) of several subtask results are obtained. |
591 |
* However, in the common (and recommended) case in which {@code |
592 |
* setRawResult} is not overridden, this effect can be obtained |
593 |
* more simply using {@code quietlyCompleteRoot();}. |
594 |
* |
595 |
* @param rawResult the raw result |
596 |
*/ |
597 |
public void complete(T rawResult) { |
598 |
CountedCompleter<?> p; |
599 |
setRawResult(rawResult); |
600 |
onCompletion(this); |
601 |
quietlyComplete(); |
602 |
if ((p = completer) != null) |
603 |
p.tryComplete(); |
604 |
} |
605 |
|
606 |
|
607 |
/** |
608 |
* If this task's pending count is zero, returns this task; |
609 |
* otherwise decrements its pending count and returns {@code |
610 |
* null}. This method is designed to be used with {@link |
611 |
* #nextComplete} in completion traversal loops. |
612 |
* |
613 |
* @return this task, if pending count was zero, else {@code null} |
614 |
*/ |
615 |
public final CountedCompleter<?> firstComplete() { |
616 |
for (int c;;) { |
617 |
if ((c = pending) == 0) |
618 |
return this; |
619 |
else if (U.compareAndSwapInt(this, PENDING, c, c - 1)) |
620 |
return null; |
621 |
} |
622 |
} |
623 |
|
624 |
/** |
625 |
* If this task does not have a completer, invokes {@link |
626 |
* ForkJoinTask#quietlyComplete} and returns {@code null}. Or, if |
627 |
* this task's pending count is non-zero, decrements its pending |
628 |
* count and returns {@code null}. Otherwise, returns the |
629 |
* completer. This method can be used as part of a completion |
630 |
* traversal loop for homogeneous task hierarchies: |
631 |
* |
632 |
* <pre> {@code |
633 |
* for (CountedCompleter<?> c = firstComplete(); |
634 |
* c != null; |
635 |
* c = c.nextComplete()) { |
636 |
* // ... process c ... |
637 |
* }}</pre> |
638 |
* |
639 |
* @return the completer, or {@code null} if none |
640 |
*/ |
641 |
public final CountedCompleter<?> nextComplete() { |
642 |
CountedCompleter<?> p; |
643 |
if ((p = completer) != null) |
644 |
return p.firstComplete(); |
645 |
else { |
646 |
quietlyComplete(); |
647 |
return null; |
648 |
} |
649 |
} |
650 |
|
651 |
/** |
652 |
* Equivalent to {@code getRoot().quietlyComplete()}. |
653 |
*/ |
654 |
public final void quietlyCompleteRoot() { |
655 |
for (CountedCompleter<?> a = this, p;;) { |
656 |
if ((p = a.completer) == null) { |
657 |
a.quietlyComplete(); |
658 |
return; |
659 |
} |
660 |
a = p; |
661 |
} |
662 |
} |
663 |
|
664 |
/** |
665 |
* Supports ForkJoinTask exception propagation. |
666 |
*/ |
667 |
void internalPropagateException(Throwable ex) { |
668 |
CountedCompleter<?> a = this, s = a; |
669 |
while (a.onExceptionalCompletion(ex, s) && |
670 |
(a = (s = a).completer) != null && a.status >= 0) |
671 |
a.recordExceptionalCompletion(ex); |
672 |
} |
673 |
|
674 |
/** |
675 |
* Implements execution conventions for CountedCompleters. |
676 |
*/ |
677 |
protected final boolean exec() { |
678 |
compute(); |
679 |
return false; |
680 |
} |
681 |
|
682 |
/** |
683 |
* Returns the result of the computation. By default |
684 |
* returns {@code null}, which is appropriate for {@code Void} |
685 |
* actions, but in other cases should be overridden, almost |
686 |
* always to return a field or function of a field that |
687 |
* holds the result upon completion. |
688 |
* |
689 |
* @return the result of the computation |
690 |
*/ |
691 |
public T getRawResult() { return null; } |
692 |
|
693 |
/** |
694 |
* A method that result-bearing CountedCompleters may optionally |
695 |
* use to help maintain result data. By default, does nothing. |
696 |
* Overrides are not recommended. However, if this method is |
697 |
* overridden to update existing objects or fields, then it must |
698 |
* in general be defined to be thread-safe. |
699 |
*/ |
700 |
protected void setRawResult(T t) { } |
701 |
|
702 |
// Unsafe mechanics |
703 |
private static final sun.misc.Unsafe U; |
704 |
private static final long PENDING; |
705 |
static { |
706 |
try { |
707 |
U = sun.misc.Unsafe.getUnsafe(); |
708 |
PENDING = U.objectFieldOffset |
709 |
(CountedCompleter.class.getDeclaredField("pending")); |
710 |
} catch (Exception e) { |
711 |
throw new Error(e); |
712 |
} |
713 |
} |
714 |
|
715 |
/** |
716 |
* Returns a sun.misc.Unsafe. Suitable for use in a 3rd party package. |
717 |
* Replace with a simple call to Unsafe.getUnsafe when integrating |
718 |
* into a jdk. |
719 |
* |
720 |
* @return a sun.misc.Unsafe |
721 |
*/ |
722 |
private static sun.misc.Unsafe getUnsafe() { |
723 |
try { |
724 |
return sun.misc.Unsafe.getUnsafe(); |
725 |
} catch (SecurityException se) { |
726 |
try { |
727 |
return java.security.AccessController.doPrivileged |
728 |
(new java.security |
729 |
.PrivilegedExceptionAction<sun.misc.Unsafe>() { |
730 |
public sun.misc.Unsafe run() throws Exception { |
731 |
java.lang.reflect.Field f = sun.misc |
732 |
.Unsafe.class.getDeclaredField("theUnsafe"); |
733 |
f.setAccessible(true); |
734 |
return (sun.misc.Unsafe) f.get(null); |
735 |
}}); |
736 |
} catch (java.security.PrivilegedActionException e) { |
737 |
throw new RuntimeException("Could not initialize intrinsics", |
738 |
e.getCause()); |
739 |
} |
740 |
} |
741 |
} |
742 |
} |