[ViewVC] Diff of: jsr166/jsr166/src/jsr166y/ForkJoinPool.java

Comparing jsr166/src/jsr166y/ForkJoinPool.java (file contents):
Revision 1.116 by dl, Fri Jan 27 17:27:28 2012 UTC vs.
Revision 1.144 by jsr166, Sun Nov 18 18:03:10 2012 UTC

#	Line 11 \| Line 11 \| import java.util.Arrays;
11		import java.util.Collection;
12		import java.util.Collections;
13		import java.util.List;
14	–	import java.util.Random;
14		import java.util.concurrent.AbstractExecutorService;
15		import java.util.concurrent.Callable;
16		import java.util.concurrent.ExecutorService;
#	Line 19 \| Line 18 \| import java.util.concurrent.Future;
18		import java.util.concurrent.RejectedExecutionException;
19		import java.util.concurrent.RunnableFuture;
20		import java.util.concurrent.TimeUnit;
22	–	import java.util.concurrent.atomic.AtomicInteger;
23	–	import java.util.concurrent.atomic.AtomicLong;
24	–	import java.util.concurrent.locks.ReentrantLock;
25	–	import java.util.concurrent.locks.Condition;
21
22		/**
23		* An {@link ExecutorService} for running {@link ForkJoinTask}s.
#	Line 42 \| Line 37 \| import java.util.concurrent.locks.Condit
37		* ForkJoinPool}s may also be appropriate for use with event-style
38		* tasks that are never joined.
39		*
40	<	* <p>A {@code ForkJoinPool} is constructed with a given target
41	<	* parallelism level; by default, equal to the number of available
42	<	* processors. The pool attempts to maintain enough active (or
43	<	* available) threads by dynamically adding, suspending, or resuming
44	<	* internal worker threads, even if some tasks are stalled waiting to
45	<	* join others. However, no such adjustments are guaranteed in the
46	<	* face of blocked IO or other unmanaged synchronization. The nested
47	<	* {@link ManagedBlocker} interface enables extension of the kinds of
40	>	* <p>A static {@link #commonPool} is available and appropriate for
41	>	* most applications. The common pool is used by any ForkJoinTask that
42	>	* is not explicitly submitted to a specified pool. Using the common
43	>	* pool normally reduces resource usage (its threads are slowly
44	>	* reclaimed during periods of non-use, and reinstated upon subsequent
45	>	* use).
46	>	*
47	>	* <p>For applications that require separate or custom pools, a {@code
48	>	* ForkJoinPool} may be constructed with a given target parallelism
49	>	* level; by default, equal to the number of available processors. The
50	>	* pool attempts to maintain enough active (or available) threads by
51	>	* dynamically adding, suspending, or resuming internal worker
52	>	* threads, even if some tasks are stalled waiting to join
53	>	* others. However, no such adjustments are guaranteed in the face of
54	>	* blocked IO or other unmanaged synchronization. The nested {@link
55	>	* ManagedBlocker} interface enables extension of the kinds of
56		* synchronization accommodated.
57		*
58		* <p>In addition to execution and lifecycle control methods, this
#	Line 59 \| Line 62 \| import java.util.concurrent.locks.Condit
62		* {@link #toString} returns indications of pool state in a
63		* convenient form for informal monitoring.
64		*
65	<	* <p> As is the case with other ExecutorServices, there are three
66	<	* main task execution methods summarized in the following
67	<	* table. These are designed to be used primarily by clients not
68	<	* already engaged in fork/join computations in the current pool. The
69	<	* main forms of these methods accept instances of {@code
70	<	* ForkJoinTask}, but overloaded forms also allow mixed execution of
71	<	* plain {@code Runnable}- or {@code Callable}- based activities as
72	<	* well. However, tasks that are already executing in a pool should
73	<	* normally instead use the within-computation forms listed in the
74	<	* table unless using async event-style tasks that are not usually
75	<	* joined, in which case there is little difference among choice of
73	<	* methods.
65	>	* <p>As is the case with other ExecutorServices, there are three
66	>	* main task execution methods summarized in the following table.
67	>	* These are designed to be used primarily by clients not already
68	>	* engaged in fork/join computations in the current pool. The main
69	>	* forms of these methods accept instances of {@code ForkJoinTask},
70	>	* but overloaded forms also allow mixed execution of plain {@code
71	>	* Runnable}- or {@code Callable}- based activities as well. However,
72	>	* tasks that are already executing in a pool should normally instead
73	>	* use the within-computation forms listed in the table unless using
74	>	* async event-style tasks that are not usually joined, in which case
75	>	* there is little difference among choice of methods.
76		*
77		* <table BORDER CELLPADDING=3 CELLSPACING=1>
78		* <tr>
#	Line 95 \| Line 97 \| import java.util.concurrent.locks.Condit
97		* </tr>
98		* </table>
99		*
100	<	* <p><b>Sample Usage.</b> Normally a single {@code ForkJoinPool} is
101	<	* used for all parallel task execution in a program or subsystem.
102	<	* Otherwise, use would not usually outweigh the construction and
103	<	* bookkeeping overhead of creating a large set of threads. For
104	<	* example, a common pool could be used for the {@code SortTasks}
105	<	* illustrated in {@link RecursiveAction}. Because {@code
106	<	* ForkJoinPool} uses threads in {@linkplain java.lang.Thread#isDaemon
107	<	* daemon} mode, there is typically no need to explicitly {@link
108	<	* #shutdown} such a pool upon program exit.
109	<	*
108	<	* <pre> {@code
109	<	* static final ForkJoinPool mainPool = new ForkJoinPool();
110	<	* ...
111	<	* public void sort(long[] array) {
112	<	* mainPool.invoke(new SortTask(array, 0, array.length));
113	<	* }}</pre>
100	>	* <p>The common pool is by default constructed with default
101	>	* parameters, but these may be controlled by setting three {@link
102	>	* System#getProperty properties} with prefix {@code
103	>	* java.util.concurrent.ForkJoinPool.common}: {@code parallelism} --
104	>	* an integer greater than zero, {@code threadFactory} -- the class
105	>	* name of a {@link ForkJoinWorkerThreadFactory}, and {@code
106	>	* exceptionHandler} -- the class name of a {@link
107	>	* java.lang.Thread.UncaughtExceptionHandler
108	>	* Thread.UncaughtExceptionHandler}. Upon any error in establishing
109	>	* these settings, default parameters are used.
110		*
111		* <p><b>Implementation notes</b>: This implementation restricts the
112		* maximum number of running threads to 32767. Attempts to create
#	Line 131 \| Line 127 \| public class ForkJoinPool extends Abstra
127		*
128		* This class and its nested classes provide the main
129		* functionality and control for a set of worker threads:
130	<	* Submissions from non-FJ threads enter into submission
131	<	* queues. Workers take these tasks and typically split them into
132	<	* subtasks that may be stolen by other workers. Preference rules
133	<	* give first priority to processing tasks from their own queues
134	<	* (LIFO or FIFO, depending on mode), then to randomized FIFO
135	<	* steals of tasks in other queues.
130	>	* Submissions from non-FJ threads enter into submission queues.
131	>	* Workers take these tasks and typically split them into subtasks
132	>	* that may be stolen by other workers. Preference rules give
133	>	* first priority to processing tasks from their own queues (LIFO
134	>	* or FIFO, depending on mode), then to randomized FIFO steals of
135	>	* tasks in other queues.
136		*
137	<	* WorkQueues.
137	>	* WorkQueues
138		* ==========
139		*
140		* Most operations occur within work-stealing queues (in nested
#	Line 156 \| Line 152 \| public class ForkJoinPool extends Abstra
152		* (http://research.sun.com/scalable/pubs/index.html) and
153		* "Idempotent work stealing" by Michael, Saraswat, and Vechev,
154		* PPoPP 2009 (http://portal.acm.org/citation.cfm?id=1504186).
155	<	* The main differences ultimately stem from gc requirements that
155	>	* The main differences ultimately stem from GC requirements that
156		* we null out taken slots as soon as we can, to maintain as small
157		* a footprint as possible even in programs generating huge
158		* numbers of tasks. To accomplish this, we shift the CAS
#	Line 178 \| Line 174 \| public class ForkJoinPool extends Abstra
174		* If an attempted steal fails, a thief always chooses a different
175		* random victim target to try next. So, in order for one thief to
176		* progress, it suffices for any in-progress poll or new push on
177	<	* any empty queue to complete.
177	>	* any empty queue to complete. (This is why we normally use
178	>	* method pollAt and its variants that try once at the apparent
179	>	* base index, else consider alternative actions, rather than
180	>	* method poll.)
181		*
182		* This approach also enables support of a user mode in which local
183		* task processing is in FIFO, not LIFO order, simply by using
#	Line 188 \| Line 187 \| public class ForkJoinPool extends Abstra
187		* rarely provide the best possible performance on a given
188		* machine, but portably provide good throughput by averaging over
189		* these factors. (Further, even if we did try to use such
190	<	* information, we do not usually have a basis for exploiting
191	<	* it. For example, some sets of tasks profit from cache
192	<	* affinities, but others are harmed by cache pollution effects.)
190	>	* information, we do not usually have a basis for exploiting it.
191	>	* For example, some sets of tasks profit from cache affinities,
192	>	* but others are harmed by cache pollution effects.)
193		*
194		* WorkQueues are also used in a similar way for tasks submitted
195		* to the pool. We cannot mix these tasks in the same queues used
196		* for work-stealing (this would contaminate lifo/fifo
197	<	* processing). Instead, we loosely associate submission queues
197	>	* processing). Instead, we randomly associate submission queues
198		* with submitting threads, using a form of hashing. The
199		* ThreadLocal Submitter class contains a value initially used as
200		* a hash code for choosing existing queues, but may be randomly
201		* repositioned upon contention with other submitters. In
202	<	* essence, submitters act like workers except that they never
203	<	* take tasks, and they are multiplexed on to a finite number of
204	<	* shared work queues. However, classes are set up so that future
205	<	* extensions could allow submitters to optionally help perform
206	<	* tasks as well. Pool submissions from internal workers are also
207	<	* allowed, but use randomized rather than thread-hashed queue
208	<	* indices to avoid imbalance. Insertion of tasks in shared mode
209	<	* requires a lock (mainly to protect in the case of resizing) but
210	<	* we use only a simple spinlock (using bits in field runState),
211	<	* because submitters encountering a busy queue try or create
212	<	* others so never block.
202	>	* essence, submitters act like workers except that they are
203	>	* restricted to executing local tasks that they submitted (or in
204	>	* the case of CountedCompleters, others with the same root task).
205	>	* However, because most shared/external queue operations are more
206	>	* expensive than internal, and because, at steady state, external
207	>	* submitters will compete for CPU with workers, ForkJoinTask.join
208	>	* and related methods disable them from repeatedly helping to
209	>	* process tasks if all workers are active. Insertion of tasks in
210	>	* shared mode requires a lock (mainly to protect in the case of
211	>	* resizing) but we use only a simple spinlock (using bits in
212	>	* field qlock), because submitters encountering a busy queue move
213	>	* on to try or create other queues -- they block only when
214	>	* creating and registering new queues.
215		*
216	<	* Management.
216	>	* Management
217		* ==========
218		*
219		* The main throughput advantages of work-stealing stem from
#	Line 222 \| Line 223 \| public class ForkJoinPool extends Abstra
223		* tactic for avoiding bottlenecks is packing nearly all
224		* essentially atomic control state into two volatile variables
225		* that are by far most often read (not written) as status and
226	<	* consistency checks
226	>	* consistency checks.
227		*
228		* Field "ctl" contains 64 bits holding all the information needed
229		* to atomically decide to add, inactivate, enqueue (on an event
#	Line 232 \| Line 233 \| public class ForkJoinPool extends Abstra
233		* and their negations (used for thresholding) to fit into 16bit
234		* fields.
235		*
236	<	* Field "runState" contains 32 bits needed to register and
237	<	* deregister WorkQueues, as well as to enable shutdown. It is
238	<	* only modified under a lock (normally briefly held, but
239	<	* occasionally protecting allocations and resizings) but even
240	<	* when locked remains available to check consistency.
236	>	* Field "plock" is a form of sequence lock with a saturating
237	>	* shutdown bit (similarly for per-queue "qlocks"), mainly
238	>	* protecting updates to the workQueues array, as well as to
239	>	* enable shutdown. When used as a lock, it is normally only very
240	>	* briefly held, so is nearly always available after at most a
241	>	* brief spin, but we use a monitor-based backup strategy to
242	>	* blocking when needed.
243		*
244		* Recording WorkQueues. WorkQueues are recorded in the
245	<	* "workQueues" array that is created upon pool construction and
246	<	* expanded if necessary. Updates to the array while recording
247	<	* new workers and unrecording terminated ones are protected from
248	<	* each other by a lock but the array is otherwise concurrently
249	<	* readable, and accessed directly. To simplify index-based
250	<	* operations, the array size is always a power of two, and all
251	<	* readers must tolerate null slots. Shared (submission) queues
252	<	* are at even indices, worker queues at odd indices. Grouping
253	<	* them together in this way simplifies and speeds up task
254	<	* scanning. To avoid flailing during start-up, the array is
255	<	* presized to hold twice #parallelism workers (which is unlikely
256	<	* to need further resizing during execution). But to avoid
257	<	* dealing with so many null slots, variable runState includes a
258	<	* mask for the nearest power of two that contains all current
256	<	* workers. All worker thread creation is on-demand, triggered by
257	<	* task submissions, replacement of terminated workers, and/or
245	>	* "workQueues" array that is created upon first use and expanded
246	>	* if necessary. Updates to the array while recording new workers
247	>	* and unrecording terminated ones are protected from each other
248	>	* by a lock but the array is otherwise concurrently readable, and
249	>	* accessed directly. To simplify index-based operations, the
250	>	* array size is always a power of two, and all readers must
251	>	* tolerate null slots. Worker queues are at odd indices Shared
252	>	* (submission) queues are at even indices, up to a maximum of 64
253	>	* slots, to limit growth even if array needs to expand to add
254	>	* more workers. Grouping them together in this way simplifies and
255	>	* speeds up task scanning.
256	>	*
257	>	* All worker thread creation is on-demand, triggered by task
258	>	* submissions, replacement of terminated workers, and/or
259		* compensation for blocked workers. However, all other support
260		* code is set up to work with other policies. To ensure that we
261		* do not hold on to worker references that would prevent GC, ALL
#	Line 267 \| Line 268 \| public class ForkJoinPool extends Abstra
268		* both index-check and null-check the IDs. All such accesses
269		* ignore bad IDs by returning out early from what they are doing,
270		* since this can only be associated with termination, in which
271	<	* case it is OK to give up.
272	<	*
273	<	* All uses of the workQueues array check that it is non-null
274	<	* (even if previously non-null). This allows nulling during
275	<	* termination, which is currently not necessary, but remains an
276	<	* option for resource-revocation-based shutdown schemes. It also
276	<	* helps reduce JIT issuance of uncommon-trap code, which tends to
271	>	* case it is OK to give up. All uses of the workQueues array
272	>	* also check that it is non-null (even if previously
273	>	* non-null). This allows nulling during termination, which is
274	>	* currently not necessary, but remains an option for
275	>	* resource-revocation-based shutdown schemes. It also helps
276	>	* reduce JIT issuance of uncommon-trap code, which tends to
277		* unnecessarily complicate control flow in some methods.
278		*
279		* Event Queuing. Unlike HPC work-stealing frameworks, we cannot
#	Line 301 \| Line 301 \| public class ForkJoinPool extends Abstra
301		* some other queued worker rather than itself, which has the same
302		* net effect. Because enqueued workers may actually be rescanning
303		* rather than waiting, we set and clear the "parker" field of
304	<	* Workqueues to reduce unnecessary calls to unpark. (This
304	>	* WorkQueues to reduce unnecessary calls to unpark. (This
305		* requires a secondary recheck to avoid missed signals.) Note
306		* the unusual conventions about Thread.interrupts surrounding
307		* parking and other blocking: Because interrupts are used solely
#	Line 313 \| Line 313 \| public class ForkJoinPool extends Abstra
313		*
314		* Signalling. We create or wake up workers only when there
315		* appears to be at least one task they might be able to find and
316	<	* execute. When a submission is added or another worker adds a
317	<	* task to a queue that previously had fewer than two tasks, they
318	<	* signal waiting workers (or trigger creation of new ones if
319	<	* fewer than the given parallelism level -- see signalWork).
320	<	* These primary signals are buttressed by signals during rescans;
321	<	* together these cover the signals needed in cases when more
322	<	* tasks are pushed but untaken, and improve performance compared
323	<	* to having one thread wake up all workers.
316	>	* execute. However, many other threads may notice the same task
317	>	* and each signal to wake up a thread that might take it. So in
318	>	* general, pools will be over-signalled. When a submission is
319	>	* added or another worker adds a task to a queue that is
320	>	* apparently empty, they signal waiting workers (or trigger
321	>	* creation of new ones if fewer than the given parallelism level
322	>	* -- see signalWork). These primary signals are buttressed by
323	>	* signals whenever other threads scan for work or do not have a
324	>	* task to process. On most platforms, signalling (unpark)
325	>	* overhead time is noticeably long, and the time between
326	>	* signalling a thread and it actually making progress can be very
327	>	* noticeably long, so it is worth offloading these delays from
328	>	* critical paths as much as possible.
329		*
330		* Trimming workers. To release resources after periods of lack of
331		* use, a worker starting to wait when the pool is quiescent will
332	<	* time out and terminate if the pool has remained quiescent for
333	<	* SHRINK_RATE nanosecs. This will slowly propagate, eventually
334	<	* terminating all workers after long periods of non-use.
332	>	* time out and terminate if the pool has remained quiescent for a
333	>	* given period -- a short period if there are more threads than
334	>	* parallelism, longer as the number of threads decreases. This
335	>	* will slowly propagate, eventually terminating all workers after
336	>	* periods of non-use.
337		*
338		* Shutdown and Termination. A call to shutdownNow atomically sets
339	<	* a runState bit and then (non-atomically) sets each workers
340	<	* runState status, cancels all unprocessed tasks, and wakes up
339	>	* a plock bit and then (non-atomically) sets each worker's
340	>	* qlock status, cancels all unprocessed tasks, and wakes up
341		* all waiting workers. Detecting whether termination should
342		* commence after a non-abrupt shutdown() call requires more work
343		* and bookkeeping. We need consensus about quiescence (i.e., that
#	Line 338 \| Line 345 \| public class ForkJoinPool extends Abstra
345		* indication but non-abrupt shutdown still requires a rechecking
346		* scan for any workers that are inactive but not queued.
347		*
348	<	* Joining Tasks.
349	<	* ==============
348	>	* Joining Tasks
349	>	* =============
350		*
351		* Any of several actions may be taken when one worker is waiting
352	<	* to join a task stolen (or always held by) another. Because we
352	>	* to join a task stolen (or always held) by another. Because we
353		* are multiplexing many tasks on to a pool of workers, we can't
354		* just let them block (as in Thread.join). We also cannot just
355		* reassign the joiner's run-time stack with another and replace
356		* it later, which would be a form of "continuation", that even if
357		* possible is not necessarily a good idea since we sometimes need
358	<	* both an unblocked task and its continuation to
359	<	* progress. Instead we combine two tactics:
358	>	* both an unblocked task and its continuation to progress.
359	>	* Instead we combine two tactics:
360		*
361		* Helping: Arranging for the joiner to execute some task that it
362		* would be running if the steal had not occurred.
#	Line 358 \| Line 365 \| public class ForkJoinPool extends Abstra
365		* method tryCompensate() may create or re-activate a spare
366		* thread to compensate for blocked joiners until they unblock.
367		*
368	<	* A third form (implemented in tryRemoveAndExec and
369	<	* tryPollForAndExec) amounts to helping a hypothetical
370	<	* compensator: If we can readily tell that a possible action of a
371	<	* compensator is to steal and execute the task being joined, the
372	<	* joining thread can do so directly, without the need for a
373	<	* compensation thread (although at the expense of larger run-time
374	<	* stacks, but the tradeoff is typically worthwhile).
368	>	* A third form (implemented in tryRemoveAndExec) amounts to
369	>	* helping a hypothetical compensator: If we can readily tell that
370	>	* a possible action of a compensator is to steal and execute the
371	>	* task being joined, the joining thread can do so directly,
372	>	* without the need for a compensation thread (although at the
373	>	* expense of larger run-time stacks, but the tradeoff is
374	>	* typically worthwhile).
375		*
376		* The ManagedBlocker extension API can't use helping so relies
377		* only on compensation in method awaitBlocker.
#	Line 384 \| Line 391 \| public class ForkJoinPool extends Abstra
391		* (http://portal.acm.org/citation.cfm?id=155354). It differs in
392		* that: (1) We only maintain dependency links across workers upon
393		* steals, rather than use per-task bookkeeping. This sometimes
394	<	* requires a linear scan of workers array to locate stealers, but
395	<	* often doesn't because stealers leave hints (that may become
394	>	* requires a linear scan of workQueues array to locate stealers,
395	>	* but often doesn't because stealers leave hints (that may become
396		* stale/wrong) of where to locate them. A stealHint is only a
397		* hint because a worker might have had multiple steals and the
398		* hint records only one of them (usually the most current).
#	Line 396 \| Line 403 \| public class ForkJoinPool extends Abstra
403		* which means that we miss links in the chain during long-lived
404		* tasks, GC stalls etc (which is OK since blocking in such cases
405		* is usually a good idea). (4) We bound the number of attempts
406	<	* to find work (see MAX_HELP_DEPTH) and fall back to suspending
407	<	* the worker and if necessary replacing it with another.
406	>	* to find work (see MAX_HELP) and fall back to suspending the
407	>	* worker and if necessary replacing it with another.
408	>	*
409	>	* Helping actions for CountedCompleters are much simpler: Method
410	>	* helpComplete can take and execute any task with the same root
411	>	* as the task being waited on. However, this still entails some
412	>	* traversal of completer chains, so is less efficient than using
413	>	* CountedCompleters without explicit joins.
414		*
415		* It is impossible to keep exactly the target parallelism number
416		* of threads running at any given time. Determining the
417		* existence of conservatively safe helping targets, the
418		* availability of already-created spares, and the apparent need
419		* to create new spares are all racy, so we rely on multiple
420	<	* retries of each. Currently, in keeping with on-demand
421	<	* signalling policy, we compensate only if blocking would leave
422	<	* less than one active (non-waiting, non-blocked) worker.
423	<	* Additionally, to avoid some false alarms due to GC, lagging
424	<	* counters, system activity, etc, compensated blocking for joins
425	<	* is only attempted after rechecks stabilize in
426	<	* ForkJoinTask.awaitJoin. (Retries are interspersed with
427	<	* Thread.yield, for good citizenship.)
428	<	*
429	<	* Style notes: There is a lot of representation-level coupling
430	<	* among classes ForkJoinPool, ForkJoinWorkerThread, and
431	<	* ForkJoinTask. The fields of WorkQueue maintain data structures
432	<	* managed by ForkJoinPool, so are directly accessed. There is
433	<	* little point trying to reduce this, since any associated future
434	<	* changes in representations will need to be accompanied by
435	<	* algorithmic changes anyway. All together, these low-level
436	<	* implementation choices produce as much as a factor of 4
437	<	* performance improvement compared to naive implementations, and
438	<	* enable the processing of billions of tasks per second, at the
439	<	* expense of some ugliness.
440	<	*
441	<	* Methods signalWork() and scan() are the main bottlenecks so are
442	<	* especially heavily micro-optimized/mangled. There are lots of
443	<	* inline assignments (of form "while ((local = field) != 0)")
444	<	* which are usually the simplest way to ensure the required read
445	<	* orderings (which are sometimes critical). This leads to a
446	<	* "C"-like style of listing declarations of these locals at the
447	<	* heads of methods or blocks. There are several occurrences of
448	<	* the unusual "do {} while (!cas...)" which is the simplest way
449	<	* to force an update of a CAS'ed variable. There are also other
450	<	* coding oddities that help some methods perform reasonably even
451	<	* when interpreted (not compiled).
452	<	*
453	<	* The order of declarations in this file is: (1) declarations of
454	<	* statics (2) fields (along with constants used when unpacking
455	<	* some of them), listed in an order that tends to reduce
456	<	* contention among them a bit under most JVMs; (3) nested
457	<	* classes; (4) internal control methods; (5) callbacks and other
458	<	* support for ForkJoinTask methods; (6) exported methods (plus a
459	<	* few little helpers); (7) static block initializing all statics
460	<	* in a minimally dependent order.
420	>	* retries of each. Compensation in the apparent absence of
421	>	* helping opportunities is challenging to control on JVMs, where
422	>	* GC and other activities can stall progress of tasks that in
423	>	* turn stall out many other dependent tasks, without us being
424	>	* able to determine whether they will ever require compensation.
425	>	* Even though work-stealing otherwise encounters little
426	>	* degradation in the presence of more threads than cores,
427	>	* aggressively adding new threads in such cases entails risk of
428	>	* unwanted positive feedback control loops in which more threads
429	>	* cause more dependent stalls (as well as delayed progress of
430	>	* unblocked threads to the point that we know they are available)
431	>	* leading to more situations requiring more threads, and so
432	>	* on. This aspect of control can be seen as an (analytically
433	>	* intractable) game with an opponent that may choose the worst
434	>	* (for us) active thread to stall at any time. We take several
435	>	* precautions to bound losses (and thus bound gains), mainly in
436	>	* methods tryCompensate and awaitJoin.
437	>	*
438	>	* Common Pool
439	>	* ===========
440	>	*
441	>	* The static commonPool always exists after static
442	>	* initialization. Since it (or any other created pool) need
443	>	* never be used, we minimize initial construction overhead and
444	>	* footprint to the setup of about a dozen fields, with no nested
445	>	* allocation. Most bootstrapping occurs within method
446	>	* fullExternalPush during the first submission to the pool.
447	>	*
448	>	* When external threads submit to the common pool, they can
449	>	* perform some subtask processing (see externalHelpJoin and
450	>	* related methods). We do not need to record whether these
451	>	* submissions are to the common pool -- if not, externalHelpJoin
452	>	* returns quickly (at the most helping to signal some common pool
453	>	* workers). These submitters would otherwise be blocked waiting
454	>	* for completion, so the extra effort (with liberally sprinkled
455	>	* task status checks) in inapplicable cases amounts to an odd
456	>	* form of limited spin-wait before blocking in ForkJoinTask.join.
457	>	*
458	>	* Style notes
459	>	* ===========
460	>	*
461	>	* There is a lot of representation-level coupling among classes
462	>	* ForkJoinPool, ForkJoinWorkerThread, and ForkJoinTask. The
463	>	* fields of WorkQueue maintain data structures managed by
464	>	* ForkJoinPool, so are directly accessed. There is little point
465	>	* trying to reduce this, since any associated future changes in
466	>	* representations will need to be accompanied by algorithmic
467	>	* changes anyway. Several methods intrinsically sprawl because
468	>	* they must accumulate sets of consistent reads of volatiles held
469	>	* in local variables. Methods signalWork() and scan() are the
470	>	* main bottlenecks, so are especially heavily
471	>	* micro-optimized/mangled. There are lots of inline assignments
472	>	* (of form "while ((local = field) != 0)") which are usually the
473	>	* simplest way to ensure the required read orderings (which are
474	>	* sometimes critical). This leads to a "C"-like style of listing
475	>	* declarations of these locals at the heads of methods or blocks.
476	>	* There are several occurrences of the unusual "do {} while
477	>	* (!cas...)" which is the simplest way to force an update of a
478	>	* CAS'ed variable. There are also other coding oddities (including
479	>	* several unnecessary-looking hoisted null checks) that help
480	>	* some methods perform reasonably even when interpreted (not
481	>	* compiled).
482	>	*
483	>	* The order of declarations in this file is:
484	>	* (1) Static utility functions
485	>	* (2) Nested (static) classes
486	>	* (3) Static fields
487	>	* (4) Fields, along with constants used when unpacking some of them
488	>	* (5) Internal control methods
489	>	* (6) Callbacks and other support for ForkJoinTask methods
490	>	* (7) Exported methods
491	>	* (8) Static block initializing statics in minimally dependent order
492	>	*/
493	>
494	>	// Static utilities
495	>
496	>	/**
497	>	* If there is a security manager, makes sure caller has
498	>	* permission to modify threads.
499		*/
500	+	private static void checkPermission() {
501	+	SecurityManager security = System.getSecurityManager();
502	+	if (security != null)
503	+	security.checkPermission(modifyThreadPermission);
504	+	}
505	+
506	+	// Nested classes
507
508		/**
509		* Factory for creating new {@link ForkJoinWorkerThread}s.
#	Line 475 \| Line 533 \| public class ForkJoinPool extends Abstra
533		}
534
535		/**
536	<	* Creates a new ForkJoinWorkerThread. This factory is used unless
537	<	* overridden in ForkJoinPool constructors.
538	<	*/
539	<	public static final ForkJoinWorkerThreadFactory
482	<	defaultForkJoinWorkerThreadFactory;
483	<
484	<	/**
485	<	* Permission required for callers of methods that may start or
486	<	* kill threads.
487	<	*/
488	<	private static final RuntimePermission modifyThreadPermission;
489	<
490	<	/**
491	<	* If there is a security manager, makes sure caller has
492	<	* permission to modify threads.
536	>	* Class for artificial tasks that are used to replace the target
537	>	* of local joins if they are removed from an interior queue slot
538	>	* in WorkQueue.tryRemoveAndExec. We don't need the proxy to
539	>	* actually do anything beyond having a unique identity.
540		*/
541	<	private static void checkPermission() {
542	<	SecurityManager security = System.getSecurityManager();
543	<	if (security != null)
544	<	security.checkPermission(modifyThreadPermission);
541	>	static final class EmptyTask extends ForkJoinTask<Void> {
542	>	private static final long serialVersionUID = -7721805057305804111L;
543	>	EmptyTask() { status = ForkJoinTask.NORMAL; } // force done
544	>	public final Void getRawResult() { return null; }
545	>	public final void setRawResult(Void x) {}
546	>	public final boolean exec() { return true; }
547		}
548
549		/**
501	–	* Generator for assigning sequence numbers as pool names.
502	–	*/
503	–	private static final AtomicInteger poolNumberGenerator;
504	–
505	–	/**
506	–	* Bits and masks for control variables
507	–	*
508	–	* Field ctl is a long packed with:
509	–	* AC: Number of active running workers minus target parallelism (16 bits)
510	–	* TC: Number of total workers minus target parallelism (16 bits)
511	–	* ST: true if pool is terminating (1 bit)
512	–	* EC: the wait count of top waiting thread (15 bits)
513	–	* ID: ~(poolIndex >>> 1) of top of Treiber stack of waiters (16 bits)
514	–	*
515	–	* When convenient, we can extract the upper 32 bits of counts and
516	–	* the lower 32 bits of queue state, u = (int)(ctl >>> 32) and e =
517	–	* (int)ctl. The ec field is never accessed alone, but always
518	–	* together with id and st. The offsets of counts by the target
519	–	* parallelism and the positionings of fields makes it possible to
520	–	* perform the most common checks via sign tests of fields: When
521	–	* ac is negative, there are not enough active workers, when tc is
522	–	* negative, there are not enough total workers, when id is
523	–	* negative, there is at least one waiting worker, and when e is
524	–	* negative, the pool is terminating. To deal with these possibly
525	–	* negative fields, we use casts in and out of "short" and/or
526	–	* signed shifts to maintain signedness.
527	–	*
528	–	* When a thread is queued (inactivated), its eventCount field is
529	–	* negative, which is the only way to tell if a worker is
530	–	* prevented from executing tasks, even though it must continue to
531	–	* scan for them to avoid queuing races.
532	–	*
533	–	* Field runState is an int packed with:
534	–	* SHUTDOWN: true if shutdown is enabled (1 bit)
535	–	* SEQ: a sequence number updated upon (de)registering workers (15 bits)
536	–	* MASK: mask (power of 2 - 1) covering all registered poolIndexes (16 bits)
537	–	*
538	–	* The combination of mask and sequence number enables simple
539	–	* consistency checks: Staleness of read-only operations on the
540	–	* workers and queues arrays can be checked by comparing runState
541	–	* before vs after the reads. The low 16 bits (i.e, anding with
542	–	* SMASK) hold (the smallest power of two covering all worker
543	–	* indices, minus one. The mask for queues (vs workers) is twice
544	–	* this value plus 1.
545	–	*/
546	–
547	–	// bit positions/shifts for fields
548	–	private static final int AC_SHIFT = 48;
549	–	private static final int TC_SHIFT = 32;
550	–	private static final int ST_SHIFT = 31;
551	–	private static final int EC_SHIFT = 16;
552	–
553	–	// bounds
554	–	private static final int MAX_ID = 0x7fff; // max poolIndex
555	–	private static final int SMASK = 0xffff; // mask short bits
556	–	private static final int SHORT_SIGN = 1 << 15;
557	–	private static final int INT_SIGN = 1 << 31;
558	–
559	–	// masks
560	–	private static final long STOP_BIT = 0x0001L << ST_SHIFT;
561	–	private static final long AC_MASK = ((long)SMASK) << AC_SHIFT;
562	–	private static final long TC_MASK = ((long)SMASK) << TC_SHIFT;
563	–
564	–	// units for incrementing and decrementing
565	–	private static final long TC_UNIT = 1L << TC_SHIFT;
566	–	private static final long AC_UNIT = 1L << AC_SHIFT;
567	–
568	–	// masks and units for dealing with u = (int)(ctl >>> 32)
569	–	private static final int UAC_SHIFT = AC_SHIFT - 32;
570	–	private static final int UTC_SHIFT = TC_SHIFT - 32;
571	–	private static final int UAC_MASK = SMASK << UAC_SHIFT;
572	–	private static final int UTC_MASK = SMASK << UTC_SHIFT;
573	–	private static final int UAC_UNIT = 1 << UAC_SHIFT;
574	–	private static final int UTC_UNIT = 1 << UTC_SHIFT;
575	–
576	–	// masks and units for dealing with e = (int)ctl
577	–	private static final int E_MASK = 0x7fffffff; // no STOP_BIT
578	–	private static final int E_SEQ = 1 << EC_SHIFT;
579	–
580	–	// runState bits
581	–	private static final int SHUTDOWN = 1 << 31;
582	–	private static final int RS_SEQ = 1 << 16;
583	–	private static final int RS_SEQ_MASK = 0x7fff0000;
584	–
585	–	// access mode for WorkQueue
586	–	static final int LIFO_QUEUE = 0;
587	–	static final int FIFO_QUEUE = 1;
588	–	static final int SHARED_QUEUE = -1;
589	–
590	–	/**
591	–	* The wakeup interval (in nanoseconds) for a worker waiting for a
592	–	* task when the pool is quiescent to instead try to shrink the
593	–	* number of workers. The exact value does not matter too
594	–	* much. It must be short enough to release resources during
595	–	* sustained periods of idleness, but not so short that threads
596	–	* are continually re-created.
597	–	*/
598	–	private static final long SHRINK_RATE =
599	–	4L * 1000L * 1000L * 1000L; // 4 seconds
600	–
601	–	/**
602	–	* The timeout value for attempted shrinkage, includes
603	–	* some slop to cope with system timer imprecision.
604	–	*/
605	–	private static final long SHRINK_TIMEOUT = SHRINK_RATE - (SHRINK_RATE / 10);
606	–
607	–	/**
608	–	* The maximum stolen->joining link depth allowed in tryHelpStealer.
609	–	* Depths for legitimate chains are unbounded, but we use a fixed
610	–	* constant to avoid (otherwise unchecked) cycles and to bound
611	–	* staleness of traversal parameters at the expense of sometimes
612	–	* blocking when we could be helping.
613	–	*/
614	–	private static final int MAX_HELP_DEPTH = 16;
615	–
616	–	/*
617	–	* Field layout order in this class tends to matter more than one
618	–	* would like. Runtime layout order is only loosely related to
619	–	* declaration order and may differ across JVMs, but the following
620	–	* empirically works OK on current JVMs.
621	–	*/
622	–
623	–	volatile long ctl; // main pool control
624	–	final int parallelism; // parallelism level
625	–	final int localMode; // per-worker scheduling mode
626	–	int nextPoolIndex; // hint used in registerWorker
627	–	volatile int runState; // shutdown status, seq, and mask
628	–	WorkQueue[] workQueues; // main registry
629	–	final ReentrantLock lock; // for registration
630	–	final Condition termination; // for awaitTermination
631	–	final ForkJoinWorkerThreadFactory factory; // factory for new workers
632	–	final Thread.UncaughtExceptionHandler ueh; // per-worker UEH
633	–	final AtomicLong stealCount; // collect counts when terminated
634	–	final AtomicInteger nextWorkerNumber; // to create worker name string
635	–	final String workerNamePrefix; // Prefix for assigning worker names
636	–
637	–	/**
550		* Queues supporting work-stealing as well as external task
551		* submission. See above for main rationale and algorithms.
552		* Implementation relies heavily on "Unsafe" intrinsics
#	Line 648 \| Line 560 \| public class ForkJoinPool extends Abstra
560		*
561		* Field "top" is the index (mod array.length) of the next queue
562		* slot to push to or pop from. It is written only by owner thread
563	<	* for push, or under lock for trySharedPush, and accessed by
564	<	* other threads only after reading (volatile) base. Both top and
565	<	* base are allowed to wrap around on overflow, but (top - base)
566	<	* (or more commonly -(base - top) to force volatile read of base
567	<	* before top) still estimates size.
563	>	* for push, or under lock for external/shared push, and accessed
564	>	* by other threads only after reading (volatile) base. Both top
565	>	* and base are allowed to wrap around on overflow, but (top -
566	>	* base) (or more commonly -(base - top) to force volatile read of
567	>	* base before top) still estimates size. The lock ("qlock") is
568	>	* forced to -1 on termination, causing all further lock attempts
569	>	* to fail. (Note: we don't need CAS for termination state because
570	>	* upon pool shutdown, all shared-queues will stop being used
571	>	* anyway.) Nearly all lock bodies are set up so that exceptions
572	>	* within lock bodies are "impossible" (modulo JVM errors that
573	>	* would cause failure anyway.)
574		*
575		* The array slots are read and written using the emulation of
576		* volatiles/atomics provided by Unsafe. Insertions must in
577		* general use putOrderedObject as a form of releasing store to
578		* ensure that all writes to the task object are ordered before
579	<	* its publication in the queue. (Although we can avoid one case
580	<	* of this when locked in trySharedPush.) All removals entail a
581	<	* CAS to null. The array is always a power of two. To ensure
582	<	* safety of Unsafe array operations, all accesses perform
665	<	* explicit null checks and implicit bounds checks via
666	<	* power-of-two masking.
579	>	* its publication in the queue. All removals entail a CAS to
580	>	* null. The array is always a power of two. To ensure safety of
581	>	* Unsafe array operations, all accesses perform explicit null
582	>	* checks and implicit bounds checks via power-of-two masking.
583		*
584		* In addition to basic queuing support, this class contains
585		* fields described elsewhere to control execution. It turns out
586	<	* to work better memory-layout-wise to include them in this
587	<	* class rather than a separate class.
586	>	* to work better memory-layout-wise to include them in this class
587	>	* rather than a separate class.
588		*
589		* Performance on most platforms is very sensitive to placement of
590		* instances of both WorkQueues and their arrays -- we absolutely
#	Line 683 \| Line 599 \| public class ForkJoinPool extends Abstra
599		* avoiding really bad worst-case access. (Until better JVM
600		* support is in place, this padding is dependent on transient
601		* properties of JVM field layout rules.) We also take care in
602	<	* allocating and sizing and resizing the array. Non-shared queue
603	<	* arrays are initialized (via method growArray) by workers before
604	<	* use. Others are allocated on first use.
602	>	* allocating, sizing and resizing the array. Non-shared queue
603	>	* arrays are initialized by workers before use. Others are
604	>	* allocated on first use.
605		*/
606		static final class WorkQueue {
607		/**
608		* Capacity of work-stealing queue array upon initialization.
609	<	* Must be a power of two; at least 4, but set larger to
610	<	* reduce cacheline sharing among queues.
609	>	* Must be a power of two; at least 4, but should be larger to
610	>	* reduce or eliminate cacheline sharing among queues.
611	>	* Currently, it is much larger, as a partial workaround for
612	>	* the fact that JVMs often place arrays in locations that
613	>	* share GC bookkeeping (especially cardmarks) such that
614	>	* per-write accesses encounter serious memory contention.
615		*/
616	<	static final int INITIAL_QUEUE_CAPACITY = 1 << 8;
616	>	static final int INITIAL_QUEUE_CAPACITY = 1 << 13;
617
618		/**
619		* Maximum size for queue arrays. Must be a power of two less
#	Line 704 \| Line 624 \| public class ForkJoinPool extends Abstra
624		*/
625		static final int MAXIMUM_QUEUE_CAPACITY = 1 << 26; // 64M
626
707	–	volatile long totalSteals; // cumulative number of steals
627		int seed; // for random scanning; initialize nonzero
628		volatile int eventCount; // encoded inactivation count; < 0 if inactive
629		int nextWait; // encoded record of next event waiter
711	–	int rescans; // remaining scans until block
712	–	int nsteals; // top-level task executions since last idle
630		final int mode; // lifo, fifo, or shared
631	+	int nsteals; // cumulative number of steals
632		int poolIndex; // index of this queue in pool (or 0)
633		int stealHint; // index of most recent known stealer
634	<	volatile int runState; // 1: locked, -1: terminate; else 0
634	>	volatile int qlock; // 1: locked, -1: terminate; else 0
635		volatile int base; // index of next slot for poll
636		int top; // index of next slot for push
637		ForkJoinTask<?>[] array; // the elements (initially unallocated)
638	+	final ForkJoinPool pool; // the containing pool (may be null)
639		final ForkJoinWorkerThread owner; // owning thread or null if shared
640		volatile Thread parker; // == owner during call to park; else null
641	<	ForkJoinTask<?> currentJoin; // task being joined in awaitJoin
641	>	volatile ForkJoinTask<?> currentJoin; // task being joined in awaitJoin
642		ForkJoinTask<?> currentSteal; // current non-local task being executed
643		// Heuristic padding to ameliorate unfortunate memory placements
644	<	Object p00, p01, p02, p03, p04, p05, p06, p07, p08, p09, p0a;
644	>	Object p00, p01, p02, p03, p04, p05, p06, p07;
645	>	Object p08, p09, p0a, p0b, p0c, p0d, p0e;
646
647	<	WorkQueue(ForkJoinWorkerThread owner, int mode) {
728	<	this.owner = owner;
647	>	WorkQueue(ForkJoinPool pool, ForkJoinWorkerThread owner, int mode) {
648		this.mode = mode;
649	+	this.pool = pool;
650	+	this.owner = owner;
651		// Place indices in the center of array (that is not yet allocated)
652		base = top = INITIAL_QUEUE_CAPACITY >>> 1;
653		}
654
655		/**
735	–	* Returns number of tasks in the queue
736	–	*/
737	–	final int queueSize() {
738	–	int n = base - top; // non-owner callers must read base first
739	–	return (n >= 0) ? 0 : -n;
740	–	}
741	–
742	–	/**
656		* Pushes a task. Call only by owner in unshared queues.
657	+	* Cases needing resizing or rejection are relayed to fullPush
658	+	* (that also handles shared queues).
659		*
660		* @param task the task. Caller must ensure non-null.
661	<	* @param p, if non-null, pool to signal if necessary
662	<	* @throw RejectedExecutionException if array cannot
663	<	* be resized
664	<	*/
665	<	final void push(ForkJoinTask<?> task, ForkJoinPool p) {
666	<	ForkJoinTask<?>[] a;
752	<	int s = top, m, n;
753	<	if ((a = array) != null) { // ignore if queue removed
661	>	* @throw RejectedExecutionException if array cannot be resized
662	>	*/
663	>	final void push(ForkJoinTask<?> task) {
664	>	ForkJoinPool p; ForkJoinTask<?>[] a;
665	>	int s = top, n;
666	>	if ((a = array) != null && a.length > (n = s + 1 - base)) {
667		U.putOrderedObject
668	<	(a, (((m = a.length - 1) & s) << ASHIFT) + ABASE, task);
669	<	if ((n = (top = s + 1) - base) <= 2) {
670	<	if (p != null)
671	<	p.signalWork();
759	<	}
760	<	else if (n >= m)
761	<	growArray(true);
668	>	(a, (((a.length - 1) & s) << ASHIFT) + ABASE, task);
669	>	top = s + 1;
670	>	if (n <= 1 && (p = pool) != null)
671	>	p.signalWork(this, 1);
672		}
673	+	else
674	+	fullPush(task, true);
675		}
676
677		/**
678		* Pushes a task if lock is free and array is either big
679	<	* enough or can be resized to be big enough.
679	>	* enough or can be resized to be big enough. Note: a
680	>	* specialization of a common fast path of this method is in
681	>	* ForkJoinPool.externalPush. When called from a FJWT queue,
682	>	* this can fail only if the pool has been shut down or
683	>	* an out of memory error.
684		*
685		* @param task the task. Caller must ensure non-null.
686	<	* @return true if submitted
686	>	* @param owned if true, throw RJE on failure
687		*/
688	<	final boolean trySharedPush(ForkJoinTask<?> task) {
689	<	boolean submitted = false;
690	<	if (runState == 0 && U.compareAndSwapInt(this, RUNSTATE, 0, 1)) {
691	<	ForkJoinTask<?>[] a = array;
692	<	int s = top, n = s - base;
693	<	try {
694	<	if ((a != null && n < a.length - 1) \|\|
695	<	(a = growArray(false)) != null) { // must presize
696	<	int j = (((a.length - 1) & s) << ASHIFT) + ABASE;
697	<	U.putObject(a, (long)j, task); // don't need "ordered"
698	<	top = s + 1;
699	<	submitted = true;
688	>	final boolean fullPush(ForkJoinTask<?> task, boolean owned) {
689	>	ForkJoinPool p; ForkJoinTask<?>[] a;
690	>	if (owned) {
691	>	if (qlock < 0) // must be shutting down
692	>	throw new RejectedExecutionException();
693	>	}
694	>	else if (!U.compareAndSwapInt(this, QLOCK, 0, 1))
695	>	return false;
696	>	try {
697	>	int s = top, oldLen, len;
698	>	if ((a = array) == null)
699	>	a = array = new ForkJoinTask<?>[len=INITIAL_QUEUE_CAPACITY];
700	>	else if ((oldLen = a.length) > s + 1 - base)
701	>	len = oldLen;
702	>	else if ((len = oldLen << 1) > MAXIMUM_QUEUE_CAPACITY)
703	>	throw new RejectedExecutionException("Capacity exceeded");
704	>	else {
705	>	int oldMask, b;
706	>	ForkJoinTask<?>[] oldA = a;
707	>	a = array = new ForkJoinTask<?>[len];
708	>	if ((oldMask = oldLen - 1) >= 0 && s - (b = base) > 0) {
709	>	int mask = len - 1;
710	>	do {
711	>	ForkJoinTask<?> x;
712	>	int oldj = ((b & oldMask) << ASHIFT) + ABASE;
713	>	int j = ((b & mask) << ASHIFT) + ABASE;
714	>	x = (ForkJoinTask<?>)
715	>	U.getObjectVolatile(oldA, oldj);
716	>	if (x != null &&
717	>	U.compareAndSwapObject(oldA, oldj, x, null))
718	>	U.putObjectVolatile(a, j, x);
719	>	} while (++b != s);
720		}
785	–	} finally {
786	–	runState = 0; // unlock
721		}
722	+	U.putOrderedObject
723	+	(a, (((len - 1) & s) << ASHIFT) + ABASE, task);
724	+	top = s + 1;
725	+	} finally {
726	+	if (!owned)
727	+	qlock = 0;
728		}
729	<	return submitted;
729	>	if ((p = pool) != null)
730	>	p.signalWork(this, 1);
731	>	return true;
732		}
733
734		/**
735	<	* Takes next task, if one exists, in FIFO order.
735	>	* Takes next task, if one exists, in LIFO order. Call only
736	>	* by owner in unshared queues.
737		*/
738	<	final ForkJoinTask<?> poll() {
739	<	ForkJoinTask<?>[] a; int b, i;
740	<	while ((b = base) - top < 0 && (a = array) != null &&
741	<	(i = (a.length - 1) & b) >= 0) {
742	<	int j = (i << ASHIFT) + ABASE;
743	<	ForkJoinTask<?> t = (ForkJoinTask<?>)U.getObjectVolatile(a, j);
744	<	if (t != null && base == b &&
738	>	final ForkJoinTask<?> pop() {
739	>	ForkJoinTask<?>[] a; ForkJoinTask<?> t; int m;
740	>	if ((a = array) != null && (m = a.length - 1) >= 0) {
741	>	for (int s; (s = top - 1) - base >= 0;) {
742	>	long j = ((m & s) << ASHIFT) + ABASE;
743	>	if ((t = (ForkJoinTask<?>)U.getObject(a, j)) == null)
744	>	break;
745	>	if (U.compareAndSwapObject(a, j, t, null)) {
746	>	top = s;
747	>	return t;
748	>	}
749	>	}
750	>	}
751	>	return null;
752	>	}
753	>
754	>	/**
755	>	* Takes a task in FIFO order if b is base of queue and a task
756	>	* can be claimed without contention. Specialized versions
757	>	* appear in ForkJoinPool methods scan and tryHelpStealer.
758	>	*/
759	>	final ForkJoinTask<?> pollAt(int b) {
760	>	ForkJoinTask<?> t; ForkJoinTask<?>[] a;
761	>	if ((a = array) != null) {
762	>	int j = (((a.length - 1) & b) << ASHIFT) + ABASE;
763	>	if ((t = (ForkJoinTask<?>)U.getObjectVolatile(a, j)) != null &&
764	>	base == b &&
765		U.compareAndSwapObject(a, j, t, null)) {
766		base = b + 1;
767		return t;
#	Line 808 \| Line 771 \| public class ForkJoinPool extends Abstra
771		}
772
773		/**
774	<	* Takes next task, if one exists, in LIFO order.
812	<	* Call only by owner in unshared queues.
774	>	* Takes next task, if one exists, in FIFO order.
775		*/
776	<	final ForkJoinTask<?> pop() {
777	<	ForkJoinTask<?> t; int m;
778	<	ForkJoinTask<?>[] a = array;
779	<	if (a != null && (m = a.length - 1) >= 0) {
780	<	for (int s; (s = top - 1) - base >= 0;) {
781	<	int j = ((m & s) << ASHIFT) + ABASE;
782	<	if ((t = (ForkJoinTask<?>)U.getObjectVolatile(a, j)) == null)
783	<	break;
784	<	if (U.compareAndSwapObject(a, j, t, null)) {
823	<	top = s;
776	>	final ForkJoinTask<?> poll() {
777	>	ForkJoinTask<?>[] a; int b; ForkJoinTask<?> t;
778	>	while ((b = base) - top < 0 && (a = array) != null) {
779	>	int j = (((a.length - 1) & b) << ASHIFT) + ABASE;
780	>	t = (ForkJoinTask<?>)U.getObjectVolatile(a, j);
781	>	if (t != null) {
782	>	if (base == b &&
783	>	U.compareAndSwapObject(a, j, t, null)) {
784	>	base = b + 1;
785		return t;
786		}
787		}
788	+	else if (base == b) {
789	+	if (b + 1 == top)
790	+	break;
791	+	Thread.yield(); // wait for lagging update (very rare)
792	+	}
793		}
794		return null;
795		}
#	Line 848 \| Line 814 \| public class ForkJoinPool extends Abstra
814		}
815
816		/**
851	–	* Returns task at index b if b is current base of queue.
852	–	*/
853	–	final ForkJoinTask<?> pollAt(int b) {
854	–	ForkJoinTask<?>[] a; int i;
855	–	ForkJoinTask<?> task = null;
856	–	if ((a = array) != null && (i = ((a.length - 1) & b)) >= 0) {
857	–	int j = (i << ASHIFT) + ABASE;
858	–	ForkJoinTask<?> t = (ForkJoinTask<?>)U.getObjectVolatile(a, j);
859	–	if (t != null && base == b &&
860	–	U.compareAndSwapObject(a, j, t, null)) {
861	–	base = b + 1;
862	–	task = t;
863	–	}
864	–	}
865	–	return task;
866	–	}
867	–
868	–	/**
817		* Pops the given task only if it is at the current top.
818	+	* (A shared version is available only via FJP.tryExternalUnpush)
819		*/
820		final boolean tryUnpush(ForkJoinTask<?> t) {
821		ForkJoinTask<?>[] a; int s;
#	Line 880 \| Line 829 \| public class ForkJoinPool extends Abstra
829		}
830
831		/**
832	<	* Polls the given task only if it is at the current base.
832	>	* Removes and cancels all known tasks, ignoring any exceptions.
833		*/
834	<	final boolean pollFor(ForkJoinTask<?> task) {
835	<	ForkJoinTask<?>[] a; int b, i;
836	<	if ((b = base) - top < 0 && (a = array) != null &&
837	<	(i = (a.length - 1) & b) >= 0) {
838	<	int j = (i << ASHIFT) + ABASE;
839	<	if (U.getObjectVolatile(a, j) == task && base == b &&
840	<	U.compareAndSwapObject(a, j, task, null)) {
841	<	base = b + 1;
842	<	return true;
834	>	final void cancelAll() {
835	>	ForkJoinTask.cancelIgnoringExceptions(currentJoin);
836	>	ForkJoinTask.cancelIgnoringExceptions(currentSteal);
837	>	for (ForkJoinTask<?> t; (t = poll()) != null; )
838	>	ForkJoinTask.cancelIgnoringExceptions(t);
839	>	}
840	>
841	>	/**
842	>	* Computes next value for random probes. Scans don't require
843	>	* a very high quality generator, but also not a crummy one.
844	>	* Marsaglia xor-shift is cheap and works well enough. Note:
845	>	* This is manually inlined in its usages in ForkJoinPool to
846	>	* avoid writes inside busy scan loops.
847	>	*/
848	>	final int nextSeed() {
849	>	int r = seed;
850	>	r ^= r << 13;
851	>	r ^= r >>> 17;
852	>	return seed = r ^= r << 5;
853	>	}
854	>
855	>	/**
856	>	* Provides a more accurate estimate of size than (top - base)
857	>	* by ordering reads and checking whether a near-empty queue
858	>	* has at least one unclaimed task.
859	>	*/
860	>	final int queueSize() {
861	>	ForkJoinTask<?>[] a; int k, s, n;
862	>	return ((n = base - (s = top)) < 0 &&
863	>	(n != -1 \|\|
864	>	((a = array) != null && (k = a.length) > 0 &&
865	>	U.getObject
866	>	(a, (long)((((k - 1) & (s - 1)) << ASHIFT) + ABASE)) != null))) ?
867	>	-n : 0;
868	>	}
869	>
870	>	// Specialized execution methods
871	>
872	>	/**
873	>	* Pops and runs tasks until empty.
874	>	*/
875	>	private void popAndExecAll() {
876	>	// A bit faster than repeated pop calls
877	>	ForkJoinTask<?>[] a; int m, s; long j; ForkJoinTask<?> t;
878	>	while ((a = array) != null && (m = a.length - 1) >= 0 &&
879	>	(s = top - 1) - base >= 0 &&
880	>	(t = ((ForkJoinTask<?>)
881	>	U.getObject(a, j = ((m & s) << ASHIFT) + ABASE)))
882	>	!= null) {
883	>	if (U.compareAndSwapObject(a, j, t, null)) {
884	>	top = s;
885	>	t.doExec();
886		}
887		}
896	–	return false;
888		}
889
890		/**
891	<	* If present, removes from queue and executes the given task, or
892	<	* any other cancelled task. Returns (true) immediately on any CAS
891	>	* Polls and runs tasks until empty.
892	>	*/
893	>	private void pollAndExecAll() {
894	>	for (ForkJoinTask<?> t; (t = poll()) != null;)
895	>	t.doExec();
896	>	}
897	>
898	>	/**
899	>	* If present, removes from queue and executes the given task,
900	>	* or any other cancelled task. Returns (true) on any CAS
901		* or consistency check failure so caller can retry.
902		*
903	<	* @return false if no progress can be made
903	>	* @return false if no progress can be made, else true;
904		*/
905		final boolean tryRemoveAndExec(ForkJoinTask<?> task) {
906	<	boolean removed = false, empty = true, progress = true;
906	>	boolean stat = true, removed = false, empty = true;
907		ForkJoinTask<?>[] a; int m, s, b, n;
908		if ((a = array) != null && (m = a.length - 1) >= 0 &&
909		(n = (s = top) - (b = base)) > 0) {
#	Line 934 \| Line 933 \| public class ForkJoinPool extends Abstra
933		}
934		if (--n == 0) {
935		if (!empty && base == b)
936	<	progress = false;
936	>	stat = false;
937		break;
938		}
939		}
940		}
941		if (removed)
942		task.doExec();
943	<	return progress;
943	>	return stat;
944		}
945
946		/**
947	<	* Initializes or doubles the capacity of array. Call either
948	<	* by owner or with lock held -- it is OK for base, but not
950	<	* top, to move while resizings are in progress.
951	<	*
952	<	* @param rejectOnFailure if true, throw exception if capacity
953	<	* exceeded (relayed ultimately to user); else return null.
947	>	* Polls for and executes the given task or any other task in
948	>	* its CountedCompleter computation
949		*/
950	<	final ForkJoinTask<?>[] growArray(boolean rejectOnFailure) {
951	<	ForkJoinTask<?>[] oldA = array;
952	<	int size = oldA != null ? oldA.length << 1 : INITIAL_QUEUE_CAPACITY;
953	<	if (size <= MAXIMUM_QUEUE_CAPACITY) {
954	<	int oldMask, t, b;
955	<	ForkJoinTask<?>[] a = array = new ForkJoinTask<?>[size];
956	<	if (oldA != null && (oldMask = oldA.length - 1) >= 0 &&
957	<	(t = top) - (b = base) > 0) {
958	<	int mask = size - 1;
959	<	do {
960	<	ForkJoinTask<?> x;
961	<	int oldj = ((b & oldMask) << ASHIFT) + ABASE;
962	<	int j = ((b & mask) << ASHIFT) + ABASE;
963	<	x = (ForkJoinTask<?>)U.getObjectVolatile(oldA, oldj);
964	<	if (x != null &&
965	<	U.compareAndSwapObject(oldA, oldj, x, null))
966	<	U.putObjectVolatile(a, j, x);
967	<	} while (++b != t);
950	>	final boolean pollAndExecCC(ForkJoinTask<?> root) {
951	>	ForkJoinTask<?>[] a; int b; Object o;
952	>	outer: while ((b = base) - top < 0 && (a = array) != null) {
953	>	long j = (((a.length - 1) & b) << ASHIFT) + ABASE;
954	>	if ((o = U.getObject(a, j)) == null \|\|
955	>	!(o instanceof CountedCompleter))
956	>	break;
957	>	for (CountedCompleter<?> t = (CountedCompleter<?>)o, r = t;;) {
958	>	if (r == root) {
959	>	if (base == b &&
960	>	U.compareAndSwapObject(a, j, t, null)) {
961	>	base = b + 1;
962	>	t.doExec();
963	>	return true;
964	>	}
965	>	else
966	>	break; // restart
967	>	}
968	>	if ((r = r.completer) == null)
969	>	break outer; // not part of root computation
970		}
974	–	return a;
975	–	}
976	–	else if (!rejectOnFailure)
977	–	return null;
978	–	else
979	–	throw new RejectedExecutionException("Queue capacity exceeded");
980	–	}
981	–
982	–	/**
983	–	* Removes and cancels all known tasks, ignoring any exceptions
984	–	*/
985	–	final void cancelAll() {
986	–	ForkJoinTask.cancelIgnoringExceptions(currentJoin);
987	–	ForkJoinTask.cancelIgnoringExceptions(currentSteal);
988	–	for (ForkJoinTask<?> t; (t = poll()) != null; )
989	–	ForkJoinTask.cancelIgnoringExceptions(t);
990	–	}
991	–
992	–	// Execution methods
993	–
994	–	/**
995	–	* Removes and runs tasks until empty, using local mode
996	–	* ordering.
997	–	*/
998	–	final void runLocalTasks() {
999	–	if (base - top < 0) {
1000	–	for (ForkJoinTask<?> t; (t = nextLocalTask()) != null; )
1001	–	t.doExec();
971		}
972	+	return false;
973		}
974
975		/**
976		* Executes a top-level task and any local tasks remaining
977		* after execution.
1008	–	*
1009	–	* @return true unless terminating
978		*/
979	<	final boolean runTask(ForkJoinTask<?> t) {
1012	<	boolean alive = true;
979	>	final void runTask(ForkJoinTask<?> t) {
980		if (t != null) {
981	<	currentSteal = t;
1015	<	t.doExec();
1016	<	runLocalTasks();
1017	<	++nsteals;
981	>	(currentSteal = t).doExec();
982		currentSteal = null;
983	+	if (++nsteals < 0) { // spill on overflow
984	+	ForkJoinPool p;
985	+	if ((p = pool) != null)
986	+	p.collectStealCount(this);
987	+	}
988	+	if (top != base) { // process remaining local tasks
989	+	if (mode == 0)
990	+	popAndExecAll();
991	+	else
992	+	pollAndExecAll();
993	+	}
994		}
1020	–	else if (runState < 0) // terminating
1021	–	alive = false;
1022	–	return alive;
995		}
996
997		/**
998	<	* Executes a non-top-level (stolen) task
998	>	* Executes a non-top-level (stolen) task.
999		*/
1000		final void runSubtask(ForkJoinTask<?> t) {
1001		if (t != null) {
1002		ForkJoinTask<?> ps = currentSteal;
1003	<	currentSteal = t;
1032	<	t.doExec();
1003	>	(currentSteal = t).doExec();
1004		currentSteal = ps;
1005		}
1006		}
1007
1008		/**
1009	<	* Computes next value for random probes. Scans don't require
1039	<	* a very high quality generator, but also not a crummy one.
1040	<	* Marsaglia xor-shift is cheap and works well enough. Note:
1041	<	* This is manually inlined in several usages in ForkJoinPool
1042	<	* to avoid writes inside busy scan loops.
1009	>	* Returns true if owned and not known to be blocked.
1010		*/
1011	<	final int nextSeed() {
1012	<	int r = seed;
1013	<	r ^= r << 13;
1014	<	r ^= r >>> 17;
1015	<	r ^= r << 5;
1016	<	return seed = r;
1011	>	final boolean isApparentlyUnblocked() {
1012	>	Thread wt; Thread.State s;
1013	>	return (eventCount >= 0 &&
1014	>	(wt = owner) != null &&
1015	>	(s = wt.getState()) != Thread.State.BLOCKED &&
1016	>	s != Thread.State.WAITING &&
1017	>	s != Thread.State.TIMED_WAITING);
1018	>	}
1019	>
1020	>	/**
1021	>	* If this owned and is not already interrupted, try to
1022	>	* interrupt and/or unpark, ignoring exceptions.
1023	>	*/
1024	>	final void interruptOwner() {
1025	>	Thread wt, p;
1026	>	if ((wt = owner) != null && !wt.isInterrupted()) {
1027	>	try {
1028	>	wt.interrupt();
1029	>	} catch (SecurityException ignore) {
1030	>	}
1031	>	}
1032	>	if ((p = parker) != null)
1033	>	U.unpark(p);
1034		}
1035
1036		// Unsafe mechanics
1037		private static final sun.misc.Unsafe U;
1038	<	private static final long RUNSTATE;
1038	>	private static final long QLOCK;
1039		private static final int ABASE;
1040		private static final int ASHIFT;
1041		static {
#	Line 1060 \| Line 1044 \| public class ForkJoinPool extends Abstra
1044		U = getUnsafe();
1045		Class<?> k = WorkQueue.class;
1046		Class<?> ak = ForkJoinTask[].class;
1047	<	RUNSTATE = U.objectFieldOffset
1048	<	(k.getDeclaredField("runState"));
1047	>	QLOCK = U.objectFieldOffset
1048	>	(k.getDeclaredField("qlock"));
1049		ABASE = U.arrayBaseOffset(ak);
1050		s = U.arrayIndexScale(ak);
1051		} catch (Exception e) {
#	Line 1074 \| Line 1058 \| public class ForkJoinPool extends Abstra
1058		}
1059
1060		/**
1061	<	* Class for artificial tasks that are used to replace the target
1062	<	* of local joins if they are removed from an interior queue slot
1063	<	* in WorkQueue.tryRemoveAndExec. We don't need the proxy to
1064	<	* actually do anything beyond having a unique identity.
1061	>	* Per-thread records for threads that submit to pools. Currently
1062	>	* holds only pseudo-random seed / index that is used to choose
1063	>	* submission queues in method externalPush. In the future, this may
1064	>	* also incorporate a means to implement different task rejection
1065	>	* and resubmission policies.
1066	>	*
1067	>	* Seeds for submitters and workers/workQueues work in basically
1068	>	* the same way but are initialized and updated using slightly
1069	>	* different mechanics. Both are initialized using the same
1070	>	* approach as in class ThreadLocal, where successive values are
1071	>	* unlikely to collide with previous values. Seeds are then
1072	>	* randomly modified upon collisions using xorshifts, which
1073	>	* requires a non-zero seed.
1074		*/
1075	<	static final class EmptyTask extends ForkJoinTask<Void> {
1076	<	EmptyTask() { status = ForkJoinTask.NORMAL; } // force done
1077	<	public Void getRawResult() { return null; }
1085	<	public void setRawResult(Void x) {}
1086	<	public boolean exec() { return true; }
1075	>	static final class Submitter {
1076	>	int seed;
1077	>	Submitter(int s) { seed = s; }
1078		}
1079
1080	<	/**
1081	<	<<<<<<< ForkJoinPool.java
1082	<	* Per-thread records for (typically non-FJ) threads that submit
1092	<	* to pools. Cureently holds only psuedo-random seed / index that
1093	<	* is used to chose submission queues in method doSubmit. In the
1094	<	* future, this may incorporate a means to implement different
1095	<	* task rejection and resubmission policies.
1096	<	*/
1097	<	static final class Submitter {
1098	<	int seed; // seed for random submission queue selection
1080	>	/** Property prefix for constructing common pool */
1081	>	private static final String propPrefix =
1082	>	"java.util.concurrent.ForkJoinPool.common.";
1083
1084	<	// Heuristic padding to ameliorate unfortunate memory placements
1101	<	int p0, p1, p2, p3, p4, p5, p6, p7, p8, p9, pa, pb, pc, pd, pe;
1084	>	// static fields (initialized in static initializer below)
1085
1086	<	Submitter() {
1087	<	// Use identityHashCode, forced negative, for seed
1088	<	seed = System.identityHashCode(Thread.currentThread()) \| (1 << 31);
1089	<	}
1086	>	/**
1087	>	* Creates a new ForkJoinWorkerThread. This factory is used unless
1088	>	* overridden in ForkJoinPool constructors.
1089	>	*/
1090	>	public static final ForkJoinWorkerThreadFactory
1091	>	defaultForkJoinWorkerThreadFactory;
1092
1093	<	/**
1094	<	* Computes next value for random probes. Like method
1095	<	* WorkQueue.nextSeed, this is manually inlined in several
1096	<	* usages to avoid writes inside busy loops.
1097	<	*/
1098	<	final int nextSeed() {
1099	<	int r = seed;
1115	<	r ^= r << 13;
1116	<	r ^= r >>> 17;
1117	<	return seed = r ^= r << 5;
1118	<	}
1119	<	}
1093	>	/**
1094	>	* Common (static) pool. Non-null for public use unless a static
1095	>	* construction exception, but internal usages null-check on use
1096	>	* to paranoically avoid potential initialization circularities
1097	>	* as well as to simplify generated code.
1098	>	*/
1099	>	static final ForkJoinPool commonPool;
1100
1101	<	/** ThreadLocal class for Submitters */
1102	<	static final class ThreadSubmitter extends ThreadLocal<Submitter> {
1103	<	public Submitter initialValue() { return new Submitter(); }
1104	<	}
1101	>	/**
1102	>	* Permission required for callers of methods that may start or
1103	>	* kill threads.
1104	>	*/
1105	>	private static final RuntimePermission modifyThreadPermission;
1106
1107		/**
1108	<	* Per-thread submission bookeeping. Shared across all pools
1108	>	* Per-thread submission bookkeeping. Shared across all pools
1109		* to reduce ThreadLocal pollution and because random motion
1110		* to avoid contention in one pool is likely to hold for others.
1111	+	* Lazily initialized on first submission (but null-checked
1112	+	* in other contexts to avoid unnecessary initialization).
1113		*/
1114	<	static final ThreadSubmitter submitters = new ThreadSubmitter();
1114	>	static final ThreadLocal<Submitter> submitters;
1115
1116		/**
1117	<	* Top-level runloop for workers
1117	>	* Common pool parallelism. Must equal commonPool.parallelism.
1118		*/
1119	<	final void runWorker(ForkJoinWorkerThread wt) {
1137	<	// Initialize queue array and seed in this thread
1138	<	WorkQueue w = wt.workQueue;
1139	<	w.growArray(false);
1140	<	// Same initial hash as Submitters
1141	<	w.seed = System.identityHashCode(Thread.currentThread()) \| (1 << 31);
1119	>	static final int commonPoolParallelism;
1120
1121	<	do {} while (w.runTask(scan(w)));
1121	>	/**
1122	>	* Sequence number for creating workerNamePrefix.
1123	>	*/
1124	>	private static int poolNumberSequence;
1125	>
1126	>	/**
1127	>	* Return the next sequence number. We don't expect this to
1128	>	* ever contend so use simple builtin sync.
1129	>	*/
1130	>	private static final synchronized int nextPoolId() {
1131	>	return ++poolNumberSequence;
1132		}
1133
1134	<	// Creating, registering and deregistering workers
1134	>	// static constants
1135
1136		/**
1137	<	* Tries to create and start a worker
1137	>	* Initial timeout value (in nanoseconds) for the thread
1138	>	* triggering quiescence to park waiting for new work. On timeout,
1139	>	* the thread will instead try to shrink the number of
1140	>	* workers. The value should be large enough to avoid overly
1141	>	* aggressive shrinkage during most transient stalls (long GCs
1142	>	* etc).
1143		*/
1144	<	private void addWorker() {
1145	<	Throwable ex = null;
1146	<	ForkJoinWorkerThread w = null;
1147	<	try {
1148	<	if ((w = factory.newThread(this)) != null) {
1149	<	w.start();
1150	<	return;
1144	>	private static final long IDLE_TIMEOUT = 2000L * 1000L * 1000L; // 2sec
1145	>
1146	>	/**
1147	>	* Timeout value when there are more threads than parallelism level
1148	>	*/
1149	>	private static final long FAST_IDLE_TIMEOUT = 200L * 1000L * 1000L;
1150	>
1151	>	/**
1152	>	* The maximum stolen->joining link depth allowed in method
1153	>	* tryHelpStealer. Must be a power of two. Depths for legitimate
1154	>	* chains are unbounded, but we use a fixed constant to avoid
1155	>	* (otherwise unchecked) cycles and to bound staleness of
1156	>	* traversal parameters at the expense of sometimes blocking when
1157	>	* we could be helping.
1158	>	*/
1159	>	private static final int MAX_HELP = 64;
1160	>
1161	>	/**
1162	>	* Increment for seed generators. See class ThreadLocal for
1163	>	* explanation.
1164	>	*/
1165	>	private static final int SEED_INCREMENT = 0x61c88647;
1166	>
1167	>	/**
1168	>	* Bits and masks for control variables
1169	>	*
1170	>	* Field ctl is a long packed with:
1171	>	* AC: Number of active running workers minus target parallelism (16 bits)
1172	>	* TC: Number of total workers minus target parallelism (16 bits)
1173	>	* ST: true if pool is terminating (1 bit)
1174	>	* EC: the wait count of top waiting thread (15 bits)
1175	>	* ID: poolIndex of top of Treiber stack of waiters (16 bits)
1176	>	*
1177	>	* When convenient, we can extract the upper 32 bits of counts and
1178	>	* the lower 32 bits of queue state, u = (int)(ctl >>> 32) and e =
1179	>	* (int)ctl. The ec field is never accessed alone, but always
1180	>	* together with id and st. The offsets of counts by the target
1181	>	* parallelism and the positionings of fields makes it possible to
1182	>	* perform the most common checks via sign tests of fields: When
1183	>	* ac is negative, there are not enough active workers, when tc is
1184	>	* negative, there are not enough total workers, and when e is
1185	>	* negative, the pool is terminating. To deal with these possibly
1186	>	* negative fields, we use casts in and out of "short" and/or
1187	>	* signed shifts to maintain signedness.
1188	>	*
1189	>	* When a thread is queued (inactivated), its eventCount field is
1190	>	* set negative, which is the only way to tell if a worker is
1191	>	* prevented from executing tasks, even though it must continue to
1192	>	* scan for them to avoid queuing races. Note however that
1193	>	* eventCount updates lag releases so usage requires care.
1194	>	*
1195	>	* Field plock is an int packed with:
1196	>	* SHUTDOWN: true if shutdown is enabled (1 bit)
1197	>	* SEQ: a sequence lock, with PL_LOCK bit set if locked (30 bits)
1198	>	* SIGNAL: set when threads may be waiting on the lock (1 bit)
1199	>	*
1200	>	* The sequence number enables simple consistency checks:
1201	>	* Staleness of read-only operations on the workQueues array can
1202	>	* be checked by comparing plock before vs after the reads.
1203	>	*/
1204	>
1205	>	// bit positions/shifts for fields
1206	>	private static final int AC_SHIFT = 48;
1207	>	private static final int TC_SHIFT = 32;
1208	>	private static final int ST_SHIFT = 31;
1209	>	private static final int EC_SHIFT = 16;
1210	>
1211	>	// bounds
1212	>	private static final int SMASK = 0xffff; // short bits
1213	>	private static final int MAX_CAP = 0x7fff; // max #workers - 1
1214	>	private static final int EVENMASK = 0xfffe; // even short bits
1215	>	private static final int SQMASK = 0x007e; // max 64 (even) slots
1216	>	private static final int SHORT_SIGN = 1 << 15;
1217	>	private static final int INT_SIGN = 1 << 31;
1218	>
1219	>	// masks
1220	>	private static final long STOP_BIT = 0x0001L << ST_SHIFT;
1221	>	private static final long AC_MASK = ((long)SMASK) << AC_SHIFT;
1222	>	private static final long TC_MASK = ((long)SMASK) << TC_SHIFT;
1223	>
1224	>	// units for incrementing and decrementing
1225	>	private static final long TC_UNIT = 1L << TC_SHIFT;
1226	>	private static final long AC_UNIT = 1L << AC_SHIFT;
1227	>
1228	>	// masks and units for dealing with u = (int)(ctl >>> 32)
1229	>	private static final int UAC_SHIFT = AC_SHIFT - 32;
1230	>	private static final int UTC_SHIFT = TC_SHIFT - 32;
1231	>	private static final int UAC_MASK = SMASK << UAC_SHIFT;
1232	>	private static final int UTC_MASK = SMASK << UTC_SHIFT;
1233	>	private static final int UAC_UNIT = 1 << UAC_SHIFT;
1234	>	private static final int UTC_UNIT = 1 << UTC_SHIFT;
1235	>
1236	>	// masks and units for dealing with e = (int)ctl
1237	>	private static final int E_MASK = 0x7fffffff; // no STOP_BIT
1238	>	private static final int E_SEQ = 1 << EC_SHIFT;
1239	>
1240	>	// plock bits
1241	>	private static final int SHUTDOWN = 1 << 31;
1242	>	private static final int PL_LOCK = 2;
1243	>	private static final int PL_SIGNAL = 1;
1244	>	private static final int PL_SPINS = 1 << 8;
1245	>
1246	>	// access mode for WorkQueue
1247	>	static final int LIFO_QUEUE = 0;
1248	>	static final int FIFO_QUEUE = 1;
1249	>	static final int SHARED_QUEUE = -1;
1250	>
1251	>	// Instance fields
1252	>
1253	>	/*
1254	>	* Field layout order in this class tends to matter more than one
1255	>	* would like. Runtime layout order is only loosely related to
1256	>	* declaration order and may differ across JVMs, but the following
1257	>	* empirically works OK on current JVMs.
1258	>	*/
1259	>	volatile long stealCount; // collects worker counts
1260	>	volatile long ctl; // main pool control
1261	>	final int parallelism; // parallelism level
1262	>	final int localMode; // per-worker scheduling mode
1263	>	volatile int indexSeed; // worker/submitter index seed
1264	>	volatile int plock; // shutdown status and seqLock
1265	>	WorkQueue[] workQueues; // main registry
1266	>	final ForkJoinWorkerThreadFactory factory; // factory for new workers
1267	>	final Thread.UncaughtExceptionHandler ueh; // per-worker UEH
1268	>	final String workerNamePrefix; // to create worker name string
1269	>
1270	>	/*
1271	>	* Acquires the plock lock to protect worker array and related
1272	>	* updates. This method is called only if an initial CAS on plock
1273	>	* fails. This acts as a spinLock for normal cases, but falls back
1274	>	* to builtin monitor to block when (rarely) needed. This would be
1275	>	* a terrible idea for a highly contended lock, but works fine as
1276	>	* a more conservative alternative to a pure spinlock. See
1277	>	* internal ConcurrentHashMap documentation for further
1278	>	* explanation of nearly the same construction.
1279	>	*/
1280	>	private int acquirePlock() {
1281	>	int spins = PL_SPINS, r = 0, ps, nps;
1282	>	for (;;) {
1283	>	if (((ps = plock) & PL_LOCK) == 0 &&
1284	>	U.compareAndSwapInt(this, PLOCK, ps, nps = ps + PL_LOCK))
1285	>	return nps;
1286	>	else if (r == 0)
1287	>	r = ThreadLocalRandom.current().nextInt(); // randomize spins
1288	>	else if (spins >= 0) {
1289	>	r ^= r << 1; r ^= r >>> 3; r ^= r << 10; // xorshift
1290	>	if (r >= 0)
1291	>	--spins;
1292	>	}
1293	>	else if (U.compareAndSwapInt(this, PLOCK, ps, ps \| PL_SIGNAL)) {
1294	>	synchronized (this) {
1295	>	if ((plock & PL_SIGNAL) != 0) {
1296	>	try {
1297	>	wait();
1298	>	} catch (InterruptedException ie) {
1299	>	try {
1300	>	Thread.currentThread().interrupt();
1301	>	} catch (SecurityException ignore) {
1302	>	}
1303	>	}
1304	>	}
1305	>	else
1306	>	notifyAll();
1307	>	}
1308		}
1159	–	} catch (Throwable e) {
1160	–	ex = e;
1309		}
1162	–	deregisterWorker(w, ex);
1310		}
1311
1312		/**
1313	<	* Callback from ForkJoinWorkerThread constructor to assign a
1314	<	* public name. This must be separate from registerWorker because
1168	<	* it is called during the "super" constructor call in
1169	<	* ForkJoinWorkerThread.
1313	>	* Unlocks and signals any thread waiting for plock. Called only
1314	>	* when CAS of seq value for unlock fails.
1315		*/
1316	<	final String nextWorkerName() {
1317	<	return workerNamePrefix.concat
1318	<	(Integer.toString(nextWorkerNumber.addAndGet(1)));
1316	>	private void releasePlock(int ps) {
1317	>	plock = ps;
1318	>	synchronized (this) { notifyAll(); }
1319		}
1320
1321	+	// Registering and deregistering workers
1322	+
1323		/**
1324	<	* Callback from ForkJoinWorkerThread constructor to establish and
1325	<	* record its WorkQueue
1324	>	* Callback from ForkJoinWorkerThread constructor to establish its
1325	>	* poolIndex and record its WorkQueue. To avoid scanning bias due
1326	>	* to packing entries in front of the workQueues array, we treat
1327	>	* the array as a simple power-of-two hash table using per-thread
1328	>	* seed as hash, expanding as needed.
1329		*
1330	<	* @param wt the worker thread
1330	>	* @param w the worker's queue
1331		*/
1332	<	final void registerWorker(ForkJoinWorkerThread wt) {
1333	<	WorkQueue w = wt.workQueue;
1334	<	ReentrantLock lock = this.lock;
1335	<	lock.lock();
1332	>	final void registerWorker(WorkQueue w) {
1333	>	int s, ps; // generate a rarely colliding candidate index seed
1334	>	do {} while (!U.compareAndSwapInt(this, INDEXSEED,
1335	>	s = indexSeed, s += SEED_INCREMENT) \|\|
1336	>	s == 0); // skip 0
1337	>	if (((ps = plock) & PL_LOCK) != 0 \|\|
1338	>	!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK))
1339	>	ps = acquirePlock();
1340	>	int nps = (ps & SHUTDOWN) \| ((ps + PL_LOCK) & ~SHUTDOWN);
1341		try {
1342	<	int k = nextPoolIndex;
1343	<	WorkQueue[] ws = workQueues;
1344	<	if (ws != null) { // ignore on shutdown
1345	<	int n = ws.length;
1346	<	if (k < 0 \|\| (k & 1) == 0 \|\| k >= n \|\| ws[k] != null) {
1347	<	for (k = 1; k < n && ws[k] != null; k += 2)
1348	<	; // workers are at odd indices
1349	<	if (k >= n) // resize
1350	<	workQueues = ws = Arrays.copyOf(ws, n << 1);
1351	<	}
1352	<	w.poolIndex = k;
1353	<	w.eventCount = ~(k >>> 1) & SMASK; // Set up wait count
1354	<	ws[k] = w; // record worker
1355	<	nextPoolIndex = k + 2;
1356	<	int rs = runState;
1357	<	int m = rs & SMASK; // recalculate runState mask
1358	<	if (k > m)
1359	<	m = (m << 1) + 1;
1205	<	runState = (rs & SHUTDOWN) \| ((rs + RS_SEQ) & RS_SEQ_MASK) \| m;
1342	>	WorkQueue[] ws;
1343	>	if (w != null && (ws = workQueues) != null) {
1344	>	w.seed = s;
1345	>	int n = ws.length, m = n - 1;
1346	>	int r = (s << 1) \| 1; // use odd-numbered indices
1347	>	if (ws[r &= m] != null) { // collision
1348	>	int probes = 0; // step by approx half size
1349	>	int step = (n <= 4) ? 2 : ((n >>> 1) & EVENMASK) + 2;
1350	>	while (ws[r = (r + step) & m] != null) {
1351	>	if (++probes >= n) {
1352	>	workQueues = ws = Arrays.copyOf(ws, n <<= 1);
1353	>	m = n - 1;
1354	>	probes = 0;
1355	>	}
1356	>	}
1357	>	}
1358	>	w.eventCount = w.poolIndex = r; // establish before recording
1359	>	ws[r] = w;
1360		}
1361		} finally {
1362	<	lock.unlock();
1362	>	if (!U.compareAndSwapInt(this, PLOCK, ps, nps))
1363	>	releasePlock(nps);
1364		}
1365		}
1366
1367		/**
1368	<	* Final callback from terminating worker, as well as failure to
1369	<	* construct or start a worker in addWorker. Removes record of
1370	<	* worker from array, and adjusts counts. If pool is shutting
1371	<	* down, tries to complete termination.
1368	>	* Final callback from terminating worker, as well as upon failure
1369	>	* to construct or start a worker. Removes record of worker from
1370	>	* array, and adjusts counts. If pool is shutting down, tries to
1371	>	* complete termination.
1372		*
1373	<	* @param wt the worker thread or null if addWorker failed
1373	>	* @param wt the worker thread or null if construction failed
1374		* @param ex the exception causing failure, or null if none
1375		*/
1376		final void deregisterWorker(ForkJoinWorkerThread wt, Throwable ex) {
1377		WorkQueue w = null;
1378		if (wt != null && (w = wt.workQueue) != null) {
1379	<	w.runState = -1; // ensure runState is set
1380	<	stealCount.getAndAdd(w.totalSteals + w.nsteals);
1381	<	int idx = w.poolIndex;
1382	<	ReentrantLock lock = this.lock;
1383	<	lock.lock();
1384	<	try { // remove record from array
1379	>	int ps;
1380	>	collectStealCount(w);
1381	>	w.qlock = -1; // ensure set
1382	>	if (((ps = plock) & PL_LOCK) != 0 \|\|
1383	>	!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK))
1384	>	ps = acquirePlock();
1385	>	int nps = (ps & SHUTDOWN) \| ((ps + PL_LOCK) & ~SHUTDOWN);
1386	>	try {
1387	>	int idx = w.poolIndex;
1388		WorkQueue[] ws = workQueues;
1389		if (ws != null && idx >= 0 && idx < ws.length && ws[idx] == w)
1390	<	ws[nextPoolIndex = idx] = null;
1390	>	ws[idx] = null;
1391		} finally {
1392	<	lock.unlock();
1392	>	if (!U.compareAndSwapInt(this, PLOCK, ps, nps))
1393	>	releasePlock(nps);
1394		}
1395		}
1396
#	Line 1241 \| Line 1400 \| public class ForkJoinPool extends Abstra
1400		((c - TC_UNIT) & TC_MASK) \|
1401		(c & ~(AC_MASK\|TC_MASK)))));
1402
1403	<	if (!tryTerminate(false) && w != null) {
1403	>	if (!tryTerminate(false, false) && w != null) {
1404		w.cancelAll(); // cancel remaining tasks
1405		if (w.array != null) // suppress signal if never ran
1406	<	signalWork(); // wake up or create replacement
1406	>	signalWork(null, 1); // wake up or create replacement
1407	>	if (ex == null) // help clean refs on way out
1408	>	ForkJoinTask.helpExpungeStaleExceptions();
1409		}
1410
1411		if (ex != null) // rethrow
1412	<	U.throwException(ex);
1412	>	ForkJoinTask.rethrow(ex);
1413		}
1414
1415		/**
1416	<	* Tries to add and register a new queue at the given index.
1417	<	*
1418	<	* @param idx the workQueues array index to register the queue
1419	<	* @return the queue, or null if could not add because could
1420	<	* not acquire lock or idx is unusable
1421	<	*/
1422	<	private WorkQueue tryAddSharedQueue(int idx) {
1423	<	WorkQueue q = null;
1424	<	ReentrantLock lock = this.lock;
1425	<	if (idx >= 0 && (idx & 1) == 0 && !lock.isLocked()) {
1426	<	// create queue outside of lock but only if apparently free
1427	<	WorkQueue nq = new WorkQueue(null, SHARED_QUEUE);
1428	<	if (lock.tryLock()) {
1268	<	try {
1269	<	WorkQueue[] ws = workQueues;
1270	<	if (ws != null && idx < ws.length) {
1271	<	if ((q = ws[idx]) == null) {
1272	<	int rs; // update runState seq
1273	<	ws[idx] = q = nq;
1274	<	runState = (((rs = runState) & SHUTDOWN) \|
1275	<	((rs + RS_SEQ) & ~SHUTDOWN));
1276	<	}
1277	<	}
1278	<	} finally {
1279	<	lock.unlock();
1280	<	}
1281	<	}
1416	>	* Collect worker steal count into total. Called on termination
1417	>	* and upon int overflow of local count. (There is a possible race
1418	>	* in the latter case vs any caller of getStealCount, which can
1419	>	* make its results less accurate than usual.)
1420	>	*/
1421	>	final void collectStealCount(WorkQueue w) {
1422	>	if (w != null) {
1423	>	long sc;
1424	>	int ns = w.nsteals;
1425	>	w.nsteals = 0; // handle overflow
1426	>	long steals = (ns >= 0) ? ns : 1L + (long)(Integer.MAX_VALUE);
1427	>	do {} while (!U.compareAndSwapLong(this, STEALCOUNT,
1428	>	sc = stealCount, sc + steals));
1429		}
1283	–	return q;
1430		}
1431
1432	<	// Maintaining ctl counts
1432	>	// Submissions
1433
1434		/**
1435	<	* Increments active count; mainly called upon return from blocking
1436	<	*/
1437	<	final void incrementActiveCount() {
1438	<	long c;
1439	<	do {} while (!U.compareAndSwapLong(this, CTL, c = ctl, c + AC_UNIT));
1435	>	* Unless shutting down, adds the given task to a submission queue
1436	>	* at submitter's current queue index (modulo submission
1437	>	* range). Only the most common path is directly handled in this
1438	>	* method. All others are relayed to fullExternalPush.
1439	>	*
1440	>	* @param task the task. Caller must ensure non-null.
1441	>	*/
1442	>	final void externalPush(ForkJoinTask<?> task) {
1443	>	WorkQueue[] ws; WorkQueue q; Submitter z; int m; ForkJoinTask<?>[] a;
1444	>	if ((z = submitters.get()) != null && plock > 0 &&
1445	>	(ws = workQueues) != null && (m = (ws.length - 1)) >= 0 &&
1446	>	(q = ws[m & z.seed & SQMASK]) != null &&
1447	>	U.compareAndSwapInt(q, QLOCK, 0, 1)) { // lock
1448	>	int s = q.top, n;
1449	>	if ((a = q.array) != null && a.length > (n = s + 1 - q.base)) {
1450	>	U.putObject(a, (long)(((a.length - 1) & s) << ASHIFT) + ABASE,
1451	>	task);
1452	>	q.top = s + 1; // push on to deque
1453	>	q.qlock = 0;
1454	>	if (n <= 1)
1455	>	signalWork(q, 1);
1456	>	return;
1457	>	}
1458	>	q.qlock = 0;
1459	>	}
1460	>	fullExternalPush(task);
1461		}
1462
1463		/**
1464	<	* Activates or creates a worker
1465	<	*/
1466	<	final void signalWork() {
1467	<	/*
1468	<	* The while condition is true if: (there is are too few total
1469	<	* workers OR there is at least one waiter) AND (there are too
1470	<	* few active workers OR the pool is terminating). The value
1471	<	* of e distinguishes the remaining cases: zero (no waiters)
1472	<	* for create, negative if terminating (in which case do
1473	<	* nothing), else release a waiter. The secondary checks for
1474	<	* release (non-null array etc) can fail if the pool begins
1475	<	* terminating after the test, and don't impose any added cost
1476	<	* because JVMs must perform null and bounds checks anyway.
1477	<	*/
1478	<	long c; int e, u;
1479	<	while ((((e = (int)(c = ctl)) \| (u = (int)(c >>> 32))) &
1480	<	(INT_SIGN\|SHORT_SIGN)) == (INT_SIGN\|SHORT_SIGN)) {
1481	<	WorkQueue[] ws = workQueues; int i; WorkQueue w; Thread p;
1482	<	if (e == 0) { // add a new worker
1483	<	if (U.compareAndSwapLong
1484	<	(this, CTL, c, (long)(((u + UTC_UNIT) & UTC_MASK) \|
1485	<	((u + UAC_UNIT) & UAC_MASK)) << 32)) {
1486	<	addWorker();
1487	<	break;
1488	<	}
1464	>	* Full version of externalPush. This method is called, among
1465	>	* other times, upon the first submission of the first task to the
1466	>	* pool, so must perform secondary initialization: creating
1467	>	* workQueue array and setting plock to a valid value. It also
1468	>	* detects first submission by an external thread by looking up
1469	>	* its ThreadLocal, and creates a new shared queue if the one at
1470	>	* index if empty or contended. The lock bodies must be
1471	>	* exception-free (so no try/finally) so we optimistically
1472	>	* allocate new queues/arrays outside the locks and throw them
1473	>	* away if (very rarely) not needed. Note that the plock seq value
1474	>	* can eventually wrap around zero, but if so harmlessly fails to
1475	>	* reinitialize.
1476	>	*/
1477	>	private void fullExternalPush(ForkJoinTask<?> task) {
1478	>	for (Submitter z = null;;) {
1479	>	WorkQueue[] ws; WorkQueue q; int ps, m, r, s;
1480	>	if ((ps = plock) < 0)
1481	>	throw new RejectedExecutionException();
1482	>	else if ((ws = workQueues) == null \|\| (m = ws.length - 1) < 0) {
1483	>	int n = parallelism - 1; n \|= n >>> 1; n \|= n >>> 2;
1484	>	n \|= n >>> 4; n \|= n >>> 8; n \|= n >>> 16;
1485	>	WorkQueue[] nws = new WorkQueue[(n + 1) << 1]; // power of two
1486	>	if ((ps & PL_LOCK) != 0 \|\|
1487	>	!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK))
1488	>	ps = acquirePlock();
1489	>	if ((ws = workQueues) == null)
1490	>	workQueues = nws;
1491	>	int nps = (ps & SHUTDOWN) \| ((ps + PL_LOCK) & ~SHUTDOWN);
1492	>	if (!U.compareAndSwapInt(this, PLOCK, ps, nps))
1493	>	releasePlock(nps);
1494	>	}
1495	>	else if (z == null && (z = submitters.get()) == null) {
1496	>	if (U.compareAndSwapInt(this, INDEXSEED,
1497	>	s = indexSeed, s += SEED_INCREMENT) &&
1498	>	s != 0) // skip 0
1499	>	submitters.set(z = new Submitter(s));
1500		}
1501	<	else if (e > 0 && ws != null &&
1502	<	(i = ((~e << 1) \| 1) & SMASK) < ws.length &&
1503	<	(w = ws[i]) != null &&
1504	<	w.eventCount == (e \| INT_SIGN)) {
1505	<	if (U.compareAndSwapLong
1506	<	(this, CTL, c, (((long)(w.nextWait & E_MASK)) \|
1507	<	((long)(u + UAC_UNIT) << 32)))) {
1508	<	w.eventCount = (e + E_SEQ) & E_MASK;
1509	<	if ((p = w.parker) != null)
1510	<	U.unpark(p); // release a waiting worker
1511	<	break;
1512	<	}
1501	>	else {
1502	>	int k = (r = z.seed) & m & SQMASK;
1503	>	if ((q = ws[k]) == null && (ps & PL_LOCK) == 0) {
1504	>	(q = new WorkQueue(this, null, SHARED_QUEUE)).poolIndex = k;
1505	>	if (((ps = plock) & PL_LOCK) != 0 \|\|
1506	>	!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK))
1507	>	ps = acquirePlock();
1508	>	WorkQueue w = null;
1509	>	if ((ws = workQueues) != null && k < ws.length &&
1510	>	(w = ws[k]) == null)
1511	>	ws[k] = q;
1512	>	else
1513	>	q = w;
1514	>	int nps = (ps & SHUTDOWN) \| ((ps + PL_LOCK) & ~SHUTDOWN);
1515	>	if (!U.compareAndSwapInt(this, PLOCK, ps, nps))
1516	>	releasePlock(nps);
1517	>	}
1518	>	if (q != null && q.qlock == 0 && q.fullPush(task, false))
1519	>	return;
1520	>	r ^= r << 13; // same xorshift as WorkQueues
1521	>	r ^= r >>> 17;
1522	>	z.seed = r ^= r << 5; // move to a different index
1523		}
1336	–	else
1337	–	break;
1524		}
1525		}
1526
1527	+	// Maintaining ctl counts
1528	+
1529		/**
1530	<	* Tries to decrement active count (sometimes implicitly) and
1343	<	* possibly release or create a compensating worker in preparation
1344	<	* for blocking. Fails on contention or termination.
1345	<	*
1346	<	* @return true if the caller can block, else should recheck and retry
1530	>	* Increments active count; mainly called upon return from blocking.
1531		*/
1532	<	final boolean tryCompensate() {
1533	<	WorkQueue[] ws; WorkQueue w; Thread p;
1534	<	int pc = parallelism, e, u, ac, tc, i;
1535	<	long c = ctl;
1532	>	final void incrementActiveCount() {
1533	>	long c;
1534	>	do {} while (!U.compareAndSwapLong(this, CTL, c = ctl, c + AC_UNIT));
1535	>	}
1536
1537	<	if ((e = (int)c) >= 0) {
1538	<	if ((ac = ((u = (int)(c >>> 32)) >> UAC_SHIFT)) <= 0 &&
1539	<	e != 0 && (ws = workQueues) != null &&
1540	<	(i = ((~e << 1) \| 1) & SMASK) < ws.length &&
1541	<	(w = ws[i]) != null) {
1542	<	if (w.eventCount == (e \| INT_SIGN) &&
1543	<	U.compareAndSwapLong
1544	<	(this, CTL, c, ((long)(w.nextWait & E_MASK) \|
1545	<	(c & (AC_MASK\|TC_MASK))))) {
1546	<	w.eventCount = (e + E_SEQ) & E_MASK;
1547	<	if ((p = w.parker) != null)
1548	<	U.unpark(p);
1549	<	return true; // release an idle worker
1537	>	/**
1538	>	* Tries to create (at most one) or activate (possibly several)
1539	>	* workers if too few are active. On contention failure, continues
1540	>	* until at least one worker is signalled or the given queue is
1541	>	* empty or all workers are active.
1542	>	*
1543	>	* @param q if non-null, the queue holding tasks to be signalled
1544	>	* @param signals the target number of signals.
1545	>	*/
1546	>	final void signalWork(WorkQueue q, int signals) {
1547	>	long c; int e, u, i; WorkQueue[] ws; WorkQueue w; Thread p;
1548	>	while ((u = (int)((c = ctl) >>> 32)) < 0) {
1549	>	if ((e = (int)c) > 0) {
1550	>	if ((ws = workQueues) != null && ws.length > (i = e & SMASK) &&
1551	>	(w = ws[i]) != null && w.eventCount == (e \| INT_SIGN)) {
1552	>	long nc = (((long)(w.nextWait & E_MASK)) \|
1553	>	((long)(u + UAC_UNIT) << 32));
1554	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1555	>	w.eventCount = (e + E_SEQ) & E_MASK;
1556	>	if ((p = w.parker) != null)
1557	>	U.unpark(p);
1558	>	if (--signals <= 0)
1559	>	break;
1560	>	}
1561	>	else
1562	>	signals = 1;
1563	>	if ((q != null && q.queueSize() == 0))
1564	>	break;
1565		}
1566	+	else
1567	+	break;
1568		}
1569	<	else if ((tc = (short)(u >>> UTC_SHIFT)) >= 0 && ac + pc > 1) {
1570	<	long nc = ((c - AC_UNIT) & AC_MASK) \| (c & ~AC_MASK);
1571	<	if (U.compareAndSwapLong(this, CTL, c, nc))
1371	<	return true; // no compensation needed
1372	<	}
1373	<	else if (tc + pc < MAX_ID) {
1374	<	long nc = ((c + TC_UNIT) & TC_MASK) \| (c & ~TC_MASK);
1569	>	else if (e == 0 && (u & SHORT_SIGN) != 0) {
1570	>	long nc = (long)(((u + UTC_UNIT) & UTC_MASK) \|
1571	>	((u + UAC_UNIT) & UAC_MASK)) << 32;
1572		if (U.compareAndSwapLong(this, CTL, c, nc)) {
1573	<	addWorker();
1574	<	return true; // create replacement
1573	>	ForkJoinWorkerThread wt = null;
1574	>	Throwable ex = null;
1575	>	boolean started = false;
1576	>	try {
1577	>	ForkJoinWorkerThreadFactory fac;
1578	>	if ((fac = factory) != null &&
1579	>	(wt = fac.newThread(this)) != null) {
1580	>	wt.start();
1581	>	started = true;
1582	>	}
1583	>	} catch (Throwable rex) {
1584	>	ex = rex;
1585	>	}
1586	>	if (!started)
1587	>	deregisterWorker(wt, ex); // adjust counts on failure
1588	>	break;
1589		}
1590		}
1591	+	else
1592	+	break;
1593		}
1381	–	return false;
1594		}
1595
1596	<	// Submissions
1596	>	// Scanning for tasks
1597
1598		/**
1599	<	* Unless shutting down, adds the given task to a submission queue
1388	<	* at submitter's current queue index. If no queue exists at the
1389	<	* index, one is created unless pool lock is busy. If the queue
1390	<	* and/or lock are busy, another index is randomly chosen.
1599	>	* Top-level runloop for workers, called by ForkJoinWorkerThread.run.
1600		*/
1601	<	private void doSubmit(ForkJoinTask<?> task) {
1602	<	if (task == null)
1603	<	throw new NullPointerException();
1604	<	Submitter s = submitters.get();
1396	<	for (int r = s.seed;;) {
1397	<	WorkQueue q; int k;
1398	<	int rs = runState, m = rs & SMASK;
1399	<	WorkQueue[] ws = workQueues;
1400	<	if (rs < 0 \|\| ws == null) // shutting down
1401	<	throw new RejectedExecutionException();
1402	<	if (ws.length > m && // k must be at index
1403	<	((q = ws[k = (r << 1) & m]) != null \|\|
1404	<	(q = tryAddSharedQueue(k)) != null) &&
1405	<	q.trySharedPush(task)) {
1406	<	signalWork();
1407	<	return;
1408	<	}
1409	<	r ^= r << 13; // xorshift seed to new position
1410	<	r ^= r >>> 17;
1411	<	if (((s.seed = r ^= r << 5) & m) == 0)
1412	<	Thread.yield(); // occasionally yield if busy
1413	<	}
1601	>	final void runWorker(WorkQueue w) {
1602	>	// initialize queue array in this thread
1603	>	w.array = new ForkJoinTask<?>[WorkQueue.INITIAL_QUEUE_CAPACITY];
1604	>	do { w.runTask(scan(w)); } while (w.qlock >= 0);
1605		}
1606
1416	–
1417	–	// Scanning for tasks
1418	–
1607		/**
1608		* Scans for and, if found, returns one task, else possibly
1609		* inactivates the worker. This method operates on single reads of
1610	<	* volatile state and is designed to be re-invoked continuously in
1611	<	* part because it returns upon detecting inconsistencies,
1610	>	* volatile state and is designed to be re-invoked continuously,
1611	>	* in part because it returns upon detecting inconsistencies,
1612		* contention, or state changes that indicate possible success on
1613		* re-invocation.
1614		*
1615	<	* The scan searches for tasks across queues, randomly selecting
1616	<	* the first #queues probes, favoring steals 2:1 over submissions
1617	<	* (by exploiting even/odd indexing), and then performing a
1618	<	* circular sweep of all queues. The scan terminates upon either
1619	<	* finding a non-empty queue, or completing a full sweep. If the
1620	<	* worker is not inactivated, it takes and returns a task from
1621	<	* this queue. On failure to find a task, we take one of the
1622	<	* following actions, after which the caller will retry calling
1623	<	* this method unless terminated.
1624	<	*
1625	<	* * If not a complete sweep, try to release a waiting worker. If
1438	<	* the scan terminated because the worker is inactivated, then the
1439	<	* released worker will often be the calling worker, and it can
1440	<	* succeed obtaining a task on the next call. Or maybe it is
1441	<	* another worker, but with same net effect. Releasing in other
1442	<	* cases as well ensures that we have enough workers running.
1443	<	*
1444	<	* * If the caller has run a task since the the last empty scan,
1445	<	* return (to allow rescan) if other workers are not also yet
1446	<	* enqueued. Field WorkQueue.rescans counts down on each scan to
1447	<	* ensure eventual inactivation, and occasional calls to
1448	<	* Thread.yield to help avoid interference with more useful
1449	<	* activities on the system.
1615	>	* The scan searches for tasks across a random permutation of
1616	>	* queues (starting at a random index and stepping by a random
1617	>	* relative prime, checking each at least once). The scan
1618	>	* terminates upon either finding a non-empty queue, or completing
1619	>	* the sweep. If the worker is not inactivated, it takes and
1620	>	* returns a task from this queue. Otherwise, if not activated, it
1621	>	* signals workers (that may include itself) and returns so caller
1622	>	* can retry. Also returns for trtry if the worker array may have
1623	>	* changed during an empty scan. On failure to find a task, we
1624	>	* take one of the following actions, after which the caller will
1625	>	* retry calling this method unless terminated.
1626		*
1627	<	* * If pool is terminating, terminate the worker
1627	>	* * If pool is terminating, terminate the worker.
1628		*
1629		* * If not already enqueued, try to inactivate and enqueue the
1630	<	* worker on wait queue.
1631	<	*
1632	<	* * If already enqueued and none of the above apply, either park
1633	<	* awaiting signal, or if this is the most recent waiter and pool
1634	<	* is quiescent, relay to idleAwaitWork to check for termination
1635	<	* and possibly shrink pool.
1630	>	* worker on wait queue. Or, if inactivating has caused the pool
1631	>	* to be quiescent, relay to idleAwaitWork to check for
1632	>	* termination and possibly shrink pool.
1633	>	*
1634	>	* * If already enqueued and none of the above apply, possibly
1635	>	* (with 1/2 probability) park awaiting signal, else lingering to
1636	>	* help scan and signal.
1637		*
1638		* @param w the worker (via its WorkQueue)
1639	<	* @return a task or null of none found
1639	>	* @return a task or null if none found
1640		*/
1641		private final ForkJoinTask<?> scan(WorkQueue w) {
1642	<	boolean swept = false; // true after full empty scan
1643	<	WorkQueue[] ws; // volatile read order matters
1644	<	int r = w.seed, ec = w.eventCount; // ec is negative if inactive
1645	<	int rs = runState, m = rs & SMASK;
1646	<	if ((ws = workQueues) != null && ws.length > m) {
1647	<	ForkJoinTask<?> task = null;
1648	<	for (int k = 0, j = -2 - m; ; ++j) {
1649	<	WorkQueue q; int b;
1650	<	if (j < 0) { // random probes while j negative
1651	<	r ^= r << 13; r ^= r >>> 17; k = (r ^= r << 5) \| (j & 1);
1652	<	} // worker (not submit) for odd j
1653	<	else // cyclic scan when j >= 0
1654	<	k += (m >>> 1) \| 1; // step by half to reduce bias
1655	<
1656	<	if ((q = ws[k & m]) != null && (b = q.base) - q.top < 0) {
1657	<	if (ec >= 0)
1658	<	task = q.pollAt(b); // steal
1659	<	break;
1642	>	WorkQueue[] ws; WorkQueue q; // first update random seed
1643	>	int r = w.seed; r ^= r << 13; r ^= r >>> 17; w.seed = r ^= r << 5;
1644	>	int ps = plock, m; // volatile read order matters
1645	>	if ((ws = workQueues) != null && (m = ws.length - 1) > 0) {
1646	>	int ec = w.eventCount; // ec is negative if inactive
1647	>	int step = (r >>> 16) \| 1; // relatively prime
1648	>	for (int j = (m + 1) << 2; ; --j, r += step) {
1649	>	ForkJoinTask<?> t; ForkJoinTask<?>[] a; int b, n;
1650	>	if ((q = ws[r & m]) != null && (b = q.base) - q.top < 0 &&
1651	>	(a = q.array) != null) { // probably nonempty
1652	>	int i = (((a.length - 1) & b) << ASHIFT) + ABASE;
1653	>	t = (ForkJoinTask<?>)U.getObjectVolatile(a, i);
1654	>	if (q.base == b && ec >= 0 && t != null &&
1655	>	U.compareAndSwapObject(a, i, t, null)) {
1656	>	if ((n = q.top - (q.base = b + 1)) > 0)
1657	>	signalWork(q, n);
1658	>	return t; // taken
1659	>	}
1660	>	if (j < m \|\| (ec < 0 && (ec = w.eventCount) < 0)) {
1661	>	if ((n = q.queueSize() - 1) > 0)
1662	>	signalWork(q, n);
1663	>	break; // let caller retry after signal
1664	>	}
1665		}
1666	<	else if (j > m) {
1667	<	if (rs == runState) // staleness check
1668	<	swept = true;
1666	>	else if (j < 0) { // end of scan
1667	>	long c = ctl; int e;
1668	>	if (plock != ps) // incomplete sweep
1669	>	break;
1670	>	if ((e = (int)c) < 0) // pool is terminating
1671	>	w.qlock = -1;
1672	>	else if (ec >= 0) { // try to enqueue/inactivate
1673	>	long nc = ((long)ec \|
1674	>	((c - AC_UNIT) & (AC_MASK\|TC_MASK)));
1675	>	w.nextWait = e;
1676	>	w.eventCount = ec \| INT_SIGN; // mark as inactive
1677	>	if (ctl != c \|\|
1678	>	!U.compareAndSwapLong(this, CTL, c, nc))
1679	>	w.eventCount = ec; // unmark on CAS failure
1680	>	else if ((int)(c >> AC_SHIFT) == 1 - parallelism)
1681	>	idleAwaitWork(w, nc, c); // quiescent
1682	>	}
1683	>	else if (w.seed >= 0 && w.eventCount < 0) {
1684	>	Thread wt = Thread.currentThread();
1685	>	Thread.interrupted(); // clear status
1686	>	U.putObject(wt, PARKBLOCKER, this);
1687	>	w.parker = wt; // emulate LockSupport.park
1688	>	if (w.eventCount < 0) // recheck
1689	>	U.park(false, 0L);
1690	>	w.parker = null;
1691	>	U.putObject(wt, PARKBLOCKER, null);
1692	>	}
1693		break;
1694		}
1695		}
1490	–	w.seed = r; // save seed for next scan
1491	–	if (task != null)
1492	–	return task;
1493	–	}
1494	–
1495	–	// Decode ctl on empty scan
1496	–	long c = ctl; int e = (int)c, a = (int)(c >> AC_SHIFT), nr, ns;
1497	–	if (!swept) { // try to release a waiter
1498	–	WorkQueue v; Thread p;
1499	–	if (e > 0 && a < 0 && ws != null &&
1500	–	(v = ws[((~e << 1) \| 1) & m]) != null &&
1501	–	v.eventCount == (e \| INT_SIGN) && U.compareAndSwapLong
1502	–	(this, CTL, c, ((long)(v.nextWait & E_MASK) \|
1503	–	((c + AC_UNIT) & (AC_MASK\|TC_MASK))))) {
1504	–	v.eventCount = (e + E_SEQ) & E_MASK;
1505	–	if ((p = v.parker) != null)
1506	–	U.unpark(p);
1507	–	}
1508	–	}
1509	–	else if ((nr = w.rescans) > 0) { // continue rescanning
1510	–	int ac = a + parallelism;
1511	–	if ((w.rescans = (ac < nr) ? ac : nr - 1) > 0 && w.seed < 0 &&
1512	–	w.eventCount == ec)
1513	–	Thread.yield(); // 1 bit randomness for yield call
1514	–	}
1515	–	else if (e < 0) // pool is terminating
1516	–	w.runState = -1;
1517	–	else if (ec >= 0) { // try to enqueue
1518	–	long nc = (long)ec \| ((c - AC_UNIT) & (AC_MASK\|TC_MASK));
1519	–	w.nextWait = e;
1520	–	w.eventCount = ec \| INT_SIGN; // mark as inactive
1521	–	if (!U.compareAndSwapLong(this, CTL, c, nc))
1522	–	w.eventCount = ec; // back out on CAS failure
1523	–	else if ((ns = w.nsteals) != 0) { // set rescans if ran task
1524	–	if (a <= 0) // ... unless too many active
1525	–	w.rescans = a + parallelism;
1526	–	w.nsteals = 0;
1527	–	w.totalSteals += ns;
1528	–	}
1529	–	}
1530	–	else{ // already queued
1531	–	if (parallelism == -a)
1532	–	idleAwaitWork(w); // quiescent
1533	–	if (w.eventCount == ec) {
1534	–	Thread.interrupted(); // clear status
1535	–	ForkJoinWorkerThread wt = w.owner;
1536	–	U.putObject(wt, PARKBLOCKER, this);
1537	–	w.parker = wt; // emulate LockSupport.park
1538	–	if (w.eventCount == ec) // recheck
1539	–	U.park(false, 0L); // block
1540	–	w.parker = null;
1541	–	U.putObject(wt, PARKBLOCKER, null);
1542	–	}
1696		}
1697		return null;
1698		}
1699
1700		/**
1701	<	* If inactivating worker w has caused pool to become quiescent,
1702	<	* check for pool termination, and, so long as this is not the
1703	<	* only worker, wait for event for up to SHRINK_RATE nanosecs On
1704	<	* timeout, if ctl has not changed, terminate the worker, which
1705	<	* will in turn wake up another worker to possibly repeat this
1706	<	* process.
1701	>	* If inactivating worker w has caused the pool to become
1702	>	* quiescent, checks for pool termination, and, so long as this is
1703	>	* not the only worker, waits for event for up to a given
1704	>	* duration. On timeout, if ctl has not changed, terminates the
1705	>	* worker, which will in turn wake up another worker to possibly
1706	>	* repeat this process.
1707		*
1708		* @param w the calling worker
1709	+	* @param currentCtl the ctl value triggering possible quiescence
1710	+	* @param prevCtl the ctl value to restore if thread is terminated
1711		*/
1712	<	private void idleAwaitWork(WorkQueue w) {
1713	<	long c; int nw, ec;
1714	<	if (!tryTerminate(false) &&
1715	<	(int)((c = ctl) >> AC_SHIFT) + parallelism == 0 &&
1716	<	(ec = w.eventCount) == ((int)c \| INT_SIGN) &&
1717	<	(nw = w.nextWait) != 0) {
1718	<	long nc = ((long)(nw & E_MASK) \| // ctl to restore on timeout
1719	<	((c + AC_UNIT) & AC_MASK) \| (c & TC_MASK));
1720	<	ForkJoinTask.helpExpungeStaleExceptions(); // help clean
1566	<	ForkJoinWorkerThread wt = w.owner;
1567	<	while (ctl == c) {
1568	<	long startTime = System.nanoTime();
1712	>	private void idleAwaitWork(WorkQueue w, long currentCtl, long prevCtl) {
1713	>	if (w.eventCount < 0 &&
1714	>	(this == commonPool \|\| !tryTerminate(false, false)) &&
1715	>	(int)prevCtl != 0) {
1716	>	int dc = -(short)(currentCtl >>> TC_SHIFT);
1717	>	long parkTime = dc < 0 ? FAST_IDLE_TIMEOUT: (dc + 1) * IDLE_TIMEOUT;
1718	>	long deadline = System.nanoTime() + parkTime - 100000L; // 1ms slop
1719	>	Thread wt = Thread.currentThread();
1720	>	while (ctl == currentCtl) {
1721		Thread.interrupted(); // timed variant of version in scan()
1722		U.putObject(wt, PARKBLOCKER, this);
1723		w.parker = wt;
1724	<	if (ctl == c)
1725	<	U.park(false, SHRINK_RATE);
1724	>	if (ctl == currentCtl)
1725	>	U.park(false, parkTime);
1726		w.parker = null;
1727		U.putObject(wt, PARKBLOCKER, null);
1728	<	if (ctl != c)
1728	>	if (ctl != currentCtl)
1729		break;
1730	<	if (System.nanoTime() - startTime >= SHRINK_TIMEOUT &&
1731	<	U.compareAndSwapLong(this, CTL, c, nc)) {
1732	<	w.runState = -1; // shrink
1733	<	w.eventCount = (ec + E_SEQ) \| E_MASK;
1730	>	if (deadline - System.nanoTime() <= 0L &&
1731	>	U.compareAndSwapLong(this, CTL, currentCtl, prevCtl)) {
1732	>	w.eventCount = (w.eventCount + E_SEQ) \| E_MASK;
1733	>	w.qlock = -1; // shrink
1734		break;
1735		}
1736		}
#	Line 1586 \| Line 1738 \| public class ForkJoinPool extends Abstra
1738		}
1739
1740		/**
1741	+	* Scans through queues looking for work while joining a task;
1742	+	* if any are present, signals.
1743	+	*
1744	+	* @param task to return early if done
1745	+	* @param origin an index to start scan
1746	+	*/
1747	+	final int helpSignal(ForkJoinTask<?> task, int origin) {
1748	+	WorkQueue[] ws; WorkQueue q; int m, n, s;
1749	+	if (task != null && (ws = workQueues) != null &&
1750	+	(m = ws.length - 1) >= 0) {
1751	+	for (int i = 0; i <= m; ++i) {
1752	+	if ((s = task.status) < 0)
1753	+	return s;
1754	+	if ((q = ws[(i + origin) & m]) != null &&
1755	+	(n = q.queueSize()) > 0) {
1756	+	signalWork(q, n);
1757	+	if ((int)(ctl >> AC_SHIFT) >= 0)
1758	+	break;
1759	+	}
1760	+	}
1761	+	}
1762	+	return 0;
1763	+	}
1764	+
1765	+	/**
1766		* Tries to locate and execute tasks for a stealer of the given
1767		* task, or in turn one of its stealers, Traces currentSteal ->
1768		* currentJoin links looking for a thread working on a descendant
#	Line 1596 \| Line 1773 \| public class ForkJoinPool extends Abstra
1773		* leaves hints in workers to speed up subsequent calls. The
1774		* implementation is very branchy to cope with potential
1775		* inconsistencies or loops encountering chains that are stale,
1776	<	* unknown, or of length greater than MAX_HELP_DEPTH links. All
1600	<	* of these cases are dealt with by just retrying by caller.
1776	>	* unknown, or so long that they are likely cyclic.
1777		*
1778		* @param joiner the joining worker
1779		* @param task the task to join
1780	<	* @return true if found or ran a task (and so is immediately retryable)
1780	>	* @return 0 if no progress can be made, negative if task
1781	>	* known complete, else positive
1782		*/
1783	<	final boolean tryHelpStealer(WorkQueue joiner, ForkJoinTask<?> task) {
1784	<	ForkJoinTask<?> subtask; // current target
1785	<	boolean progress = false;
1786	<	int depth = 0; // current chain depth
1787	<	int m = runState & SMASK;
1788	<	WorkQueue[] ws = workQueues;
1789	<
1790	<	if (ws != null && ws.length > m && (subtask = task).status >= 0) {
1791	<	outer:for (WorkQueue j = joiner;;) {
1792	<	// Try to find the stealer of subtask, by first using hint
1793	<	WorkQueue stealer = null;
1794	<	WorkQueue v = ws[j.stealHint & m];
1795	<	if (v != null && v.currentSteal == subtask)
1796	<	stealer = v;
1797	<	else {
1798	<	for (int i = 1; i <= m; i += 2) {
1799	<	if ((v = ws[i]) != null && v.currentSteal == subtask) {
1800	<	stealer = v;
1801	<	j.stealHint = i; // save hint
1802	<	break;
1783	>	private int tryHelpStealer(WorkQueue joiner, ForkJoinTask<?> task) {
1784	>	int stat = 0, steps = 0; // bound to avoid cycles
1785	>	if (joiner != null && task != null) { // hoist null checks
1786	>	restart: for (;;) {
1787	>	ForkJoinTask<?> subtask = task; // current target
1788	>	for (WorkQueue j = joiner, v;;) { // v is stealer of subtask
1789	>	WorkQueue[] ws; int m, s, h;
1790	>	if ((s = task.status) < 0) {
1791	>	stat = s;
1792	>	break restart;
1793	>	}
1794	>	if ((ws = workQueues) == null \|\| (m = ws.length - 1) <= 0)
1795	>	break restart; // shutting down
1796	>	if ((v = ws[h = (j.stealHint \| 1) & m]) == null \|\|
1797	>	v.currentSteal != subtask) {
1798	>	for (int origin = h;;) { // find stealer
1799	>	if (((h = (h + 2) & m) & 15) == 1 &&
1800	>	(subtask.status < 0 \|\| j.currentJoin != subtask))
1801	>	continue restart; // occasional staleness check
1802	>	if ((v = ws[h]) != null &&
1803	>	v.currentSteal == subtask) {
1804	>	j.stealHint = h; // save hint
1805	>	break;
1806	>	}
1807	>	if (h == origin)
1808	>	break restart; // cannot find stealer
1809		}
1810		}
1811	<	if (stealer == null)
1811	>	for (;;) { // help stealer or descend to its stealer
1812	>	ForkJoinTask[] a; int b;
1813	>	if (subtask.status < 0) // surround probes with
1814	>	continue restart; // consistency checks
1815	>	if ((b = v.base) - v.top < 0 && (a = v.array) != null) {
1816	>	int i = (((a.length - 1) & b) << ASHIFT) + ABASE;
1817	>	ForkJoinTask<?> t =
1818	>	(ForkJoinTask<?>)U.getObjectVolatile(a, i);
1819	>	if (subtask.status < 0 \|\| j.currentJoin != subtask \|\|
1820	>	v.currentSteal != subtask)
1821	>	continue restart; // stale
1822	>	stat = 1; // apparent progress
1823	>	if (t != null && v.base == b &&
1824	>	U.compareAndSwapObject(a, i, t, null)) {
1825	>	v.base = b + 1; // help stealer
1826	>	joiner.runSubtask(t);
1827	>	}
1828	>	else if (v.base == b && ++steps == MAX_HELP)
1829	>	break restart; // v apparently stalled
1830	>	}
1831	>	else { // empty -- try to descend
1832	>	ForkJoinTask<?> next = v.currentJoin;
1833	>	if (subtask.status < 0 \|\| j.currentJoin != subtask \|\|
1834	>	v.currentSteal != subtask)
1835	>	continue restart; // stale
1836	>	else if (next == null \|\| ++steps == MAX_HELP)
1837	>	break restart; // dead-end or maybe cyclic
1838	>	else {
1839	>	subtask = next;
1840	>	j = v;
1841	>	break;
1842	>	}
1843	>	}
1844	>	}
1845	>	}
1846	>	}
1847	>	}
1848	>	return stat;
1849	>	}
1850	>
1851	>	/**
1852	>	* Analog of tryHelpStealer for CountedCompleters. Tries to steal
1853	>	* and run tasks within the target's computation
1854	>	*
1855	>	* @param task the task to join
1856	>	* @param mode if shared, exit upon completing any task
1857	>	* if all workers are active
1858	>	*
1859	>	*/
1860	>	private int helpComplete(ForkJoinTask<?> task, int mode) {
1861	>	WorkQueue[] ws; WorkQueue q; int m, n, s;
1862	>	if (task != null && (ws = workQueues) != null &&
1863	>	(m = ws.length - 1) >= 0) {
1864	>	for (int j = 1, origin = j;;) {
1865	>	if ((s = task.status) < 0)
1866	>	return s;
1867	>	if ((q = ws[j & m]) != null && q.pollAndExecCC(task)) {
1868	>	origin = j;
1869	>	if (mode == SHARED_QUEUE && (int)(ctl >> AC_SHIFT) >= 0)
1870		break;
1871		}
1872	+	else if ((j = (j + 2) & m) == origin)
1873	+	break;
1874	+	}
1875	+	}
1876	+	return 0;
1877	+	}
1878
1879	<	for (WorkQueue q = stealer;;) { // Try to help stealer
1880	<	ForkJoinTask<?> t; int b;
1881	<	if (task.status < 0)
1882	<	break outer;
1883	<	if ((b = q.base) - q.top < 0) {
1884	<	progress = true;
1885	<	if (subtask.status < 0)
1886	<	break outer; // stale
1887	<	if ((t = q.pollAt(b)) != null) {
1888	<	stealer.stealHint = joiner.poolIndex;
1889	<	joiner.runSubtask(t);
1879	>	/**
1880	>	* Tries to decrement active count (sometimes implicitly) and
1881	>	* possibly release or create a compensating worker in preparation
1882	>	* for blocking. Fails on contention or termination. Otherwise,
1883	>	* adds a new thread if no idle workers are available and pool
1884	>	* may become starved.
1885	>	*/
1886	>	final boolean tryCompensate() {
1887	>	int pc = parallelism, e, u, i, tc; long c;
1888	>	WorkQueue[] ws; WorkQueue w; Thread p;
1889	>	if ((e = (int)(c = ctl)) >= 0 && (ws = workQueues) != null) {
1890	>	if (e != 0 && (i = e & SMASK) < ws.length &&
1891	>	(w = ws[i]) != null && w.eventCount == (e \| INT_SIGN)) {
1892	>	long nc = ((long)(w.nextWait & E_MASK) \|
1893	>	(c & (AC_MASK\|TC_MASK)));
1894	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1895	>	w.eventCount = (e + E_SEQ) & E_MASK;
1896	>	if ((p = w.parker) != null)
1897	>	U.unpark(p);
1898	>	return true; // replace with idle worker
1899	>	}
1900	>	}
1901	>	else if ((short)((u = (int)(c >>> 32)) >>> UTC_SHIFT) >= 0 &&
1902	>	(u >> UAC_SHIFT) + pc > 1) {
1903	>	long nc = ((c - AC_UNIT) & AC_MASK) \| (c & ~AC_MASK);
1904	>	if (U.compareAndSwapLong(this, CTL, c, nc))
1905	>	return true; // no compensation
1906	>	}
1907	>	else if ((tc = u + pc) < MAX_CAP) {
1908	>	long nc = ((c + TC_UNIT) & TC_MASK) \| (c & ~TC_MASK);
1909	>	if (U.compareAndSwapLong(this, CTL, c, nc)) {
1910	>	Throwable ex = null;
1911	>	ForkJoinWorkerThread wt = null;
1912	>	try {
1913	>	ForkJoinWorkerThreadFactory fac;
1914	>	if ((fac = factory) != null &&
1915	>	(wt = fac.newThread(this)) != null) {
1916	>	wt.start();
1917	>	return true;
1918		}
1919	+	} catch (Throwable rex) {
1920	+	ex = rex;
1921		}
1922	<	else { // empty - try to descend to find stealer's stealer
1646	<	ForkJoinTask<?> next = stealer.currentJoin;
1647	<	if (++depth == MAX_HELP_DEPTH \|\| subtask.status < 0 \|\|
1648	<	next == null \|\| next == subtask)
1649	<	break outer; // max depth, stale, dead-end, cyclic
1650	<	subtask = next;
1651	<	j = stealer;
1652	<	break;
1653	<	}
1922	>	deregisterWorker(wt, ex); // adjust counts etc
1923		}
1924		}
1925		}
1926	<	return progress;
1926	>	return false;
1927		}
1928
1929		/**
1930	<	* If task is at base of some steal queue, steals and executes it.
1930	>	* Helps and/or blocks until the given task is done.
1931		*
1932		* @param joiner the joining worker
1933		* @param task the task
1934	+	* @return task status on exit
1935		*/
1936	<	final void tryPollForAndExec(WorkQueue joiner, ForkJoinTask<?> task) {
1937	<	WorkQueue[] ws;
1938	<	int m = runState & SMASK;
1939	<	if ((ws = workQueues) != null && ws.length > m) {
1940	<	for (int j = 1; j <= m && task.status >= 0; j += 2) {
1941	<	WorkQueue q = ws[j];
1942	<	if (q != null && q.pollFor(task)) {
1943	<	joiner.runSubtask(task);
1944	<	break;
1936	>	final int awaitJoin(WorkQueue joiner, ForkJoinTask<?> task) {
1937	>	int s = 0;
1938	>	if (joiner != null && task != null && (s = task.status) >= 0) {
1939	>	ForkJoinTask<?> prevJoin = joiner.currentJoin;
1940	>	joiner.currentJoin = task;
1941	>	do {} while ((s = task.status) >= 0 &&
1942	>	joiner.queueSize() > 0 &&
1943	>	joiner.tryRemoveAndExec(task)); // process local tasks
1944	>	if (s >= 0 && (s = task.status) >= 0 &&
1945	>	(s = helpSignal(task, joiner.poolIndex)) >= 0 &&
1946	>	(task instanceof CountedCompleter))
1947	>	s = helpComplete(task, LIFO_QUEUE);
1948	>	while (s >= 0 && (s = task.status) >= 0) {
1949	>	if ((joiner.queueSize() > 0 \|\| // try helping
1950	>	(s = tryHelpStealer(joiner, task)) == 0) &&
1951	>	(s = task.status) >= 0 && tryCompensate()) {
1952	>	if (task.trySetSignal() && (s = task.status) >= 0) {
1953	>	synchronized (task) {
1954	>	if (task.status >= 0) {
1955	>	try { // see ForkJoinTask
1956	>	task.wait(); // for explanation
1957	>	} catch (InterruptedException ie) {
1958	>	}
1959	>	}
1960	>	else
1961	>	task.notifyAll();
1962	>	}
1963	>	}
1964	>	long c; // re-activate
1965	>	do {} while (!U.compareAndSwapLong
1966	>	(this, CTL, c = ctl, c + AC_UNIT));
1967		}
1968		}
1969	+	joiner.currentJoin = prevJoin;
1970	+	}
1971	+	return s;
1972	+	}
1973	+
1974	+	/**
1975	+	* Stripped-down variant of awaitJoin used by timed joins. Tries
1976	+	* to help join only while there is continuous progress. (Caller
1977	+	* will then enter a timed wait.)
1978	+	*
1979	+	* @param joiner the joining worker
1980	+	* @param task the task
1981	+	*/
1982	+	final void helpJoinOnce(WorkQueue joiner, ForkJoinTask<?> task) {
1983	+	int s;
1984	+	if (joiner != null && task != null && (s = task.status) >= 0) {
1985	+	ForkJoinTask<?> prevJoin = joiner.currentJoin;
1986	+	joiner.currentJoin = task;
1987	+	do {} while ((s = task.status) >= 0 &&
1988	+	joiner.queueSize() > 0 &&
1989	+	joiner.tryRemoveAndExec(task));
1990	+	if (s >= 0 && (s = task.status) >= 0 &&
1991	+	(s = helpSignal(task, joiner.poolIndex)) >= 0 &&
1992	+	(task instanceof CountedCompleter))
1993	+	s = helpComplete(task, LIFO_QUEUE);
1994	+	if (s >= 0 && joiner.queueSize() == 0) {
1995	+	do {} while (task.status >= 0 &&
1996	+	tryHelpStealer(joiner, task) > 0);
1997	+	}
1998	+	joiner.currentJoin = prevJoin;
1999		}
2000		}
2001
2002		/**
2003	<	* Returns a non-empty steal queue, if one is found during a random,
2004	<	* then cyclic scan, else null. This method must be retried by
2005	<	* caller if, by the time it tries to use the queue, it is empty.
2003	>	* Returns a (probably) non-empty steal queue, if one is found
2004	>	* during a random, then cyclic scan, else null. This method must
2005	>	* be retried by caller if, by the time it tries to use the queue,
2006	>	* it is empty.
2007	>	* @param r a (random) seed for scanning
2008		*/
2009	<	private WorkQueue findNonEmptyStealQueue(WorkQueue w) {
2010	<	int r = w.seed; // Same idea as scan(), but ignoring submissions
2009	>	private WorkQueue findNonEmptyStealQueue(int r) {
2010	>	int step = (r >>> 16) \| 1;
2011		for (WorkQueue[] ws;;) {
2012	<	int m = runState & SMASK;
2013	<	if ((ws = workQueues) == null)
2012	>	int ps = plock, m;
2013	>	if ((ws = workQueues) == null \|\| (m = ws.length - 1) < 1)
2014		return null;
2015	<	if (ws.length > m) {
2016	<	WorkQueue q;
2017	<	for (int n = m << 2, k = r, j = -n;;) {
2018	<	r ^= r << 13; r ^= r >>> 17; r ^= r << 5;
2019	<	if ((q = ws[(k \| 1) & m]) != null && q.base - q.top < 0) {
2020	<	w.seed = r;
1697	<	return q;
1698	<	}
1699	<	else if (j > n)
2015	>	for (int j = (m + 1) << 2; ; r += step) {
2016	>	WorkQueue q = ws[((r << 1) \| 1) & m];
2017	>	if (q != null && q.queueSize() > 0)
2018	>	return q;
2019	>	else if (--j < 0) {
2020	>	if (plock == ps)
2021		return null;
2022	<	else
1702	<	k = (j++ < 0) ? r : k + ((m >>> 1) \| 1);
1703	<
2022	>	break;
2023		}
2024		}
2025		}
#	Line 1714 \| Line 2033 \| public class ForkJoinPool extends Abstra
2033		*/
2034		final void helpQuiescePool(WorkQueue w) {
2035		for (boolean active = true;;) {
2036	<	w.runLocalTasks(); // exhaust local queue
2037	<	WorkQueue q = findNonEmptyStealQueue(w);
2036	>	ForkJoinTask<?> localTask; // exhaust local queue
2037	>	while ((localTask = w.nextLocalTask()) != null)
2038	>	localTask.doExec();
2039	>	// Similar to loop in scan(), but ignoring submissions
2040	>	WorkQueue q = findNonEmptyStealQueue(w.nextSeed());
2041		if (q != null) {
2042	<	ForkJoinTask<?> t;
2042	>	ForkJoinTask<?> t; int b;
2043		if (!active) { // re-establish active count
2044		long c;
2045		active = true;
2046		do {} while (!U.compareAndSwapLong
2047		(this, CTL, c = ctl, c + AC_UNIT));
2048		}
2049	<	if ((t = q.poll()) != null)
2049	>	if ((b = q.base) - q.top < 0 && (t = q.pollAt(b)) != null)
2050		w.runSubtask(t);
2051		}
2052		else {
#	Line 1746 \| Line 2068 \| public class ForkJoinPool extends Abstra
2068		}
2069
2070		/**
2071	<	* Gets and removes a local or stolen task for the given worker
2071	>	* Gets and removes a local or stolen task for the given worker.
2072		*
2073		* @return a task, if available
2074		*/
2075		final ForkJoinTask<?> nextTaskFor(WorkQueue w) {
2076		for (ForkJoinTask<?> t;;) {
2077	<	WorkQueue q;
2077	>	WorkQueue q; int b;
2078		if ((t = w.nextLocalTask()) != null)
2079		return t;
2080	<	if ((q = findNonEmptyStealQueue(w)) == null)
2080	>	if ((q = findNonEmptyStealQueue(w.nextSeed())) == null)
2081		return null;
2082	<	if ((t = q.poll()) != null)
2082	>	if ((b = q.base) - q.top < 0 && (t = q.pollAt(b)) != null)
2083		return t;
2084		}
2085		}
2086
2087		/**
2088	<	* Returns the approximate (non-atomic) number of idle threads per
2089	<	* active thread to offset steal queue size for method
2090	<	* ForkJoinTask.getSurplusQueuedTaskCount().
2091	<	*/
2092	<	final int idlePerActive() {
2093	<	// Approximate at powers of two for small values, saturate past 4
2094	<	int p = parallelism;
2095	<	int a = p + (int)(ctl >> AC_SHIFT);
2096	<	return (a > (p >>>= 1) ? 0 :
2097	<	a > (p >>>= 1) ? 1 :
2098	<	a > (p >>>= 1) ? 2 :
2099	<	a > (p >>>= 1) ? 4 :
2100	<	8);
2101	<	}
2102	<
2103	<	// Termination
2104	<
2105	<	/**
2106	<	* Sets SHUTDOWN bit of runState under lock
2107	<	*/
2108	<	private void enableShutdown() {
2109	<	ReentrantLock lock = this.lock;
2110	<	if (runState >= 0) {
2111	<	lock.lock(); // don't need try/finally
2112	<	runState \|= SHUTDOWN;
2113	<	lock.unlock();
2088	>	* Returns a cheap heuristic guide for task partitioning when
2089	>	* programmers, frameworks, tools, or languages have little or no
2090	>	* idea about task granularity. In essence by offering this
2091	>	* method, we ask users only about tradeoffs in overhead vs
2092	>	* expected throughput and its variance, rather than how finely to
2093	>	* partition tasks.
2094	>	*
2095	>	* In a steady state strict (tree-structured) computation, each
2096	>	* thread makes available for stealing enough tasks for other
2097	>	* threads to remain active. Inductively, if all threads play by
2098	>	* the same rules, each thread should make available only a
2099	>	* constant number of tasks.
2100	>	*
2101	>	* The minimum useful constant is just 1. But using a value of 1
2102	>	* would require immediate replenishment upon each steal to
2103	>	* maintain enough tasks, which is infeasible. Further,
2104	>	* partitionings/granularities of offered tasks should minimize
2105	>	* steal rates, which in general means that threads nearer the top
2106	>	* of computation tree should generate more than those nearer the
2107	>	* bottom. In perfect steady state, each thread is at
2108	>	* approximately the same level of computation tree. However,
2109	>	* producing extra tasks amortizes the uncertainty of progress and
2110	>	* diffusion assumptions.
2111	>	*
2112	>	* So, users will want to use values larger, but not much larger
2113	>	* than 1 to both smooth over transient shortages and hedge
2114	>	* against uneven progress; as traded off against the cost of
2115	>	* extra task overhead. We leave the user to pick a threshold
2116	>	* value to compare with the results of this call to guide
2117	>	* decisions, but recommend values such as 3.
2118	>	*
2119	>	* When all threads are active, it is on average OK to estimate
2120	>	* surplus strictly locally. In steady-state, if one thread is
2121	>	* maintaining say 2 surplus tasks, then so are others. So we can
2122	>	* just use estimated queue length. However, this strategy alone
2123	>	* leads to serious mis-estimates in some non-steady-state
2124	>	* conditions (ramp-up, ramp-down, other stalls). We can detect
2125	>	* many of these by further considering the number of "idle"
2126	>	* threads, that are known to have zero queued tasks, so
2127	>	* compensate by a factor of (#idle/#active) threads.
2128	>	*
2129	>	* Note: The approximation of #busy workers as #active workers is
2130	>	* not very good under current signalling scheme, and should be
2131	>	* improved.
2132	>	*/
2133	>	static int getSurplusQueuedTaskCount() {
2134	>	Thread t; ForkJoinWorkerThread wt; ForkJoinPool pool; WorkQueue q;
2135	>	if (((t = Thread.currentThread()) instanceof ForkJoinWorkerThread)) {
2136	>	int b = (q = (wt = (ForkJoinWorkerThread)t).workQueue).base;
2137	>	int p = (pool = wt.pool).parallelism;
2138	>	int a = (int)(pool.ctl >> AC_SHIFT) + p;
2139	>	return q.top - b - (a > (p >>>= 1) ? 0 :
2140	>	a > (p >>>= 1) ? 1 :
2141	>	a > (p >>>= 1) ? 2 :
2142	>	a > (p >>>= 1) ? 4 :
2143	>	8);
2144		}
2145	+	return 0;
2146		}
2147
2148	+	// Termination
2149	+
2150		/**
2151	<	* Possibly initiates and/or completes termination. Upon
2152	<	* termination, cancels all queued tasks and then
2151	>	* Possibly initiates and/or completes termination. The caller
2152	>	* triggering termination runs three passes through workQueues:
2153	>	* (0) Setting termination status, followed by wakeups of queued
2154	>	* workers; (1) cancelling all tasks; (2) interrupting lagging
2155	>	* threads (likely in external tasks, but possibly also blocked in
2156	>	* joins). Each pass repeats previous steps because of potential
2157	>	* lagging thread creation.
2158		*
2159		* @param now if true, unconditionally terminate, else only
2160		* if no work and no active workers
2161	+	* @param enable if true, enable shutdown when next possible
2162		* @return true if now terminating or terminated
2163		*/
2164	<	private boolean tryTerminate(boolean now) {
2164	>	private boolean tryTerminate(boolean now, boolean enable) {
2165	>	if (this == commonPool) // cannot shut down
2166	>	return false;
2167		for (long c;;) {
2168		if (((c = ctl) & STOP_BIT) != 0) { // already terminating
2169		if ((short)(c >>> TC_SHIFT) == -parallelism) {
2170	<	ReentrantLock lock = this.lock; // signal when no workers
2171	<	lock.lock(); // don't need try/finally
2172	<	termination.signalAll(); // signal when 0 workers
1810	<	lock.unlock();
2170	>	synchronized (this) {
2171	>	notifyAll(); // signal when 0 workers
2172	>	}
2173		}
2174		return true;
2175		}
2176	<	if (!now) {
2177	<	if ((int)(c >> AC_SHIFT) != -parallelism \|\| runState >= 0 \|\|
2176	>	if (plock >= 0) { // not yet enabled
2177	>	int ps;
2178	>	if (!enable)
2179	>	return false;
2180	>	if (((ps = plock) & PL_LOCK) != 0 \|\|
2181	>	!U.compareAndSwapInt(this, PLOCK, ps, ps += PL_LOCK))
2182	>	ps = acquirePlock();
2183	>	int nps = SHUTDOWN;
2184	>	if (!U.compareAndSwapInt(this, PLOCK, ps, nps))
2185	>	releasePlock(nps);
2186	>	}
2187	>	if (!now) { // check if idle & no tasks
2188	>	if ((int)(c >> AC_SHIFT) != -parallelism \|\|
2189		hasQueuedSubmissions())
2190		return false;
2191		// Check for unqueued inactive workers. One pass suffices.
2192		WorkQueue[] ws = workQueues; WorkQueue w;
2193		if (ws != null) {
2194	<	int n = ws.length;
1822	<	for (int i = 1; i < n; i += 2) {
2194	>	for (int i = 1; i < ws.length; i += 2) {
2195		if ((w = ws[i]) != null && w.eventCount >= 0)
2196		return false;
2197		}
2198		}
2199		}
2200	<	if (U.compareAndSwapLong(this, CTL, c, c \| STOP_BIT))
2201	<	startTerminating();
2200	>	if (U.compareAndSwapLong(this, CTL, c, c \| STOP_BIT)) {
2201	>	for (int pass = 0; pass < 3; ++pass) {
2202	>	WorkQueue[] ws = workQueues;
2203	>	if (ws != null) {
2204	>	WorkQueue w;
2205	>	int n = ws.length;
2206	>	for (int i = 0; i < n; ++i) {
2207	>	if ((w = ws[i]) != null) {
2208	>	w.qlock = -1;
2209	>	if (pass > 0) {
2210	>	w.cancelAll();
2211	>	if (pass > 1)
2212	>	w.interruptOwner();
2213	>	}
2214	>	}
2215	>	}
2216	>	// Wake up workers parked on event queue
2217	>	int i, e; long cc; Thread p;
2218	>	while ((e = (int)(cc = ctl) & E_MASK) != 0 &&
2219	>	(i = e & SMASK) < n &&
2220	>	(w = ws[i]) != null) {
2221	>	long nc = ((long)(w.nextWait & E_MASK) \|
2222	>	((cc + AC_UNIT) & AC_MASK) \|
2223	>	(cc & (TC_MASK\|STOP_BIT)));
2224	>	if (w.eventCount == (e \| INT_SIGN) &&
2225	>	U.compareAndSwapLong(this, CTL, cc, nc)) {
2226	>	w.eventCount = (e + E_SEQ) & E_MASK;
2227	>	w.qlock = -1;
2228	>	if ((p = w.parker) != null)
2229	>	U.unpark(p);
2230	>	}
2231	>	}
2232	>	}
2233	>	}
2234	>	}
2235		}
2236		}
2237
2238	+	// external operations on common pool
2239	+
2240		/**
2241	<	* Initiates termination: Runs three passes through workQueues:
2242	<	* (0) Setting termination status, followed by wakeups of queued
1836	<	* workers; (1) cancelling all tasks; (2) interrupting lagging
1837	<	* threads (likely in external tasks, but possibly also blocked in
1838	<	* joins). Each pass repeats previous steps because of potential
1839	<	* lagging thread creation.
2241	>	* Returns common pool queue for a thread that has submitted at
2242	>	* least one task.
2243		*/
2244	<	private void startTerminating() {
2245	<	for (int pass = 0; pass < 3; ++pass) {
2246	<	WorkQueue[] ws = workQueues;
2247	<	if (ws != null) {
2248	<	WorkQueue w; Thread wt;
2249	<	int n = ws.length;
2250	<	for (int j = 0; j < n; ++j) {
2251	<	if ((w = ws[j]) != null) {
2252	<	w.runState = -1;
2253	<	if (pass > 0) {
2254	<	w.cancelAll();
2255	<	if (pass > 1 && (wt = w.owner) != null &&
2256	<	!wt.isInterrupted()) {
2257	<	try {
2258	<	wt.interrupt();
2259	<	} catch (SecurityException ignore) {
2244	>	static WorkQueue commonSubmitterQueue() {
2245	>	ForkJoinPool p; WorkQueue[] ws; int m; Submitter z;
2246	>	return ((z = submitters.get()) != null &&
2247	>	(p = commonPool) != null &&
2248	>	(ws = p.workQueues) != null &&
2249	>	(m = ws.length - 1) >= 0) ?
2250	>	ws[m & z.seed & SQMASK] : null;
2251	>	}
2252	>
2253	>	/**
2254	>	* Tries to pop the given task from submitter's queue in common pool.
2255	>	*/
2256	>	static boolean tryExternalUnpush(ForkJoinTask<?> t) {
2257	>	ForkJoinPool p; WorkQueue[] ws; WorkQueue q; Submitter z;
2258	>	ForkJoinTask<?>[] a; int m, s; long j;
2259	>	if ((z = submitters.get()) != null &&
2260	>	(p = commonPool) != null &&
2261	>	(ws = p.workQueues) != null &&
2262	>	(m = ws.length - 1) >= 0 &&
2263	>	(q = ws[m & z.seed & SQMASK]) != null &&
2264	>	(s = q.top) != q.base &&
2265	>	(a = q.array) != null &&
2266	>	U.getObjectVolatile
2267	>	(a, j = (((a.length - 1) & (s - 1)) << ASHIFT) + ABASE) == t &&
2268	>	U.compareAndSwapInt(q, QLOCK, 0, 1)) {
2269	>	if (q.array == a && q.top == s && // recheck
2270	>	U.compareAndSwapObject(a, j, t, null)) {
2271	>	q.top = s - 1;
2272	>	q.qlock = 0;
2273	>	return true;
2274	>	}
2275	>	q.qlock = 0;
2276	>	}
2277	>	return false;
2278	>	}
2279	>
2280	>	/**
2281	>	* Tries to pop and run local tasks within the same computation
2282	>	* as the given root. On failure, tries to help complete from
2283	>	* other queues via helpComplete.
2284	>	*/
2285	>	private void externalHelpComplete(WorkQueue q, ForkJoinTask<?> root) {
2286	>	ForkJoinTask<?>[] a; int m;
2287	>	if (q != null && (a = q.array) != null && (m = (a.length - 1)) >= 0 &&
2288	>	root != null && root.status >= 0) {
2289	>	for (;;) {
2290	>	int s; Object o; CountedCompleter<?> task = null;
2291	>	if ((s = q.top) - q.base > 0) {
2292	>	long j = ((m & (s - 1)) << ASHIFT) + ABASE;
2293	>	if ((o = U.getObject(a, j)) != null &&
2294	>	(o instanceof CountedCompleter)) {
2295	>	CountedCompleter<?> t = (CountedCompleter<?>)o, r = t;
2296	>	do {
2297	>	if (r == root) {
2298	>	if (U.compareAndSwapInt(q, QLOCK, 0, 1)) {
2299	>	if (q.array == a && q.top == s &&
2300	>	U.compareAndSwapObject(a, j, t, null)) {
2301	>	q.top = s - 1;
2302	>	task = t;
2303	>	}
2304	>	q.qlock = 0;
2305		}
2306	+	break;
2307		}
2308	<	}
2308	>	} while ((r = r.completer) != null);
2309		}
2310		}
2311	<	// Wake up workers parked on event queue
2312	<	int i, e; long c; Thread p;
2313	<	while ((i = ((~(e = (int)(c = ctl)) << 1) \| 1) & SMASK) < n &&
2314	<	(w = ws[i]) != null &&
2315	<	w.eventCount == (e \| INT_SIGN)) {
2316	<	long nc = ((long)(w.nextWait & E_MASK) \|
2317	<	((c + AC_UNIT) & AC_MASK) \|
2318	<	(c & (TC_MASK\|STOP_BIT)));
2319	<	if (U.compareAndSwapLong(this, CTL, c, nc)) {
2320	<	w.eventCount = (e + E_SEQ) & E_MASK;
2321	<	if ((p = w.parker) != null)
2322	<	U.unpark(p);
2323	<	}
2311	>	if (task != null)
2312	>	task.doExec();
2313	>	if (root.status < 0 \|\| (int)(ctl >> AC_SHIFT) >= 0)
2314	>	break;
2315	>	if (task == null) {
2316	>	if (helpSignal(root, q.poolIndex) >= 0)
2317	>	helpComplete(root, SHARED_QUEUE);
2318	>	break;
2319	>	}
2320	>	}
2321	>	}
2322	>	}
2323	>
2324	>	/**
2325	>	* Tries to help execute or signal availability of the given task
2326	>	* from submitter's queue in common pool.
2327	>	*/
2328	>	static void externalHelpJoin(ForkJoinTask<?> t) {
2329	>	// Some hard-to-avoid overlap with tryExternalUnpush
2330	>	ForkJoinPool p; WorkQueue[] ws; WorkQueue q, w; Submitter z;
2331	>	ForkJoinTask<?>[] a; int m, s, n; long j;
2332	>	if (t != null && t.status >= 0 &&
2333	>	(z = submitters.get()) != null &&
2334	>	(p = commonPool) != null &&
2335	>	(ws = p.workQueues) != null &&
2336	>	(m = ws.length - 1) >= 0 &&
2337	>	(q = ws[m & z.seed & SQMASK]) != null &&
2338	>	(a = q.array) != null) {
2339	>	if ((s = q.top) != q.base &&
2340	>	U.getObjectVolatile
2341	>	(a, j = (((a.length - 1) & (s - 1)) << ASHIFT) + ABASE) == t &&
2342	>	U.compareAndSwapInt(q, QLOCK, 0, 1)) {
2343	>	if (q.array == a && q.top == s &&
2344	>	U.compareAndSwapObject(a, j, t, null)) {
2345	>	q.top = s - 1;
2346	>	q.qlock = 0;
2347	>	t.doExec();
2348		}
2349	+	else
2350	+	q.qlock = 0;
2351	+	}
2352	+	if (t.status >= 0) {
2353	+	if (t instanceof CountedCompleter)
2354	+	p.externalHelpComplete(q, t);
2355	+	else
2356	+	p.helpSignal(t, q.poolIndex);
2357		}
2358		}
2359		}
2360
2361	+	/**
2362	+	* Restricted version of helpQuiescePool for external callers
2363	+	*/
2364	+	static void externalHelpQuiescePool() {
2365	+	ForkJoinPool p; ForkJoinTask<?> t; WorkQueue q; int b;
2366	+	int r = ThreadLocalRandom.current().nextInt();
2367	+	if ((p = commonPool) != null &&
2368	+	(q = p.findNonEmptyStealQueue(r)) != null &&
2369	+	(b = q.base) - q.top < 0 &&
2370	+	(t = q.pollAt(b)) != null)
2371	+	t.doExec();
2372	+	}
2373	+
2374		// Exported methods
2375
2376		// Constructors
#	Line 1946 \| Line 2440 \| public class ForkJoinPool extends Abstra
2440		checkPermission();
2441		if (factory == null)
2442		throw new NullPointerException();
2443	<	if (parallelism <= 0 \|\| parallelism > MAX_ID)
2443	>	if (parallelism <= 0 \|\| parallelism > MAX_CAP)
2444		throw new IllegalArgumentException();
2445		this.parallelism = parallelism;
2446		this.factory = factory;
2447		this.ueh = handler;
2448		this.localMode = asyncMode ? FIFO_QUEUE : LIFO_QUEUE;
1955	–	this.nextPoolIndex = 1;
2449		long np = (long)(-parallelism); // offset ctl counts
2450		this.ctl = ((np << AC_SHIFT) & AC_MASK) \| ((np << TC_SHIFT) & TC_MASK);
2451	<	// initialize workQueues array with room for 2*parallelism if possible
1959	<	int n = parallelism << 1;
1960	<	if (n >= MAX_ID)
1961	<	n = MAX_ID;
1962	<	else { // See Hackers Delight, sec 3.2, where n < (1 << 16)
1963	<	n \|= n >>> 1; n \|= n >>> 2; n \|= n >>> 4; n \|= n >>> 8;
1964	<	}
1965	<	this.workQueues = new WorkQueue[(n + 1) << 1];
1966	<	ReentrantLock lck = this.lock = new ReentrantLock();
1967	<	this.termination = lck.newCondition();
1968	<	this.stealCount = new AtomicLong();
1969	<	this.nextWorkerNumber = new AtomicInteger();
2451	>	int pn = nextPoolId();
2452		StringBuilder sb = new StringBuilder("ForkJoinPool-");
2453	<	sb.append(poolNumberGenerator.incrementAndGet());
2453	>	sb.append(Integer.toString(pn));
2454		sb.append("-worker-");
2455		this.workerNamePrefix = sb.toString();
2456	<	// Create initial submission queue
2457	<	WorkQueue sq = tryAddSharedQueue(0);
2458	<	if (sq != null)
2459	<	sq.growArray(false);
2456	>	}
2457	>
2458	>	/**
2459	>	* Constructor for common pool, suitable only for static initialization.
2460	>	* Basically the same as above, but uses smallest possible initial footprint.
2461	>	*/
2462	>	ForkJoinPool(int parallelism, long ctl,
2463	>	ForkJoinWorkerThreadFactory factory,
2464	>	Thread.UncaughtExceptionHandler handler) {
2465	>	this.parallelism = parallelism;
2466	>	this.ctl = ctl;
2467	>	this.factory = factory;
2468	>	this.ueh = handler;
2469	>	this.localMode = LIFO_QUEUE;
2470	>	this.workerNamePrefix = "ForkJoinPool.commonPool-worker-";
2471	>	}
2472	>
2473	>	/**
2474	>	* Returns the common pool instance.
2475	>	*
2476	>	* @return the common pool instance
2477	>	*/
2478	>	public static ForkJoinPool commonPool() {
2479	>	return commonPool; // cannot be null (if so, a static init error)
2480		}
2481
2482		// Execution methods
#	Line 1996 \| Line 2498 \| public class ForkJoinPool extends Abstra
2498		* scheduled for execution
2499		*/
2500		public <T> T invoke(ForkJoinTask<T> task) {
2501	<	doSubmit(task);
2501	>	if (task == null)
2502	>	throw new NullPointerException();
2503	>	externalPush(task);
2504		return task.join();
2505		}
2506
#	Line 2009 \| Line 2513 \| public class ForkJoinPool extends Abstra
2513		* scheduled for execution
2514		*/
2515		public void execute(ForkJoinTask<?> task) {
2516	<	doSubmit(task);
2516	>	if (task == null)
2517	>	throw new NullPointerException();
2518	>	externalPush(task);
2519		}
2520
2521		// AbstractExecutorService methods
#	Line 2026 \| Line 2532 \| public class ForkJoinPool extends Abstra
2532		if (task instanceof ForkJoinTask<?>) // avoid re-wrap
2533		job = (ForkJoinTask<?>) task;
2534		else
2535	<	job = ForkJoinTask.adapt(task, null);
2536	<	doSubmit(job);
2535	>	job = new ForkJoinTask.AdaptedRunnableAction(task);
2536	>	externalPush(job);
2537		}
2538
2539		/**
#	Line 2040 \| Line 2546 \| public class ForkJoinPool extends Abstra
2546		* scheduled for execution
2547		*/
2548		public <T> ForkJoinTask<T> submit(ForkJoinTask<T> task) {
2549	<	doSubmit(task);
2549	>	if (task == null)
2550	>	throw new NullPointerException();
2551	>	externalPush(task);
2552		return task;
2553		}
2554
#	Line 2050 \| Line 2558 \| public class ForkJoinPool extends Abstra
2558		* scheduled for execution
2559		*/
2560		public <T> ForkJoinTask<T> submit(Callable<T> task) {
2561	<	if (task == null)
2562	<	throw new NullPointerException();
2055	<	ForkJoinTask<T> job = ForkJoinTask.adapt(task);
2056	<	doSubmit(job);
2561	>	ForkJoinTask<T> job = new ForkJoinTask.AdaptedCallable<T>(task);
2562	>	externalPush(job);
2563		return job;
2564		}
2565
#	Line 2063 \| Line 2569 \| public class ForkJoinPool extends Abstra
2569		* scheduled for execution
2570		*/
2571		public <T> ForkJoinTask<T> submit(Runnable task, T result) {
2572	<	if (task == null)
2573	<	throw new NullPointerException();
2068	<	ForkJoinTask<T> job = ForkJoinTask.adapt(task, result);
2069	<	doSubmit(job);
2572	>	ForkJoinTask<T> job = new ForkJoinTask.AdaptedRunnable<T>(task, result);
2573	>	externalPush(job);
2574		return job;
2575		}
2576
#	Line 2082 \| Line 2586 \| public class ForkJoinPool extends Abstra
2586		if (task instanceof ForkJoinTask<?>) // avoid re-wrap
2587		job = (ForkJoinTask<?>) task;
2588		else
2589	<	job = ForkJoinTask.adapt(task, null);
2590	<	doSubmit(job);
2589	>	job = new ForkJoinTask.AdaptedRunnableAction(task);
2590	>	externalPush(job);
2591		return job;
2592		}
2593
#	Line 2092 \| Line 2596 \| public class ForkJoinPool extends Abstra
2596		* @throws RejectedExecutionException {@inheritDoc}
2597		*/
2598		public <T> List<Future<T>> invokeAll(Collection<? extends Callable<T>> tasks) {
2599	<	ArrayList<ForkJoinTask<T>> forkJoinTasks =
2600	<	new ArrayList<ForkJoinTask<T>>(tasks.size());
2601	<	for (Callable<T> task : tasks)
2602	<	forkJoinTasks.add(ForkJoinTask.adapt(task));
2603	<	invoke(new InvokeAll<T>(forkJoinTasks));
2604	<
2599	>	// In previous versions of this class, this method constructed
2600	>	// a task to run ForkJoinTask.invokeAll, but now external
2601	>	// invocation of multiple tasks is at least as efficient.
2602	>	List<ForkJoinTask<T>> fs = new ArrayList<ForkJoinTask<T>>(tasks.size());
2603	>	// Workaround needed because method wasn't declared with
2604	>	// wildcards in return type but should have been.
2605		@SuppressWarnings({"unchecked", "rawtypes"})
2606	<	List<Future<T>> futures = (List<Future<T>>) (List) forkJoinTasks;
2103	<	return futures;
2104	<	}
2606	>	List<Future<T>> futures = (List<Future<T>>) (List) fs;
2607
2608	<	static final class InvokeAll<T> extends RecursiveAction {
2609	<	final ArrayList<ForkJoinTask<T>> tasks;
2610	<	InvokeAll(ArrayList<ForkJoinTask<T>> tasks) { this.tasks = tasks; }
2611	<	public void compute() {
2612	<	try { invokeAll(tasks); }
2613	<	catch (Exception ignore) {}
2608	>	boolean done = false;
2609	>	try {
2610	>	for (Callable<T> t : tasks) {
2611	>	ForkJoinTask<T> f = new ForkJoinTask.AdaptedCallable<T>(t);
2612	>	externalPush(f);
2613	>	fs.add(f);
2614	>	}
2615	>	for (ForkJoinTask<T> f : fs)
2616	>	f.quietlyJoin();
2617	>	done = true;
2618	>	return futures;
2619	>	} finally {
2620	>	if (!done)
2621	>	for (ForkJoinTask<T> f : fs)
2622	>	f.cancel(false);
2623		}
2113	–	private static final long serialVersionUID = -7914297376763021607L;
2624		}
2625
2626		/**
#	Line 2142 \| Line 2652 \| public class ForkJoinPool extends Abstra
2652		}
2653
2654		/**
2655	+	* Returns the targeted parallelism level of the common pool.
2656	+	*
2657	+	* @return the targeted parallelism level of the common pool
2658	+	*/
2659	+	public static int getCommonPoolParallelism() {
2660	+	return commonPoolParallelism;
2661	+	}
2662	+
2663	+	/**
2664		* Returns the number of worker threads that have started but not
2665		* yet terminated. The result returned by this method may differ
2666		* from {@link #getParallelism} when threads are created to
#	Line 2175 \| Line 2694 \| public class ForkJoinPool extends Abstra
2694		int rc = 0;
2695		WorkQueue[] ws; WorkQueue w;
2696		if ((ws = workQueues) != null) {
2697	<	int n = ws.length;
2698	<	for (int i = 1; i < n; i += 2) {
2180	<	Thread.State s; ForkJoinWorkerThread wt;
2181	<	if ((w = ws[i]) != null && (wt = w.owner) != null &&
2182	<	w.eventCount >= 0 &&
2183	<	(s = wt.getState()) != Thread.State.BLOCKED &&
2184	<	s != Thread.State.WAITING &&
2185	<	s != Thread.State.TIMED_WAITING)
2697	>	for (int i = 1; i < ws.length; i += 2) {
2698	>	if ((w = ws[i]) != null && w.isApparentlyUnblocked())
2699		++rc;
2700		}
2701		}
#	Line 2228 \| Line 2741 \| public class ForkJoinPool extends Abstra
2741		* @return the number of steals
2742		*/
2743		public long getStealCount() {
2744	<	long count = stealCount.get();
2744	>	long count = stealCount;
2745		WorkQueue[] ws; WorkQueue w;
2746		if ((ws = workQueues) != null) {
2747	<	int n = ws.length;
2235	<	for (int i = 1; i < n; i += 2) {
2747	>	for (int i = 1; i < ws.length; i += 2) {
2748		if ((w = ws[i]) != null)
2749	<	count += w.totalSteals;
2749	>	count += w.nsteals;
2750		}
2751		}
2752		return count;
#	Line 2254 \| Line 2766 \| public class ForkJoinPool extends Abstra
2766		long count = 0;
2767		WorkQueue[] ws; WorkQueue w;
2768		if ((ws = workQueues) != null) {
2769	<	int n = ws.length;
2258	<	for (int i = 1; i < n; i += 2) {
2769	>	for (int i = 1; i < ws.length; i += 2) {
2770		if ((w = ws[i]) != null)
2771		count += w.queueSize();
2772		}
#	Line 2274 \| Line 2785 \| public class ForkJoinPool extends Abstra
2785		int count = 0;
2786		WorkQueue[] ws; WorkQueue w;
2787		if ((ws = workQueues) != null) {
2788	<	int n = ws.length;
2278	<	for (int i = 0; i < n; i += 2) {
2788	>	for (int i = 0; i < ws.length; i += 2) {
2789		if ((w = ws[i]) != null)
2790		count += w.queueSize();
2791		}
#	Line 2292 \| Line 2802 \| public class ForkJoinPool extends Abstra
2802		public boolean hasQueuedSubmissions() {
2803		WorkQueue[] ws; WorkQueue w;
2804		if ((ws = workQueues) != null) {
2805	<	int n = ws.length;
2296	<	for (int i = 0; i < n; i += 2) {
2805	>	for (int i = 0; i < ws.length; i += 2) {
2806		if ((w = ws[i]) != null && w.queueSize() != 0)
2807		return true;
2808		}
#	Line 2311 \| Line 2820 \| public class ForkJoinPool extends Abstra
2820		protected ForkJoinTask<?> pollSubmission() {
2821		WorkQueue[] ws; WorkQueue w; ForkJoinTask<?> t;
2822		if ((ws = workQueues) != null) {
2823	<	int n = ws.length;
2315	<	for (int i = 0; i < n; i += 2) {
2823	>	for (int i = 0; i < ws.length; i += 2) {
2824		if ((w = ws[i]) != null && (t = w.poll()) != null)
2825		return t;
2826		}
#	Line 2341 \| Line 2849 \| public class ForkJoinPool extends Abstra
2849		int count = 0;
2850		WorkQueue[] ws; WorkQueue w; ForkJoinTask<?> t;
2851		if ((ws = workQueues) != null) {
2852	<	int n = ws.length;
2345	<	for (int i = 0; i < n; ++i) {
2852	>	for (int i = 0; i < ws.length; ++i) {
2853		if ((w = ws[i]) != null) {
2854		while ((t = w.poll()) != null) {
2855		c.add(t);
#	Line 2362 \| Line 2869 \| public class ForkJoinPool extends Abstra
2869		* @return a string identifying this pool, as well as its state
2870		*/
2871		public String toString() {
2872	<	long st = getStealCount();
2873	<	long qt = getQueuedTaskCount();
2874	<	long qs = getQueuedSubmissionCount();
2368	<	int rc = getRunningThreadCount();
2369	<	int pc = parallelism;
2872	>	// Use a single pass through workQueues to collect counts
2873	>	long qt = 0L, qs = 0L; int rc = 0;
2874	>	long st = stealCount;
2875		long c = ctl;
2876	+	WorkQueue[] ws; WorkQueue w;
2877	+	if ((ws = workQueues) != null) {
2878	+	for (int i = 0; i < ws.length; ++i) {
2879	+	if ((w = ws[i]) != null) {
2880	+	int size = w.queueSize();
2881	+	if ((i & 1) == 0)
2882	+	qs += size;
2883	+	else {
2884	+	qt += size;
2885	+	st += w.nsteals;
2886	+	if (w.isApparentlyUnblocked())
2887	+	++rc;
2888	+	}
2889	+	}
2890	+	}
2891	+	}
2892	+	int pc = parallelism;
2893		int tc = pc + (short)(c >>> TC_SHIFT);
2894		int ac = pc + (int)(c >> AC_SHIFT);
2895		if (ac < 0) // ignore transient negative
#	Line 2376 \| Line 2898 \| public class ForkJoinPool extends Abstra
2898		if ((c & STOP_BIT) != 0)
2899		level = (tc == 0) ? "Terminated" : "Terminating";
2900		else
2901	<	level = runState < 0 ? "Shutting down" : "Running";
2901	>	level = plock < 0 ? "Shutting down" : "Running";
2902		return super.toString() +
2903		"[" + level +
2904		", parallelism = " + pc +
#	Line 2390 \| Line 2912 \| public class ForkJoinPool extends Abstra
2912		}
2913
2914		/**
2915	<	* Initiates an orderly shutdown in which previously submitted
2916	<	* tasks are executed, but no new tasks will be accepted.
2917	<	* Invocation has no additional effect if already shut down.
2918	<	* Tasks that are in the process of being submitted concurrently
2919	<	* during the course of this method may or may not be rejected.
2915	>	* Possibly initiates an orderly shutdown in which previously
2916	>	* submitted tasks are executed, but no new tasks will be
2917	>	* accepted. Invocation has no effect on execution state if this
2918	>	* is the {@link #commonPool}, and no additional effect if
2919	>	* already shut down. Tasks that are in the process of being
2920	>	* submitted concurrently during the course of this method may or
2921	>	* may not be rejected.
2922		*
2923		* @throws SecurityException if a security manager exists and
2924		* the caller is not permitted to modify threads
#	Line 2403 \| Line 2927 \| public class ForkJoinPool extends Abstra
2927		*/
2928		public void shutdown() {
2929		checkPermission();
2930	<	enableShutdown();
2407	<	tryTerminate(false);
2930	>	tryTerminate(false, true);
2931		}
2932
2933		/**
2934	<	* Attempts to cancel and/or stop all tasks, and reject all
2935	<	* subsequently submitted tasks. Tasks that are in the process of
2936	<	* being submitted or executed concurrently during the course of
2937	<	* this method may or may not be rejected. This method cancels
2938	<	* both existing and unexecuted tasks, in order to permit
2939	<	* termination in the presence of task dependencies. So the method
2940	<	* always returns an empty list (unlike the case for some other
2941	<	* Executors).
2934	>	* Possibly attempts to cancel and/or stop all tasks, and reject
2935	>	* all subsequently submitted tasks. Invocation has no effect on
2936	>	* execution state if this is the {@link #commonPool}, and no
2937	>	* additional effect if already shut down. Otherwise, tasks that
2938	>	* are in the process of being submitted or executed concurrently
2939	>	* during the course of this method may or may not be
2940	>	* rejected. This method cancels both existing and unexecuted
2941	>	* tasks, in order to permit termination in the presence of task
2942	>	* dependencies. So the method always returns an empty list
2943	>	* (unlike the case for some other Executors).
2944		*
2945		* @return an empty list
2946		* @throws SecurityException if a security manager exists and
#	Line 2425 \| Line 2950 \| public class ForkJoinPool extends Abstra
2950		*/
2951		public List<Runnable> shutdownNow() {
2952		checkPermission();
2953	<	enableShutdown();
2429	<	tryTerminate(true);
2953	>	tryTerminate(true, true);
2954		return Collections.emptyList();
2955		}
2956
#	Line 2466 \| Line 2990 \| public class ForkJoinPool extends Abstra
2990		* @return {@code true} if this pool has been shut down
2991		*/
2992		public boolean isShutdown() {
2993	<	return runState < 0;
2993	>	return plock < 0;
2994		}
2995
2996		/**
2997	<	* Blocks until all tasks have completed execution after a shutdown
2998	<	* request, or the timeout occurs, or the current thread is
2999	<	* interrupted, whichever happens first.
2997	>	* Blocks until all tasks have completed execution after a
2998	>	* shutdown request, or the timeout occurs, or the current thread
2999	>	* is interrupted, whichever happens first. Note that the {@link
3000	>	* #commonPool()} never terminates until program shutdown so
3001	>	* this method will always time out.
3002		*
3003		* @param timeout the maximum time to wait
3004		* @param unit the time unit of the timeout argument
#	Line 2483 \| Line 3009 \| public class ForkJoinPool extends Abstra
3009		public boolean awaitTermination(long timeout, TimeUnit unit)
3010		throws InterruptedException {
3011		long nanos = unit.toNanos(timeout);
3012	<	final ReentrantLock lock = this.lock;
3013	<	lock.lock();
3014	<	try {
3015	<	for (;;) {
3016	<	if (isTerminated())
3017	<	return true;
3018	<	if (nanos <= 0)
3019	<	return false;
3020	<	nanos = termination.awaitNanos(nanos);
3012	>	if (isTerminated())
3013	>	return true;
3014	>	long startTime = System.nanoTime();
3015	>	boolean terminated = false;
3016	>	synchronized (this) {
3017	>	for (long waitTime = nanos, millis = 0L;;) {
3018	>	if (terminated = isTerminated() \|\|
3019	>	waitTime <= 0L \|\|
3020	>	(millis = unit.toMillis(waitTime)) <= 0L)
3021	>	break;
3022	>	wait(millis);
3023	>	waitTime = nanos - (System.nanoTime() - startTime);
3024		}
2496	–	} finally {
2497	–	lock.unlock();
3025		}
3026	+	return terminated;
3027		}
3028
3029		/**
#	Line 2594 \| Line 3122 \| public class ForkJoinPool extends Abstra
3122		public static void managedBlock(ManagedBlocker blocker)
3123		throws InterruptedException {
3124		Thread t = Thread.currentThread();
3125	<	ForkJoinPool p = ((t instanceof ForkJoinWorkerThread) ?
3126	<	((ForkJoinWorkerThread)t).pool : null);
3127	<	while (!blocker.isReleasable()) {
3128	<	if (p == null \|\| p.tryCompensate()) {
3129	<	try {
3130	<	do {} while (!blocker.isReleasable() && !blocker.block());
3131	<	} finally {
3132	<	if (p != null)
3125	>	if (t instanceof ForkJoinWorkerThread) {
3126	>	ForkJoinPool p = ((ForkJoinWorkerThread)t).pool;
3127	>	while (!blocker.isReleasable()) { // variant of helpSignal
3128	>	WorkQueue[] ws; WorkQueue q; int m, n;
3129	>	if ((ws = p.workQueues) != null && (m = ws.length - 1) >= 0) {
3130	>	for (int i = 0; i <= m; ++i) {
3131	>	if (blocker.isReleasable())
3132	>	return;
3133	>	if ((q = ws[i]) != null && (n = q.queueSize()) > 0) {
3134	>	p.signalWork(q, n);
3135	>	if ((int)(p.ctl >> AC_SHIFT) >= 0)
3136	>	break;
3137	>	}
3138	>	}
3139	>	}
3140	>	if (p.tryCompensate()) {
3141	>	try {
3142	>	do {} while (!blocker.isReleasable() &&
3143	>	!blocker.block());
3144	>	} finally {
3145		p.incrementActiveCount();
3146	+	}
3147	+	break;
3148		}
2607	–	break;
3149		}
3150		}
3151	+	else {
3152	+	do {} while (!blocker.isReleasable() &&
3153	+	!blocker.block());
3154	+	}
3155		}
3156
3157		// AbstractExecutorService overrides. These rely on undocumented
#	Line 2614 \| Line 3159 \| public class ForkJoinPool extends Abstra
3159		// implement RunnableFuture.
3160
3161		protected <T> RunnableFuture<T> newTaskFor(Runnable runnable, T value) {
3162	<	return (RunnableFuture<T>) ForkJoinTask.adapt(runnable, value);
3162	>	return new ForkJoinTask.AdaptedRunnable<T>(runnable, value);
3163		}
3164
3165		protected <T> RunnableFuture<T> newTaskFor(Callable<T> callable) {
3166	<	return (RunnableFuture<T>) ForkJoinTask.adapt(callable);
3166	>	return new ForkJoinTask.AdaptedCallable<T>(callable);
3167		}
3168
3169		// Unsafe mechanics
3170		private static final sun.misc.Unsafe U;
3171		private static final long CTL;
2627	–	private static final long RUNSTATE;
3172		private static final long PARKBLOCKER;
3173	+	private static final int ABASE;
3174	+	private static final int ASHIFT;
3175	+	private static final long STEALCOUNT;
3176	+	private static final long PLOCK;
3177	+	private static final long INDEXSEED;
3178	+	private static final long QLOCK;
3179
3180		static {
3181	<	poolNumberGenerator = new AtomicInteger();
3182	<	modifyThreadPermission = new RuntimePermission("modifyThread");
3183	<	defaultForkJoinWorkerThreadFactory =
3184	<	new DefaultForkJoinWorkerThreadFactory();
3185	<	int s;
3181	>	// Establish common pool parameters
3182	>	// TBD: limit or report ignored exceptions?
3183	>
3184	>	int par = 0;
3185	>	ForkJoinWorkerThreadFactory fac = null;
3186	>	Thread.UncaughtExceptionHandler handler = null;
3187	>	try {
3188	>	String pp = System.getProperty(propPrefix + "parallelism");
3189	>	String hp = System.getProperty(propPrefix + "exceptionHandler");
3190	>	String fp = System.getProperty(propPrefix + "threadFactory");
3191	>	if (fp != null)
3192	>	fac = ((ForkJoinWorkerThreadFactory)ClassLoader.
3193	>	getSystemClassLoader().loadClass(fp).newInstance());
3194	>	if (hp != null)
3195	>	handler = ((Thread.UncaughtExceptionHandler)ClassLoader.
3196	>	getSystemClassLoader().loadClass(hp).newInstance());
3197	>	if (pp != null)
3198	>	par = Integer.parseInt(pp);
3199	>	} catch (Exception ignore) {
3200	>	}
3201	>
3202	>	int s; // initialize field offsets for CAS etc
3203		try {
3204		U = getUnsafe();
3205		Class<?> k = ForkJoinPool.class;
2639	–	Class<?> tk = Thread.class;
3206		CTL = U.objectFieldOffset
3207		(k.getDeclaredField("ctl"));
3208	<	RUNSTATE = U.objectFieldOffset
3209	<	(k.getDeclaredField("runState"));
3208	>	STEALCOUNT = U.objectFieldOffset
3209	>	(k.getDeclaredField("stealCount"));
3210	>	PLOCK = U.objectFieldOffset
3211	>	(k.getDeclaredField("plock"));
3212	>	INDEXSEED = U.objectFieldOffset
3213	>	(k.getDeclaredField("indexSeed"));
3214	>	Class<?> tk = Thread.class;
3215		PARKBLOCKER = U.objectFieldOffset
3216		(tk.getDeclaredField("parkBlocker"));
3217	+	Class<?> wk = WorkQueue.class;
3218	+	QLOCK = U.objectFieldOffset
3219	+	(wk.getDeclaredField("qlock"));
3220	+	Class<?> ak = ForkJoinTask[].class;
3221	+	ABASE = U.arrayBaseOffset(ak);
3222	+	s = U.arrayIndexScale(ak);
3223	+	ASHIFT = 31 - Integer.numberOfLeadingZeros(s);
3224		} catch (Exception e) {
3225		throw new Error(e);
3226		}
3227	+	if ((s & (s-1)) != 0)
3228	+	throw new Error("data type scale not a power of two");
3229	+
3230	+	/*
3231	+	* For extra caution, computations to set up pool state are
3232	+	* here; the constructor just assigns these values to fields.
3233	+	*/
3234	+	ForkJoinWorkerThreadFactory defaultFac =
3235	+	defaultForkJoinWorkerThreadFactory =
3236	+	new DefaultForkJoinWorkerThreadFactory();
3237	+	if (fac == null)
3238	+	fac = defaultFac;
3239	+	if (par <= 0)
3240	+	par = Runtime.getRuntime().availableProcessors();
3241	+	if (par > MAX_CAP)
3242	+	par = MAX_CAP;
3243	+	long np = (long)(-par); // precompute initial ctl value
3244	+	long ct = ((np << AC_SHIFT) & AC_MASK) \| ((np << TC_SHIFT) & TC_MASK);
3245	+
3246	+	commonPoolParallelism = par;
3247	+	commonPool = new ForkJoinPool(par, ct, fac, handler);
3248	+	modifyThreadPermission = new RuntimePermission("modifyThread");
3249	+	submitters = new ThreadLocal<Submitter>();
3250		}
3251
3252		/**

Diff Legend

-–
+Removed lines
-+
+Added lines
-<
+Changed lines
->
+Changed lines

Comparing jsr166/src/jsr166y/ForkJoinPool.java (file contents): Revision 1.116 by dl, Fri Jan 27 17:27:28 2012 UTC vs. Revision 1.144 by jsr166, Sun Nov 18 18:03:10 2012 UTC

Diff Legend

Comparing jsr166/src/jsr166y/ForkJoinPool.java (file contents):
Revision 1.116 by dl, Fri Jan 27 17:27:28 2012 UTC vs.
Revision 1.144 by jsr166, Sun Nov 18 18:03:10 2012 UTC