21 #include "kmp_wait_release.h" 22 #include "kmp_wrapper_getpid.h" 24 #include "tsan_annotations.h" 27 #include <sys/syscall.h> 45 void __kmp_validate_locks(
void) {
50 x = ~((kmp_uint32)0) - 2;
53 for (i = 0; i < 8; ++i, ++x, ++y) {
54 kmp_uint32 z = (x - y);
58 KMP_ASSERT(offsetof(kmp_base_queuing_lock, tail_id) % 8 == 0);
72 static kmp_int32 __kmp_get_tas_lock_owner(kmp_tas_lock_t *lck) {
73 return KMP_LOCK_STRIP(KMP_ATOMIC_LD_RLX(&lck->lk.poll)) - 1;
76 static inline bool __kmp_is_tas_lock_nestable(kmp_tas_lock_t *lck) {
77 return lck->lk.depth_locked != -1;
80 __forceinline
static int 81 __kmp_acquire_tas_lock_timed_template(kmp_tas_lock_t *lck, kmp_int32 gtid) {
84 #ifdef USE_LOCK_PROFILE 85 kmp_uint32 curr = KMP_LOCK_STRIP(lck->lk.poll);
86 if ((curr != 0) && (curr != gtid + 1))
87 __kmp_printf(
"LOCK CONTENTION: %p\n", lck);
91 kmp_int32 tas_free = KMP_LOCK_FREE(tas);
92 kmp_int32 tas_busy = KMP_LOCK_BUSY(gtid + 1, tas);
94 if (KMP_ATOMIC_LD_RLX(&lck->lk.poll) == tas_free &&
95 __kmp_atomic_compare_store_acq(&lck->lk.poll, tas_free, tas_busy)) {
96 KMP_FSYNC_ACQUIRED(lck);
97 return KMP_LOCK_ACQUIRED_FIRST;
101 KMP_FSYNC_PREPARE(lck);
102 KMP_INIT_YIELD(spins);
103 kmp_backoff_t backoff = __kmp_spin_backoff_params;
105 __kmp_spin_backoff(&backoff);
106 KMP_YIELD_OVERSUB_ELSE_SPIN(spins);
107 }
while (KMP_ATOMIC_LD_RLX(&lck->lk.poll) != tas_free ||
108 !__kmp_atomic_compare_store_acq(&lck->lk.poll, tas_free, tas_busy));
109 KMP_FSYNC_ACQUIRED(lck);
110 return KMP_LOCK_ACQUIRED_FIRST;
113 int __kmp_acquire_tas_lock(kmp_tas_lock_t *lck, kmp_int32 gtid) {
114 int retval = __kmp_acquire_tas_lock_timed_template(lck, gtid);
115 ANNOTATE_TAS_ACQUIRED(lck);
119 static int __kmp_acquire_tas_lock_with_checks(kmp_tas_lock_t *lck,
121 char const *
const func =
"omp_set_lock";
122 if ((
sizeof(kmp_tas_lock_t) <= OMP_LOCK_T_SIZE) &&
123 __kmp_is_tas_lock_nestable(lck)) {
124 KMP_FATAL(LockNestableUsedAsSimple, func);
126 if ((gtid >= 0) && (__kmp_get_tas_lock_owner(lck) == gtid)) {
127 KMP_FATAL(LockIsAlreadyOwned, func);
129 return __kmp_acquire_tas_lock(lck, gtid);
132 int __kmp_test_tas_lock(kmp_tas_lock_t *lck, kmp_int32 gtid) {
133 kmp_int32 tas_free = KMP_LOCK_FREE(tas);
134 kmp_int32 tas_busy = KMP_LOCK_BUSY(gtid + 1, tas);
135 if (KMP_ATOMIC_LD_RLX(&lck->lk.poll) == tas_free &&
136 __kmp_atomic_compare_store_acq(&lck->lk.poll, tas_free, tas_busy)) {
137 KMP_FSYNC_ACQUIRED(lck);
143 static int __kmp_test_tas_lock_with_checks(kmp_tas_lock_t *lck,
145 char const *
const func =
"omp_test_lock";
146 if ((
sizeof(kmp_tas_lock_t) <= OMP_LOCK_T_SIZE) &&
147 __kmp_is_tas_lock_nestable(lck)) {
148 KMP_FATAL(LockNestableUsedAsSimple, func);
150 return __kmp_test_tas_lock(lck, gtid);
153 int __kmp_release_tas_lock(kmp_tas_lock_t *lck, kmp_int32 gtid) {
156 KMP_FSYNC_RELEASING(lck);
157 ANNOTATE_TAS_RELEASED(lck);
158 KMP_ATOMIC_ST_REL(&lck->lk.poll, KMP_LOCK_FREE(tas));
162 return KMP_LOCK_RELEASED;
165 static int __kmp_release_tas_lock_with_checks(kmp_tas_lock_t *lck,
167 char const *
const func =
"omp_unset_lock";
169 if ((
sizeof(kmp_tas_lock_t) <= OMP_LOCK_T_SIZE) &&
170 __kmp_is_tas_lock_nestable(lck)) {
171 KMP_FATAL(LockNestableUsedAsSimple, func);
173 if (__kmp_get_tas_lock_owner(lck) == -1) {
174 KMP_FATAL(LockUnsettingFree, func);
176 if ((gtid >= 0) && (__kmp_get_tas_lock_owner(lck) >= 0) &&
177 (__kmp_get_tas_lock_owner(lck) != gtid)) {
178 KMP_FATAL(LockUnsettingSetByAnother, func);
180 return __kmp_release_tas_lock(lck, gtid);
183 void __kmp_init_tas_lock(kmp_tas_lock_t *lck) {
184 lck->lk.poll = KMP_LOCK_FREE(tas);
187 void __kmp_destroy_tas_lock(kmp_tas_lock_t *lck) { lck->lk.poll = 0; }
189 static void __kmp_destroy_tas_lock_with_checks(kmp_tas_lock_t *lck) {
190 char const *
const func =
"omp_destroy_lock";
191 if ((
sizeof(kmp_tas_lock_t) <= OMP_LOCK_T_SIZE) &&
192 __kmp_is_tas_lock_nestable(lck)) {
193 KMP_FATAL(LockNestableUsedAsSimple, func);
195 if (__kmp_get_tas_lock_owner(lck) != -1) {
196 KMP_FATAL(LockStillOwned, func);
198 __kmp_destroy_tas_lock(lck);
203 int __kmp_acquire_nested_tas_lock(kmp_tas_lock_t *lck, kmp_int32 gtid) {
204 KMP_DEBUG_ASSERT(gtid >= 0);
206 if (__kmp_get_tas_lock_owner(lck) == gtid) {
207 lck->lk.depth_locked += 1;
208 return KMP_LOCK_ACQUIRED_NEXT;
210 __kmp_acquire_tas_lock_timed_template(lck, gtid);
211 ANNOTATE_TAS_ACQUIRED(lck);
212 lck->lk.depth_locked = 1;
213 return KMP_LOCK_ACQUIRED_FIRST;
217 static int __kmp_acquire_nested_tas_lock_with_checks(kmp_tas_lock_t *lck,
219 char const *
const func =
"omp_set_nest_lock";
220 if (!__kmp_is_tas_lock_nestable(lck)) {
221 KMP_FATAL(LockSimpleUsedAsNestable, func);
223 return __kmp_acquire_nested_tas_lock(lck, gtid);
226 int __kmp_test_nested_tas_lock(kmp_tas_lock_t *lck, kmp_int32 gtid) {
229 KMP_DEBUG_ASSERT(gtid >= 0);
231 if (__kmp_get_tas_lock_owner(lck) == gtid) {
232 retval = ++lck->lk.depth_locked;
233 }
else if (!__kmp_test_tas_lock(lck, gtid)) {
237 retval = lck->lk.depth_locked = 1;
242 static int __kmp_test_nested_tas_lock_with_checks(kmp_tas_lock_t *lck,
244 char const *
const func =
"omp_test_nest_lock";
245 if (!__kmp_is_tas_lock_nestable(lck)) {
246 KMP_FATAL(LockSimpleUsedAsNestable, func);
248 return __kmp_test_nested_tas_lock(lck, gtid);
251 int __kmp_release_nested_tas_lock(kmp_tas_lock_t *lck, kmp_int32 gtid) {
252 KMP_DEBUG_ASSERT(gtid >= 0);
255 if (--(lck->lk.depth_locked) == 0) {
256 __kmp_release_tas_lock(lck, gtid);
257 return KMP_LOCK_RELEASED;
259 return KMP_LOCK_STILL_HELD;
262 static int __kmp_release_nested_tas_lock_with_checks(kmp_tas_lock_t *lck,
264 char const *
const func =
"omp_unset_nest_lock";
266 if (!__kmp_is_tas_lock_nestable(lck)) {
267 KMP_FATAL(LockSimpleUsedAsNestable, func);
269 if (__kmp_get_tas_lock_owner(lck) == -1) {
270 KMP_FATAL(LockUnsettingFree, func);
272 if (__kmp_get_tas_lock_owner(lck) != gtid) {
273 KMP_FATAL(LockUnsettingSetByAnother, func);
275 return __kmp_release_nested_tas_lock(lck, gtid);
278 void __kmp_init_nested_tas_lock(kmp_tas_lock_t *lck) {
279 __kmp_init_tas_lock(lck);
280 lck->lk.depth_locked = 0;
283 void __kmp_destroy_nested_tas_lock(kmp_tas_lock_t *lck) {
284 __kmp_destroy_tas_lock(lck);
285 lck->lk.depth_locked = 0;
288 static void __kmp_destroy_nested_tas_lock_with_checks(kmp_tas_lock_t *lck) {
289 char const *
const func =
"omp_destroy_nest_lock";
290 if (!__kmp_is_tas_lock_nestable(lck)) {
291 KMP_FATAL(LockSimpleUsedAsNestable, func);
293 if (__kmp_get_tas_lock_owner(lck) != -1) {
294 KMP_FATAL(LockStillOwned, func);
296 __kmp_destroy_nested_tas_lock(lck);
309 static kmp_int32 __kmp_get_futex_lock_owner(kmp_futex_lock_t *lck) {
310 return KMP_LOCK_STRIP((TCR_4(lck->lk.poll) >> 1)) - 1;
313 static inline bool __kmp_is_futex_lock_nestable(kmp_futex_lock_t *lck) {
314 return lck->lk.depth_locked != -1;
317 __forceinline
static int 318 __kmp_acquire_futex_lock_timed_template(kmp_futex_lock_t *lck, kmp_int32 gtid) {
319 kmp_int32 gtid_code = (gtid + 1) << 1;
323 #ifdef USE_LOCK_PROFILE 324 kmp_uint32 curr = KMP_LOCK_STRIP(TCR_4(lck->lk.poll));
325 if ((curr != 0) && (curr != gtid_code))
326 __kmp_printf(
"LOCK CONTENTION: %p\n", lck);
330 KMP_FSYNC_PREPARE(lck);
331 KA_TRACE(1000, (
"__kmp_acquire_futex_lock: lck:%p(0x%x), T#%d entering\n",
332 lck, lck->lk.poll, gtid));
336 while ((poll_val = KMP_COMPARE_AND_STORE_RET32(
337 &(lck->lk.poll), KMP_LOCK_FREE(futex),
338 KMP_LOCK_BUSY(gtid_code, futex))) != KMP_LOCK_FREE(futex)) {
340 kmp_int32 cond = KMP_LOCK_STRIP(poll_val) & 1;
343 (
"__kmp_acquire_futex_lock: lck:%p, T#%d poll_val = 0x%x cond = 0x%x\n",
344 lck, gtid, poll_val, cond));
355 if (!KMP_COMPARE_AND_STORE_REL32(&(lck->lk.poll), poll_val,
356 poll_val | KMP_LOCK_BUSY(1, futex))) {
359 (
"__kmp_acquire_futex_lock: lck:%p(0x%x), T#%d can't set bit 0\n",
360 lck, lck->lk.poll, gtid));
363 poll_val |= KMP_LOCK_BUSY(1, futex);
366 (
"__kmp_acquire_futex_lock: lck:%p(0x%x), T#%d bit 0 set\n", lck,
367 lck->lk.poll, gtid));
372 (
"__kmp_acquire_futex_lock: lck:%p, T#%d before futex_wait(0x%x)\n",
373 lck, gtid, poll_val));
376 if ((rc = syscall(__NR_futex, &(lck->lk.poll), FUTEX_WAIT, poll_val, NULL,
378 KA_TRACE(1000, (
"__kmp_acquire_futex_lock: lck:%p, T#%d futex_wait(0x%x) " 379 "failed (rc=%d errno=%d)\n",
380 lck, gtid, poll_val, rc, errno));
385 (
"__kmp_acquire_futex_lock: lck:%p, T#%d after futex_wait(0x%x)\n",
386 lck, gtid, poll_val));
393 KMP_FSYNC_ACQUIRED(lck);
394 KA_TRACE(1000, (
"__kmp_acquire_futex_lock: lck:%p(0x%x), T#%d exiting\n", lck,
395 lck->lk.poll, gtid));
396 return KMP_LOCK_ACQUIRED_FIRST;
399 int __kmp_acquire_futex_lock(kmp_futex_lock_t *lck, kmp_int32 gtid) {
400 int retval = __kmp_acquire_futex_lock_timed_template(lck, gtid);
401 ANNOTATE_FUTEX_ACQUIRED(lck);
405 static int __kmp_acquire_futex_lock_with_checks(kmp_futex_lock_t *lck,
407 char const *
const func =
"omp_set_lock";
408 if ((
sizeof(kmp_futex_lock_t) <= OMP_LOCK_T_SIZE) &&
409 __kmp_is_futex_lock_nestable(lck)) {
410 KMP_FATAL(LockNestableUsedAsSimple, func);
412 if ((gtid >= 0) && (__kmp_get_futex_lock_owner(lck) == gtid)) {
413 KMP_FATAL(LockIsAlreadyOwned, func);
415 return __kmp_acquire_futex_lock(lck, gtid);
418 int __kmp_test_futex_lock(kmp_futex_lock_t *lck, kmp_int32 gtid) {
419 if (KMP_COMPARE_AND_STORE_ACQ32(&(lck->lk.poll), KMP_LOCK_FREE(futex),
420 KMP_LOCK_BUSY((gtid + 1) << 1, futex))) {
421 KMP_FSYNC_ACQUIRED(lck);
427 static int __kmp_test_futex_lock_with_checks(kmp_futex_lock_t *lck,
429 char const *
const func =
"omp_test_lock";
430 if ((
sizeof(kmp_futex_lock_t) <= OMP_LOCK_T_SIZE) &&
431 __kmp_is_futex_lock_nestable(lck)) {
432 KMP_FATAL(LockNestableUsedAsSimple, func);
434 return __kmp_test_futex_lock(lck, gtid);
437 int __kmp_release_futex_lock(kmp_futex_lock_t *lck, kmp_int32 gtid) {
440 KA_TRACE(1000, (
"__kmp_release_futex_lock: lck:%p(0x%x), T#%d entering\n",
441 lck, lck->lk.poll, gtid));
443 KMP_FSYNC_RELEASING(lck);
444 ANNOTATE_FUTEX_RELEASED(lck);
446 kmp_int32 poll_val = KMP_XCHG_FIXED32(&(lck->lk.poll), KMP_LOCK_FREE(futex));
449 (
"__kmp_release_futex_lock: lck:%p, T#%d released poll_val = 0x%x\n",
450 lck, gtid, poll_val));
452 if (KMP_LOCK_STRIP(poll_val) & 1) {
454 (
"__kmp_release_futex_lock: lck:%p, T#%d futex_wake 1 thread\n",
456 syscall(__NR_futex, &(lck->lk.poll), FUTEX_WAKE, KMP_LOCK_BUSY(1, futex),
462 KA_TRACE(1000, (
"__kmp_release_futex_lock: lck:%p(0x%x), T#%d exiting\n", lck,
463 lck->lk.poll, gtid));
466 return KMP_LOCK_RELEASED;
469 static int __kmp_release_futex_lock_with_checks(kmp_futex_lock_t *lck,
471 char const *
const func =
"omp_unset_lock";
473 if ((
sizeof(kmp_futex_lock_t) <= OMP_LOCK_T_SIZE) &&
474 __kmp_is_futex_lock_nestable(lck)) {
475 KMP_FATAL(LockNestableUsedAsSimple, func);
477 if (__kmp_get_futex_lock_owner(lck) == -1) {
478 KMP_FATAL(LockUnsettingFree, func);
480 if ((gtid >= 0) && (__kmp_get_futex_lock_owner(lck) >= 0) &&
481 (__kmp_get_futex_lock_owner(lck) != gtid)) {
482 KMP_FATAL(LockUnsettingSetByAnother, func);
484 return __kmp_release_futex_lock(lck, gtid);
487 void __kmp_init_futex_lock(kmp_futex_lock_t *lck) {
488 TCW_4(lck->lk.poll, KMP_LOCK_FREE(futex));
491 void __kmp_destroy_futex_lock(kmp_futex_lock_t *lck) { lck->lk.poll = 0; }
493 static void __kmp_destroy_futex_lock_with_checks(kmp_futex_lock_t *lck) {
494 char const *
const func =
"omp_destroy_lock";
495 if ((
sizeof(kmp_futex_lock_t) <= OMP_LOCK_T_SIZE) &&
496 __kmp_is_futex_lock_nestable(lck)) {
497 KMP_FATAL(LockNestableUsedAsSimple, func);
499 if (__kmp_get_futex_lock_owner(lck) != -1) {
500 KMP_FATAL(LockStillOwned, func);
502 __kmp_destroy_futex_lock(lck);
507 int __kmp_acquire_nested_futex_lock(kmp_futex_lock_t *lck, kmp_int32 gtid) {
508 KMP_DEBUG_ASSERT(gtid >= 0);
510 if (__kmp_get_futex_lock_owner(lck) == gtid) {
511 lck->lk.depth_locked += 1;
512 return KMP_LOCK_ACQUIRED_NEXT;
514 __kmp_acquire_futex_lock_timed_template(lck, gtid);
515 ANNOTATE_FUTEX_ACQUIRED(lck);
516 lck->lk.depth_locked = 1;
517 return KMP_LOCK_ACQUIRED_FIRST;
521 static int __kmp_acquire_nested_futex_lock_with_checks(kmp_futex_lock_t *lck,
523 char const *
const func =
"omp_set_nest_lock";
524 if (!__kmp_is_futex_lock_nestable(lck)) {
525 KMP_FATAL(LockSimpleUsedAsNestable, func);
527 return __kmp_acquire_nested_futex_lock(lck, gtid);
530 int __kmp_test_nested_futex_lock(kmp_futex_lock_t *lck, kmp_int32 gtid) {
533 KMP_DEBUG_ASSERT(gtid >= 0);
535 if (__kmp_get_futex_lock_owner(lck) == gtid) {
536 retval = ++lck->lk.depth_locked;
537 }
else if (!__kmp_test_futex_lock(lck, gtid)) {
541 retval = lck->lk.depth_locked = 1;
546 static int __kmp_test_nested_futex_lock_with_checks(kmp_futex_lock_t *lck,
548 char const *
const func =
"omp_test_nest_lock";
549 if (!__kmp_is_futex_lock_nestable(lck)) {
550 KMP_FATAL(LockSimpleUsedAsNestable, func);
552 return __kmp_test_nested_futex_lock(lck, gtid);
555 int __kmp_release_nested_futex_lock(kmp_futex_lock_t *lck, kmp_int32 gtid) {
556 KMP_DEBUG_ASSERT(gtid >= 0);
559 if (--(lck->lk.depth_locked) == 0) {
560 __kmp_release_futex_lock(lck, gtid);
561 return KMP_LOCK_RELEASED;
563 return KMP_LOCK_STILL_HELD;
566 static int __kmp_release_nested_futex_lock_with_checks(kmp_futex_lock_t *lck,
568 char const *
const func =
"omp_unset_nest_lock";
570 if (!__kmp_is_futex_lock_nestable(lck)) {
571 KMP_FATAL(LockSimpleUsedAsNestable, func);
573 if (__kmp_get_futex_lock_owner(lck) == -1) {
574 KMP_FATAL(LockUnsettingFree, func);
576 if (__kmp_get_futex_lock_owner(lck) != gtid) {
577 KMP_FATAL(LockUnsettingSetByAnother, func);
579 return __kmp_release_nested_futex_lock(lck, gtid);
582 void __kmp_init_nested_futex_lock(kmp_futex_lock_t *lck) {
583 __kmp_init_futex_lock(lck);
584 lck->lk.depth_locked = 0;
587 void __kmp_destroy_nested_futex_lock(kmp_futex_lock_t *lck) {
588 __kmp_destroy_futex_lock(lck);
589 lck->lk.depth_locked = 0;
592 static void __kmp_destroy_nested_futex_lock_with_checks(kmp_futex_lock_t *lck) {
593 char const *
const func =
"omp_destroy_nest_lock";
594 if (!__kmp_is_futex_lock_nestable(lck)) {
595 KMP_FATAL(LockSimpleUsedAsNestable, func);
597 if (__kmp_get_futex_lock_owner(lck) != -1) {
598 KMP_FATAL(LockStillOwned, func);
600 __kmp_destroy_nested_futex_lock(lck);
603 #endif // KMP_USE_FUTEX 608 static kmp_int32 __kmp_get_ticket_lock_owner(kmp_ticket_lock_t *lck) {
609 return std::atomic_load_explicit(&lck->lk.owner_id,
610 std::memory_order_relaxed) -
614 static inline bool __kmp_is_ticket_lock_nestable(kmp_ticket_lock_t *lck) {
615 return std::atomic_load_explicit(&lck->lk.depth_locked,
616 std::memory_order_relaxed) != -1;
619 static kmp_uint32 __kmp_bakery_check(
void *now_serving, kmp_uint32 my_ticket) {
620 return std::atomic_load_explicit((std::atomic<unsigned> *)now_serving,
621 std::memory_order_acquire) == my_ticket;
624 __forceinline
static int 625 __kmp_acquire_ticket_lock_timed_template(kmp_ticket_lock_t *lck,
627 kmp_uint32 my_ticket = std::atomic_fetch_add_explicit(
628 &lck->lk.next_ticket, 1U, std::memory_order_relaxed);
630 #ifdef USE_LOCK_PROFILE 631 if (std::atomic_load_explicit(&lck->lk.now_serving,
632 std::memory_order_relaxed) != my_ticket)
633 __kmp_printf(
"LOCK CONTENTION: %p\n", lck);
637 if (std::atomic_load_explicit(&lck->lk.now_serving,
638 std::memory_order_acquire) == my_ticket) {
639 return KMP_LOCK_ACQUIRED_FIRST;
641 KMP_WAIT_PTR(&lck->lk.now_serving, my_ticket, __kmp_bakery_check, lck);
642 return KMP_LOCK_ACQUIRED_FIRST;
645 int __kmp_acquire_ticket_lock(kmp_ticket_lock_t *lck, kmp_int32 gtid) {
646 int retval = __kmp_acquire_ticket_lock_timed_template(lck, gtid);
647 ANNOTATE_TICKET_ACQUIRED(lck);
651 static int __kmp_acquire_ticket_lock_with_checks(kmp_ticket_lock_t *lck,
653 char const *
const func =
"omp_set_lock";
655 if (!std::atomic_load_explicit(&lck->lk.initialized,
656 std::memory_order_relaxed)) {
657 KMP_FATAL(LockIsUninitialized, func);
659 if (lck->lk.self != lck) {
660 KMP_FATAL(LockIsUninitialized, func);
662 if (__kmp_is_ticket_lock_nestable(lck)) {
663 KMP_FATAL(LockNestableUsedAsSimple, func);
665 if ((gtid >= 0) && (__kmp_get_ticket_lock_owner(lck) == gtid)) {
666 KMP_FATAL(LockIsAlreadyOwned, func);
669 __kmp_acquire_ticket_lock(lck, gtid);
671 std::atomic_store_explicit(&lck->lk.owner_id, gtid + 1,
672 std::memory_order_relaxed);
673 return KMP_LOCK_ACQUIRED_FIRST;
676 int __kmp_test_ticket_lock(kmp_ticket_lock_t *lck, kmp_int32 gtid) {
677 kmp_uint32 my_ticket = std::atomic_load_explicit(&lck->lk.next_ticket,
678 std::memory_order_relaxed);
680 if (std::atomic_load_explicit(&lck->lk.now_serving,
681 std::memory_order_relaxed) == my_ticket) {
682 kmp_uint32 next_ticket = my_ticket + 1;
683 if (std::atomic_compare_exchange_strong_explicit(
684 &lck->lk.next_ticket, &my_ticket, next_ticket,
685 std::memory_order_acquire, std::memory_order_acquire)) {
692 static int __kmp_test_ticket_lock_with_checks(kmp_ticket_lock_t *lck,
694 char const *
const func =
"omp_test_lock";
696 if (!std::atomic_load_explicit(&lck->lk.initialized,
697 std::memory_order_relaxed)) {
698 KMP_FATAL(LockIsUninitialized, func);
700 if (lck->lk.self != lck) {
701 KMP_FATAL(LockIsUninitialized, func);
703 if (__kmp_is_ticket_lock_nestable(lck)) {
704 KMP_FATAL(LockNestableUsedAsSimple, func);
707 int retval = __kmp_test_ticket_lock(lck, gtid);
710 std::atomic_store_explicit(&lck->lk.owner_id, gtid + 1,
711 std::memory_order_relaxed);
716 int __kmp_release_ticket_lock(kmp_ticket_lock_t *lck, kmp_int32 gtid) {
717 kmp_uint32 distance = std::atomic_load_explicit(&lck->lk.next_ticket,
718 std::memory_order_relaxed) -
719 std::atomic_load_explicit(&lck->lk.now_serving,
720 std::memory_order_relaxed);
722 ANNOTATE_TICKET_RELEASED(lck);
723 std::atomic_fetch_add_explicit(&lck->lk.now_serving, 1U,
724 std::memory_order_release);
727 (kmp_uint32)(__kmp_avail_proc ? __kmp_avail_proc : __kmp_xproc));
728 return KMP_LOCK_RELEASED;
731 static int __kmp_release_ticket_lock_with_checks(kmp_ticket_lock_t *lck,
733 char const *
const func =
"omp_unset_lock";
735 if (!std::atomic_load_explicit(&lck->lk.initialized,
736 std::memory_order_relaxed)) {
737 KMP_FATAL(LockIsUninitialized, func);
739 if (lck->lk.self != lck) {
740 KMP_FATAL(LockIsUninitialized, func);
742 if (__kmp_is_ticket_lock_nestable(lck)) {
743 KMP_FATAL(LockNestableUsedAsSimple, func);
745 if (__kmp_get_ticket_lock_owner(lck) == -1) {
746 KMP_FATAL(LockUnsettingFree, func);
748 if ((gtid >= 0) && (__kmp_get_ticket_lock_owner(lck) >= 0) &&
749 (__kmp_get_ticket_lock_owner(lck) != gtid)) {
750 KMP_FATAL(LockUnsettingSetByAnother, func);
752 std::atomic_store_explicit(&lck->lk.owner_id, 0, std::memory_order_relaxed);
753 return __kmp_release_ticket_lock(lck, gtid);
756 void __kmp_init_ticket_lock(kmp_ticket_lock_t *lck) {
757 lck->lk.location = NULL;
759 std::atomic_store_explicit(&lck->lk.next_ticket, 0U,
760 std::memory_order_relaxed);
761 std::atomic_store_explicit(&lck->lk.now_serving, 0U,
762 std::memory_order_relaxed);
763 std::atomic_store_explicit(
764 &lck->lk.owner_id, 0,
765 std::memory_order_relaxed);
766 std::atomic_store_explicit(
767 &lck->lk.depth_locked, -1,
768 std::memory_order_relaxed);
769 std::atomic_store_explicit(&lck->lk.initialized,
true,
770 std::memory_order_release);
773 void __kmp_destroy_ticket_lock(kmp_ticket_lock_t *lck) {
774 std::atomic_store_explicit(&lck->lk.initialized,
false,
775 std::memory_order_release);
777 lck->lk.location = NULL;
778 std::atomic_store_explicit(&lck->lk.next_ticket, 0U,
779 std::memory_order_relaxed);
780 std::atomic_store_explicit(&lck->lk.now_serving, 0U,
781 std::memory_order_relaxed);
782 std::atomic_store_explicit(&lck->lk.owner_id, 0, std::memory_order_relaxed);
783 std::atomic_store_explicit(&lck->lk.depth_locked, -1,
784 std::memory_order_relaxed);
787 static void __kmp_destroy_ticket_lock_with_checks(kmp_ticket_lock_t *lck) {
788 char const *
const func =
"omp_destroy_lock";
790 if (!std::atomic_load_explicit(&lck->lk.initialized,
791 std::memory_order_relaxed)) {
792 KMP_FATAL(LockIsUninitialized, func);
794 if (lck->lk.self != lck) {
795 KMP_FATAL(LockIsUninitialized, func);
797 if (__kmp_is_ticket_lock_nestable(lck)) {
798 KMP_FATAL(LockNestableUsedAsSimple, func);
800 if (__kmp_get_ticket_lock_owner(lck) != -1) {
801 KMP_FATAL(LockStillOwned, func);
803 __kmp_destroy_ticket_lock(lck);
808 int __kmp_acquire_nested_ticket_lock(kmp_ticket_lock_t *lck, kmp_int32 gtid) {
809 KMP_DEBUG_ASSERT(gtid >= 0);
811 if (__kmp_get_ticket_lock_owner(lck) == gtid) {
812 std::atomic_fetch_add_explicit(&lck->lk.depth_locked, 1,
813 std::memory_order_relaxed);
814 return KMP_LOCK_ACQUIRED_NEXT;
816 __kmp_acquire_ticket_lock_timed_template(lck, gtid);
817 ANNOTATE_TICKET_ACQUIRED(lck);
818 std::atomic_store_explicit(&lck->lk.depth_locked, 1,
819 std::memory_order_relaxed);
820 std::atomic_store_explicit(&lck->lk.owner_id, gtid + 1,
821 std::memory_order_relaxed);
822 return KMP_LOCK_ACQUIRED_FIRST;
826 static int __kmp_acquire_nested_ticket_lock_with_checks(kmp_ticket_lock_t *lck,
828 char const *
const func =
"omp_set_nest_lock";
830 if (!std::atomic_load_explicit(&lck->lk.initialized,
831 std::memory_order_relaxed)) {
832 KMP_FATAL(LockIsUninitialized, func);
834 if (lck->lk.self != lck) {
835 KMP_FATAL(LockIsUninitialized, func);
837 if (!__kmp_is_ticket_lock_nestable(lck)) {
838 KMP_FATAL(LockSimpleUsedAsNestable, func);
840 return __kmp_acquire_nested_ticket_lock(lck, gtid);
843 int __kmp_test_nested_ticket_lock(kmp_ticket_lock_t *lck, kmp_int32 gtid) {
846 KMP_DEBUG_ASSERT(gtid >= 0);
848 if (__kmp_get_ticket_lock_owner(lck) == gtid) {
849 retval = std::atomic_fetch_add_explicit(&lck->lk.depth_locked, 1,
850 std::memory_order_relaxed) +
852 }
else if (!__kmp_test_ticket_lock(lck, gtid)) {
855 std::atomic_store_explicit(&lck->lk.depth_locked, 1,
856 std::memory_order_relaxed);
857 std::atomic_store_explicit(&lck->lk.owner_id, gtid + 1,
858 std::memory_order_relaxed);
864 static int __kmp_test_nested_ticket_lock_with_checks(kmp_ticket_lock_t *lck,
866 char const *
const func =
"omp_test_nest_lock";
868 if (!std::atomic_load_explicit(&lck->lk.initialized,
869 std::memory_order_relaxed)) {
870 KMP_FATAL(LockIsUninitialized, func);
872 if (lck->lk.self != lck) {
873 KMP_FATAL(LockIsUninitialized, func);
875 if (!__kmp_is_ticket_lock_nestable(lck)) {
876 KMP_FATAL(LockSimpleUsedAsNestable, func);
878 return __kmp_test_nested_ticket_lock(lck, gtid);
881 int __kmp_release_nested_ticket_lock(kmp_ticket_lock_t *lck, kmp_int32 gtid) {
882 KMP_DEBUG_ASSERT(gtid >= 0);
884 if ((std::atomic_fetch_add_explicit(&lck->lk.depth_locked, -1,
885 std::memory_order_relaxed) -
887 std::atomic_store_explicit(&lck->lk.owner_id, 0, std::memory_order_relaxed);
888 __kmp_release_ticket_lock(lck, gtid);
889 return KMP_LOCK_RELEASED;
891 return KMP_LOCK_STILL_HELD;
894 static int __kmp_release_nested_ticket_lock_with_checks(kmp_ticket_lock_t *lck,
896 char const *
const func =
"omp_unset_nest_lock";
898 if (!std::atomic_load_explicit(&lck->lk.initialized,
899 std::memory_order_relaxed)) {
900 KMP_FATAL(LockIsUninitialized, func);
902 if (lck->lk.self != lck) {
903 KMP_FATAL(LockIsUninitialized, func);
905 if (!__kmp_is_ticket_lock_nestable(lck)) {
906 KMP_FATAL(LockSimpleUsedAsNestable, func);
908 if (__kmp_get_ticket_lock_owner(lck) == -1) {
909 KMP_FATAL(LockUnsettingFree, func);
911 if (__kmp_get_ticket_lock_owner(lck) != gtid) {
912 KMP_FATAL(LockUnsettingSetByAnother, func);
914 return __kmp_release_nested_ticket_lock(lck, gtid);
917 void __kmp_init_nested_ticket_lock(kmp_ticket_lock_t *lck) {
918 __kmp_init_ticket_lock(lck);
919 std::atomic_store_explicit(&lck->lk.depth_locked, 0,
920 std::memory_order_relaxed);
924 void __kmp_destroy_nested_ticket_lock(kmp_ticket_lock_t *lck) {
925 __kmp_destroy_ticket_lock(lck);
926 std::atomic_store_explicit(&lck->lk.depth_locked, 0,
927 std::memory_order_relaxed);
931 __kmp_destroy_nested_ticket_lock_with_checks(kmp_ticket_lock_t *lck) {
932 char const *
const func =
"omp_destroy_nest_lock";
934 if (!std::atomic_load_explicit(&lck->lk.initialized,
935 std::memory_order_relaxed)) {
936 KMP_FATAL(LockIsUninitialized, func);
938 if (lck->lk.self != lck) {
939 KMP_FATAL(LockIsUninitialized, func);
941 if (!__kmp_is_ticket_lock_nestable(lck)) {
942 KMP_FATAL(LockSimpleUsedAsNestable, func);
944 if (__kmp_get_ticket_lock_owner(lck) != -1) {
945 KMP_FATAL(LockStillOwned, func);
947 __kmp_destroy_nested_ticket_lock(lck);
952 static const ident_t *__kmp_get_ticket_lock_location(kmp_ticket_lock_t *lck) {
953 return lck->lk.location;
956 static void __kmp_set_ticket_lock_location(kmp_ticket_lock_t *lck,
958 lck->lk.location = loc;
961 static kmp_lock_flags_t __kmp_get_ticket_lock_flags(kmp_ticket_lock_t *lck) {
962 return lck->lk.flags;
965 static void __kmp_set_ticket_lock_flags(kmp_ticket_lock_t *lck,
966 kmp_lock_flags_t flags) {
967 lck->lk.flags = flags;
1025 #ifdef DEBUG_QUEUING_LOCKS 1028 #define TRACE_BUF_ELE 1024 1029 static char traces[TRACE_BUF_ELE][128] = {0};
1031 #define TRACE_LOCK(X, Y) \ 1032 KMP_SNPRINTF(traces[tc++ % TRACE_BUF_ELE], 128, "t%d at %s\n", X, Y); 1033 #define TRACE_LOCK_T(X, Y, Z) \ 1034 KMP_SNPRINTF(traces[tc++ % TRACE_BUF_ELE], 128, "t%d at %s%d\n", X, Y, Z); 1035 #define TRACE_LOCK_HT(X, Y, Z, Q) \ 1036 KMP_SNPRINTF(traces[tc++ % TRACE_BUF_ELE], 128, "t%d at %s %d,%d\n", X, Y, \ 1039 static void __kmp_dump_queuing_lock(kmp_info_t *this_thr, kmp_int32 gtid,
1040 kmp_queuing_lock_t *lck, kmp_int32 head_id,
1041 kmp_int32 tail_id) {
1044 __kmp_printf_no_lock(
"\n__kmp_dump_queuing_lock: TRACE BEGINS HERE! \n");
1046 i = tc % TRACE_BUF_ELE;
1047 __kmp_printf_no_lock(
"%s\n", traces[i]);
1048 i = (i + 1) % TRACE_BUF_ELE;
1049 while (i != (tc % TRACE_BUF_ELE)) {
1050 __kmp_printf_no_lock(
"%s", traces[i]);
1051 i = (i + 1) % TRACE_BUF_ELE;
1053 __kmp_printf_no_lock(
"\n");
1055 __kmp_printf_no_lock(
"\n__kmp_dump_queuing_lock: gtid+1:%d, spin_here:%d, " 1056 "next_wait:%d, head_id:%d, tail_id:%d\n",
1057 gtid + 1, this_thr->th.th_spin_here,
1058 this_thr->th.th_next_waiting, head_id, tail_id);
1060 __kmp_printf_no_lock(
"\t\thead: %d ", lck->lk.head_id);
1062 if (lck->lk.head_id >= 1) {
1063 t = __kmp_threads[lck->lk.head_id - 1]->th.th_next_waiting;
1065 __kmp_printf_no_lock(
"-> %d ", t);
1066 t = __kmp_threads[t - 1]->th.th_next_waiting;
1069 __kmp_printf_no_lock(
"; tail: %d ", lck->lk.tail_id);
1070 __kmp_printf_no_lock(
"\n\n");
1075 static kmp_int32 __kmp_get_queuing_lock_owner(kmp_queuing_lock_t *lck) {
1076 return TCR_4(lck->lk.owner_id) - 1;
1079 static inline bool __kmp_is_queuing_lock_nestable(kmp_queuing_lock_t *lck) {
1080 return lck->lk.depth_locked != -1;
1084 template <
bool takeTime>
1087 __forceinline
static int 1088 __kmp_acquire_queuing_lock_timed_template(kmp_queuing_lock_t *lck,
1090 kmp_info_t *this_thr = __kmp_thread_from_gtid(gtid);
1091 volatile kmp_int32 *head_id_p = &lck->lk.head_id;
1092 volatile kmp_int32 *tail_id_p = &lck->lk.tail_id;
1093 volatile kmp_uint32 *spin_here_p;
1094 kmp_int32 need_mf = 1;
1097 ompt_state_t prev_state = ompt_state_undefined;
1101 (
"__kmp_acquire_queuing_lock: lck:%p, T#%d entering\n", lck, gtid));
1103 KMP_FSYNC_PREPARE(lck);
1104 KMP_DEBUG_ASSERT(this_thr != NULL);
1105 spin_here_p = &this_thr->th.th_spin_here;
1107 #ifdef DEBUG_QUEUING_LOCKS 1108 TRACE_LOCK(gtid + 1,
"acq ent");
1110 __kmp_dump_queuing_lock(this_thr, gtid, lck, *head_id_p, *tail_id_p);
1111 if (this_thr->th.th_next_waiting != 0)
1112 __kmp_dump_queuing_lock(this_thr, gtid, lck, *head_id_p, *tail_id_p);
1114 KMP_DEBUG_ASSERT(!*spin_here_p);
1115 KMP_DEBUG_ASSERT(this_thr->th.th_next_waiting == 0);
1124 *spin_here_p = TRUE;
1136 #ifdef DEBUG_QUEUING_LOCKS 1138 TRACE_LOCK_HT(gtid + 1,
"acq read: ", head, tail);
1147 enqueued = KMP_COMPARE_AND_STORE_ACQ64((
volatile kmp_int64 *)tail_id_p,
1149 KMP_PACK_64(gtid + 1, gtid + 1));
1150 #ifdef DEBUG_QUEUING_LOCKS 1152 TRACE_LOCK(gtid + 1,
"acq enq: (-1,0)->(tid,tid)");
1158 KMP_DEBUG_ASSERT(tail != gtid + 1);
1160 #ifdef DEBUG_QUEUING_LOCKS 1161 TRACE_LOCK_HT(gtid + 1,
"acq read: ", head, tail);
1169 enqueued = KMP_COMPARE_AND_STORE_ACQ32(tail_id_p, tail, gtid + 1);
1171 #ifdef DEBUG_QUEUING_LOCKS 1173 TRACE_LOCK(gtid + 1,
"acq enq: (h,t)->(h,tid)");
1180 kmp_int32 grabbed_lock;
1182 #ifdef DEBUG_QUEUING_LOCKS 1184 TRACE_LOCK_HT(gtid + 1,
"acq read: ", head, tail);
1190 grabbed_lock = KMP_COMPARE_AND_STORE_ACQ32(head_id_p, 0, -1);
1194 *spin_here_p = FALSE;
1198 (
"__kmp_acquire_queuing_lock: lck:%p, T#%d exiting: no queuing\n",
1200 #ifdef DEBUG_QUEUING_LOCKS 1201 TRACE_LOCK_HT(gtid + 1,
"acq exit: ", head, 0);
1205 if (ompt_enabled.enabled && prev_state != ompt_state_undefined) {
1207 this_thr->th.ompt_thread_info.state = prev_state;
1208 this_thr->th.ompt_thread_info.wait_id = 0;
1212 KMP_FSYNC_ACQUIRED(lck);
1213 return KMP_LOCK_ACQUIRED_FIRST;
1220 if (ompt_enabled.enabled && prev_state == ompt_state_undefined) {
1222 prev_state = this_thr->th.ompt_thread_info.state;
1223 this_thr->th.ompt_thread_info.wait_id = (uint64_t)lck;
1224 this_thr->th.ompt_thread_info.state = ompt_state_wait_lock;
1230 kmp_info_t *tail_thr = __kmp_thread_from_gtid(tail - 1);
1231 KMP_ASSERT(tail_thr != NULL);
1232 tail_thr->th.th_next_waiting = gtid + 1;
1236 (
"__kmp_acquire_queuing_lock: lck:%p, T#%d waiting for lock\n",
1241 KMP_WAIT(spin_here_p, FALSE, KMP_EQ, lck);
1243 #ifdef DEBUG_QUEUING_LOCKS 1244 TRACE_LOCK(gtid + 1,
"acq spin");
1246 if (this_thr->th.th_next_waiting != 0)
1247 __kmp_dump_queuing_lock(this_thr, gtid, lck, *head_id_p, *tail_id_p);
1249 KMP_DEBUG_ASSERT(this_thr->th.th_next_waiting == 0);
1250 KA_TRACE(1000, (
"__kmp_acquire_queuing_lock: lck:%p, T#%d exiting: after " 1251 "waiting on queue\n",
1254 #ifdef DEBUG_QUEUING_LOCKS 1255 TRACE_LOCK(gtid + 1,
"acq exit 2");
1260 this_thr->th.ompt_thread_info.state = prev_state;
1261 this_thr->th.ompt_thread_info.wait_id = 0;
1265 return KMP_LOCK_ACQUIRED_FIRST;
1271 KMP_YIELD_OVERSUB();
1273 #ifdef DEBUG_QUEUING_LOCKS 1274 TRACE_LOCK(gtid + 1,
"acq retry");
1277 KMP_ASSERT2(0,
"should not get here");
1278 return KMP_LOCK_ACQUIRED_FIRST;
1281 int __kmp_acquire_queuing_lock(kmp_queuing_lock_t *lck, kmp_int32 gtid) {
1282 KMP_DEBUG_ASSERT(gtid >= 0);
1284 int retval = __kmp_acquire_queuing_lock_timed_template<false>(lck, gtid);
1285 ANNOTATE_QUEUING_ACQUIRED(lck);
1289 static int __kmp_acquire_queuing_lock_with_checks(kmp_queuing_lock_t *lck,
1291 char const *
const func =
"omp_set_lock";
1292 if (lck->lk.initialized != lck) {
1293 KMP_FATAL(LockIsUninitialized, func);
1295 if (__kmp_is_queuing_lock_nestable(lck)) {
1296 KMP_FATAL(LockNestableUsedAsSimple, func);
1298 if (__kmp_get_queuing_lock_owner(lck) == gtid) {
1299 KMP_FATAL(LockIsAlreadyOwned, func);
1302 __kmp_acquire_queuing_lock(lck, gtid);
1304 lck->lk.owner_id = gtid + 1;
1305 return KMP_LOCK_ACQUIRED_FIRST;
1308 int __kmp_test_queuing_lock(kmp_queuing_lock_t *lck, kmp_int32 gtid) {
1309 volatile kmp_int32 *head_id_p = &lck->lk.head_id;
1312 kmp_info_t *this_thr;
1315 KA_TRACE(1000, (
"__kmp_test_queuing_lock: T#%d entering\n", gtid));
1316 KMP_DEBUG_ASSERT(gtid >= 0);
1318 this_thr = __kmp_thread_from_gtid(gtid);
1319 KMP_DEBUG_ASSERT(this_thr != NULL);
1320 KMP_DEBUG_ASSERT(!this_thr->th.th_spin_here);
1327 if (KMP_COMPARE_AND_STORE_ACQ32(head_id_p, 0, -1)) {
1329 (
"__kmp_test_queuing_lock: T#%d exiting: holding lock\n", gtid));
1330 KMP_FSYNC_ACQUIRED(lck);
1331 ANNOTATE_QUEUING_ACQUIRED(lck);
1337 (
"__kmp_test_queuing_lock: T#%d exiting: without lock\n", gtid));
1341 static int __kmp_test_queuing_lock_with_checks(kmp_queuing_lock_t *lck,
1343 char const *
const func =
"omp_test_lock";
1344 if (lck->lk.initialized != lck) {
1345 KMP_FATAL(LockIsUninitialized, func);
1347 if (__kmp_is_queuing_lock_nestable(lck)) {
1348 KMP_FATAL(LockNestableUsedAsSimple, func);
1351 int retval = __kmp_test_queuing_lock(lck, gtid);
1354 lck->lk.owner_id = gtid + 1;
1359 int __kmp_release_queuing_lock(kmp_queuing_lock_t *lck, kmp_int32 gtid) {
1360 kmp_info_t *this_thr;
1361 volatile kmp_int32 *head_id_p = &lck->lk.head_id;
1362 volatile kmp_int32 *tail_id_p = &lck->lk.tail_id;
1365 (
"__kmp_release_queuing_lock: lck:%p, T#%d entering\n", lck, gtid));
1366 KMP_DEBUG_ASSERT(gtid >= 0);
1367 this_thr = __kmp_thread_from_gtid(gtid);
1368 KMP_DEBUG_ASSERT(this_thr != NULL);
1369 #ifdef DEBUG_QUEUING_LOCKS 1370 TRACE_LOCK(gtid + 1,
"rel ent");
1372 if (this_thr->th.th_spin_here)
1373 __kmp_dump_queuing_lock(this_thr, gtid, lck, *head_id_p, *tail_id_p);
1374 if (this_thr->th.th_next_waiting != 0)
1375 __kmp_dump_queuing_lock(this_thr, gtid, lck, *head_id_p, *tail_id_p);
1377 KMP_DEBUG_ASSERT(!this_thr->th.th_spin_here);
1378 KMP_DEBUG_ASSERT(this_thr->th.th_next_waiting == 0);
1380 KMP_FSYNC_RELEASING(lck);
1381 ANNOTATE_QUEUING_RELEASED(lck);
1390 #ifdef DEBUG_QUEUING_LOCKS 1392 TRACE_LOCK_HT(gtid + 1,
"rel read: ", head, tail);
1394 __kmp_dump_queuing_lock(this_thr, gtid, lck, head, tail);
1396 KMP_DEBUG_ASSERT(head !=
1401 if (KMP_COMPARE_AND_STORE_REL32(head_id_p, -1, 0)) {
1404 (
"__kmp_release_queuing_lock: lck:%p, T#%d exiting: queue empty\n",
1406 #ifdef DEBUG_QUEUING_LOCKS 1407 TRACE_LOCK_HT(gtid + 1,
"rel exit: ", 0, 0);
1413 return KMP_LOCK_RELEASED;
1420 #ifdef DEBUG_QUEUING_LOCKS 1422 __kmp_dump_queuing_lock(this_thr, gtid, lck, head, tail);
1424 KMP_DEBUG_ASSERT(head > 0);
1427 dequeued = KMP_COMPARE_AND_STORE_REL64(
1428 RCAST(
volatile kmp_int64 *, tail_id_p), KMP_PACK_64(head, head),
1429 KMP_PACK_64(-1, 0));
1430 #ifdef DEBUG_QUEUING_LOCKS 1431 TRACE_LOCK(gtid + 1,
"rel deq: (h,h)->(-1,0)");
1435 volatile kmp_int32 *waiting_id_p;
1436 kmp_info_t *head_thr = __kmp_thread_from_gtid(head - 1);
1437 KMP_DEBUG_ASSERT(head_thr != NULL);
1438 waiting_id_p = &head_thr->th.th_next_waiting;
1441 #ifdef DEBUG_QUEUING_LOCKS 1442 if (head <= 0 || tail <= 0)
1443 __kmp_dump_queuing_lock(this_thr, gtid, lck, head, tail);
1445 KMP_DEBUG_ASSERT(head > 0 && tail > 0);
1452 KMP_WAIT((
volatile kmp_uint32 *)waiting_id_p, 0, KMP_NEQ, NULL);
1453 #ifdef DEBUG_QUEUING_LOCKS 1454 TRACE_LOCK(gtid + 1,
"rel deq: (h,t)->(h',t)");
1461 kmp_info_t *head_thr = __kmp_thread_from_gtid(head - 1);
1462 KMP_DEBUG_ASSERT(head_thr != NULL);
1465 #ifdef DEBUG_QUEUING_LOCKS 1466 if (head <= 0 || tail <= 0)
1467 __kmp_dump_queuing_lock(this_thr, gtid, lck, head, tail);
1469 KMP_DEBUG_ASSERT(head > 0 && tail > 0);
1473 head_thr->th.th_next_waiting = 0;
1474 #ifdef DEBUG_QUEUING_LOCKS 1475 TRACE_LOCK_T(gtid + 1,
"rel nw=0 for t=", head);
1480 head_thr->th.th_spin_here = FALSE;
1482 KA_TRACE(1000, (
"__kmp_release_queuing_lock: lck:%p, T#%d exiting: after " 1485 #ifdef DEBUG_QUEUING_LOCKS 1486 TRACE_LOCK(gtid + 1,
"rel exit 2");
1488 return KMP_LOCK_RELEASED;
1493 #ifdef DEBUG_QUEUING_LOCKS 1494 TRACE_LOCK(gtid + 1,
"rel retry");
1498 KMP_ASSERT2(0,
"should not get here");
1499 return KMP_LOCK_RELEASED;
1502 static int __kmp_release_queuing_lock_with_checks(kmp_queuing_lock_t *lck,
1504 char const *
const func =
"omp_unset_lock";
1506 if (lck->lk.initialized != lck) {
1507 KMP_FATAL(LockIsUninitialized, func);
1509 if (__kmp_is_queuing_lock_nestable(lck)) {
1510 KMP_FATAL(LockNestableUsedAsSimple, func);
1512 if (__kmp_get_queuing_lock_owner(lck) == -1) {
1513 KMP_FATAL(LockUnsettingFree, func);
1515 if (__kmp_get_queuing_lock_owner(lck) != gtid) {
1516 KMP_FATAL(LockUnsettingSetByAnother, func);
1518 lck->lk.owner_id = 0;
1519 return __kmp_release_queuing_lock(lck, gtid);
1522 void __kmp_init_queuing_lock(kmp_queuing_lock_t *lck) {
1523 lck->lk.location = NULL;
1524 lck->lk.head_id = 0;
1525 lck->lk.tail_id = 0;
1526 lck->lk.next_ticket = 0;
1527 lck->lk.now_serving = 0;
1528 lck->lk.owner_id = 0;
1529 lck->lk.depth_locked = -1;
1530 lck->lk.initialized = lck;
1532 KA_TRACE(1000, (
"__kmp_init_queuing_lock: lock %p initialized\n", lck));
1535 void __kmp_destroy_queuing_lock(kmp_queuing_lock_t *lck) {
1536 lck->lk.initialized = NULL;
1537 lck->lk.location = NULL;
1538 lck->lk.head_id = 0;
1539 lck->lk.tail_id = 0;
1540 lck->lk.next_ticket = 0;
1541 lck->lk.now_serving = 0;
1542 lck->lk.owner_id = 0;
1543 lck->lk.depth_locked = -1;
1546 static void __kmp_destroy_queuing_lock_with_checks(kmp_queuing_lock_t *lck) {
1547 char const *
const func =
"omp_destroy_lock";
1548 if (lck->lk.initialized != lck) {
1549 KMP_FATAL(LockIsUninitialized, func);
1551 if (__kmp_is_queuing_lock_nestable(lck)) {
1552 KMP_FATAL(LockNestableUsedAsSimple, func);
1554 if (__kmp_get_queuing_lock_owner(lck) != -1) {
1555 KMP_FATAL(LockStillOwned, func);
1557 __kmp_destroy_queuing_lock(lck);
1562 int __kmp_acquire_nested_queuing_lock(kmp_queuing_lock_t *lck, kmp_int32 gtid) {
1563 KMP_DEBUG_ASSERT(gtid >= 0);
1565 if (__kmp_get_queuing_lock_owner(lck) == gtid) {
1566 lck->lk.depth_locked += 1;
1567 return KMP_LOCK_ACQUIRED_NEXT;
1569 __kmp_acquire_queuing_lock_timed_template<false>(lck, gtid);
1570 ANNOTATE_QUEUING_ACQUIRED(lck);
1572 lck->lk.depth_locked = 1;
1574 lck->lk.owner_id = gtid + 1;
1575 return KMP_LOCK_ACQUIRED_FIRST;
1580 __kmp_acquire_nested_queuing_lock_with_checks(kmp_queuing_lock_t *lck,
1582 char const *
const func =
"omp_set_nest_lock";
1583 if (lck->lk.initialized != lck) {
1584 KMP_FATAL(LockIsUninitialized, func);
1586 if (!__kmp_is_queuing_lock_nestable(lck)) {
1587 KMP_FATAL(LockSimpleUsedAsNestable, func);
1589 return __kmp_acquire_nested_queuing_lock(lck, gtid);
1592 int __kmp_test_nested_queuing_lock(kmp_queuing_lock_t *lck, kmp_int32 gtid) {
1595 KMP_DEBUG_ASSERT(gtid >= 0);
1597 if (__kmp_get_queuing_lock_owner(lck) == gtid) {
1598 retval = ++lck->lk.depth_locked;
1599 }
else if (!__kmp_test_queuing_lock(lck, gtid)) {
1603 retval = lck->lk.depth_locked = 1;
1605 lck->lk.owner_id = gtid + 1;
1610 static int __kmp_test_nested_queuing_lock_with_checks(kmp_queuing_lock_t *lck,
1612 char const *
const func =
"omp_test_nest_lock";
1613 if (lck->lk.initialized != lck) {
1614 KMP_FATAL(LockIsUninitialized, func);
1616 if (!__kmp_is_queuing_lock_nestable(lck)) {
1617 KMP_FATAL(LockSimpleUsedAsNestable, func);
1619 return __kmp_test_nested_queuing_lock(lck, gtid);
1622 int __kmp_release_nested_queuing_lock(kmp_queuing_lock_t *lck, kmp_int32 gtid) {
1623 KMP_DEBUG_ASSERT(gtid >= 0);
1626 if (--(lck->lk.depth_locked) == 0) {
1628 lck->lk.owner_id = 0;
1629 __kmp_release_queuing_lock(lck, gtid);
1630 return KMP_LOCK_RELEASED;
1632 return KMP_LOCK_STILL_HELD;
1636 __kmp_release_nested_queuing_lock_with_checks(kmp_queuing_lock_t *lck,
1638 char const *
const func =
"omp_unset_nest_lock";
1640 if (lck->lk.initialized != lck) {
1641 KMP_FATAL(LockIsUninitialized, func);
1643 if (!__kmp_is_queuing_lock_nestable(lck)) {
1644 KMP_FATAL(LockSimpleUsedAsNestable, func);
1646 if (__kmp_get_queuing_lock_owner(lck) == -1) {
1647 KMP_FATAL(LockUnsettingFree, func);
1649 if (__kmp_get_queuing_lock_owner(lck) != gtid) {
1650 KMP_FATAL(LockUnsettingSetByAnother, func);
1652 return __kmp_release_nested_queuing_lock(lck, gtid);
1655 void __kmp_init_nested_queuing_lock(kmp_queuing_lock_t *lck) {
1656 __kmp_init_queuing_lock(lck);
1657 lck->lk.depth_locked = 0;
1660 void __kmp_destroy_nested_queuing_lock(kmp_queuing_lock_t *lck) {
1661 __kmp_destroy_queuing_lock(lck);
1662 lck->lk.depth_locked = 0;
1666 __kmp_destroy_nested_queuing_lock_with_checks(kmp_queuing_lock_t *lck) {
1667 char const *
const func =
"omp_destroy_nest_lock";
1668 if (lck->lk.initialized != lck) {
1669 KMP_FATAL(LockIsUninitialized, func);
1671 if (!__kmp_is_queuing_lock_nestable(lck)) {
1672 KMP_FATAL(LockSimpleUsedAsNestable, func);
1674 if (__kmp_get_queuing_lock_owner(lck) != -1) {
1675 KMP_FATAL(LockStillOwned, func);
1677 __kmp_destroy_nested_queuing_lock(lck);
1682 static const ident_t *__kmp_get_queuing_lock_location(kmp_queuing_lock_t *lck) {
1683 return lck->lk.location;
1686 static void __kmp_set_queuing_lock_location(kmp_queuing_lock_t *lck,
1688 lck->lk.location = loc;
1691 static kmp_lock_flags_t __kmp_get_queuing_lock_flags(kmp_queuing_lock_t *lck) {
1692 return lck->lk.flags;
1695 static void __kmp_set_queuing_lock_flags(kmp_queuing_lock_t *lck,
1696 kmp_lock_flags_t flags) {
1697 lck->lk.flags = flags;
1700 #if KMP_USE_ADAPTIVE_LOCKS 1704 #if (KMP_COMPILER_ICC && __INTEL_COMPILER >= 1300) || \ 1705 (KMP_COMPILER_MSVC && _MSC_VER >= 1700) || \ 1706 (KMP_COMPILER_CLANG && KMP_MSVC_COMPAT) 1708 #include <immintrin.h> 1709 #define SOFT_ABORT_MASK (_XABORT_RETRY | _XABORT_CONFLICT | _XABORT_EXPLICIT) 1714 #define _XBEGIN_STARTED (~0u) 1715 #define _XABORT_EXPLICIT (1 << 0) 1716 #define _XABORT_RETRY (1 << 1) 1717 #define _XABORT_CONFLICT (1 << 2) 1718 #define _XABORT_CAPACITY (1 << 3) 1719 #define _XABORT_DEBUG (1 << 4) 1720 #define _XABORT_NESTED (1 << 5) 1721 #define _XABORT_CODE(x) ((unsigned char)(((x) >> 24) & 0xFF)) 1724 #define SOFT_ABORT_MASK (_XABORT_RETRY | _XABORT_CONFLICT | _XABORT_EXPLICIT) 1726 #define STRINGIZE_INTERNAL(arg) #arg 1727 #define STRINGIZE(arg) STRINGIZE_INTERNAL(arg) 1733 static __inline
int _xbegin() {
1761 #endif // KMP_ARCH_X86_64 1770 __asm__
volatile(
"1: .byte 0xC7; .byte 0xF8;\n" 1773 "1: movl %%eax,%0\n" 1775 :
"+r"(res)::
"memory",
"%eax");
1776 #endif // KMP_OS_WINDOWS 1781 static __inline
void _xend() {
1789 __asm__
volatile(
".byte 0x0f; .byte 0x01; .byte 0xd5" :::
"memory");
1798 #define _xabort(ARG) _asm _emit 0xc6 _asm _emit 0xf8 _asm _emit ARG 1800 #define _xabort(ARG) \ 1801 __asm__ volatile(".byte 0xC6; .byte 0xF8; .byte " STRINGIZE(ARG):::"memory"); 1804 #endif // KMP_COMPILER_ICC && __INTEL_COMPILER >= 1300 1807 #if KMP_DEBUG_ADAPTIVE_LOCKS 1812 static kmp_adaptive_lock_statistics_t destroyedStats;
1815 static kmp_adaptive_lock_info_t liveLocks;
1818 static kmp_bootstrap_lock_t chain_lock =
1819 KMP_BOOTSTRAP_LOCK_INITIALIZER(chain_lock);
1822 void __kmp_init_speculative_stats() {
1823 kmp_adaptive_lock_info_t *lck = &liveLocks;
1825 memset(CCAST(kmp_adaptive_lock_statistics_t *, &(lck->stats)), 0,
1826 sizeof(lck->stats));
1827 lck->stats.next = lck;
1828 lck->stats.prev = lck;
1830 KMP_ASSERT(lck->stats.next->stats.prev == lck);
1831 KMP_ASSERT(lck->stats.prev->stats.next == lck);
1833 __kmp_init_bootstrap_lock(&chain_lock);
1837 static void __kmp_remember_lock(kmp_adaptive_lock_info_t *lck) {
1838 __kmp_acquire_bootstrap_lock(&chain_lock);
1840 lck->stats.next = liveLocks.stats.next;
1841 lck->stats.prev = &liveLocks;
1843 liveLocks.stats.next = lck;
1844 lck->stats.next->stats.prev = lck;
1846 KMP_ASSERT(lck->stats.next->stats.prev == lck);
1847 KMP_ASSERT(lck->stats.prev->stats.next == lck);
1849 __kmp_release_bootstrap_lock(&chain_lock);
1852 static void __kmp_forget_lock(kmp_adaptive_lock_info_t *lck) {
1853 KMP_ASSERT(lck->stats.next->stats.prev == lck);
1854 KMP_ASSERT(lck->stats.prev->stats.next == lck);
1856 kmp_adaptive_lock_info_t *n = lck->stats.next;
1857 kmp_adaptive_lock_info_t *p = lck->stats.prev;
1863 static void __kmp_zero_speculative_stats(kmp_adaptive_lock_info_t *lck) {
1864 memset(CCAST(kmp_adaptive_lock_statistics_t *, &lck->stats), 0,
1865 sizeof(lck->stats));
1866 __kmp_remember_lock(lck);
1869 static void __kmp_add_stats(kmp_adaptive_lock_statistics_t *t,
1870 kmp_adaptive_lock_info_t *lck) {
1871 kmp_adaptive_lock_statistics_t
volatile *s = &lck->stats;
1873 t->nonSpeculativeAcquireAttempts += lck->acquire_attempts;
1874 t->successfulSpeculations += s->successfulSpeculations;
1875 t->hardFailedSpeculations += s->hardFailedSpeculations;
1876 t->softFailedSpeculations += s->softFailedSpeculations;
1877 t->nonSpeculativeAcquires += s->nonSpeculativeAcquires;
1878 t->lemmingYields += s->lemmingYields;
1881 static void __kmp_accumulate_speculative_stats(kmp_adaptive_lock_info_t *lck) {
1882 __kmp_acquire_bootstrap_lock(&chain_lock);
1884 __kmp_add_stats(&destroyedStats, lck);
1885 __kmp_forget_lock(lck);
1887 __kmp_release_bootstrap_lock(&chain_lock);
1890 static float percent(kmp_uint32 count, kmp_uint32 total) {
1891 return (total == 0) ? 0.0 : (100.0 * count) / total;
1894 static FILE *__kmp_open_stats_file() {
1895 if (strcmp(__kmp_speculative_statsfile,
"-") == 0)
1898 size_t buffLen = KMP_STRLEN(__kmp_speculative_statsfile) + 20;
1899 char buffer[buffLen];
1900 KMP_SNPRINTF(&buffer[0], buffLen, __kmp_speculative_statsfile,
1901 (kmp_int32)getpid());
1902 FILE *result = fopen(&buffer[0],
"w");
1905 return result ? result : stdout;
1908 void __kmp_print_speculative_stats() {
1909 kmp_adaptive_lock_statistics_t total = destroyedStats;
1910 kmp_adaptive_lock_info_t *lck;
1912 for (lck = liveLocks.stats.next; lck != &liveLocks; lck = lck->stats.next) {
1913 __kmp_add_stats(&total, lck);
1915 kmp_adaptive_lock_statistics_t *t = &total;
1916 kmp_uint32 totalSections =
1917 t->nonSpeculativeAcquires + t->successfulSpeculations;
1918 kmp_uint32 totalSpeculations = t->successfulSpeculations +
1919 t->hardFailedSpeculations +
1920 t->softFailedSpeculations;
1921 if (totalSections <= 0)
1924 FILE *statsFile = __kmp_open_stats_file();
1926 fprintf(statsFile,
"Speculative lock statistics (all approximate!)\n");
1927 fprintf(statsFile,
" Lock parameters: \n" 1928 " max_soft_retries : %10d\n" 1929 " max_badness : %10d\n",
1930 __kmp_adaptive_backoff_params.max_soft_retries,
1931 __kmp_adaptive_backoff_params.max_badness);
1932 fprintf(statsFile,
" Non-speculative acquire attempts : %10d\n",
1933 t->nonSpeculativeAcquireAttempts);
1934 fprintf(statsFile,
" Total critical sections : %10d\n",
1936 fprintf(statsFile,
" Successful speculations : %10d (%5.1f%%)\n",
1937 t->successfulSpeculations,
1938 percent(t->successfulSpeculations, totalSections));
1939 fprintf(statsFile,
" Non-speculative acquires : %10d (%5.1f%%)\n",
1940 t->nonSpeculativeAcquires,
1941 percent(t->nonSpeculativeAcquires, totalSections));
1942 fprintf(statsFile,
" Lemming yields : %10d\n\n",
1945 fprintf(statsFile,
" Speculative acquire attempts : %10d\n",
1947 fprintf(statsFile,
" Successes : %10d (%5.1f%%)\n",
1948 t->successfulSpeculations,
1949 percent(t->successfulSpeculations, totalSpeculations));
1950 fprintf(statsFile,
" Soft failures : %10d (%5.1f%%)\n",
1951 t->softFailedSpeculations,
1952 percent(t->softFailedSpeculations, totalSpeculations));
1953 fprintf(statsFile,
" Hard failures : %10d (%5.1f%%)\n",
1954 t->hardFailedSpeculations,
1955 percent(t->hardFailedSpeculations, totalSpeculations));
1957 if (statsFile != stdout)
1961 #define KMP_INC_STAT(lck, stat) (lck->lk.adaptive.stats.stat++) 1963 #define KMP_INC_STAT(lck, stat) 1965 #endif // KMP_DEBUG_ADAPTIVE_LOCKS 1967 static inline bool __kmp_is_unlocked_queuing_lock(kmp_queuing_lock_t *lck) {
1970 bool res = lck->lk.head_id == 0;
1974 #if KMP_COMPILER_ICC 1977 __sync_synchronize();
1984 static __inline
void 1985 __kmp_update_badness_after_success(kmp_adaptive_lock_t *lck) {
1987 lck->lk.adaptive.badness = 0;
1988 KMP_INC_STAT(lck, successfulSpeculations);
1992 static __inline
void __kmp_step_badness(kmp_adaptive_lock_t *lck) {
1993 kmp_uint32 newBadness = (lck->lk.adaptive.badness << 1) | 1;
1994 if (newBadness > lck->lk.adaptive.max_badness) {
1997 lck->lk.adaptive.badness = newBadness;
2002 static __inline
int __kmp_should_speculate(kmp_adaptive_lock_t *lck,
2004 kmp_uint32 badness = lck->lk.adaptive.badness;
2005 kmp_uint32 attempts = lck->lk.adaptive.acquire_attempts;
2006 int res = (attempts & badness) == 0;
2012 static int __kmp_test_adaptive_lock_only(kmp_adaptive_lock_t *lck,
2014 int retries = lck->lk.adaptive.max_soft_retries;
2021 kmp_uint32 status = _xbegin();
2026 if (status == _XBEGIN_STARTED) {
2031 if (!__kmp_is_unlocked_queuing_lock(GET_QLK_PTR(lck))) {
2035 KMP_ASSERT2(0,
"should not get here");
2040 if (status & SOFT_ABORT_MASK) {
2041 KMP_INC_STAT(lck, softFailedSpeculations);
2044 KMP_INC_STAT(lck, hardFailedSpeculations);
2049 }
while (retries--);
2053 __kmp_step_badness(lck);
2060 static int __kmp_test_adaptive_lock(kmp_adaptive_lock_t *lck, kmp_int32 gtid) {
2062 if (__kmp_should_speculate(lck, gtid) &&
2063 __kmp_test_adaptive_lock_only(lck, gtid))
2068 lck->lk.adaptive.acquire_attempts++;
2071 if (__kmp_test_queuing_lock(GET_QLK_PTR(lck), gtid)) {
2072 KMP_INC_STAT(lck, nonSpeculativeAcquires);
2079 static int __kmp_test_adaptive_lock_with_checks(kmp_adaptive_lock_t *lck,
2081 char const *
const func =
"omp_test_lock";
2082 if (lck->lk.qlk.initialized != GET_QLK_PTR(lck)) {
2083 KMP_FATAL(LockIsUninitialized, func);
2086 int retval = __kmp_test_adaptive_lock(lck, gtid);
2089 lck->lk.qlk.owner_id = gtid + 1;
2105 static void __kmp_acquire_adaptive_lock(kmp_adaptive_lock_t *lck,
2107 if (__kmp_should_speculate(lck, gtid)) {
2108 if (__kmp_is_unlocked_queuing_lock(GET_QLK_PTR(lck))) {
2109 if (__kmp_test_adaptive_lock_only(lck, gtid))
2118 while (!__kmp_is_unlocked_queuing_lock(GET_QLK_PTR(lck))) {
2119 KMP_INC_STAT(lck, lemmingYields);
2123 if (__kmp_test_adaptive_lock_only(lck, gtid))
2130 lck->lk.adaptive.acquire_attempts++;
2132 __kmp_acquire_queuing_lock_timed_template<FALSE>(GET_QLK_PTR(lck), gtid);
2134 KMP_INC_STAT(lck, nonSpeculativeAcquires);
2135 ANNOTATE_QUEUING_ACQUIRED(lck);
2138 static void __kmp_acquire_adaptive_lock_with_checks(kmp_adaptive_lock_t *lck,
2140 char const *
const func =
"omp_set_lock";
2141 if (lck->lk.qlk.initialized != GET_QLK_PTR(lck)) {
2142 KMP_FATAL(LockIsUninitialized, func);
2144 if (__kmp_get_queuing_lock_owner(GET_QLK_PTR(lck)) == gtid) {
2145 KMP_FATAL(LockIsAlreadyOwned, func);
2148 __kmp_acquire_adaptive_lock(lck, gtid);
2150 lck->lk.qlk.owner_id = gtid + 1;
2153 static int __kmp_release_adaptive_lock(kmp_adaptive_lock_t *lck,
2155 if (__kmp_is_unlocked_queuing_lock(GET_QLK_PTR(
2160 __kmp_update_badness_after_success(lck);
2163 __kmp_release_queuing_lock(GET_QLK_PTR(lck), gtid);
2165 return KMP_LOCK_RELEASED;
2168 static int __kmp_release_adaptive_lock_with_checks(kmp_adaptive_lock_t *lck,
2170 char const *
const func =
"omp_unset_lock";
2172 if (lck->lk.qlk.initialized != GET_QLK_PTR(lck)) {
2173 KMP_FATAL(LockIsUninitialized, func);
2175 if (__kmp_get_queuing_lock_owner(GET_QLK_PTR(lck)) == -1) {
2176 KMP_FATAL(LockUnsettingFree, func);
2178 if (__kmp_get_queuing_lock_owner(GET_QLK_PTR(lck)) != gtid) {
2179 KMP_FATAL(LockUnsettingSetByAnother, func);
2181 lck->lk.qlk.owner_id = 0;
2182 __kmp_release_adaptive_lock(lck, gtid);
2183 return KMP_LOCK_RELEASED;
2186 static void __kmp_init_adaptive_lock(kmp_adaptive_lock_t *lck) {
2187 __kmp_init_queuing_lock(GET_QLK_PTR(lck));
2188 lck->lk.adaptive.badness = 0;
2189 lck->lk.adaptive.acquire_attempts = 0;
2190 lck->lk.adaptive.max_soft_retries =
2191 __kmp_adaptive_backoff_params.max_soft_retries;
2192 lck->lk.adaptive.max_badness = __kmp_adaptive_backoff_params.max_badness;
2193 #if KMP_DEBUG_ADAPTIVE_LOCKS 2194 __kmp_zero_speculative_stats(&lck->lk.adaptive);
2196 KA_TRACE(1000, (
"__kmp_init_adaptive_lock: lock %p initialized\n", lck));
2199 static void __kmp_destroy_adaptive_lock(kmp_adaptive_lock_t *lck) {
2200 #if KMP_DEBUG_ADAPTIVE_LOCKS 2201 __kmp_accumulate_speculative_stats(&lck->lk.adaptive);
2203 __kmp_destroy_queuing_lock(GET_QLK_PTR(lck));
2207 static void __kmp_destroy_adaptive_lock_with_checks(kmp_adaptive_lock_t *lck) {
2208 char const *
const func =
"omp_destroy_lock";
2209 if (lck->lk.qlk.initialized != GET_QLK_PTR(lck)) {
2210 KMP_FATAL(LockIsUninitialized, func);
2212 if (__kmp_get_queuing_lock_owner(GET_QLK_PTR(lck)) != -1) {
2213 KMP_FATAL(LockStillOwned, func);
2215 __kmp_destroy_adaptive_lock(lck);
2218 #endif // KMP_USE_ADAPTIVE_LOCKS 2224 static kmp_int32 __kmp_get_drdpa_lock_owner(kmp_drdpa_lock_t *lck) {
2225 return lck->lk.owner_id - 1;
2228 static inline bool __kmp_is_drdpa_lock_nestable(kmp_drdpa_lock_t *lck) {
2229 return lck->lk.depth_locked != -1;
2232 __forceinline
static int 2233 __kmp_acquire_drdpa_lock_timed_template(kmp_drdpa_lock_t *lck, kmp_int32 gtid) {
2234 kmp_uint64 ticket = KMP_ATOMIC_INC(&lck->lk.next_ticket);
2235 kmp_uint64 mask = lck->lk.mask;
2236 std::atomic<kmp_uint64> *polls = lck->lk.polls;
2238 #ifdef USE_LOCK_PROFILE 2239 if (polls[ticket & mask] != ticket)
2240 __kmp_printf(
"LOCK CONTENTION: %p\n", lck);
2252 KMP_FSYNC_PREPARE(lck);
2253 KMP_INIT_YIELD(spins);
2254 while (polls[ticket & mask] < ticket) {
2255 KMP_YIELD_OVERSUB_ELSE_SPIN(spins);
2263 mask = lck->lk.mask;
2264 polls = lck->lk.polls;
2268 KMP_FSYNC_ACQUIRED(lck);
2269 KA_TRACE(1000, (
"__kmp_acquire_drdpa_lock: ticket #%lld acquired lock %p\n",
2271 lck->lk.now_serving = ticket;
2278 if ((lck->lk.old_polls != NULL) && (ticket >= lck->lk.cleanup_ticket)) {
2279 __kmp_free(lck->lk.old_polls);
2280 lck->lk.old_polls = NULL;
2281 lck->lk.cleanup_ticket = 0;
2287 if (lck->lk.old_polls == NULL) {
2288 bool reconfigure =
false;
2289 std::atomic<kmp_uint64> *old_polls = polls;
2290 kmp_uint32 num_polls = TCR_4(lck->lk.num_polls);
2292 if (TCR_4(__kmp_nth) >
2293 (__kmp_avail_proc ? __kmp_avail_proc : __kmp_xproc)) {
2296 if (num_polls > 1) {
2298 num_polls = TCR_4(lck->lk.num_polls);
2301 polls = (std::atomic<kmp_uint64> *)__kmp_allocate(num_polls *
2309 kmp_uint64 num_waiting = TCR_8(lck->lk.next_ticket) - ticket - 1;
2310 if (num_waiting > num_polls) {
2311 kmp_uint32 old_num_polls = num_polls;
2314 mask = (mask << 1) | 1;
2316 }
while (num_polls <= num_waiting);
2322 polls = (std::atomic<kmp_uint64> *)__kmp_allocate(num_polls *
2325 for (i = 0; i < old_num_polls; i++) {
2326 polls[i].store(old_polls[i]);
2341 KA_TRACE(1000, (
"__kmp_acquire_drdpa_lock: ticket #%lld reconfiguring " 2342 "lock %p to %d polls\n",
2343 ticket, lck, num_polls));
2345 lck->lk.old_polls = old_polls;
2346 lck->lk.polls = polls;
2350 lck->lk.num_polls = num_polls;
2351 lck->lk.mask = mask;
2359 lck->lk.cleanup_ticket = lck->lk.next_ticket;
2362 return KMP_LOCK_ACQUIRED_FIRST;
2365 int __kmp_acquire_drdpa_lock(kmp_drdpa_lock_t *lck, kmp_int32 gtid) {
2366 int retval = __kmp_acquire_drdpa_lock_timed_template(lck, gtid);
2367 ANNOTATE_DRDPA_ACQUIRED(lck);
2371 static int __kmp_acquire_drdpa_lock_with_checks(kmp_drdpa_lock_t *lck,
2373 char const *
const func =
"omp_set_lock";
2374 if (lck->lk.initialized != lck) {
2375 KMP_FATAL(LockIsUninitialized, func);
2377 if (__kmp_is_drdpa_lock_nestable(lck)) {
2378 KMP_FATAL(LockNestableUsedAsSimple, func);
2380 if ((gtid >= 0) && (__kmp_get_drdpa_lock_owner(lck) == gtid)) {
2381 KMP_FATAL(LockIsAlreadyOwned, func);
2384 __kmp_acquire_drdpa_lock(lck, gtid);
2386 lck->lk.owner_id = gtid + 1;
2387 return KMP_LOCK_ACQUIRED_FIRST;
2390 int __kmp_test_drdpa_lock(kmp_drdpa_lock_t *lck, kmp_int32 gtid) {
2393 kmp_uint64 ticket = lck->lk.next_ticket;
2394 std::atomic<kmp_uint64> *polls = lck->lk.polls;
2395 kmp_uint64 mask = lck->lk.mask;
2396 if (polls[ticket & mask] == ticket) {
2397 kmp_uint64 next_ticket = ticket + 1;
2398 if (__kmp_atomic_compare_store_acq(&lck->lk.next_ticket, ticket,
2400 KMP_FSYNC_ACQUIRED(lck);
2401 KA_TRACE(1000, (
"__kmp_test_drdpa_lock: ticket #%lld acquired lock %p\n",
2403 lck->lk.now_serving = ticket;
2417 static int __kmp_test_drdpa_lock_with_checks(kmp_drdpa_lock_t *lck,
2419 char const *
const func =
"omp_test_lock";
2420 if (lck->lk.initialized != lck) {
2421 KMP_FATAL(LockIsUninitialized, func);
2423 if (__kmp_is_drdpa_lock_nestable(lck)) {
2424 KMP_FATAL(LockNestableUsedAsSimple, func);
2427 int retval = __kmp_test_drdpa_lock(lck, gtid);
2430 lck->lk.owner_id = gtid + 1;
2435 int __kmp_release_drdpa_lock(kmp_drdpa_lock_t *lck, kmp_int32 gtid) {
2438 kmp_uint64 ticket = lck->lk.now_serving + 1;
2439 std::atomic<kmp_uint64> *polls = lck->lk.polls;
2440 kmp_uint64 mask = lck->lk.mask;
2441 KA_TRACE(1000, (
"__kmp_release_drdpa_lock: ticket #%lld released lock %p\n",
2443 KMP_FSYNC_RELEASING(lck);
2444 ANNOTATE_DRDPA_RELEASED(lck);
2445 polls[ticket & mask] = ticket;
2446 return KMP_LOCK_RELEASED;
2449 static int __kmp_release_drdpa_lock_with_checks(kmp_drdpa_lock_t *lck,
2451 char const *
const func =
"omp_unset_lock";
2453 if (lck->lk.initialized != lck) {
2454 KMP_FATAL(LockIsUninitialized, func);
2456 if (__kmp_is_drdpa_lock_nestable(lck)) {
2457 KMP_FATAL(LockNestableUsedAsSimple, func);
2459 if (__kmp_get_drdpa_lock_owner(lck) == -1) {
2460 KMP_FATAL(LockUnsettingFree, func);
2462 if ((gtid >= 0) && (__kmp_get_drdpa_lock_owner(lck) >= 0) &&
2463 (__kmp_get_drdpa_lock_owner(lck) != gtid)) {
2464 KMP_FATAL(LockUnsettingSetByAnother, func);
2466 lck->lk.owner_id = 0;
2467 return __kmp_release_drdpa_lock(lck, gtid);
2470 void __kmp_init_drdpa_lock(kmp_drdpa_lock_t *lck) {
2471 lck->lk.location = NULL;
2473 lck->lk.num_polls = 1;
2474 lck->lk.polls = (std::atomic<kmp_uint64> *)__kmp_allocate(
2475 lck->lk.num_polls *
sizeof(*(lck->lk.polls)));
2476 lck->lk.cleanup_ticket = 0;
2477 lck->lk.old_polls = NULL;
2478 lck->lk.next_ticket = 0;
2479 lck->lk.now_serving = 0;
2480 lck->lk.owner_id = 0;
2481 lck->lk.depth_locked = -1;
2482 lck->lk.initialized = lck;
2484 KA_TRACE(1000, (
"__kmp_init_drdpa_lock: lock %p initialized\n", lck));
2487 void __kmp_destroy_drdpa_lock(kmp_drdpa_lock_t *lck) {
2488 lck->lk.initialized = NULL;
2489 lck->lk.location = NULL;
2490 if (lck->lk.polls.load() != NULL) {
2491 __kmp_free(lck->lk.polls.load());
2492 lck->lk.polls = NULL;
2494 if (lck->lk.old_polls != NULL) {
2495 __kmp_free(lck->lk.old_polls);
2496 lck->lk.old_polls = NULL;
2499 lck->lk.num_polls = 0;
2500 lck->lk.cleanup_ticket = 0;
2501 lck->lk.next_ticket = 0;
2502 lck->lk.now_serving = 0;
2503 lck->lk.owner_id = 0;
2504 lck->lk.depth_locked = -1;
2507 static void __kmp_destroy_drdpa_lock_with_checks(kmp_drdpa_lock_t *lck) {
2508 char const *
const func =
"omp_destroy_lock";
2509 if (lck->lk.initialized != lck) {
2510 KMP_FATAL(LockIsUninitialized, func);
2512 if (__kmp_is_drdpa_lock_nestable(lck)) {
2513 KMP_FATAL(LockNestableUsedAsSimple, func);
2515 if (__kmp_get_drdpa_lock_owner(lck) != -1) {
2516 KMP_FATAL(LockStillOwned, func);
2518 __kmp_destroy_drdpa_lock(lck);
2523 int __kmp_acquire_nested_drdpa_lock(kmp_drdpa_lock_t *lck, kmp_int32 gtid) {
2524 KMP_DEBUG_ASSERT(gtid >= 0);
2526 if (__kmp_get_drdpa_lock_owner(lck) == gtid) {
2527 lck->lk.depth_locked += 1;
2528 return KMP_LOCK_ACQUIRED_NEXT;
2530 __kmp_acquire_drdpa_lock_timed_template(lck, gtid);
2531 ANNOTATE_DRDPA_ACQUIRED(lck);
2533 lck->lk.depth_locked = 1;
2535 lck->lk.owner_id = gtid + 1;
2536 return KMP_LOCK_ACQUIRED_FIRST;
2540 static void __kmp_acquire_nested_drdpa_lock_with_checks(kmp_drdpa_lock_t *lck,
2542 char const *
const func =
"omp_set_nest_lock";
2543 if (lck->lk.initialized != lck) {
2544 KMP_FATAL(LockIsUninitialized, func);
2546 if (!__kmp_is_drdpa_lock_nestable(lck)) {
2547 KMP_FATAL(LockSimpleUsedAsNestable, func);
2549 __kmp_acquire_nested_drdpa_lock(lck, gtid);
2552 int __kmp_test_nested_drdpa_lock(kmp_drdpa_lock_t *lck, kmp_int32 gtid) {
2555 KMP_DEBUG_ASSERT(gtid >= 0);
2557 if (__kmp_get_drdpa_lock_owner(lck) == gtid) {
2558 retval = ++lck->lk.depth_locked;
2559 }
else if (!__kmp_test_drdpa_lock(lck, gtid)) {
2563 retval = lck->lk.depth_locked = 1;
2565 lck->lk.owner_id = gtid + 1;
2570 static int __kmp_test_nested_drdpa_lock_with_checks(kmp_drdpa_lock_t *lck,
2572 char const *
const func =
"omp_test_nest_lock";
2573 if (lck->lk.initialized != lck) {
2574 KMP_FATAL(LockIsUninitialized, func);
2576 if (!__kmp_is_drdpa_lock_nestable(lck)) {
2577 KMP_FATAL(LockSimpleUsedAsNestable, func);
2579 return __kmp_test_nested_drdpa_lock(lck, gtid);
2582 int __kmp_release_nested_drdpa_lock(kmp_drdpa_lock_t *lck, kmp_int32 gtid) {
2583 KMP_DEBUG_ASSERT(gtid >= 0);
2586 if (--(lck->lk.depth_locked) == 0) {
2588 lck->lk.owner_id = 0;
2589 __kmp_release_drdpa_lock(lck, gtid);
2590 return KMP_LOCK_RELEASED;
2592 return KMP_LOCK_STILL_HELD;
2595 static int __kmp_release_nested_drdpa_lock_with_checks(kmp_drdpa_lock_t *lck,
2597 char const *
const func =
"omp_unset_nest_lock";
2599 if (lck->lk.initialized != lck) {
2600 KMP_FATAL(LockIsUninitialized, func);
2602 if (!__kmp_is_drdpa_lock_nestable(lck)) {
2603 KMP_FATAL(LockSimpleUsedAsNestable, func);
2605 if (__kmp_get_drdpa_lock_owner(lck) == -1) {
2606 KMP_FATAL(LockUnsettingFree, func);
2608 if (__kmp_get_drdpa_lock_owner(lck) != gtid) {
2609 KMP_FATAL(LockUnsettingSetByAnother, func);
2611 return __kmp_release_nested_drdpa_lock(lck, gtid);
2614 void __kmp_init_nested_drdpa_lock(kmp_drdpa_lock_t *lck) {
2615 __kmp_init_drdpa_lock(lck);
2616 lck->lk.depth_locked = 0;
2619 void __kmp_destroy_nested_drdpa_lock(kmp_drdpa_lock_t *lck) {
2620 __kmp_destroy_drdpa_lock(lck);
2621 lck->lk.depth_locked = 0;
2624 static void __kmp_destroy_nested_drdpa_lock_with_checks(kmp_drdpa_lock_t *lck) {
2625 char const *
const func =
"omp_destroy_nest_lock";
2626 if (lck->lk.initialized != lck) {
2627 KMP_FATAL(LockIsUninitialized, func);
2629 if (!__kmp_is_drdpa_lock_nestable(lck)) {
2630 KMP_FATAL(LockSimpleUsedAsNestable, func);
2632 if (__kmp_get_drdpa_lock_owner(lck) != -1) {
2633 KMP_FATAL(LockStillOwned, func);
2635 __kmp_destroy_nested_drdpa_lock(lck);
2640 static const ident_t *__kmp_get_drdpa_lock_location(kmp_drdpa_lock_t *lck) {
2641 return lck->lk.location;
2644 static void __kmp_set_drdpa_lock_location(kmp_drdpa_lock_t *lck,
2646 lck->lk.location = loc;
2649 static kmp_lock_flags_t __kmp_get_drdpa_lock_flags(kmp_drdpa_lock_t *lck) {
2650 return lck->lk.flags;
2653 static void __kmp_set_drdpa_lock_flags(kmp_drdpa_lock_t *lck,
2654 kmp_lock_flags_t flags) {
2655 lck->lk.flags = flags;
2659 #if KMP_ARCH_X86 || KMP_ARCH_X86_64 2660 #define __kmp_tsc() __kmp_hardware_timestamp() 2662 kmp_backoff_t __kmp_spin_backoff_params = {1, 4096, 100};
2665 extern kmp_uint64 __kmp_now_nsec();
2666 kmp_backoff_t __kmp_spin_backoff_params = {1, 256, 100};
2667 #define __kmp_tsc() __kmp_now_nsec() 2677 static inline bool before(kmp_uint64 a, kmp_uint64 b) {
2678 return ((kmp_int64)b - (kmp_int64)a) > 0;
2682 void __kmp_spin_backoff(kmp_backoff_t *boff) {
2685 for (i = boff->step; i > 0; i--) {
2686 kmp_uint64 goal = __kmp_tsc() + boff->min_tick;
2689 }
while (before(__kmp_tsc(), goal));
2691 boff->step = (boff->step << 1 | 1) & (boff->max_backoff - 1);
2694 #if KMP_USE_DYNAMIC_LOCK 2698 static void __kmp_init_direct_lock(kmp_dyna_lock_t *lck,
2699 kmp_dyna_lockseq_t seq) {
2700 TCW_4(*lck, KMP_GET_D_TAG(seq));
2703 (
"__kmp_init_direct_lock: initialized direct lock with type#%d\n", seq));
2709 #define HLE_ACQUIRE ".byte 0xf2;" 2710 #define HLE_RELEASE ".byte 0xf3;" 2712 static inline kmp_uint32 swap4(kmp_uint32
volatile *p, kmp_uint32 v) {
2713 __asm__
volatile(HLE_ACQUIRE
"xchg %1,%0" :
"+r"(v),
"+m"(*p) : :
"memory");
2717 static void __kmp_destroy_hle_lock(kmp_dyna_lock_t *lck) { TCW_4(*lck, 0); }
2719 static void __kmp_destroy_hle_lock_with_checks(kmp_dyna_lock_t *lck) {
2723 static void __kmp_acquire_hle_lock(kmp_dyna_lock_t *lck, kmp_int32 gtid) {
2725 if (swap4(lck, KMP_LOCK_BUSY(1, hle)) != KMP_LOCK_FREE(hle)) {
2728 while (*(kmp_uint32
volatile *)lck != KMP_LOCK_FREE(hle)) {
2729 for (
int i = delay; i != 0; --i)
2731 delay = ((delay << 1) | 1) & 7;
2733 }
while (swap4(lck, KMP_LOCK_BUSY(1, hle)) != KMP_LOCK_FREE(hle));
2737 static void __kmp_acquire_hle_lock_with_checks(kmp_dyna_lock_t *lck,
2739 __kmp_acquire_hle_lock(lck, gtid);
2742 static int __kmp_release_hle_lock(kmp_dyna_lock_t *lck, kmp_int32 gtid) {
2743 __asm__
volatile(HLE_RELEASE
"movl %1,%0" 2745 :
"r"(KMP_LOCK_FREE(hle))
2747 return KMP_LOCK_RELEASED;
2750 static int __kmp_release_hle_lock_with_checks(kmp_dyna_lock_t *lck,
2752 return __kmp_release_hle_lock(lck, gtid);
2755 static int __kmp_test_hle_lock(kmp_dyna_lock_t *lck, kmp_int32 gtid) {
2756 return swap4(lck, KMP_LOCK_BUSY(1, hle)) == KMP_LOCK_FREE(hle);
2759 static int __kmp_test_hle_lock_with_checks(kmp_dyna_lock_t *lck,
2761 return __kmp_test_hle_lock(lck, gtid);
2764 static void __kmp_init_rtm_lock(kmp_queuing_lock_t *lck) {
2765 __kmp_init_queuing_lock(lck);
2768 static void __kmp_destroy_rtm_lock(kmp_queuing_lock_t *lck) {
2769 __kmp_destroy_queuing_lock(lck);
2772 static void __kmp_destroy_rtm_lock_with_checks(kmp_queuing_lock_t *lck) {
2773 __kmp_destroy_queuing_lock_with_checks(lck);
2776 static void __kmp_acquire_rtm_lock(kmp_queuing_lock_t *lck, kmp_int32 gtid) {
2777 unsigned retries = 3, status;
2780 if (status == _XBEGIN_STARTED) {
2781 if (__kmp_is_unlocked_queuing_lock(lck))
2785 if ((status & _XABORT_EXPLICIT) && _XABORT_CODE(status) == 0xff) {
2787 while (!__kmp_is_unlocked_queuing_lock(lck)) {
2790 }
else if (!(status & _XABORT_RETRY))
2792 }
while (retries--);
2795 __kmp_acquire_queuing_lock(lck, gtid);
2798 static void __kmp_acquire_rtm_lock_with_checks(kmp_queuing_lock_t *lck,
2800 __kmp_acquire_rtm_lock(lck, gtid);
2803 static int __kmp_release_rtm_lock(kmp_queuing_lock_t *lck, kmp_int32 gtid) {
2804 if (__kmp_is_unlocked_queuing_lock(lck)) {
2809 __kmp_release_queuing_lock(lck, gtid);
2811 return KMP_LOCK_RELEASED;
2814 static int __kmp_release_rtm_lock_with_checks(kmp_queuing_lock_t *lck,
2816 return __kmp_release_rtm_lock(lck, gtid);
2819 static int __kmp_test_rtm_lock(kmp_queuing_lock_t *lck, kmp_int32 gtid) {
2820 unsigned retries = 3, status;
2823 if (status == _XBEGIN_STARTED && __kmp_is_unlocked_queuing_lock(lck)) {
2826 if (!(status & _XABORT_RETRY))
2828 }
while (retries--);
2830 return (__kmp_is_unlocked_queuing_lock(lck)) ? 1 : 0;
2833 static int __kmp_test_rtm_lock_with_checks(kmp_queuing_lock_t *lck,
2835 return __kmp_test_rtm_lock(lck, gtid);
2838 #endif // KMP_USE_TSX 2841 static void __kmp_init_indirect_lock(kmp_dyna_lock_t *l,
2842 kmp_dyna_lockseq_t tag);
2843 static void __kmp_destroy_indirect_lock(kmp_dyna_lock_t *lock);
2844 static int __kmp_set_indirect_lock(kmp_dyna_lock_t *lock, kmp_int32);
2845 static int __kmp_unset_indirect_lock(kmp_dyna_lock_t *lock, kmp_int32);
2846 static int __kmp_test_indirect_lock(kmp_dyna_lock_t *lock, kmp_int32);
2847 static int __kmp_set_indirect_lock_with_checks(kmp_dyna_lock_t *lock,
2849 static int __kmp_unset_indirect_lock_with_checks(kmp_dyna_lock_t *lock,
2851 static int __kmp_test_indirect_lock_with_checks(kmp_dyna_lock_t *lock,
2855 #define KMP_FOREACH_LOCK_KIND(m, a) m(ticket, a) m(queuing, a) m(drdpa, a) 2857 #define expand1(lk, op) \ 2858 static void __kmp_##op##_##lk##_##lock(kmp_user_lock_p lock) { \ 2859 __kmp_##op##_##lk##_##lock(&lock->lk); \ 2861 #define expand2(lk, op) \ 2862 static int __kmp_##op##_##lk##_##lock(kmp_user_lock_p lock, \ 2864 return __kmp_##op##_##lk##_##lock(&lock->lk, gtid); \ 2866 #define expand3(lk, op) \ 2867 static void __kmp_set_##lk##_##lock_flags(kmp_user_lock_p lock, \ 2868 kmp_lock_flags_t flags) { \ 2869 __kmp_set_##lk##_lock_flags(&lock->lk, flags); \ 2871 #define expand4(lk, op) \ 2872 static void __kmp_set_##lk##_##lock_location(kmp_user_lock_p lock, \ 2873 const ident_t *loc) { \ 2874 __kmp_set_##lk##_lock_location(&lock->lk, loc); \ 2877 KMP_FOREACH_LOCK_KIND(expand1, init)
2878 KMP_FOREACH_LOCK_KIND(expand1, init_nested)
2879 KMP_FOREACH_LOCK_KIND(expand1, destroy)
2880 KMP_FOREACH_LOCK_KIND(expand1, destroy_nested)
2881 KMP_FOREACH_LOCK_KIND(expand2, acquire)
2882 KMP_FOREACH_LOCK_KIND(expand2, acquire_nested)
2883 KMP_FOREACH_LOCK_KIND(expand2, release)
2884 KMP_FOREACH_LOCK_KIND(expand2, release_nested)
2885 KMP_FOREACH_LOCK_KIND(expand2, test)
2886 KMP_FOREACH_LOCK_KIND(expand2, test_nested)
2887 KMP_FOREACH_LOCK_KIND(expand3, )
2888 KMP_FOREACH_LOCK_KIND(expand4, )
2899 #define expand(l, op) 0, __kmp_init_direct_lock, 2900 void (*__kmp_direct_init[])(kmp_dyna_lock_t *, kmp_dyna_lockseq_t) = {
2901 __kmp_init_indirect_lock, 0, KMP_FOREACH_D_LOCK(expand, init)};
2905 #define expand(l, op) 0, (void (*)(kmp_dyna_lock_t *))__kmp_##op##_##l##_lock, 2906 static void (*direct_destroy[])(kmp_dyna_lock_t *) = {
2907 __kmp_destroy_indirect_lock, 0, KMP_FOREACH_D_LOCK(expand, destroy)};
2909 #define expand(l, op) \ 2910 0, (void (*)(kmp_dyna_lock_t *))__kmp_destroy_##l##_lock_with_checks, 2911 static void (*direct_destroy_check[])(kmp_dyna_lock_t *) = {
2912 __kmp_destroy_indirect_lock, 0, KMP_FOREACH_D_LOCK(expand, destroy)};
2916 #define expand(l, op) \ 2917 0, (int (*)(kmp_dyna_lock_t *, kmp_int32))__kmp_##op##_##l##_lock, 2918 static int (*direct_set[])(kmp_dyna_lock_t *, kmp_int32) = {
2919 __kmp_set_indirect_lock, 0, KMP_FOREACH_D_LOCK(expand, acquire)};
2921 #define expand(l, op) \ 2922 0, (int (*)(kmp_dyna_lock_t *, kmp_int32))__kmp_##op##_##l##_lock_with_checks, 2923 static int (*direct_set_check[])(kmp_dyna_lock_t *, kmp_int32) = {
2924 __kmp_set_indirect_lock_with_checks, 0,
2925 KMP_FOREACH_D_LOCK(expand, acquire)};
2929 #define expand(l, op) \ 2930 0, (int (*)(kmp_dyna_lock_t *, kmp_int32))__kmp_##op##_##l##_lock, 2931 static int (*direct_unset[])(kmp_dyna_lock_t *, kmp_int32) = {
2932 __kmp_unset_indirect_lock, 0, KMP_FOREACH_D_LOCK(expand, release)};
2933 static int (*direct_test[])(kmp_dyna_lock_t *, kmp_int32) = {
2934 __kmp_test_indirect_lock, 0, KMP_FOREACH_D_LOCK(expand, test)};
2936 #define expand(l, op) \ 2937 0, (int (*)(kmp_dyna_lock_t *, kmp_int32))__kmp_##op##_##l##_lock_with_checks, 2938 static int (*direct_unset_check[])(kmp_dyna_lock_t *, kmp_int32) = {
2939 __kmp_unset_indirect_lock_with_checks, 0,
2940 KMP_FOREACH_D_LOCK(expand, release)};
2941 static int (*direct_test_check[])(kmp_dyna_lock_t *, kmp_int32) = {
2942 __kmp_test_indirect_lock_with_checks, 0, KMP_FOREACH_D_LOCK(expand, test)};
2946 void (*(*__kmp_direct_destroy))(kmp_dyna_lock_t *) = 0;
2947 int (*(*__kmp_direct_set))(kmp_dyna_lock_t *, kmp_int32) = 0;
2948 int (*(*__kmp_direct_unset))(kmp_dyna_lock_t *, kmp_int32) = 0;
2949 int (*(*__kmp_direct_test))(kmp_dyna_lock_t *, kmp_int32) = 0;
2952 #define expand(l, op) (void (*)(kmp_user_lock_p)) __kmp_##op##_##l##_##lock, 2953 void (*__kmp_indirect_init[])(kmp_user_lock_p) = {
2954 KMP_FOREACH_I_LOCK(expand, init)};
2957 #define expand(l, op) (void (*)(kmp_user_lock_p)) __kmp_##op##_##l##_##lock, 2958 static void (*indirect_destroy[])(kmp_user_lock_p) = {
2959 KMP_FOREACH_I_LOCK(expand, destroy)};
2961 #define expand(l, op) \ 2962 (void (*)(kmp_user_lock_p)) __kmp_##op##_##l##_##lock_with_checks, 2963 static void (*indirect_destroy_check[])(kmp_user_lock_p) = {
2964 KMP_FOREACH_I_LOCK(expand, destroy)};
2968 #define expand(l, op) \ 2969 (int (*)(kmp_user_lock_p, kmp_int32)) __kmp_##op##_##l##_##lock, 2970 static int (*indirect_set[])(kmp_user_lock_p,
2971 kmp_int32) = {KMP_FOREACH_I_LOCK(expand, acquire)};
2973 #define expand(l, op) \ 2974 (int (*)(kmp_user_lock_p, kmp_int32)) __kmp_##op##_##l##_##lock_with_checks, 2975 static int (*indirect_set_check[])(kmp_user_lock_p, kmp_int32) = {
2976 KMP_FOREACH_I_LOCK(expand, acquire)};
2980 #define expand(l, op) \ 2981 (int (*)(kmp_user_lock_p, kmp_int32)) __kmp_##op##_##l##_##lock, 2982 static int (*indirect_unset[])(kmp_user_lock_p, kmp_int32) = {
2983 KMP_FOREACH_I_LOCK(expand, release)};
2984 static int (*indirect_test[])(kmp_user_lock_p,
2985 kmp_int32) = {KMP_FOREACH_I_LOCK(expand, test)};
2987 #define expand(l, op) \ 2988 (int (*)(kmp_user_lock_p, kmp_int32)) __kmp_##op##_##l##_##lock_with_checks, 2989 static int (*indirect_unset_check[])(kmp_user_lock_p, kmp_int32) = {
2990 KMP_FOREACH_I_LOCK(expand, release)};
2991 static int (*indirect_test_check[])(kmp_user_lock_p, kmp_int32) = {
2992 KMP_FOREACH_I_LOCK(expand, test)};
2996 void (*(*__kmp_indirect_destroy))(kmp_user_lock_p) = 0;
2997 int (*(*__kmp_indirect_set))(kmp_user_lock_p, kmp_int32) = 0;
2998 int (*(*__kmp_indirect_unset))(kmp_user_lock_p, kmp_int32) = 0;
2999 int (*(*__kmp_indirect_test))(kmp_user_lock_p, kmp_int32) = 0;
3002 kmp_indirect_lock_table_t __kmp_i_lock_table;
3005 static kmp_uint32 __kmp_indirect_lock_size[KMP_NUM_I_LOCKS] = {0};
3008 void (*__kmp_indirect_set_location[KMP_NUM_I_LOCKS])(kmp_user_lock_p,
3010 void (*__kmp_indirect_set_flags[KMP_NUM_I_LOCKS])(kmp_user_lock_p,
3011 kmp_lock_flags_t) = {0};
3012 const ident_t *(*__kmp_indirect_get_location[KMP_NUM_I_LOCKS])(
3013 kmp_user_lock_p) = {0};
3014 kmp_lock_flags_t (*__kmp_indirect_get_flags[KMP_NUM_I_LOCKS])(
3015 kmp_user_lock_p) = {0};
3018 static kmp_indirect_lock_t *__kmp_indirect_lock_pool[KMP_NUM_I_LOCKS] = {0};
3025 kmp_indirect_lock_t *__kmp_allocate_indirect_lock(
void **user_lock,
3027 kmp_indirect_locktag_t tag) {
3028 kmp_indirect_lock_t *lck;
3029 kmp_lock_index_t idx;
3031 __kmp_acquire_lock(&__kmp_global_lock, gtid);
3033 if (__kmp_indirect_lock_pool[tag] != NULL) {
3035 lck = __kmp_indirect_lock_pool[tag];
3036 if (OMP_LOCK_T_SIZE <
sizeof(
void *))
3037 idx = lck->lock->pool.index;
3038 __kmp_indirect_lock_pool[tag] = (kmp_indirect_lock_t *)lck->lock->pool.next;
3039 KA_TRACE(20, (
"__kmp_allocate_indirect_lock: reusing an existing lock %p\n",
3042 idx = __kmp_i_lock_table.next;
3044 if (idx == __kmp_i_lock_table.size) {
3046 int row = __kmp_i_lock_table.size / KMP_I_LOCK_CHUNK;
3047 kmp_indirect_lock_t **new_table = (kmp_indirect_lock_t **)__kmp_allocate(
3048 2 * row *
sizeof(kmp_indirect_lock_t *));
3049 KMP_MEMCPY(new_table, __kmp_i_lock_table.table,
3050 row *
sizeof(kmp_indirect_lock_t *));
3051 kmp_indirect_lock_t **old_table = __kmp_i_lock_table.table;
3052 __kmp_i_lock_table.table = new_table;
3053 __kmp_free(old_table);
3055 for (
int i = row; i < 2 * row; ++i)
3056 *(__kmp_i_lock_table.table + i) = (kmp_indirect_lock_t *)__kmp_allocate(
3057 KMP_I_LOCK_CHUNK *
sizeof(kmp_indirect_lock_t));
3058 __kmp_i_lock_table.size = 2 * idx;
3060 __kmp_i_lock_table.next++;
3061 lck = KMP_GET_I_LOCK(idx);
3063 lck->lock = (kmp_user_lock_p)__kmp_allocate(__kmp_indirect_lock_size[tag]);
3065 (
"__kmp_allocate_indirect_lock: allocated a new lock %p\n", lck));
3068 __kmp_release_lock(&__kmp_global_lock, gtid);
3072 if (OMP_LOCK_T_SIZE <
sizeof(
void *)) {
3073 *((kmp_lock_index_t *)user_lock) = idx
3076 *((kmp_indirect_lock_t **)user_lock) = lck;
3083 static __forceinline kmp_indirect_lock_t *
3084 __kmp_lookup_indirect_lock(
void **user_lock,
const char *func) {
3085 if (__kmp_env_consistency_check) {
3086 kmp_indirect_lock_t *lck = NULL;
3087 if (user_lock == NULL) {
3088 KMP_FATAL(LockIsUninitialized, func);
3090 if (OMP_LOCK_T_SIZE <
sizeof(
void *)) {
3091 kmp_lock_index_t idx = KMP_EXTRACT_I_INDEX(user_lock);
3092 if (idx >= __kmp_i_lock_table.size) {
3093 KMP_FATAL(LockIsUninitialized, func);
3095 lck = KMP_GET_I_LOCK(idx);
3097 lck = *((kmp_indirect_lock_t **)user_lock);
3100 KMP_FATAL(LockIsUninitialized, func);
3104 if (OMP_LOCK_T_SIZE <
sizeof(
void *)) {
3105 return KMP_GET_I_LOCK(KMP_EXTRACT_I_INDEX(user_lock));
3107 return *((kmp_indirect_lock_t **)user_lock);
3112 static void __kmp_init_indirect_lock(kmp_dyna_lock_t *lock,
3113 kmp_dyna_lockseq_t seq) {
3114 #if KMP_USE_ADAPTIVE_LOCKS 3115 if (seq == lockseq_adaptive && !__kmp_cpuinfo.rtm) {
3116 KMP_WARNING(AdaptiveNotSupported,
"kmp_lockseq_t",
"adaptive");
3117 seq = lockseq_queuing;
3121 if (seq == lockseq_rtm && !__kmp_cpuinfo.rtm) {
3122 seq = lockseq_queuing;
3125 kmp_indirect_locktag_t tag = KMP_GET_I_TAG(seq);
3126 kmp_indirect_lock_t *l =
3127 __kmp_allocate_indirect_lock((
void **)lock, __kmp_entry_gtid(), tag);
3128 KMP_I_LOCK_FUNC(l, init)(l->lock);
3130 20, (
"__kmp_init_indirect_lock: initialized indirect lock with type#%d\n",
3134 static void __kmp_destroy_indirect_lock(kmp_dyna_lock_t *lock) {
3135 kmp_uint32 gtid = __kmp_entry_gtid();
3136 kmp_indirect_lock_t *l =
3137 __kmp_lookup_indirect_lock((
void **)lock,
"omp_destroy_lock");
3138 KMP_I_LOCK_FUNC(l, destroy)(l->lock);
3139 kmp_indirect_locktag_t tag = l->type;
3141 __kmp_acquire_lock(&__kmp_global_lock, gtid);
3144 l->lock->pool.next = (kmp_user_lock_p)__kmp_indirect_lock_pool[tag];
3145 if (OMP_LOCK_T_SIZE <
sizeof(
void *)) {
3146 l->lock->pool.index = KMP_EXTRACT_I_INDEX(lock);
3148 __kmp_indirect_lock_pool[tag] = l;
3150 __kmp_release_lock(&__kmp_global_lock, gtid);
3153 static int __kmp_set_indirect_lock(kmp_dyna_lock_t *lock, kmp_int32 gtid) {
3154 kmp_indirect_lock_t *l = KMP_LOOKUP_I_LOCK(lock);
3155 return KMP_I_LOCK_FUNC(l,
set)(l->lock, gtid);
3158 static int __kmp_unset_indirect_lock(kmp_dyna_lock_t *lock, kmp_int32 gtid) {
3159 kmp_indirect_lock_t *l = KMP_LOOKUP_I_LOCK(lock);
3160 return KMP_I_LOCK_FUNC(l, unset)(l->lock, gtid);
3163 static int __kmp_test_indirect_lock(kmp_dyna_lock_t *lock, kmp_int32 gtid) {
3164 kmp_indirect_lock_t *l = KMP_LOOKUP_I_LOCK(lock);
3165 return KMP_I_LOCK_FUNC(l, test)(l->lock, gtid);
3168 static int __kmp_set_indirect_lock_with_checks(kmp_dyna_lock_t *lock,
3170 kmp_indirect_lock_t *l =
3171 __kmp_lookup_indirect_lock((
void **)lock,
"omp_set_lock");
3172 return KMP_I_LOCK_FUNC(l,
set)(l->lock, gtid);
3175 static int __kmp_unset_indirect_lock_with_checks(kmp_dyna_lock_t *lock,
3177 kmp_indirect_lock_t *l =
3178 __kmp_lookup_indirect_lock((
void **)lock,
"omp_unset_lock");
3179 return KMP_I_LOCK_FUNC(l, unset)(l->lock, gtid);
3182 static int __kmp_test_indirect_lock_with_checks(kmp_dyna_lock_t *lock,
3184 kmp_indirect_lock_t *l =
3185 __kmp_lookup_indirect_lock((
void **)lock,
"omp_test_lock");
3186 return KMP_I_LOCK_FUNC(l, test)(l->lock, gtid);
3189 kmp_dyna_lockseq_t __kmp_user_lock_seq = lockseq_queuing;
3192 kmp_int32 __kmp_get_user_lock_owner(kmp_user_lock_p lck, kmp_uint32 seq) {
3195 case lockseq_nested_tas:
3196 return __kmp_get_tas_lock_owner((kmp_tas_lock_t *)lck);
3199 case lockseq_nested_futex:
3200 return __kmp_get_futex_lock_owner((kmp_futex_lock_t *)lck);
3202 case lockseq_ticket:
3203 case lockseq_nested_ticket:
3204 return __kmp_get_ticket_lock_owner((kmp_ticket_lock_t *)lck);
3205 case lockseq_queuing:
3206 case lockseq_nested_queuing:
3207 #if KMP_USE_ADAPTIVE_LOCKS 3208 case lockseq_adaptive:
3210 return __kmp_get_queuing_lock_owner((kmp_queuing_lock_t *)lck);
3212 case lockseq_nested_drdpa:
3213 return __kmp_get_drdpa_lock_owner((kmp_drdpa_lock_t *)lck);
3220 void __kmp_init_dynamic_user_locks() {
3222 if (__kmp_env_consistency_check) {
3223 __kmp_direct_set = direct_set_check;
3224 __kmp_direct_unset = direct_unset_check;
3225 __kmp_direct_test = direct_test_check;
3226 __kmp_direct_destroy = direct_destroy_check;
3227 __kmp_indirect_set = indirect_set_check;
3228 __kmp_indirect_unset = indirect_unset_check;
3229 __kmp_indirect_test = indirect_test_check;
3230 __kmp_indirect_destroy = indirect_destroy_check;
3232 __kmp_direct_set = direct_set;
3233 __kmp_direct_unset = direct_unset;
3234 __kmp_direct_test = direct_test;
3235 __kmp_direct_destroy = direct_destroy;
3236 __kmp_indirect_set = indirect_set;
3237 __kmp_indirect_unset = indirect_unset;
3238 __kmp_indirect_test = indirect_test;
3239 __kmp_indirect_destroy = indirect_destroy;
3244 if (__kmp_init_user_locks)
3248 __kmp_i_lock_table.size = KMP_I_LOCK_CHUNK;
3249 __kmp_i_lock_table.table =
3250 (kmp_indirect_lock_t **)__kmp_allocate(
sizeof(kmp_indirect_lock_t *));
3251 *(__kmp_i_lock_table.table) = (kmp_indirect_lock_t *)__kmp_allocate(
3252 KMP_I_LOCK_CHUNK *
sizeof(kmp_indirect_lock_t));
3253 __kmp_i_lock_table.next = 0;
3256 __kmp_indirect_lock_size[locktag_ticket] =
sizeof(kmp_ticket_lock_t);
3257 __kmp_indirect_lock_size[locktag_queuing] =
sizeof(kmp_queuing_lock_t);
3258 #if KMP_USE_ADAPTIVE_LOCKS 3259 __kmp_indirect_lock_size[locktag_adaptive] =
sizeof(kmp_adaptive_lock_t);
3261 __kmp_indirect_lock_size[locktag_drdpa] =
sizeof(kmp_drdpa_lock_t);
3263 __kmp_indirect_lock_size[locktag_rtm] =
sizeof(kmp_queuing_lock_t);
3265 __kmp_indirect_lock_size[locktag_nested_tas] =
sizeof(kmp_tas_lock_t);
3267 __kmp_indirect_lock_size[locktag_nested_futex] =
sizeof(kmp_futex_lock_t);
3269 __kmp_indirect_lock_size[locktag_nested_ticket] =
sizeof(kmp_ticket_lock_t);
3270 __kmp_indirect_lock_size[locktag_nested_queuing] =
sizeof(kmp_queuing_lock_t);
3271 __kmp_indirect_lock_size[locktag_nested_drdpa] =
sizeof(kmp_drdpa_lock_t);
3274 #define fill_jumps(table, expand, sep) \ 3276 table[locktag##sep##ticket] = expand(ticket); \ 3277 table[locktag##sep##queuing] = expand(queuing); \ 3278 table[locktag##sep##drdpa] = expand(drdpa); \ 3281 #if KMP_USE_ADAPTIVE_LOCKS 3282 #define fill_table(table, expand) \ 3284 fill_jumps(table, expand, _); \ 3285 table[locktag_adaptive] = expand(queuing); \ 3286 fill_jumps(table, expand, _nested_); \ 3289 #define fill_table(table, expand) \ 3291 fill_jumps(table, expand, _); \ 3292 fill_jumps(table, expand, _nested_); \ 3294 #endif // KMP_USE_ADAPTIVE_LOCKS 3297 (void (*)(kmp_user_lock_p, const ident_t *)) __kmp_set_##l##_lock_location 3298 fill_table(__kmp_indirect_set_location, expand);
3301 (void (*)(kmp_user_lock_p, kmp_lock_flags_t)) __kmp_set_##l##_lock_flags 3302 fill_table(__kmp_indirect_set_flags, expand);
3305 (const ident_t *(*)(kmp_user_lock_p)) __kmp_get_##l##_lock_location 3306 fill_table(__kmp_indirect_get_location, expand);
3309 (kmp_lock_flags_t(*)(kmp_user_lock_p)) __kmp_get_##l##_lock_flags 3310 fill_table(__kmp_indirect_get_flags, expand);
3313 __kmp_init_user_locks = TRUE;
3317 void __kmp_cleanup_indirect_user_locks() {
3323 for (k = 0; k < KMP_NUM_I_LOCKS; ++k) {
3324 kmp_indirect_lock_t *l = __kmp_indirect_lock_pool[k];
3326 kmp_indirect_lock_t *ll = l;
3327 l = (kmp_indirect_lock_t *)l->lock->pool.next;
3328 KA_TRACE(20, (
"__kmp_cleanup_indirect_user_locks: freeing %p from pool\n",
3330 __kmp_free(ll->lock);
3333 __kmp_indirect_lock_pool[k] = NULL;
3336 for (i = 0; i < __kmp_i_lock_table.next; i++) {
3337 kmp_indirect_lock_t *l = KMP_GET_I_LOCK(i);
3338 if (l->lock != NULL) {
3340 KMP_I_LOCK_FUNC(l, destroy)(l->lock);
3343 (
"__kmp_cleanup_indirect_user_locks: destroy/freeing %p from table\n",
3345 __kmp_free(l->lock);
3349 for (i = 0; i < __kmp_i_lock_table.size / KMP_I_LOCK_CHUNK; i++)
3350 __kmp_free(__kmp_i_lock_table.table[i]);
3351 __kmp_free(__kmp_i_lock_table.table);
3353 __kmp_init_user_locks = FALSE;
3356 enum kmp_lock_kind __kmp_user_lock_kind = lk_default;
3357 int __kmp_num_locks_in_block = 1;
3359 #else // KMP_USE_DYNAMIC_LOCK 3361 static void __kmp_init_tas_lock_with_checks(kmp_tas_lock_t *lck) {
3362 __kmp_init_tas_lock(lck);
3365 static void __kmp_init_nested_tas_lock_with_checks(kmp_tas_lock_t *lck) {
3366 __kmp_init_nested_tas_lock(lck);
3370 static void __kmp_init_futex_lock_with_checks(kmp_futex_lock_t *lck) {
3371 __kmp_init_futex_lock(lck);
3374 static void __kmp_init_nested_futex_lock_with_checks(kmp_futex_lock_t *lck) {
3375 __kmp_init_nested_futex_lock(lck);
3379 static int __kmp_is_ticket_lock_initialized(kmp_ticket_lock_t *lck) {
3380 return lck == lck->lk.self;
3383 static void __kmp_init_ticket_lock_with_checks(kmp_ticket_lock_t *lck) {
3384 __kmp_init_ticket_lock(lck);
3387 static void __kmp_init_nested_ticket_lock_with_checks(kmp_ticket_lock_t *lck) {
3388 __kmp_init_nested_ticket_lock(lck);
3391 static int __kmp_is_queuing_lock_initialized(kmp_queuing_lock_t *lck) {
3392 return lck == lck->lk.initialized;
3395 static void __kmp_init_queuing_lock_with_checks(kmp_queuing_lock_t *lck) {
3396 __kmp_init_queuing_lock(lck);
3400 __kmp_init_nested_queuing_lock_with_checks(kmp_queuing_lock_t *lck) {
3401 __kmp_init_nested_queuing_lock(lck);
3404 #if KMP_USE_ADAPTIVE_LOCKS 3405 static void __kmp_init_adaptive_lock_with_checks(kmp_adaptive_lock_t *lck) {
3406 __kmp_init_adaptive_lock(lck);
3410 static int __kmp_is_drdpa_lock_initialized(kmp_drdpa_lock_t *lck) {
3411 return lck == lck->lk.initialized;
3414 static void __kmp_init_drdpa_lock_with_checks(kmp_drdpa_lock_t *lck) {
3415 __kmp_init_drdpa_lock(lck);
3418 static void __kmp_init_nested_drdpa_lock_with_checks(kmp_drdpa_lock_t *lck) {
3419 __kmp_init_nested_drdpa_lock(lck);
3426 enum kmp_lock_kind __kmp_user_lock_kind = lk_default;
3428 size_t __kmp_base_user_lock_size = 0;
3429 size_t __kmp_user_lock_size = 0;
3431 kmp_int32 (*__kmp_get_user_lock_owner_)(kmp_user_lock_p lck) = NULL;
3432 int (*__kmp_acquire_user_lock_with_checks_)(kmp_user_lock_p lck,
3433 kmp_int32 gtid) = NULL;
3435 int (*__kmp_test_user_lock_with_checks_)(kmp_user_lock_p lck,
3436 kmp_int32 gtid) = NULL;
3437 int (*__kmp_release_user_lock_with_checks_)(kmp_user_lock_p lck,
3438 kmp_int32 gtid) = NULL;
3439 void (*__kmp_init_user_lock_with_checks_)(kmp_user_lock_p lck) = NULL;
3440 void (*__kmp_destroy_user_lock_)(kmp_user_lock_p lck) = NULL;
3441 void (*__kmp_destroy_user_lock_with_checks_)(kmp_user_lock_p lck) = NULL;
3442 int (*__kmp_acquire_nested_user_lock_with_checks_)(kmp_user_lock_p lck,
3443 kmp_int32 gtid) = NULL;
3445 int (*__kmp_test_nested_user_lock_with_checks_)(kmp_user_lock_p lck,
3446 kmp_int32 gtid) = NULL;
3447 int (*__kmp_release_nested_user_lock_with_checks_)(kmp_user_lock_p lck,
3448 kmp_int32 gtid) = NULL;
3449 void (*__kmp_init_nested_user_lock_with_checks_)(kmp_user_lock_p lck) = NULL;
3450 void (*__kmp_destroy_nested_user_lock_with_checks_)(kmp_user_lock_p lck) = NULL;
3452 int (*__kmp_is_user_lock_initialized_)(kmp_user_lock_p lck) = NULL;
3453 const ident_t *(*__kmp_get_user_lock_location_)(kmp_user_lock_p lck) = NULL;
3454 void (*__kmp_set_user_lock_location_)(kmp_user_lock_p lck,
3456 kmp_lock_flags_t (*__kmp_get_user_lock_flags_)(kmp_user_lock_p lck) = NULL;
3457 void (*__kmp_set_user_lock_flags_)(kmp_user_lock_p lck,
3458 kmp_lock_flags_t flags) = NULL;
3460 void __kmp_set_user_lock_vptrs(kmp_lock_kind_t user_lock_kind) {
3461 switch (user_lock_kind) {
3467 __kmp_base_user_lock_size =
sizeof(kmp_base_tas_lock_t);
3468 __kmp_user_lock_size =
sizeof(kmp_tas_lock_t);
3470 __kmp_get_user_lock_owner_ =
3471 (kmp_int32(*)(kmp_user_lock_p))(&__kmp_get_tas_lock_owner);
3473 if (__kmp_env_consistency_check) {
3474 KMP_BIND_USER_LOCK_WITH_CHECKS(tas);
3475 KMP_BIND_NESTED_USER_LOCK_WITH_CHECKS(tas);
3477 KMP_BIND_USER_LOCK(tas);
3478 KMP_BIND_NESTED_USER_LOCK(tas);
3481 __kmp_destroy_user_lock_ =
3482 (void (*)(kmp_user_lock_p))(&__kmp_destroy_tas_lock);
3484 __kmp_is_user_lock_initialized_ = (int (*)(kmp_user_lock_p))NULL;
3486 __kmp_get_user_lock_location_ = (
const ident_t *(*)(kmp_user_lock_p))NULL;
3488 __kmp_set_user_lock_location_ =
3489 (void (*)(kmp_user_lock_p,
const ident_t *))NULL;
3491 __kmp_get_user_lock_flags_ = (kmp_lock_flags_t(*)(kmp_user_lock_p))NULL;
3493 __kmp_set_user_lock_flags_ =
3494 (void (*)(kmp_user_lock_p, kmp_lock_flags_t))NULL;
3500 __kmp_base_user_lock_size =
sizeof(kmp_base_futex_lock_t);
3501 __kmp_user_lock_size =
sizeof(kmp_futex_lock_t);
3503 __kmp_get_user_lock_owner_ =
3504 (kmp_int32(*)(kmp_user_lock_p))(&__kmp_get_futex_lock_owner);
3506 if (__kmp_env_consistency_check) {
3507 KMP_BIND_USER_LOCK_WITH_CHECKS(futex);
3508 KMP_BIND_NESTED_USER_LOCK_WITH_CHECKS(futex);
3510 KMP_BIND_USER_LOCK(futex);
3511 KMP_BIND_NESTED_USER_LOCK(futex);
3514 __kmp_destroy_user_lock_ =
3515 (void (*)(kmp_user_lock_p))(&__kmp_destroy_futex_lock);
3517 __kmp_is_user_lock_initialized_ = (int (*)(kmp_user_lock_p))NULL;
3519 __kmp_get_user_lock_location_ = (
const ident_t *(*)(kmp_user_lock_p))NULL;
3521 __kmp_set_user_lock_location_ =
3522 (void (*)(kmp_user_lock_p,
const ident_t *))NULL;
3524 __kmp_get_user_lock_flags_ = (kmp_lock_flags_t(*)(kmp_user_lock_p))NULL;
3526 __kmp_set_user_lock_flags_ =
3527 (void (*)(kmp_user_lock_p, kmp_lock_flags_t))NULL;
3530 #endif // KMP_USE_FUTEX 3533 __kmp_base_user_lock_size =
sizeof(kmp_base_ticket_lock_t);
3534 __kmp_user_lock_size =
sizeof(kmp_ticket_lock_t);
3536 __kmp_get_user_lock_owner_ =
3537 (kmp_int32(*)(kmp_user_lock_p))(&__kmp_get_ticket_lock_owner);
3539 if (__kmp_env_consistency_check) {
3540 KMP_BIND_USER_LOCK_WITH_CHECKS(ticket);
3541 KMP_BIND_NESTED_USER_LOCK_WITH_CHECKS(ticket);
3543 KMP_BIND_USER_LOCK(ticket);
3544 KMP_BIND_NESTED_USER_LOCK(ticket);
3547 __kmp_destroy_user_lock_ =
3548 (void (*)(kmp_user_lock_p))(&__kmp_destroy_ticket_lock);
3550 __kmp_is_user_lock_initialized_ =
3551 (int (*)(kmp_user_lock_p))(&__kmp_is_ticket_lock_initialized);
3553 __kmp_get_user_lock_location_ =
3554 (
const ident_t *(*)(kmp_user_lock_p))(&__kmp_get_ticket_lock_location);
3556 __kmp_set_user_lock_location_ = (void (*)(
3557 kmp_user_lock_p,
const ident_t *))(&__kmp_set_ticket_lock_location);
3559 __kmp_get_user_lock_flags_ =
3560 (kmp_lock_flags_t(*)(kmp_user_lock_p))(&__kmp_get_ticket_lock_flags);
3562 __kmp_set_user_lock_flags_ = (void (*)(kmp_user_lock_p, kmp_lock_flags_t))(
3563 &__kmp_set_ticket_lock_flags);
3567 __kmp_base_user_lock_size =
sizeof(kmp_base_queuing_lock_t);
3568 __kmp_user_lock_size =
sizeof(kmp_queuing_lock_t);
3570 __kmp_get_user_lock_owner_ =
3571 (kmp_int32(*)(kmp_user_lock_p))(&__kmp_get_queuing_lock_owner);
3573 if (__kmp_env_consistency_check) {
3574 KMP_BIND_USER_LOCK_WITH_CHECKS(queuing);
3575 KMP_BIND_NESTED_USER_LOCK_WITH_CHECKS(queuing);
3577 KMP_BIND_USER_LOCK(queuing);
3578 KMP_BIND_NESTED_USER_LOCK(queuing);
3581 __kmp_destroy_user_lock_ =
3582 (void (*)(kmp_user_lock_p))(&__kmp_destroy_queuing_lock);
3584 __kmp_is_user_lock_initialized_ =
3585 (int (*)(kmp_user_lock_p))(&__kmp_is_queuing_lock_initialized);
3587 __kmp_get_user_lock_location_ =
3588 (
const ident_t *(*)(kmp_user_lock_p))(&__kmp_get_queuing_lock_location);
3590 __kmp_set_user_lock_location_ = (void (*)(
3591 kmp_user_lock_p,
const ident_t *))(&__kmp_set_queuing_lock_location);
3593 __kmp_get_user_lock_flags_ =
3594 (kmp_lock_flags_t(*)(kmp_user_lock_p))(&__kmp_get_queuing_lock_flags);
3596 __kmp_set_user_lock_flags_ = (void (*)(kmp_user_lock_p, kmp_lock_flags_t))(
3597 &__kmp_set_queuing_lock_flags);
3600 #if KMP_USE_ADAPTIVE_LOCKS 3602 __kmp_base_user_lock_size =
sizeof(kmp_base_adaptive_lock_t);
3603 __kmp_user_lock_size =
sizeof(kmp_adaptive_lock_t);
3605 __kmp_get_user_lock_owner_ =
3606 (kmp_int32(*)(kmp_user_lock_p))(&__kmp_get_queuing_lock_owner);
3608 if (__kmp_env_consistency_check) {
3609 KMP_BIND_USER_LOCK_WITH_CHECKS(adaptive);
3611 KMP_BIND_USER_LOCK(adaptive);
3614 __kmp_destroy_user_lock_ =
3615 (void (*)(kmp_user_lock_p))(&__kmp_destroy_adaptive_lock);
3617 __kmp_is_user_lock_initialized_ =
3618 (int (*)(kmp_user_lock_p))(&__kmp_is_queuing_lock_initialized);
3620 __kmp_get_user_lock_location_ =
3621 (
const ident_t *(*)(kmp_user_lock_p))(&__kmp_get_queuing_lock_location);
3623 __kmp_set_user_lock_location_ = (void (*)(
3624 kmp_user_lock_p,
const ident_t *))(&__kmp_set_queuing_lock_location);
3626 __kmp_get_user_lock_flags_ =
3627 (kmp_lock_flags_t(*)(kmp_user_lock_p))(&__kmp_get_queuing_lock_flags);
3629 __kmp_set_user_lock_flags_ = (void (*)(kmp_user_lock_p, kmp_lock_flags_t))(
3630 &__kmp_set_queuing_lock_flags);
3633 #endif // KMP_USE_ADAPTIVE_LOCKS 3636 __kmp_base_user_lock_size =
sizeof(kmp_base_drdpa_lock_t);
3637 __kmp_user_lock_size =
sizeof(kmp_drdpa_lock_t);
3639 __kmp_get_user_lock_owner_ =
3640 (kmp_int32(*)(kmp_user_lock_p))(&__kmp_get_drdpa_lock_owner);
3642 if (__kmp_env_consistency_check) {
3643 KMP_BIND_USER_LOCK_WITH_CHECKS(drdpa);
3644 KMP_BIND_NESTED_USER_LOCK_WITH_CHECKS(drdpa);
3646 KMP_BIND_USER_LOCK(drdpa);
3647 KMP_BIND_NESTED_USER_LOCK(drdpa);
3650 __kmp_destroy_user_lock_ =
3651 (void (*)(kmp_user_lock_p))(&__kmp_destroy_drdpa_lock);
3653 __kmp_is_user_lock_initialized_ =
3654 (int (*)(kmp_user_lock_p))(&__kmp_is_drdpa_lock_initialized);
3656 __kmp_get_user_lock_location_ =
3657 (
const ident_t *(*)(kmp_user_lock_p))(&__kmp_get_drdpa_lock_location);
3659 __kmp_set_user_lock_location_ = (void (*)(
3660 kmp_user_lock_p,
const ident_t *))(&__kmp_set_drdpa_lock_location);
3662 __kmp_get_user_lock_flags_ =
3663 (kmp_lock_flags_t(*)(kmp_user_lock_p))(&__kmp_get_drdpa_lock_flags);
3665 __kmp_set_user_lock_flags_ = (void (*)(kmp_user_lock_p, kmp_lock_flags_t))(
3666 &__kmp_set_drdpa_lock_flags);
3674 kmp_lock_table_t __kmp_user_lock_table = {1, 0, NULL};
3675 kmp_user_lock_p __kmp_lock_pool = NULL;
3678 kmp_block_of_locks *__kmp_lock_blocks = NULL;
3679 int __kmp_num_locks_in_block = 1;
3681 static kmp_lock_index_t __kmp_lock_table_insert(kmp_user_lock_p lck) {
3683 kmp_lock_index_t index;
3684 if (__kmp_user_lock_table.used >= __kmp_user_lock_table.allocated) {
3685 kmp_lock_index_t size;
3686 kmp_user_lock_p *table;
3688 if (__kmp_user_lock_table.allocated == 0) {
3691 size = __kmp_user_lock_table.allocated * 2;
3693 table = (kmp_user_lock_p *)__kmp_allocate(
sizeof(kmp_user_lock_p) * size);
3694 KMP_MEMCPY(table + 1, __kmp_user_lock_table.table + 1,
3695 sizeof(kmp_user_lock_p) * (__kmp_user_lock_table.used - 1));
3696 table[0] = (kmp_user_lock_p)__kmp_user_lock_table.table;
3701 __kmp_user_lock_table.table = table;
3702 __kmp_user_lock_table.allocated = size;
3704 KMP_DEBUG_ASSERT(__kmp_user_lock_table.used <
3705 __kmp_user_lock_table.allocated);
3706 index = __kmp_user_lock_table.used;
3707 __kmp_user_lock_table.table[index] = lck;
3708 ++__kmp_user_lock_table.used;
3712 static kmp_user_lock_p __kmp_lock_block_allocate() {
3714 static int last_index = 0;
3715 if ((last_index >= __kmp_num_locks_in_block) || (__kmp_lock_blocks == NULL)) {
3719 KMP_DEBUG_ASSERT(__kmp_user_lock_size > 0);
3720 size_t space_for_locks = __kmp_user_lock_size * __kmp_num_locks_in_block;
3722 (
char *)__kmp_allocate(space_for_locks +
sizeof(kmp_block_of_locks));
3724 kmp_block_of_locks *new_block =
3725 (kmp_block_of_locks *)(&buffer[space_for_locks]);
3726 new_block->next_block = __kmp_lock_blocks;
3727 new_block->locks = (
void *)buffer;
3730 __kmp_lock_blocks = new_block;
3732 kmp_user_lock_p ret = (kmp_user_lock_p)(&(
3733 ((
char *)(__kmp_lock_blocks->locks))[last_index * __kmp_user_lock_size]));
3740 kmp_user_lock_p __kmp_user_lock_allocate(
void **user_lock, kmp_int32 gtid,
3741 kmp_lock_flags_t flags) {
3742 kmp_user_lock_p lck;
3743 kmp_lock_index_t index;
3744 KMP_DEBUG_ASSERT(user_lock);
3746 __kmp_acquire_lock(&__kmp_global_lock, gtid);
3748 if (__kmp_lock_pool == NULL) {
3753 ANNOTATE_IGNORE_WRITES_BEGIN();
3754 if (__kmp_num_locks_in_block <= 1) {
3755 lck = (kmp_user_lock_p)__kmp_allocate(__kmp_user_lock_size);
3757 lck = __kmp_lock_block_allocate();
3759 ANNOTATE_IGNORE_WRITES_END();
3763 index = __kmp_lock_table_insert(lck);
3766 lck = __kmp_lock_pool;
3767 index = __kmp_lock_pool->pool.index;
3768 __kmp_lock_pool = __kmp_lock_pool->pool.next;
3773 if (OMP_LOCK_T_SIZE <
sizeof(
void *)) {
3774 *((kmp_lock_index_t *)user_lock) = index;
3776 *((kmp_user_lock_p *)user_lock) = lck;
3780 __kmp_set_user_lock_flags(lck, flags);
3782 __kmp_release_lock(&__kmp_global_lock, gtid);
3788 void __kmp_user_lock_free(
void **user_lock, kmp_int32 gtid,
3789 kmp_user_lock_p lck) {
3790 KMP_DEBUG_ASSERT(user_lock != NULL);
3791 KMP_DEBUG_ASSERT(lck != NULL);
3793 __kmp_acquire_lock(&__kmp_global_lock, gtid);
3795 lck->pool.next = __kmp_lock_pool;
3796 __kmp_lock_pool = lck;
3797 if (OMP_LOCK_T_SIZE <
sizeof(
void *)) {
3798 kmp_lock_index_t index = *((kmp_lock_index_t *)user_lock);
3799 KMP_DEBUG_ASSERT(0 < index && index <= __kmp_user_lock_table.used);
3800 lck->pool.index = index;
3803 __kmp_release_lock(&__kmp_global_lock, gtid);
3806 kmp_user_lock_p __kmp_lookup_user_lock(
void **user_lock,
char const *func) {
3807 kmp_user_lock_p lck = NULL;
3809 if (__kmp_env_consistency_check) {
3810 if (user_lock == NULL) {
3811 KMP_FATAL(LockIsUninitialized, func);
3815 if (OMP_LOCK_T_SIZE <
sizeof(
void *)) {
3816 kmp_lock_index_t index = *((kmp_lock_index_t *)user_lock);
3817 if (__kmp_env_consistency_check) {
3818 if (!(0 < index && index < __kmp_user_lock_table.used)) {
3819 KMP_FATAL(LockIsUninitialized, func);
3822 KMP_DEBUG_ASSERT(0 < index && index < __kmp_user_lock_table.used);
3823 KMP_DEBUG_ASSERT(__kmp_user_lock_size > 0);
3824 lck = __kmp_user_lock_table.table[index];
3826 lck = *((kmp_user_lock_p *)user_lock);
3829 if (__kmp_env_consistency_check) {
3831 KMP_FATAL(LockIsUninitialized, func);
3838 void __kmp_cleanup_user_locks(
void) {
3841 __kmp_lock_pool = NULL;
3843 #define IS_CRITICAL(lck) \ 3844 ((__kmp_get_user_lock_flags_ != NULL) && \ 3845 ((*__kmp_get_user_lock_flags_)(lck)&kmp_lf_critical_section)) 3870 while (__kmp_user_lock_table.used > 1) {
3875 kmp_user_lock_p lck =
3876 __kmp_user_lock_table.table[--__kmp_user_lock_table.used];
3878 if ((__kmp_is_user_lock_initialized_ != NULL) &&
3879 (*__kmp_is_user_lock_initialized_)(lck)) {
3883 if (__kmp_env_consistency_check && (!IS_CRITICAL(lck)) &&
3884 ((loc = __kmp_get_user_lock_location(lck)) != NULL) &&
3886 kmp_str_loc_t str_loc = __kmp_str_loc_init(loc->
psource, 0);
3887 KMP_WARNING(CnsLockNotDestroyed, str_loc.file, str_loc.line);
3888 __kmp_str_loc_free(&str_loc);
3892 if (IS_CRITICAL(lck)) {
3895 (
"__kmp_cleanup_user_locks: free critical section lock %p (%p)\n",
3896 lck, *(
void **)lck));
3898 KA_TRACE(20, (
"__kmp_cleanup_user_locks: free lock %p (%p)\n", lck,
3904 __kmp_destroy_user_lock(lck);
3908 if (__kmp_lock_blocks == NULL) {
3916 kmp_user_lock_p *table_ptr = __kmp_user_lock_table.table;
3917 __kmp_user_lock_table.table = NULL;
3918 __kmp_user_lock_table.allocated = 0;
3920 while (table_ptr != NULL) {
3923 kmp_user_lock_p *next = (kmp_user_lock_p *)(table_ptr[0]);
3924 __kmp_free(table_ptr);
3929 kmp_block_of_locks_t *block_ptr = __kmp_lock_blocks;
3930 __kmp_lock_blocks = NULL;
3932 while (block_ptr != NULL) {
3933 kmp_block_of_locks_t *next = block_ptr->next_block;
3934 __kmp_free(block_ptr->locks);
3939 TCW_4(__kmp_init_user_locks, FALSE);
3942 #endif // KMP_USE_DYNAMIC_LOCK