1 /*
2  * kmp_lock.cpp -- lock-related functions
3  */
4 
5 //===----------------------------------------------------------------------===//
6 //
7 //                     The LLVM Compiler Infrastructure
8 //
9 // This file is dual licensed under the MIT and the University of Illinois Open
10 // Source Licenses. See LICENSE.txt for details.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include <stddef.h>
15 #include <atomic>
16 
17 #include "kmp.h"
18 #include "kmp_i18n.h"
19 #include "kmp_io.h"
20 #include "kmp_itt.h"
21 #include "kmp_lock.h"
22 #include "kmp_wait_release.h"
23 
24 #include "tsan_annotations.h"
25 
26 #if KMP_USE_FUTEX
27 #include <sys/syscall.h>
28 #include <unistd.h>
29 // We should really include <futex.h>, but that causes compatibility problems on
30 // different Linux* OS distributions that either require that you include (or
31 // break when you try to include) <pci/types.h>. Since all we need is the two
32 // macros below (which are part of the kernel ABI, so can't change) we just
33 // define the constants here and don't include <futex.h>
34 #ifndef FUTEX_WAIT
35 #define FUTEX_WAIT 0
36 #endif
37 #ifndef FUTEX_WAKE
38 #define FUTEX_WAKE 1
39 #endif
40 #endif
41 
42 /* Implement spin locks for internal library use.             */
43 /* The algorithm implemented is Lamport's bakery lock [1974]. */
44 
45 void __kmp_validate_locks(void) {
46   int i;
47   kmp_uint32 x, y;
48 
49   /* Check to make sure unsigned arithmetic does wraps properly */
50   x = ~((kmp_uint32)0) - 2;
51   y = x - 2;
52 
53   for (i = 0; i < 8; ++i, ++x, ++y) {
54     kmp_uint32 z = (x - y);
55     KMP_ASSERT(z == 2);
56   }
57 
58   KMP_ASSERT(offsetof(kmp_base_queuing_lock, tail_id) % 8 == 0);
59 }
60 
61 /* ------------------------------------------------------------------------ */
62 /* test and set locks */
63 
64 // For the non-nested locks, we can only assume that the first 4 bytes were
65 // allocated, since gcc only allocates 4 bytes for omp_lock_t, and the Intel
66 // compiler only allocates a 4 byte pointer on IA-32 architecture.  On
67 // Windows* OS on Intel(R) 64, we can assume that all 8 bytes were allocated.
68 //
69 // gcc reserves >= 8 bytes for nested locks, so we can assume that the
70 // entire 8 bytes were allocated for nested locks on all 64-bit platforms.
71 
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;
74 }
75 
76 static inline bool __kmp_is_tas_lock_nestable(kmp_tas_lock_t *lck) {
77   return lck->lk.depth_locked != -1;
78 }
79 
80 __forceinline static int
81 __kmp_acquire_tas_lock_timed_template(kmp_tas_lock_t *lck, kmp_int32 gtid) {
82   KMP_MB();
83 
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);
88 /* else __kmp_printf( "." );*/
89 #endif /* USE_LOCK_PROFILE */
90 
91   kmp_int32 tas_free = KMP_LOCK_FREE(tas);
92   kmp_int32 tas_busy = KMP_LOCK_BUSY(gtid + 1, tas);
93 
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;
98   }
99 
100   kmp_uint32 spins;
101   KMP_FSYNC_PREPARE(lck);
102   KMP_INIT_YIELD(spins);
103   if (TCR_4(__kmp_nth) > (__kmp_avail_proc ? __kmp_avail_proc : __kmp_xproc)) {
104     KMP_YIELD(TRUE);
105   } else {
106     KMP_YIELD_SPIN(spins);
107   }
108 
109   kmp_backoff_t backoff = __kmp_spin_backoff_params;
110   while (KMP_ATOMIC_LD_RLX(&lck->lk.poll) != tas_free ||
111          !__kmp_atomic_compare_store_acq(&lck->lk.poll, tas_free, tas_busy)) {
112     __kmp_spin_backoff(&backoff);
113     if (TCR_4(__kmp_nth) >
114         (__kmp_avail_proc ? __kmp_avail_proc : __kmp_xproc)) {
115       KMP_YIELD(TRUE);
116     } else {
117       KMP_YIELD_SPIN(spins);
118     }
119   }
120   KMP_FSYNC_ACQUIRED(lck);
121   return KMP_LOCK_ACQUIRED_FIRST;
122 }
123 
124 int __kmp_acquire_tas_lock(kmp_tas_lock_t *lck, kmp_int32 gtid) {
125   int retval = __kmp_acquire_tas_lock_timed_template(lck, gtid);
126   ANNOTATE_TAS_ACQUIRED(lck);
127   return retval;
128 }
129 
130 static int __kmp_acquire_tas_lock_with_checks(kmp_tas_lock_t *lck,
131                                               kmp_int32 gtid) {
132   char const *const func = "omp_set_lock";
133   if ((sizeof(kmp_tas_lock_t) <= OMP_LOCK_T_SIZE) &&
134       __kmp_is_tas_lock_nestable(lck)) {
135     KMP_FATAL(LockNestableUsedAsSimple, func);
136   }
137   if ((gtid >= 0) && (__kmp_get_tas_lock_owner(lck) == gtid)) {
138     KMP_FATAL(LockIsAlreadyOwned, func);
139   }
140   return __kmp_acquire_tas_lock(lck, gtid);
141 }
142 
143 int __kmp_test_tas_lock(kmp_tas_lock_t *lck, kmp_int32 gtid) {
144   kmp_int32 tas_free = KMP_LOCK_FREE(tas);
145   kmp_int32 tas_busy = KMP_LOCK_BUSY(gtid + 1, tas);
146   if (KMP_ATOMIC_LD_RLX(&lck->lk.poll) == tas_free &&
147       __kmp_atomic_compare_store_acq(&lck->lk.poll, tas_free, tas_busy)) {
148     KMP_FSYNC_ACQUIRED(lck);
149     return TRUE;
150   }
151   return FALSE;
152 }
153 
154 static int __kmp_test_tas_lock_with_checks(kmp_tas_lock_t *lck,
155                                            kmp_int32 gtid) {
156   char const *const func = "omp_test_lock";
157   if ((sizeof(kmp_tas_lock_t) <= OMP_LOCK_T_SIZE) &&
158       __kmp_is_tas_lock_nestable(lck)) {
159     KMP_FATAL(LockNestableUsedAsSimple, func);
160   }
161   return __kmp_test_tas_lock(lck, gtid);
162 }
163 
164 int __kmp_release_tas_lock(kmp_tas_lock_t *lck, kmp_int32 gtid) {
165   KMP_MB(); /* Flush all pending memory write invalidates.  */
166 
167   KMP_FSYNC_RELEASING(lck);
168   ANNOTATE_TAS_RELEASED(lck);
169   KMP_ATOMIC_ST_REL(&lck->lk.poll, KMP_LOCK_FREE(tas));
170   KMP_MB(); /* Flush all pending memory write invalidates.  */
171 
172   KMP_YIELD(TCR_4(__kmp_nth) >
173             (__kmp_avail_proc ? __kmp_avail_proc : __kmp_xproc));
174   return KMP_LOCK_RELEASED;
175 }
176 
177 static int __kmp_release_tas_lock_with_checks(kmp_tas_lock_t *lck,
178                                               kmp_int32 gtid) {
179   char const *const func = "omp_unset_lock";
180   KMP_MB(); /* in case another processor initialized lock */
181   if ((sizeof(kmp_tas_lock_t) <= OMP_LOCK_T_SIZE) &&
182       __kmp_is_tas_lock_nestable(lck)) {
183     KMP_FATAL(LockNestableUsedAsSimple, func);
184   }
185   if (__kmp_get_tas_lock_owner(lck) == -1) {
186     KMP_FATAL(LockUnsettingFree, func);
187   }
188   if ((gtid >= 0) && (__kmp_get_tas_lock_owner(lck) >= 0) &&
189       (__kmp_get_tas_lock_owner(lck) != gtid)) {
190     KMP_FATAL(LockUnsettingSetByAnother, func);
191   }
192   return __kmp_release_tas_lock(lck, gtid);
193 }
194 
195 void __kmp_init_tas_lock(kmp_tas_lock_t *lck) {
196   lck->lk.poll = KMP_LOCK_FREE(tas);
197 }
198 
199 static void __kmp_init_tas_lock_with_checks(kmp_tas_lock_t *lck) {
200   __kmp_init_tas_lock(lck);
201 }
202 
203 void __kmp_destroy_tas_lock(kmp_tas_lock_t *lck) { lck->lk.poll = 0; }
204 
205 static void __kmp_destroy_tas_lock_with_checks(kmp_tas_lock_t *lck) {
206   char const *const func = "omp_destroy_lock";
207   if ((sizeof(kmp_tas_lock_t) <= OMP_LOCK_T_SIZE) &&
208       __kmp_is_tas_lock_nestable(lck)) {
209     KMP_FATAL(LockNestableUsedAsSimple, func);
210   }
211   if (__kmp_get_tas_lock_owner(lck) != -1) {
212     KMP_FATAL(LockStillOwned, func);
213   }
214   __kmp_destroy_tas_lock(lck);
215 }
216 
217 // nested test and set locks
218 
219 int __kmp_acquire_nested_tas_lock(kmp_tas_lock_t *lck, kmp_int32 gtid) {
220   KMP_DEBUG_ASSERT(gtid >= 0);
221 
222   if (__kmp_get_tas_lock_owner(lck) == gtid) {
223     lck->lk.depth_locked += 1;
224     return KMP_LOCK_ACQUIRED_NEXT;
225   } else {
226     __kmp_acquire_tas_lock_timed_template(lck, gtid);
227     ANNOTATE_TAS_ACQUIRED(lck);
228     lck->lk.depth_locked = 1;
229     return KMP_LOCK_ACQUIRED_FIRST;
230   }
231 }
232 
233 static int __kmp_acquire_nested_tas_lock_with_checks(kmp_tas_lock_t *lck,
234                                                      kmp_int32 gtid) {
235   char const *const func = "omp_set_nest_lock";
236   if (!__kmp_is_tas_lock_nestable(lck)) {
237     KMP_FATAL(LockSimpleUsedAsNestable, func);
238   }
239   return __kmp_acquire_nested_tas_lock(lck, gtid);
240 }
241 
242 int __kmp_test_nested_tas_lock(kmp_tas_lock_t *lck, kmp_int32 gtid) {
243   int retval;
244 
245   KMP_DEBUG_ASSERT(gtid >= 0);
246 
247   if (__kmp_get_tas_lock_owner(lck) == gtid) {
248     retval = ++lck->lk.depth_locked;
249   } else if (!__kmp_test_tas_lock(lck, gtid)) {
250     retval = 0;
251   } else {
252     KMP_MB();
253     retval = lck->lk.depth_locked = 1;
254   }
255   return retval;
256 }
257 
258 static int __kmp_test_nested_tas_lock_with_checks(kmp_tas_lock_t *lck,
259                                                   kmp_int32 gtid) {
260   char const *const func = "omp_test_nest_lock";
261   if (!__kmp_is_tas_lock_nestable(lck)) {
262     KMP_FATAL(LockSimpleUsedAsNestable, func);
263   }
264   return __kmp_test_nested_tas_lock(lck, gtid);
265 }
266 
267 int __kmp_release_nested_tas_lock(kmp_tas_lock_t *lck, kmp_int32 gtid) {
268   KMP_DEBUG_ASSERT(gtid >= 0);
269 
270   KMP_MB();
271   if (--(lck->lk.depth_locked) == 0) {
272     __kmp_release_tas_lock(lck, gtid);
273     return KMP_LOCK_RELEASED;
274   }
275   return KMP_LOCK_STILL_HELD;
276 }
277 
278 static int __kmp_release_nested_tas_lock_with_checks(kmp_tas_lock_t *lck,
279                                                      kmp_int32 gtid) {
280   char const *const func = "omp_unset_nest_lock";
281   KMP_MB(); /* in case another processor initialized lock */
282   if (!__kmp_is_tas_lock_nestable(lck)) {
283     KMP_FATAL(LockSimpleUsedAsNestable, func);
284   }
285   if (__kmp_get_tas_lock_owner(lck) == -1) {
286     KMP_FATAL(LockUnsettingFree, func);
287   }
288   if (__kmp_get_tas_lock_owner(lck) != gtid) {
289     KMP_FATAL(LockUnsettingSetByAnother, func);
290   }
291   return __kmp_release_nested_tas_lock(lck, gtid);
292 }
293 
294 void __kmp_init_nested_tas_lock(kmp_tas_lock_t *lck) {
295   __kmp_init_tas_lock(lck);
296   lck->lk.depth_locked = 0; // >= 0 for nestable locks, -1 for simple locks
297 }
298 
299 static void __kmp_init_nested_tas_lock_with_checks(kmp_tas_lock_t *lck) {
300   __kmp_init_nested_tas_lock(lck);
301 }
302 
303 void __kmp_destroy_nested_tas_lock(kmp_tas_lock_t *lck) {
304   __kmp_destroy_tas_lock(lck);
305   lck->lk.depth_locked = 0;
306 }
307 
308 static void __kmp_destroy_nested_tas_lock_with_checks(kmp_tas_lock_t *lck) {
309   char const *const func = "omp_destroy_nest_lock";
310   if (!__kmp_is_tas_lock_nestable(lck)) {
311     KMP_FATAL(LockSimpleUsedAsNestable, func);
312   }
313   if (__kmp_get_tas_lock_owner(lck) != -1) {
314     KMP_FATAL(LockStillOwned, func);
315   }
316   __kmp_destroy_nested_tas_lock(lck);
317 }
318 
319 #if KMP_USE_FUTEX
320 
321 /* ------------------------------------------------------------------------ */
322 /* futex locks */
323 
324 // futex locks are really just test and set locks, with a different method
325 // of handling contention.  They take the same amount of space as test and
326 // set locks, and are allocated the same way (i.e. use the area allocated by
327 // the compiler for non-nested locks / allocate nested locks on the heap).
328 
329 static kmp_int32 __kmp_get_futex_lock_owner(kmp_futex_lock_t *lck) {
330   return KMP_LOCK_STRIP((TCR_4(lck->lk.poll) >> 1)) - 1;
331 }
332 
333 static inline bool __kmp_is_futex_lock_nestable(kmp_futex_lock_t *lck) {
334   return lck->lk.depth_locked != -1;
335 }
336 
337 __forceinline static int
338 __kmp_acquire_futex_lock_timed_template(kmp_futex_lock_t *lck, kmp_int32 gtid) {
339   kmp_int32 gtid_code = (gtid + 1) << 1;
340 
341   KMP_MB();
342 
343 #ifdef USE_LOCK_PROFILE
344   kmp_uint32 curr = KMP_LOCK_STRIP(TCR_4(lck->lk.poll));
345   if ((curr != 0) && (curr != gtid_code))
346     __kmp_printf("LOCK CONTENTION: %p\n", lck);
347 /* else __kmp_printf( "." );*/
348 #endif /* USE_LOCK_PROFILE */
349 
350   KMP_FSYNC_PREPARE(lck);
351   KA_TRACE(1000, ("__kmp_acquire_futex_lock: lck:%p(0x%x), T#%d entering\n",
352                   lck, lck->lk.poll, gtid));
353 
354   kmp_int32 poll_val;
355 
356   while ((poll_val = KMP_COMPARE_AND_STORE_RET32(
357               &(lck->lk.poll), KMP_LOCK_FREE(futex),
358               KMP_LOCK_BUSY(gtid_code, futex))) != KMP_LOCK_FREE(futex)) {
359 
360     kmp_int32 cond = KMP_LOCK_STRIP(poll_val) & 1;
361     KA_TRACE(
362         1000,
363         ("__kmp_acquire_futex_lock: lck:%p, T#%d poll_val = 0x%x cond = 0x%x\n",
364          lck, gtid, poll_val, cond));
365 
366     // NOTE: if you try to use the following condition for this branch
367     //
368     // if ( poll_val & 1 == 0 )
369     //
370     // Then the 12.0 compiler has a bug where the following block will
371     // always be skipped, regardless of the value of the LSB of poll_val.
372     if (!cond) {
373       // Try to set the lsb in the poll to indicate to the owner
374       // thread that they need to wake this thread up.
375       if (!KMP_COMPARE_AND_STORE_REL32(&(lck->lk.poll), poll_val,
376                                        poll_val | KMP_LOCK_BUSY(1, futex))) {
377         KA_TRACE(
378             1000,
379             ("__kmp_acquire_futex_lock: lck:%p(0x%x), T#%d can't set bit 0\n",
380              lck, lck->lk.poll, gtid));
381         continue;
382       }
383       poll_val |= KMP_LOCK_BUSY(1, futex);
384 
385       KA_TRACE(1000,
386                ("__kmp_acquire_futex_lock: lck:%p(0x%x), T#%d bit 0 set\n", lck,
387                 lck->lk.poll, gtid));
388     }
389 
390     KA_TRACE(
391         1000,
392         ("__kmp_acquire_futex_lock: lck:%p, T#%d before futex_wait(0x%x)\n",
393          lck, gtid, poll_val));
394 
395     kmp_int32 rc;
396     if ((rc = syscall(__NR_futex, &(lck->lk.poll), FUTEX_WAIT, poll_val, NULL,
397                       NULL, 0)) != 0) {
398       KA_TRACE(1000, ("__kmp_acquire_futex_lock: lck:%p, T#%d futex_wait(0x%x) "
399                       "failed (rc=%d errno=%d)\n",
400                       lck, gtid, poll_val, rc, errno));
401       continue;
402     }
403 
404     KA_TRACE(1000,
405              ("__kmp_acquire_futex_lock: lck:%p, T#%d after futex_wait(0x%x)\n",
406               lck, gtid, poll_val));
407     // This thread has now done a successful futex wait call and was entered on
408     // the OS futex queue.  We must now perform a futex wake call when releasing
409     // the lock, as we have no idea how many other threads are in the queue.
410     gtid_code |= 1;
411   }
412 
413   KMP_FSYNC_ACQUIRED(lck);
414   KA_TRACE(1000, ("__kmp_acquire_futex_lock: lck:%p(0x%x), T#%d exiting\n", lck,
415                   lck->lk.poll, gtid));
416   return KMP_LOCK_ACQUIRED_FIRST;
417 }
418 
419 int __kmp_acquire_futex_lock(kmp_futex_lock_t *lck, kmp_int32 gtid) {
420   int retval = __kmp_acquire_futex_lock_timed_template(lck, gtid);
421   ANNOTATE_FUTEX_ACQUIRED(lck);
422   return retval;
423 }
424 
425 static int __kmp_acquire_futex_lock_with_checks(kmp_futex_lock_t *lck,
426                                                 kmp_int32 gtid) {
427   char const *const func = "omp_set_lock";
428   if ((sizeof(kmp_futex_lock_t) <= OMP_LOCK_T_SIZE) &&
429       __kmp_is_futex_lock_nestable(lck)) {
430     KMP_FATAL(LockNestableUsedAsSimple, func);
431   }
432   if ((gtid >= 0) && (__kmp_get_futex_lock_owner(lck) == gtid)) {
433     KMP_FATAL(LockIsAlreadyOwned, func);
434   }
435   return __kmp_acquire_futex_lock(lck, gtid);
436 }
437 
438 int __kmp_test_futex_lock(kmp_futex_lock_t *lck, kmp_int32 gtid) {
439   if (KMP_COMPARE_AND_STORE_ACQ32(&(lck->lk.poll), KMP_LOCK_FREE(futex),
440                                   KMP_LOCK_BUSY((gtid + 1) << 1, futex))) {
441     KMP_FSYNC_ACQUIRED(lck);
442     return TRUE;
443   }
444   return FALSE;
445 }
446 
447 static int __kmp_test_futex_lock_with_checks(kmp_futex_lock_t *lck,
448                                              kmp_int32 gtid) {
449   char const *const func = "omp_test_lock";
450   if ((sizeof(kmp_futex_lock_t) <= OMP_LOCK_T_SIZE) &&
451       __kmp_is_futex_lock_nestable(lck)) {
452     KMP_FATAL(LockNestableUsedAsSimple, func);
453   }
454   return __kmp_test_futex_lock(lck, gtid);
455 }
456 
457 int __kmp_release_futex_lock(kmp_futex_lock_t *lck, kmp_int32 gtid) {
458   KMP_MB(); /* Flush all pending memory write invalidates.  */
459 
460   KA_TRACE(1000, ("__kmp_release_futex_lock: lck:%p(0x%x), T#%d entering\n",
461                   lck, lck->lk.poll, gtid));
462 
463   KMP_FSYNC_RELEASING(lck);
464   ANNOTATE_FUTEX_RELEASED(lck);
465 
466   kmp_int32 poll_val = KMP_XCHG_FIXED32(&(lck->lk.poll), KMP_LOCK_FREE(futex));
467 
468   KA_TRACE(1000,
469            ("__kmp_release_futex_lock: lck:%p, T#%d released poll_val = 0x%x\n",
470             lck, gtid, poll_val));
471 
472   if (KMP_LOCK_STRIP(poll_val) & 1) {
473     KA_TRACE(1000,
474              ("__kmp_release_futex_lock: lck:%p, T#%d futex_wake 1 thread\n",
475               lck, gtid));
476     syscall(__NR_futex, &(lck->lk.poll), FUTEX_WAKE, KMP_LOCK_BUSY(1, futex),
477             NULL, NULL, 0);
478   }
479 
480   KMP_MB(); /* Flush all pending memory write invalidates.  */
481 
482   KA_TRACE(1000, ("__kmp_release_futex_lock: lck:%p(0x%x), T#%d exiting\n", lck,
483                   lck->lk.poll, gtid));
484 
485   KMP_YIELD(TCR_4(__kmp_nth) >
486             (__kmp_avail_proc ? __kmp_avail_proc : __kmp_xproc));
487   return KMP_LOCK_RELEASED;
488 }
489 
490 static int __kmp_release_futex_lock_with_checks(kmp_futex_lock_t *lck,
491                                                 kmp_int32 gtid) {
492   char const *const func = "omp_unset_lock";
493   KMP_MB(); /* in case another processor initialized lock */
494   if ((sizeof(kmp_futex_lock_t) <= OMP_LOCK_T_SIZE) &&
495       __kmp_is_futex_lock_nestable(lck)) {
496     KMP_FATAL(LockNestableUsedAsSimple, func);
497   }
498   if (__kmp_get_futex_lock_owner(lck) == -1) {
499     KMP_FATAL(LockUnsettingFree, func);
500   }
501   if ((gtid >= 0) && (__kmp_get_futex_lock_owner(lck) >= 0) &&
502       (__kmp_get_futex_lock_owner(lck) != gtid)) {
503     KMP_FATAL(LockUnsettingSetByAnother, func);
504   }
505   return __kmp_release_futex_lock(lck, gtid);
506 }
507 
508 void __kmp_init_futex_lock(kmp_futex_lock_t *lck) {
509   TCW_4(lck->lk.poll, KMP_LOCK_FREE(futex));
510 }
511 
512 static void __kmp_init_futex_lock_with_checks(kmp_futex_lock_t *lck) {
513   __kmp_init_futex_lock(lck);
514 }
515 
516 void __kmp_destroy_futex_lock(kmp_futex_lock_t *lck) { lck->lk.poll = 0; }
517 
518 static void __kmp_destroy_futex_lock_with_checks(kmp_futex_lock_t *lck) {
519   char const *const func = "omp_destroy_lock";
520   if ((sizeof(kmp_futex_lock_t) <= OMP_LOCK_T_SIZE) &&
521       __kmp_is_futex_lock_nestable(lck)) {
522     KMP_FATAL(LockNestableUsedAsSimple, func);
523   }
524   if (__kmp_get_futex_lock_owner(lck) != -1) {
525     KMP_FATAL(LockStillOwned, func);
526   }
527   __kmp_destroy_futex_lock(lck);
528 }
529 
530 // nested futex locks
531 
532 int __kmp_acquire_nested_futex_lock(kmp_futex_lock_t *lck, kmp_int32 gtid) {
533   KMP_DEBUG_ASSERT(gtid >= 0);
534 
535   if (__kmp_get_futex_lock_owner(lck) == gtid) {
536     lck->lk.depth_locked += 1;
537     return KMP_LOCK_ACQUIRED_NEXT;
538   } else {
539     __kmp_acquire_futex_lock_timed_template(lck, gtid);
540     ANNOTATE_FUTEX_ACQUIRED(lck);
541     lck->lk.depth_locked = 1;
542     return KMP_LOCK_ACQUIRED_FIRST;
543   }
544 }
545 
546 static int __kmp_acquire_nested_futex_lock_with_checks(kmp_futex_lock_t *lck,
547                                                        kmp_int32 gtid) {
548   char const *const func = "omp_set_nest_lock";
549   if (!__kmp_is_futex_lock_nestable(lck)) {
550     KMP_FATAL(LockSimpleUsedAsNestable, func);
551   }
552   return __kmp_acquire_nested_futex_lock(lck, gtid);
553 }
554 
555 int __kmp_test_nested_futex_lock(kmp_futex_lock_t *lck, kmp_int32 gtid) {
556   int retval;
557 
558   KMP_DEBUG_ASSERT(gtid >= 0);
559 
560   if (__kmp_get_futex_lock_owner(lck) == gtid) {
561     retval = ++lck->lk.depth_locked;
562   } else if (!__kmp_test_futex_lock(lck, gtid)) {
563     retval = 0;
564   } else {
565     KMP_MB();
566     retval = lck->lk.depth_locked = 1;
567   }
568   return retval;
569 }
570 
571 static int __kmp_test_nested_futex_lock_with_checks(kmp_futex_lock_t *lck,
572                                                     kmp_int32 gtid) {
573   char const *const func = "omp_test_nest_lock";
574   if (!__kmp_is_futex_lock_nestable(lck)) {
575     KMP_FATAL(LockSimpleUsedAsNestable, func);
576   }
577   return __kmp_test_nested_futex_lock(lck, gtid);
578 }
579 
580 int __kmp_release_nested_futex_lock(kmp_futex_lock_t *lck, kmp_int32 gtid) {
581   KMP_DEBUG_ASSERT(gtid >= 0);
582 
583   KMP_MB();
584   if (--(lck->lk.depth_locked) == 0) {
585     __kmp_release_futex_lock(lck, gtid);
586     return KMP_LOCK_RELEASED;
587   }
588   return KMP_LOCK_STILL_HELD;
589 }
590 
591 static int __kmp_release_nested_futex_lock_with_checks(kmp_futex_lock_t *lck,
592                                                        kmp_int32 gtid) {
593   char const *const func = "omp_unset_nest_lock";
594   KMP_MB(); /* in case another processor initialized lock */
595   if (!__kmp_is_futex_lock_nestable(lck)) {
596     KMP_FATAL(LockSimpleUsedAsNestable, func);
597   }
598   if (__kmp_get_futex_lock_owner(lck) == -1) {
599     KMP_FATAL(LockUnsettingFree, func);
600   }
601   if (__kmp_get_futex_lock_owner(lck) != gtid) {
602     KMP_FATAL(LockUnsettingSetByAnother, func);
603   }
604   return __kmp_release_nested_futex_lock(lck, gtid);
605 }
606 
607 void __kmp_init_nested_futex_lock(kmp_futex_lock_t *lck) {
608   __kmp_init_futex_lock(lck);
609   lck->lk.depth_locked = 0; // >= 0 for nestable locks, -1 for simple locks
610 }
611 
612 static void __kmp_init_nested_futex_lock_with_checks(kmp_futex_lock_t *lck) {
613   __kmp_init_nested_futex_lock(lck);
614 }
615 
616 void __kmp_destroy_nested_futex_lock(kmp_futex_lock_t *lck) {
617   __kmp_destroy_futex_lock(lck);
618   lck->lk.depth_locked = 0;
619 }
620 
621 static void __kmp_destroy_nested_futex_lock_with_checks(kmp_futex_lock_t *lck) {
622   char const *const func = "omp_destroy_nest_lock";
623   if (!__kmp_is_futex_lock_nestable(lck)) {
624     KMP_FATAL(LockSimpleUsedAsNestable, func);
625   }
626   if (__kmp_get_futex_lock_owner(lck) != -1) {
627     KMP_FATAL(LockStillOwned, func);
628   }
629   __kmp_destroy_nested_futex_lock(lck);
630 }
631 
632 #endif // KMP_USE_FUTEX
633 
634 /* ------------------------------------------------------------------------ */
635 /* ticket (bakery) locks */
636 
637 static kmp_int32 __kmp_get_ticket_lock_owner(kmp_ticket_lock_t *lck) {
638   return std::atomic_load_explicit(&lck->lk.owner_id,
639                                    std::memory_order_relaxed) -
640          1;
641 }
642 
643 static inline bool __kmp_is_ticket_lock_nestable(kmp_ticket_lock_t *lck) {
644   return std::atomic_load_explicit(&lck->lk.depth_locked,
645                                    std::memory_order_relaxed) != -1;
646 }
647 
648 static kmp_uint32 __kmp_bakery_check(void *now_serving, kmp_uint32 my_ticket) {
649   return std::atomic_load_explicit((std::atomic<unsigned> *)now_serving,
650                                    std::memory_order_acquire) == my_ticket;
651 }
652 
653 __forceinline static int
654 __kmp_acquire_ticket_lock_timed_template(kmp_ticket_lock_t *lck,
655                                          kmp_int32 gtid) {
656   kmp_uint32 my_ticket = std::atomic_fetch_add_explicit(
657       &lck->lk.next_ticket, 1U, std::memory_order_relaxed);
658 
659 #ifdef USE_LOCK_PROFILE
660   if (std::atomic_load_explicit(&lck->lk.now_serving,
661                                 std::memory_order_relaxed) != my_ticket)
662     __kmp_printf("LOCK CONTENTION: %p\n", lck);
663 /* else __kmp_printf( "." );*/
664 #endif /* USE_LOCK_PROFILE */
665 
666   if (std::atomic_load_explicit(&lck->lk.now_serving,
667                                 std::memory_order_acquire) == my_ticket) {
668     return KMP_LOCK_ACQUIRED_FIRST;
669   }
670   KMP_WAIT_YIELD_PTR(&lck->lk.now_serving, my_ticket, __kmp_bakery_check, lck);
671   return KMP_LOCK_ACQUIRED_FIRST;
672 }
673 
674 int __kmp_acquire_ticket_lock(kmp_ticket_lock_t *lck, kmp_int32 gtid) {
675   int retval = __kmp_acquire_ticket_lock_timed_template(lck, gtid);
676   ANNOTATE_TICKET_ACQUIRED(lck);
677   return retval;
678 }
679 
680 static int __kmp_acquire_ticket_lock_with_checks(kmp_ticket_lock_t *lck,
681                                                  kmp_int32 gtid) {
682   char const *const func = "omp_set_lock";
683 
684   if (!std::atomic_load_explicit(&lck->lk.initialized,
685                                  std::memory_order_relaxed)) {
686     KMP_FATAL(LockIsUninitialized, func);
687   }
688   if (lck->lk.self != lck) {
689     KMP_FATAL(LockIsUninitialized, func);
690   }
691   if (__kmp_is_ticket_lock_nestable(lck)) {
692     KMP_FATAL(LockNestableUsedAsSimple, func);
693   }
694   if ((gtid >= 0) && (__kmp_get_ticket_lock_owner(lck) == gtid)) {
695     KMP_FATAL(LockIsAlreadyOwned, func);
696   }
697 
698   __kmp_acquire_ticket_lock(lck, gtid);
699 
700   std::atomic_store_explicit(&lck->lk.owner_id, gtid + 1,
701                              std::memory_order_relaxed);
702   return KMP_LOCK_ACQUIRED_FIRST;
703 }
704 
705 int __kmp_test_ticket_lock(kmp_ticket_lock_t *lck, kmp_int32 gtid) {
706   kmp_uint32 my_ticket = std::atomic_load_explicit(&lck->lk.next_ticket,
707                                                    std::memory_order_relaxed);
708 
709   if (std::atomic_load_explicit(&lck->lk.now_serving,
710                                 std::memory_order_relaxed) == my_ticket) {
711     kmp_uint32 next_ticket = my_ticket + 1;
712     if (std::atomic_compare_exchange_strong_explicit(
713             &lck->lk.next_ticket, &my_ticket, next_ticket,
714             std::memory_order_acquire, std::memory_order_acquire)) {
715       return TRUE;
716     }
717   }
718   return FALSE;
719 }
720 
721 static int __kmp_test_ticket_lock_with_checks(kmp_ticket_lock_t *lck,
722                                               kmp_int32 gtid) {
723   char const *const func = "omp_test_lock";
724 
725   if (!std::atomic_load_explicit(&lck->lk.initialized,
726                                  std::memory_order_relaxed)) {
727     KMP_FATAL(LockIsUninitialized, func);
728   }
729   if (lck->lk.self != lck) {
730     KMP_FATAL(LockIsUninitialized, func);
731   }
732   if (__kmp_is_ticket_lock_nestable(lck)) {
733     KMP_FATAL(LockNestableUsedAsSimple, func);
734   }
735 
736   int retval = __kmp_test_ticket_lock(lck, gtid);
737 
738   if (retval) {
739     std::atomic_store_explicit(&lck->lk.owner_id, gtid + 1,
740                                std::memory_order_relaxed);
741   }
742   return retval;
743 }
744 
745 int __kmp_release_ticket_lock(kmp_ticket_lock_t *lck, kmp_int32 gtid) {
746   kmp_uint32 distance = std::atomic_load_explicit(&lck->lk.next_ticket,
747                                                   std::memory_order_relaxed) -
748                         std::atomic_load_explicit(&lck->lk.now_serving,
749                                                   std::memory_order_relaxed);
750 
751   ANNOTATE_TICKET_RELEASED(lck);
752   std::atomic_fetch_add_explicit(&lck->lk.now_serving, 1U,
753                                  std::memory_order_release);
754 
755   KMP_YIELD(distance >
756             (kmp_uint32)(__kmp_avail_proc ? __kmp_avail_proc : __kmp_xproc));
757   return KMP_LOCK_RELEASED;
758 }
759 
760 static int __kmp_release_ticket_lock_with_checks(kmp_ticket_lock_t *lck,
761                                                  kmp_int32 gtid) {
762   char const *const func = "omp_unset_lock";
763 
764   if (!std::atomic_load_explicit(&lck->lk.initialized,
765                                  std::memory_order_relaxed)) {
766     KMP_FATAL(LockIsUninitialized, func);
767   }
768   if (lck->lk.self != lck) {
769     KMP_FATAL(LockIsUninitialized, func);
770   }
771   if (__kmp_is_ticket_lock_nestable(lck)) {
772     KMP_FATAL(LockNestableUsedAsSimple, func);
773   }
774   if (__kmp_get_ticket_lock_owner(lck) == -1) {
775     KMP_FATAL(LockUnsettingFree, func);
776   }
777   if ((gtid >= 0) && (__kmp_get_ticket_lock_owner(lck) >= 0) &&
778       (__kmp_get_ticket_lock_owner(lck) != gtid)) {
779     KMP_FATAL(LockUnsettingSetByAnother, func);
780   }
781   std::atomic_store_explicit(&lck->lk.owner_id, 0, std::memory_order_relaxed);
782   return __kmp_release_ticket_lock(lck, gtid);
783 }
784 
785 void __kmp_init_ticket_lock(kmp_ticket_lock_t *lck) {
786   lck->lk.location = NULL;
787   lck->lk.self = lck;
788   std::atomic_store_explicit(&lck->lk.next_ticket, 0U,
789                              std::memory_order_relaxed);
790   std::atomic_store_explicit(&lck->lk.now_serving, 0U,
791                              std::memory_order_relaxed);
792   std::atomic_store_explicit(
793       &lck->lk.owner_id, 0,
794       std::memory_order_relaxed); // no thread owns the lock.
795   std::atomic_store_explicit(
796       &lck->lk.depth_locked, -1,
797       std::memory_order_relaxed); // -1 => not a nested lock.
798   std::atomic_store_explicit(&lck->lk.initialized, true,
799                              std::memory_order_release);
800 }
801 
802 static void __kmp_init_ticket_lock_with_checks(kmp_ticket_lock_t *lck) {
803   __kmp_init_ticket_lock(lck);
804 }
805 
806 void __kmp_destroy_ticket_lock(kmp_ticket_lock_t *lck) {
807   std::atomic_store_explicit(&lck->lk.initialized, false,
808                              std::memory_order_release);
809   lck->lk.self = NULL;
810   lck->lk.location = NULL;
811   std::atomic_store_explicit(&lck->lk.next_ticket, 0U,
812                              std::memory_order_relaxed);
813   std::atomic_store_explicit(&lck->lk.now_serving, 0U,
814                              std::memory_order_relaxed);
815   std::atomic_store_explicit(&lck->lk.owner_id, 0, std::memory_order_relaxed);
816   std::atomic_store_explicit(&lck->lk.depth_locked, -1,
817                              std::memory_order_relaxed);
818 }
819 
820 static void __kmp_destroy_ticket_lock_with_checks(kmp_ticket_lock_t *lck) {
821   char const *const func = "omp_destroy_lock";
822 
823   if (!std::atomic_load_explicit(&lck->lk.initialized,
824                                  std::memory_order_relaxed)) {
825     KMP_FATAL(LockIsUninitialized, func);
826   }
827   if (lck->lk.self != lck) {
828     KMP_FATAL(LockIsUninitialized, func);
829   }
830   if (__kmp_is_ticket_lock_nestable(lck)) {
831     KMP_FATAL(LockNestableUsedAsSimple, func);
832   }
833   if (__kmp_get_ticket_lock_owner(lck) != -1) {
834     KMP_FATAL(LockStillOwned, func);
835   }
836   __kmp_destroy_ticket_lock(lck);
837 }
838 
839 // nested ticket locks
840 
841 int __kmp_acquire_nested_ticket_lock(kmp_ticket_lock_t *lck, kmp_int32 gtid) {
842   KMP_DEBUG_ASSERT(gtid >= 0);
843 
844   if (__kmp_get_ticket_lock_owner(lck) == gtid) {
845     std::atomic_fetch_add_explicit(&lck->lk.depth_locked, 1,
846                                    std::memory_order_relaxed);
847     return KMP_LOCK_ACQUIRED_NEXT;
848   } else {
849     __kmp_acquire_ticket_lock_timed_template(lck, gtid);
850     ANNOTATE_TICKET_ACQUIRED(lck);
851     std::atomic_store_explicit(&lck->lk.depth_locked, 1,
852                                std::memory_order_relaxed);
853     std::atomic_store_explicit(&lck->lk.owner_id, gtid + 1,
854                                std::memory_order_relaxed);
855     return KMP_LOCK_ACQUIRED_FIRST;
856   }
857 }
858 
859 static int __kmp_acquire_nested_ticket_lock_with_checks(kmp_ticket_lock_t *lck,
860                                                         kmp_int32 gtid) {
861   char const *const func = "omp_set_nest_lock";
862 
863   if (!std::atomic_load_explicit(&lck->lk.initialized,
864                                  std::memory_order_relaxed)) {
865     KMP_FATAL(LockIsUninitialized, func);
866   }
867   if (lck->lk.self != lck) {
868     KMP_FATAL(LockIsUninitialized, func);
869   }
870   if (!__kmp_is_ticket_lock_nestable(lck)) {
871     KMP_FATAL(LockSimpleUsedAsNestable, func);
872   }
873   return __kmp_acquire_nested_ticket_lock(lck, gtid);
874 }
875 
876 int __kmp_test_nested_ticket_lock(kmp_ticket_lock_t *lck, kmp_int32 gtid) {
877   int retval;
878 
879   KMP_DEBUG_ASSERT(gtid >= 0);
880 
881   if (__kmp_get_ticket_lock_owner(lck) == gtid) {
882     retval = std::atomic_fetch_add_explicit(&lck->lk.depth_locked, 1,
883                                             std::memory_order_relaxed) +
884              1;
885   } else if (!__kmp_test_ticket_lock(lck, gtid)) {
886     retval = 0;
887   } else {
888     std::atomic_store_explicit(&lck->lk.depth_locked, 1,
889                                std::memory_order_relaxed);
890     std::atomic_store_explicit(&lck->lk.owner_id, gtid + 1,
891                                std::memory_order_relaxed);
892     retval = 1;
893   }
894   return retval;
895 }
896 
897 static int __kmp_test_nested_ticket_lock_with_checks(kmp_ticket_lock_t *lck,
898                                                      kmp_int32 gtid) {
899   char const *const func = "omp_test_nest_lock";
900 
901   if (!std::atomic_load_explicit(&lck->lk.initialized,
902                                  std::memory_order_relaxed)) {
903     KMP_FATAL(LockIsUninitialized, func);
904   }
905   if (lck->lk.self != lck) {
906     KMP_FATAL(LockIsUninitialized, func);
907   }
908   if (!__kmp_is_ticket_lock_nestable(lck)) {
909     KMP_FATAL(LockSimpleUsedAsNestable, func);
910   }
911   return __kmp_test_nested_ticket_lock(lck, gtid);
912 }
913 
914 int __kmp_release_nested_ticket_lock(kmp_ticket_lock_t *lck, kmp_int32 gtid) {
915   KMP_DEBUG_ASSERT(gtid >= 0);
916 
917   if ((std::atomic_fetch_add_explicit(&lck->lk.depth_locked, -1,
918                                       std::memory_order_relaxed) -
919        1) == 0) {
920     std::atomic_store_explicit(&lck->lk.owner_id, 0, std::memory_order_relaxed);
921     __kmp_release_ticket_lock(lck, gtid);
922     return KMP_LOCK_RELEASED;
923   }
924   return KMP_LOCK_STILL_HELD;
925 }
926 
927 static int __kmp_release_nested_ticket_lock_with_checks(kmp_ticket_lock_t *lck,
928                                                         kmp_int32 gtid) {
929   char const *const func = "omp_unset_nest_lock";
930 
931   if (!std::atomic_load_explicit(&lck->lk.initialized,
932                                  std::memory_order_relaxed)) {
933     KMP_FATAL(LockIsUninitialized, func);
934   }
935   if (lck->lk.self != lck) {
936     KMP_FATAL(LockIsUninitialized, func);
937   }
938   if (!__kmp_is_ticket_lock_nestable(lck)) {
939     KMP_FATAL(LockSimpleUsedAsNestable, func);
940   }
941   if (__kmp_get_ticket_lock_owner(lck) == -1) {
942     KMP_FATAL(LockUnsettingFree, func);
943   }
944   if (__kmp_get_ticket_lock_owner(lck) != gtid) {
945     KMP_FATAL(LockUnsettingSetByAnother, func);
946   }
947   return __kmp_release_nested_ticket_lock(lck, gtid);
948 }
949 
950 void __kmp_init_nested_ticket_lock(kmp_ticket_lock_t *lck) {
951   __kmp_init_ticket_lock(lck);
952   std::atomic_store_explicit(&lck->lk.depth_locked, 0,
953                              std::memory_order_relaxed);
954   // >= 0 for nestable locks, -1 for simple locks
955 }
956 
957 static void __kmp_init_nested_ticket_lock_with_checks(kmp_ticket_lock_t *lck) {
958   __kmp_init_nested_ticket_lock(lck);
959 }
960 
961 void __kmp_destroy_nested_ticket_lock(kmp_ticket_lock_t *lck) {
962   __kmp_destroy_ticket_lock(lck);
963   std::atomic_store_explicit(&lck->lk.depth_locked, 0,
964                              std::memory_order_relaxed);
965 }
966 
967 static void
968 __kmp_destroy_nested_ticket_lock_with_checks(kmp_ticket_lock_t *lck) {
969   char const *const func = "omp_destroy_nest_lock";
970 
971   if (!std::atomic_load_explicit(&lck->lk.initialized,
972                                  std::memory_order_relaxed)) {
973     KMP_FATAL(LockIsUninitialized, func);
974   }
975   if (lck->lk.self != lck) {
976     KMP_FATAL(LockIsUninitialized, func);
977   }
978   if (!__kmp_is_ticket_lock_nestable(lck)) {
979     KMP_FATAL(LockSimpleUsedAsNestable, func);
980   }
981   if (__kmp_get_ticket_lock_owner(lck) != -1) {
982     KMP_FATAL(LockStillOwned, func);
983   }
984   __kmp_destroy_nested_ticket_lock(lck);
985 }
986 
987 // access functions to fields which don't exist for all lock kinds.
988 
989 static int __kmp_is_ticket_lock_initialized(kmp_ticket_lock_t *lck) {
990   return std::atomic_load_explicit(&lck->lk.initialized,
991                                    std::memory_order_relaxed) &&
992          (lck->lk.self == lck);
993 }
994 
995 static const ident_t *__kmp_get_ticket_lock_location(kmp_ticket_lock_t *lck) {
996   return lck->lk.location;
997 }
998 
999 static void __kmp_set_ticket_lock_location(kmp_ticket_lock_t *lck,
1000                                            const ident_t *loc) {
1001   lck->lk.location = loc;
1002 }
1003 
1004 static kmp_lock_flags_t __kmp_get_ticket_lock_flags(kmp_ticket_lock_t *lck) {
1005   return lck->lk.flags;
1006 }
1007 
1008 static void __kmp_set_ticket_lock_flags(kmp_ticket_lock_t *lck,
1009                                         kmp_lock_flags_t flags) {
1010   lck->lk.flags = flags;
1011 }
1012 
1013 /* ------------------------------------------------------------------------ */
1014 /* queuing locks */
1015 
1016 /* First the states
1017    (head,tail) =              0, 0  means lock is unheld, nobody on queue
1018                  UINT_MAX or -1, 0  means lock is held, nobody on queue
1019                               h, h  means lock held or about to transition,
1020                                     1 element on queue
1021                               h, t  h <> t, means lock is held or about to
1022                                     transition, >1 elements on queue
1023 
1024    Now the transitions
1025       Acquire(0,0)  = -1 ,0
1026       Release(0,0)  = Error
1027       Acquire(-1,0) =  h ,h    h > 0
1028       Release(-1,0) =  0 ,0
1029       Acquire(h,h)  =  h ,t    h > 0, t > 0, h <> t
1030       Release(h,h)  = -1 ,0    h > 0
1031       Acquire(h,t)  =  h ,t'   h > 0, t > 0, t' > 0, h <> t, h <> t', t <> t'
1032       Release(h,t)  =  h',t    h > 0, t > 0, h <> t, h <> h', h' maybe = t
1033 
1034    And pictorially
1035 
1036            +-----+
1037            | 0, 0|------- release -------> Error
1038            +-----+
1039              |  ^
1040       acquire|  |release
1041              |  |
1042              |  |
1043              v  |
1044            +-----+
1045            |-1, 0|
1046            +-----+
1047              |  ^
1048       acquire|  |release
1049              |  |
1050              |  |
1051              v  |
1052            +-----+
1053            | h, h|
1054            +-----+
1055              |  ^
1056       acquire|  |release
1057              |  |
1058              |  |
1059              v  |
1060            +-----+
1061            | h, t|----- acquire, release loopback ---+
1062            +-----+                                   |
1063                 ^                                    |
1064                 |                                    |
1065                 +------------------------------------+
1066  */
1067 
1068 #ifdef DEBUG_QUEUING_LOCKS
1069 
1070 /* Stuff for circular trace buffer */
1071 #define TRACE_BUF_ELE 1024
1072 static char traces[TRACE_BUF_ELE][128] = {0};
1073 static int tc = 0;
1074 #define TRACE_LOCK(X, Y)                                                       \
1075   KMP_SNPRINTF(traces[tc++ % TRACE_BUF_ELE], 128, "t%d at %s\n", X, Y);
1076 #define TRACE_LOCK_T(X, Y, Z)                                                  \
1077   KMP_SNPRINTF(traces[tc++ % TRACE_BUF_ELE], 128, "t%d at %s%d\n", X, Y, Z);
1078 #define TRACE_LOCK_HT(X, Y, Z, Q)                                              \
1079   KMP_SNPRINTF(traces[tc++ % TRACE_BUF_ELE], 128, "t%d at %s %d,%d\n", X, Y,   \
1080                Z, Q);
1081 
1082 static void __kmp_dump_queuing_lock(kmp_info_t *this_thr, kmp_int32 gtid,
1083                                     kmp_queuing_lock_t *lck, kmp_int32 head_id,
1084                                     kmp_int32 tail_id) {
1085   kmp_int32 t, i;
1086 
1087   __kmp_printf_no_lock("\n__kmp_dump_queuing_lock: TRACE BEGINS HERE! \n");
1088 
1089   i = tc % TRACE_BUF_ELE;
1090   __kmp_printf_no_lock("%s\n", traces[i]);
1091   i = (i + 1) % TRACE_BUF_ELE;
1092   while (i != (tc % TRACE_BUF_ELE)) {
1093     __kmp_printf_no_lock("%s", traces[i]);
1094     i = (i + 1) % TRACE_BUF_ELE;
1095   }
1096   __kmp_printf_no_lock("\n");
1097 
1098   __kmp_printf_no_lock("\n__kmp_dump_queuing_lock: gtid+1:%d, spin_here:%d, "
1099                        "next_wait:%d, head_id:%d, tail_id:%d\n",
1100                        gtid + 1, this_thr->th.th_spin_here,
1101                        this_thr->th.th_next_waiting, head_id, tail_id);
1102 
1103   __kmp_printf_no_lock("\t\thead: %d ", lck->lk.head_id);
1104 
1105   if (lck->lk.head_id >= 1) {
1106     t = __kmp_threads[lck->lk.head_id - 1]->th.th_next_waiting;
1107     while (t > 0) {
1108       __kmp_printf_no_lock("-> %d ", t);
1109       t = __kmp_threads[t - 1]->th.th_next_waiting;
1110     }
1111   }
1112   __kmp_printf_no_lock(";  tail: %d ", lck->lk.tail_id);
1113   __kmp_printf_no_lock("\n\n");
1114 }
1115 
1116 #endif /* DEBUG_QUEUING_LOCKS */
1117 
1118 static kmp_int32 __kmp_get_queuing_lock_owner(kmp_queuing_lock_t *lck) {
1119   return TCR_4(lck->lk.owner_id) - 1;
1120 }
1121 
1122 static inline bool __kmp_is_queuing_lock_nestable(kmp_queuing_lock_t *lck) {
1123   return lck->lk.depth_locked != -1;
1124 }
1125 
1126 /* Acquire a lock using a the queuing lock implementation */
1127 template <bool takeTime>
1128 /* [TLW] The unused template above is left behind because of what BEB believes
1129    is a potential compiler problem with __forceinline. */
1130 __forceinline static int
1131 __kmp_acquire_queuing_lock_timed_template(kmp_queuing_lock_t *lck,
1132                                           kmp_int32 gtid) {
1133   kmp_info_t *this_thr = __kmp_thread_from_gtid(gtid);
1134   volatile kmp_int32 *head_id_p = &lck->lk.head_id;
1135   volatile kmp_int32 *tail_id_p = &lck->lk.tail_id;
1136   volatile kmp_uint32 *spin_here_p;
1137   kmp_int32 need_mf = 1;
1138 
1139 #if OMPT_SUPPORT
1140   omp_state_t prev_state = omp_state_undefined;
1141 #endif
1142 
1143   KA_TRACE(1000,
1144            ("__kmp_acquire_queuing_lock: lck:%p, T#%d entering\n", lck, gtid));
1145 
1146   KMP_FSYNC_PREPARE(lck);
1147   KMP_DEBUG_ASSERT(this_thr != NULL);
1148   spin_here_p = &this_thr->th.th_spin_here;
1149 
1150 #ifdef DEBUG_QUEUING_LOCKS
1151   TRACE_LOCK(gtid + 1, "acq ent");
1152   if (*spin_here_p)
1153     __kmp_dump_queuing_lock(this_thr, gtid, lck, *head_id_p, *tail_id_p);
1154   if (this_thr->th.th_next_waiting != 0)
1155     __kmp_dump_queuing_lock(this_thr, gtid, lck, *head_id_p, *tail_id_p);
1156 #endif
1157   KMP_DEBUG_ASSERT(!*spin_here_p);
1158   KMP_DEBUG_ASSERT(this_thr->th.th_next_waiting == 0);
1159 
1160   /* The following st.rel to spin_here_p needs to precede the cmpxchg.acq to
1161      head_id_p that may follow, not just in execution order, but also in
1162      visibility order. This way, when a releasing thread observes the changes to
1163      the queue by this thread, it can rightly assume that spin_here_p has
1164      already been set to TRUE, so that when it sets spin_here_p to FALSE, it is
1165      not premature.  If the releasing thread sets spin_here_p to FALSE before
1166      this thread sets it to TRUE, this thread will hang. */
1167   *spin_here_p = TRUE; /* before enqueuing to prevent race */
1168 
1169   while (1) {
1170     kmp_int32 enqueued;
1171     kmp_int32 head;
1172     kmp_int32 tail;
1173 
1174     head = *head_id_p;
1175 
1176     switch (head) {
1177 
1178     case -1: {
1179 #ifdef DEBUG_QUEUING_LOCKS
1180       tail = *tail_id_p;
1181       TRACE_LOCK_HT(gtid + 1, "acq read: ", head, tail);
1182 #endif
1183       tail = 0; /* to make sure next link asynchronously read is not set
1184                 accidentally; this assignment prevents us from entering the
1185                 if ( t > 0 ) condition in the enqueued case below, which is not
1186                 necessary for this state transition */
1187 
1188       need_mf = 0;
1189       /* try (-1,0)->(tid,tid) */
1190       enqueued = KMP_COMPARE_AND_STORE_ACQ64((volatile kmp_int64 *)tail_id_p,
1191                                              KMP_PACK_64(-1, 0),
1192                                              KMP_PACK_64(gtid + 1, gtid + 1));
1193 #ifdef DEBUG_QUEUING_LOCKS
1194       if (enqueued)
1195         TRACE_LOCK(gtid + 1, "acq enq: (-1,0)->(tid,tid)");
1196 #endif
1197     } break;
1198 
1199     default: {
1200       tail = *tail_id_p;
1201       KMP_DEBUG_ASSERT(tail != gtid + 1);
1202 
1203 #ifdef DEBUG_QUEUING_LOCKS
1204       TRACE_LOCK_HT(gtid + 1, "acq read: ", head, tail);
1205 #endif
1206 
1207       if (tail == 0) {
1208         enqueued = FALSE;
1209       } else {
1210         need_mf = 0;
1211         /* try (h,t) or (h,h)->(h,tid) */
1212         enqueued = KMP_COMPARE_AND_STORE_ACQ32(tail_id_p, tail, gtid + 1);
1213 
1214 #ifdef DEBUG_QUEUING_LOCKS
1215         if (enqueued)
1216           TRACE_LOCK(gtid + 1, "acq enq: (h,t)->(h,tid)");
1217 #endif
1218       }
1219     } break;
1220 
1221     case 0: /* empty queue */
1222     {
1223       kmp_int32 grabbed_lock;
1224 
1225 #ifdef DEBUG_QUEUING_LOCKS
1226       tail = *tail_id_p;
1227       TRACE_LOCK_HT(gtid + 1, "acq read: ", head, tail);
1228 #endif
1229       /* try (0,0)->(-1,0) */
1230 
1231       /* only legal transition out of head = 0 is head = -1 with no change to
1232        * tail */
1233       grabbed_lock = KMP_COMPARE_AND_STORE_ACQ32(head_id_p, 0, -1);
1234 
1235       if (grabbed_lock) {
1236 
1237         *spin_here_p = FALSE;
1238 
1239         KA_TRACE(
1240             1000,
1241             ("__kmp_acquire_queuing_lock: lck:%p, T#%d exiting: no queuing\n",
1242              lck, gtid));
1243 #ifdef DEBUG_QUEUING_LOCKS
1244         TRACE_LOCK_HT(gtid + 1, "acq exit: ", head, 0);
1245 #endif
1246 
1247 #if OMPT_SUPPORT
1248         if (ompt_enabled.enabled && prev_state != omp_state_undefined) {
1249           /* change the state before clearing wait_id */
1250           this_thr->th.ompt_thread_info.state = prev_state;
1251           this_thr->th.ompt_thread_info.wait_id = 0;
1252         }
1253 #endif
1254 
1255         KMP_FSYNC_ACQUIRED(lck);
1256         return KMP_LOCK_ACQUIRED_FIRST; /* lock holder cannot be on queue */
1257       }
1258       enqueued = FALSE;
1259     } break;
1260     }
1261 
1262 #if OMPT_SUPPORT
1263     if (ompt_enabled.enabled && prev_state == omp_state_undefined) {
1264       /* this thread will spin; set wait_id before entering wait state */
1265       prev_state = this_thr->th.ompt_thread_info.state;
1266       this_thr->th.ompt_thread_info.wait_id = (uint64_t)lck;
1267       this_thr->th.ompt_thread_info.state = omp_state_wait_lock;
1268     }
1269 #endif
1270 
1271     if (enqueued) {
1272       if (tail > 0) {
1273         kmp_info_t *tail_thr = __kmp_thread_from_gtid(tail - 1);
1274         KMP_ASSERT(tail_thr != NULL);
1275         tail_thr->th.th_next_waiting = gtid + 1;
1276         /* corresponding wait for this write in release code */
1277       }
1278       KA_TRACE(1000,
1279                ("__kmp_acquire_queuing_lock: lck:%p, T#%d waiting for lock\n",
1280                 lck, gtid));
1281 
1282       /* ToDo: May want to consider using __kmp_wait_sleep  or something that
1283          sleeps for throughput only here. */
1284       KMP_MB();
1285       KMP_WAIT_YIELD(spin_here_p, FALSE, KMP_EQ, lck);
1286 
1287 #ifdef DEBUG_QUEUING_LOCKS
1288       TRACE_LOCK(gtid + 1, "acq spin");
1289 
1290       if (this_thr->th.th_next_waiting != 0)
1291         __kmp_dump_queuing_lock(this_thr, gtid, lck, *head_id_p, *tail_id_p);
1292 #endif
1293       KMP_DEBUG_ASSERT(this_thr->th.th_next_waiting == 0);
1294       KA_TRACE(1000, ("__kmp_acquire_queuing_lock: lck:%p, T#%d exiting: after "
1295                       "waiting on queue\n",
1296                       lck, gtid));
1297 
1298 #ifdef DEBUG_QUEUING_LOCKS
1299       TRACE_LOCK(gtid + 1, "acq exit 2");
1300 #endif
1301 
1302 #if OMPT_SUPPORT
1303       /* change the state before clearing wait_id */
1304       this_thr->th.ompt_thread_info.state = prev_state;
1305       this_thr->th.ompt_thread_info.wait_id = 0;
1306 #endif
1307 
1308       /* got lock, we were dequeued by the thread that released lock */
1309       return KMP_LOCK_ACQUIRED_FIRST;
1310     }
1311 
1312     /* Yield if number of threads > number of logical processors */
1313     /* ToDo: Not sure why this should only be in oversubscription case,
1314        maybe should be traditional YIELD_INIT/YIELD_WHEN loop */
1315     KMP_YIELD(TCR_4(__kmp_nth) >
1316               (__kmp_avail_proc ? __kmp_avail_proc : __kmp_xproc));
1317 #ifdef DEBUG_QUEUING_LOCKS
1318     TRACE_LOCK(gtid + 1, "acq retry");
1319 #endif
1320   }
1321   KMP_ASSERT2(0, "should not get here");
1322   return KMP_LOCK_ACQUIRED_FIRST;
1323 }
1324 
1325 int __kmp_acquire_queuing_lock(kmp_queuing_lock_t *lck, kmp_int32 gtid) {
1326   KMP_DEBUG_ASSERT(gtid >= 0);
1327 
1328   int retval = __kmp_acquire_queuing_lock_timed_template<false>(lck, gtid);
1329   ANNOTATE_QUEUING_ACQUIRED(lck);
1330   return retval;
1331 }
1332 
1333 static int __kmp_acquire_queuing_lock_with_checks(kmp_queuing_lock_t *lck,
1334                                                   kmp_int32 gtid) {
1335   char const *const func = "omp_set_lock";
1336   if (lck->lk.initialized != lck) {
1337     KMP_FATAL(LockIsUninitialized, func);
1338   }
1339   if (__kmp_is_queuing_lock_nestable(lck)) {
1340     KMP_FATAL(LockNestableUsedAsSimple, func);
1341   }
1342   if (__kmp_get_queuing_lock_owner(lck) == gtid) {
1343     KMP_FATAL(LockIsAlreadyOwned, func);
1344   }
1345 
1346   __kmp_acquire_queuing_lock(lck, gtid);
1347 
1348   lck->lk.owner_id = gtid + 1;
1349   return KMP_LOCK_ACQUIRED_FIRST;
1350 }
1351 
1352 int __kmp_test_queuing_lock(kmp_queuing_lock_t *lck, kmp_int32 gtid) {
1353   volatile kmp_int32 *head_id_p = &lck->lk.head_id;
1354   kmp_int32 head;
1355 #ifdef KMP_DEBUG
1356   kmp_info_t *this_thr;
1357 #endif
1358 
1359   KA_TRACE(1000, ("__kmp_test_queuing_lock: T#%d entering\n", gtid));
1360   KMP_DEBUG_ASSERT(gtid >= 0);
1361 #ifdef KMP_DEBUG
1362   this_thr = __kmp_thread_from_gtid(gtid);
1363   KMP_DEBUG_ASSERT(this_thr != NULL);
1364   KMP_DEBUG_ASSERT(!this_thr->th.th_spin_here);
1365 #endif
1366 
1367   head = *head_id_p;
1368 
1369   if (head == 0) { /* nobody on queue, nobody holding */
1370     /* try (0,0)->(-1,0) */
1371     if (KMP_COMPARE_AND_STORE_ACQ32(head_id_p, 0, -1)) {
1372       KA_TRACE(1000,
1373                ("__kmp_test_queuing_lock: T#%d exiting: holding lock\n", gtid));
1374       KMP_FSYNC_ACQUIRED(lck);
1375       ANNOTATE_QUEUING_ACQUIRED(lck);
1376       return TRUE;
1377     }
1378   }
1379 
1380   KA_TRACE(1000,
1381            ("__kmp_test_queuing_lock: T#%d exiting: without lock\n", gtid));
1382   return FALSE;
1383 }
1384 
1385 static int __kmp_test_queuing_lock_with_checks(kmp_queuing_lock_t *lck,
1386                                                kmp_int32 gtid) {
1387   char const *const func = "omp_test_lock";
1388   if (lck->lk.initialized != lck) {
1389     KMP_FATAL(LockIsUninitialized, func);
1390   }
1391   if (__kmp_is_queuing_lock_nestable(lck)) {
1392     KMP_FATAL(LockNestableUsedAsSimple, func);
1393   }
1394 
1395   int retval = __kmp_test_queuing_lock(lck, gtid);
1396 
1397   if (retval) {
1398     lck->lk.owner_id = gtid + 1;
1399   }
1400   return retval;
1401 }
1402 
1403 int __kmp_release_queuing_lock(kmp_queuing_lock_t *lck, kmp_int32 gtid) {
1404   kmp_info_t *this_thr;
1405   volatile kmp_int32 *head_id_p = &lck->lk.head_id;
1406   volatile kmp_int32 *tail_id_p = &lck->lk.tail_id;
1407 
1408   KA_TRACE(1000,
1409            ("__kmp_release_queuing_lock: lck:%p, T#%d entering\n", lck, gtid));
1410   KMP_DEBUG_ASSERT(gtid >= 0);
1411   this_thr = __kmp_thread_from_gtid(gtid);
1412   KMP_DEBUG_ASSERT(this_thr != NULL);
1413 #ifdef DEBUG_QUEUING_LOCKS
1414   TRACE_LOCK(gtid + 1, "rel ent");
1415 
1416   if (this_thr->th.th_spin_here)
1417     __kmp_dump_queuing_lock(this_thr, gtid, lck, *head_id_p, *tail_id_p);
1418   if (this_thr->th.th_next_waiting != 0)
1419     __kmp_dump_queuing_lock(this_thr, gtid, lck, *head_id_p, *tail_id_p);
1420 #endif
1421   KMP_DEBUG_ASSERT(!this_thr->th.th_spin_here);
1422   KMP_DEBUG_ASSERT(this_thr->th.th_next_waiting == 0);
1423 
1424   KMP_FSYNC_RELEASING(lck);
1425   ANNOTATE_QUEUING_RELEASED(lck);
1426 
1427   while (1) {
1428     kmp_int32 dequeued;
1429     kmp_int32 head;
1430     kmp_int32 tail;
1431 
1432     head = *head_id_p;
1433 
1434 #ifdef DEBUG_QUEUING_LOCKS
1435     tail = *tail_id_p;
1436     TRACE_LOCK_HT(gtid + 1, "rel read: ", head, tail);
1437     if (head == 0)
1438       __kmp_dump_queuing_lock(this_thr, gtid, lck, head, tail);
1439 #endif
1440     KMP_DEBUG_ASSERT(head !=
1441                      0); /* holding the lock, head must be -1 or queue head */
1442 
1443     if (head == -1) { /* nobody on queue */
1444       /* try (-1,0)->(0,0) */
1445       if (KMP_COMPARE_AND_STORE_REL32(head_id_p, -1, 0)) {
1446         KA_TRACE(
1447             1000,
1448             ("__kmp_release_queuing_lock: lck:%p, T#%d exiting: queue empty\n",
1449              lck, gtid));
1450 #ifdef DEBUG_QUEUING_LOCKS
1451         TRACE_LOCK_HT(gtid + 1, "rel exit: ", 0, 0);
1452 #endif
1453 
1454 #if OMPT_SUPPORT
1455 /* nothing to do - no other thread is trying to shift blame */
1456 #endif
1457         return KMP_LOCK_RELEASED;
1458       }
1459       dequeued = FALSE;
1460     } else {
1461       KMP_MB();
1462       tail = *tail_id_p;
1463       if (head == tail) { /* only one thread on the queue */
1464 #ifdef DEBUG_QUEUING_LOCKS
1465         if (head <= 0)
1466           __kmp_dump_queuing_lock(this_thr, gtid, lck, head, tail);
1467 #endif
1468         KMP_DEBUG_ASSERT(head > 0);
1469 
1470         /* try (h,h)->(-1,0) */
1471         dequeued = KMP_COMPARE_AND_STORE_REL64(
1472             RCAST(volatile kmp_int64 *, tail_id_p), KMP_PACK_64(head, head),
1473             KMP_PACK_64(-1, 0));
1474 #ifdef DEBUG_QUEUING_LOCKS
1475         TRACE_LOCK(gtid + 1, "rel deq: (h,h)->(-1,0)");
1476 #endif
1477 
1478       } else {
1479         volatile kmp_int32 *waiting_id_p;
1480         kmp_info_t *head_thr = __kmp_thread_from_gtid(head - 1);
1481         KMP_DEBUG_ASSERT(head_thr != NULL);
1482         waiting_id_p = &head_thr->th.th_next_waiting;
1483 
1484 /* Does this require synchronous reads? */
1485 #ifdef DEBUG_QUEUING_LOCKS
1486         if (head <= 0 || tail <= 0)
1487           __kmp_dump_queuing_lock(this_thr, gtid, lck, head, tail);
1488 #endif
1489         KMP_DEBUG_ASSERT(head > 0 && tail > 0);
1490 
1491         /* try (h,t)->(h',t) or (t,t) */
1492         KMP_MB();
1493         /* make sure enqueuing thread has time to update next waiting thread
1494          * field */
1495         *head_id_p = KMP_WAIT_YIELD((volatile kmp_uint32 *)waiting_id_p, 0,
1496                                     KMP_NEQ, NULL);
1497 #ifdef DEBUG_QUEUING_LOCKS
1498         TRACE_LOCK(gtid + 1, "rel deq: (h,t)->(h',t)");
1499 #endif
1500         dequeued = TRUE;
1501       }
1502     }
1503 
1504     if (dequeued) {
1505       kmp_info_t *head_thr = __kmp_thread_from_gtid(head - 1);
1506       KMP_DEBUG_ASSERT(head_thr != NULL);
1507 
1508 /* Does this require synchronous reads? */
1509 #ifdef DEBUG_QUEUING_LOCKS
1510       if (head <= 0 || tail <= 0)
1511         __kmp_dump_queuing_lock(this_thr, gtid, lck, head, tail);
1512 #endif
1513       KMP_DEBUG_ASSERT(head > 0 && tail > 0);
1514 
1515       /* For clean code only. Thread not released until next statement prevents
1516          race with acquire code. */
1517       head_thr->th.th_next_waiting = 0;
1518 #ifdef DEBUG_QUEUING_LOCKS
1519       TRACE_LOCK_T(gtid + 1, "rel nw=0 for t=", head);
1520 #endif
1521 
1522       KMP_MB();
1523       /* reset spin value */
1524       head_thr->th.th_spin_here = FALSE;
1525 
1526       KA_TRACE(1000, ("__kmp_release_queuing_lock: lck:%p, T#%d exiting: after "
1527                       "dequeuing\n",
1528                       lck, gtid));
1529 #ifdef DEBUG_QUEUING_LOCKS
1530       TRACE_LOCK(gtid + 1, "rel exit 2");
1531 #endif
1532       return KMP_LOCK_RELEASED;
1533     }
1534 /* KMP_CPU_PAUSE(); don't want to make releasing thread hold up acquiring
1535    threads */
1536 
1537 #ifdef DEBUG_QUEUING_LOCKS
1538     TRACE_LOCK(gtid + 1, "rel retry");
1539 #endif
1540 
1541   } /* while */
1542   KMP_ASSERT2(0, "should not get here");
1543   return KMP_LOCK_RELEASED;
1544 }
1545 
1546 static int __kmp_release_queuing_lock_with_checks(kmp_queuing_lock_t *lck,
1547                                                   kmp_int32 gtid) {
1548   char const *const func = "omp_unset_lock";
1549   KMP_MB(); /* in case another processor initialized lock */
1550   if (lck->lk.initialized != lck) {
1551     KMP_FATAL(LockIsUninitialized, func);
1552   }
1553   if (__kmp_is_queuing_lock_nestable(lck)) {
1554     KMP_FATAL(LockNestableUsedAsSimple, func);
1555   }
1556   if (__kmp_get_queuing_lock_owner(lck) == -1) {
1557     KMP_FATAL(LockUnsettingFree, func);
1558   }
1559   if (__kmp_get_queuing_lock_owner(lck) != gtid) {
1560     KMP_FATAL(LockUnsettingSetByAnother, func);
1561   }
1562   lck->lk.owner_id = 0;
1563   return __kmp_release_queuing_lock(lck, gtid);
1564 }
1565 
1566 void __kmp_init_queuing_lock(kmp_queuing_lock_t *lck) {
1567   lck->lk.location = NULL;
1568   lck->lk.head_id = 0;
1569   lck->lk.tail_id = 0;
1570   lck->lk.next_ticket = 0;
1571   lck->lk.now_serving = 0;
1572   lck->lk.owner_id = 0; // no thread owns the lock.
1573   lck->lk.depth_locked = -1; // >= 0 for nestable locks, -1 for simple locks.
1574   lck->lk.initialized = lck;
1575 
1576   KA_TRACE(1000, ("__kmp_init_queuing_lock: lock %p initialized\n", lck));
1577 }
1578 
1579 static void __kmp_init_queuing_lock_with_checks(kmp_queuing_lock_t *lck) {
1580   __kmp_init_queuing_lock(lck);
1581 }
1582 
1583 void __kmp_destroy_queuing_lock(kmp_queuing_lock_t *lck) {
1584   lck->lk.initialized = NULL;
1585   lck->lk.location = NULL;
1586   lck->lk.head_id = 0;
1587   lck->lk.tail_id = 0;
1588   lck->lk.next_ticket = 0;
1589   lck->lk.now_serving = 0;
1590   lck->lk.owner_id = 0;
1591   lck->lk.depth_locked = -1;
1592 }
1593 
1594 static void __kmp_destroy_queuing_lock_with_checks(kmp_queuing_lock_t *lck) {
1595   char const *const func = "omp_destroy_lock";
1596   if (lck->lk.initialized != lck) {
1597     KMP_FATAL(LockIsUninitialized, func);
1598   }
1599   if (__kmp_is_queuing_lock_nestable(lck)) {
1600     KMP_FATAL(LockNestableUsedAsSimple, func);
1601   }
1602   if (__kmp_get_queuing_lock_owner(lck) != -1) {
1603     KMP_FATAL(LockStillOwned, func);
1604   }
1605   __kmp_destroy_queuing_lock(lck);
1606 }
1607 
1608 // nested queuing locks
1609 
1610 int __kmp_acquire_nested_queuing_lock(kmp_queuing_lock_t *lck, kmp_int32 gtid) {
1611   KMP_DEBUG_ASSERT(gtid >= 0);
1612 
1613   if (__kmp_get_queuing_lock_owner(lck) == gtid) {
1614     lck->lk.depth_locked += 1;
1615     return KMP_LOCK_ACQUIRED_NEXT;
1616   } else {
1617     __kmp_acquire_queuing_lock_timed_template<false>(lck, gtid);
1618     ANNOTATE_QUEUING_ACQUIRED(lck);
1619     KMP_MB();
1620     lck->lk.depth_locked = 1;
1621     KMP_MB();
1622     lck->lk.owner_id = gtid + 1;
1623     return KMP_LOCK_ACQUIRED_FIRST;
1624   }
1625 }
1626 
1627 static int
1628 __kmp_acquire_nested_queuing_lock_with_checks(kmp_queuing_lock_t *lck,
1629                                               kmp_int32 gtid) {
1630   char const *const func = "omp_set_nest_lock";
1631   if (lck->lk.initialized != lck) {
1632     KMP_FATAL(LockIsUninitialized, func);
1633   }
1634   if (!__kmp_is_queuing_lock_nestable(lck)) {
1635     KMP_FATAL(LockSimpleUsedAsNestable, func);
1636   }
1637   return __kmp_acquire_nested_queuing_lock(lck, gtid);
1638 }
1639 
1640 int __kmp_test_nested_queuing_lock(kmp_queuing_lock_t *lck, kmp_int32 gtid) {
1641   int retval;
1642 
1643   KMP_DEBUG_ASSERT(gtid >= 0);
1644 
1645   if (__kmp_get_queuing_lock_owner(lck) == gtid) {
1646     retval = ++lck->lk.depth_locked;
1647   } else if (!__kmp_test_queuing_lock(lck, gtid)) {
1648     retval = 0;
1649   } else {
1650     KMP_MB();
1651     retval = lck->lk.depth_locked = 1;
1652     KMP_MB();
1653     lck->lk.owner_id = gtid + 1;
1654   }
1655   return retval;
1656 }
1657 
1658 static int __kmp_test_nested_queuing_lock_with_checks(kmp_queuing_lock_t *lck,
1659                                                       kmp_int32 gtid) {
1660   char const *const func = "omp_test_nest_lock";
1661   if (lck->lk.initialized != lck) {
1662     KMP_FATAL(LockIsUninitialized, func);
1663   }
1664   if (!__kmp_is_queuing_lock_nestable(lck)) {
1665     KMP_FATAL(LockSimpleUsedAsNestable, func);
1666   }
1667   return __kmp_test_nested_queuing_lock(lck, gtid);
1668 }
1669 
1670 int __kmp_release_nested_queuing_lock(kmp_queuing_lock_t *lck, kmp_int32 gtid) {
1671   KMP_DEBUG_ASSERT(gtid >= 0);
1672 
1673   KMP_MB();
1674   if (--(lck->lk.depth_locked) == 0) {
1675     KMP_MB();
1676     lck->lk.owner_id = 0;
1677     __kmp_release_queuing_lock(lck, gtid);
1678     return KMP_LOCK_RELEASED;
1679   }
1680   return KMP_LOCK_STILL_HELD;
1681 }
1682 
1683 static int
1684 __kmp_release_nested_queuing_lock_with_checks(kmp_queuing_lock_t *lck,
1685                                               kmp_int32 gtid) {
1686   char const *const func = "omp_unset_nest_lock";
1687   KMP_MB(); /* in case another processor initialized lock */
1688   if (lck->lk.initialized != lck) {
1689     KMP_FATAL(LockIsUninitialized, func);
1690   }
1691   if (!__kmp_is_queuing_lock_nestable(lck)) {
1692     KMP_FATAL(LockSimpleUsedAsNestable, func);
1693   }
1694   if (__kmp_get_queuing_lock_owner(lck) == -1) {
1695     KMP_FATAL(LockUnsettingFree, func);
1696   }
1697   if (__kmp_get_queuing_lock_owner(lck) != gtid) {
1698     KMP_FATAL(LockUnsettingSetByAnother, func);
1699   }
1700   return __kmp_release_nested_queuing_lock(lck, gtid);
1701 }
1702 
1703 void __kmp_init_nested_queuing_lock(kmp_queuing_lock_t *lck) {
1704   __kmp_init_queuing_lock(lck);
1705   lck->lk.depth_locked = 0; // >= 0 for nestable locks, -1 for simple locks
1706 }
1707 
1708 static void
1709 __kmp_init_nested_queuing_lock_with_checks(kmp_queuing_lock_t *lck) {
1710   __kmp_init_nested_queuing_lock(lck);
1711 }
1712 
1713 void __kmp_destroy_nested_queuing_lock(kmp_queuing_lock_t *lck) {
1714   __kmp_destroy_queuing_lock(lck);
1715   lck->lk.depth_locked = 0;
1716 }
1717 
1718 static void
1719 __kmp_destroy_nested_queuing_lock_with_checks(kmp_queuing_lock_t *lck) {
1720   char const *const func = "omp_destroy_nest_lock";
1721   if (lck->lk.initialized != lck) {
1722     KMP_FATAL(LockIsUninitialized, func);
1723   }
1724   if (!__kmp_is_queuing_lock_nestable(lck)) {
1725     KMP_FATAL(LockSimpleUsedAsNestable, func);
1726   }
1727   if (__kmp_get_queuing_lock_owner(lck) != -1) {
1728     KMP_FATAL(LockStillOwned, func);
1729   }
1730   __kmp_destroy_nested_queuing_lock(lck);
1731 }
1732 
1733 // access functions to fields which don't exist for all lock kinds.
1734 
1735 static int __kmp_is_queuing_lock_initialized(kmp_queuing_lock_t *lck) {
1736   return lck == lck->lk.initialized;
1737 }
1738 
1739 static const ident_t *__kmp_get_queuing_lock_location(kmp_queuing_lock_t *lck) {
1740   return lck->lk.location;
1741 }
1742 
1743 static void __kmp_set_queuing_lock_location(kmp_queuing_lock_t *lck,
1744                                             const ident_t *loc) {
1745   lck->lk.location = loc;
1746 }
1747 
1748 static kmp_lock_flags_t __kmp_get_queuing_lock_flags(kmp_queuing_lock_t *lck) {
1749   return lck->lk.flags;
1750 }
1751 
1752 static void __kmp_set_queuing_lock_flags(kmp_queuing_lock_t *lck,
1753                                          kmp_lock_flags_t flags) {
1754   lck->lk.flags = flags;
1755 }
1756 
1757 #if KMP_USE_ADAPTIVE_LOCKS
1758 
1759 /* RTM Adaptive locks */
1760 
1761 #if KMP_COMPILER_ICC && __INTEL_COMPILER >= 1300
1762 
1763 #include <immintrin.h>
1764 #define SOFT_ABORT_MASK (_XABORT_RETRY | _XABORT_CONFLICT | _XABORT_EXPLICIT)
1765 
1766 #else
1767 
1768 // Values from the status register after failed speculation.
1769 #define _XBEGIN_STARTED (~0u)
1770 #define _XABORT_EXPLICIT (1 << 0)
1771 #define _XABORT_RETRY (1 << 1)
1772 #define _XABORT_CONFLICT (1 << 2)
1773 #define _XABORT_CAPACITY (1 << 3)
1774 #define _XABORT_DEBUG (1 << 4)
1775 #define _XABORT_NESTED (1 << 5)
1776 #define _XABORT_CODE(x) ((unsigned char)(((x) >> 24) & 0xFF))
1777 
1778 // Aborts for which it's worth trying again immediately
1779 #define SOFT_ABORT_MASK (_XABORT_RETRY | _XABORT_CONFLICT | _XABORT_EXPLICIT)
1780 
1781 #define STRINGIZE_INTERNAL(arg) #arg
1782 #define STRINGIZE(arg) STRINGIZE_INTERNAL(arg)
1783 
1784 // Access to RTM instructions
1785 /*A version of XBegin which returns -1 on speculation, and the value of EAX on
1786   an abort. This is the same definition as the compiler intrinsic that will be
1787   supported at some point. */
1788 static __inline int _xbegin() {
1789   int res = -1;
1790 
1791 #if KMP_OS_WINDOWS
1792 #if KMP_ARCH_X86_64
1793   _asm {
1794         _emit 0xC7
1795         _emit 0xF8
1796         _emit 2
1797         _emit 0
1798         _emit 0
1799         _emit 0
1800         jmp   L2
1801         mov   res, eax
1802     L2:
1803   }
1804 #else /* IA32 */
1805   _asm {
1806         _emit 0xC7
1807         _emit 0xF8
1808         _emit 2
1809         _emit 0
1810         _emit 0
1811         _emit 0
1812         jmp   L2
1813         mov   res, eax
1814     L2:
1815   }
1816 #endif // KMP_ARCH_X86_64
1817 #else
1818   /* Note that %eax must be noted as killed (clobbered), because the XSR is
1819      returned in %eax(%rax) on abort.  Other register values are restored, so
1820      don't need to be killed.
1821 
1822      We must also mark 'res' as an input and an output, since otherwise
1823      'res=-1' may be dropped as being dead, whereas we do need the assignment on
1824      the successful (i.e., non-abort) path. */
1825   __asm__ volatile("1: .byte  0xC7; .byte 0xF8;\n"
1826                    "   .long  1f-1b-6\n"
1827                    "    jmp   2f\n"
1828                    "1:  movl  %%eax,%0\n"
1829                    "2:"
1830                    : "+r"(res)::"memory", "%eax");
1831 #endif // KMP_OS_WINDOWS
1832   return res;
1833 }
1834 
1835 /* Transaction end */
1836 static __inline void _xend() {
1837 #if KMP_OS_WINDOWS
1838   __asm {
1839         _emit 0x0f
1840         _emit 0x01
1841         _emit 0xd5
1842   }
1843 #else
1844   __asm__ volatile(".byte 0x0f; .byte 0x01; .byte 0xd5" ::: "memory");
1845 #endif
1846 }
1847 
1848 /* This is a macro, the argument must be a single byte constant which can be
1849    evaluated by the inline assembler, since it is emitted as a byte into the
1850    assembly code. */
1851 // clang-format off
1852 #if KMP_OS_WINDOWS
1853 #define _xabort(ARG) _asm _emit 0xc6 _asm _emit 0xf8 _asm _emit ARG
1854 #else
1855 #define _xabort(ARG)                                                           \
1856   __asm__ volatile(".byte 0xC6; .byte 0xF8; .byte " STRINGIZE(ARG):::"memory");
1857 #endif
1858 // clang-format on
1859 #endif // KMP_COMPILER_ICC && __INTEL_COMPILER >= 1300
1860 
1861 // Statistics is collected for testing purpose
1862 #if KMP_DEBUG_ADAPTIVE_LOCKS
1863 
1864 // We accumulate speculative lock statistics when the lock is destroyed. We
1865 // keep locks that haven't been destroyed in the liveLocks list so that we can
1866 // grab their statistics too.
1867 static kmp_adaptive_lock_statistics_t destroyedStats;
1868 
1869 // To hold the list of live locks.
1870 static kmp_adaptive_lock_info_t liveLocks;
1871 
1872 // A lock so we can safely update the list of locks.
1873 static kmp_bootstrap_lock_t chain_lock;
1874 
1875 // Initialize the list of stats.
1876 void __kmp_init_speculative_stats() {
1877   kmp_adaptive_lock_info_t *lck = &liveLocks;
1878 
1879   memset((void *)&(lck->stats), 0, sizeof(lck->stats));
1880   lck->stats.next = lck;
1881   lck->stats.prev = lck;
1882 
1883   KMP_ASSERT(lck->stats.next->stats.prev == lck);
1884   KMP_ASSERT(lck->stats.prev->stats.next == lck);
1885 
1886   __kmp_init_bootstrap_lock(&chain_lock);
1887 }
1888 
1889 // Insert the lock into the circular list
1890 static void __kmp_remember_lock(kmp_adaptive_lock_info_t *lck) {
1891   __kmp_acquire_bootstrap_lock(&chain_lock);
1892 
1893   lck->stats.next = liveLocks.stats.next;
1894   lck->stats.prev = &liveLocks;
1895 
1896   liveLocks.stats.next = lck;
1897   lck->stats.next->stats.prev = lck;
1898 
1899   KMP_ASSERT(lck->stats.next->stats.prev == lck);
1900   KMP_ASSERT(lck->stats.prev->stats.next == lck);
1901 
1902   __kmp_release_bootstrap_lock(&chain_lock);
1903 }
1904 
1905 static void __kmp_forget_lock(kmp_adaptive_lock_info_t *lck) {
1906   KMP_ASSERT(lck->stats.next->stats.prev == lck);
1907   KMP_ASSERT(lck->stats.prev->stats.next == lck);
1908 
1909   kmp_adaptive_lock_info_t *n = lck->stats.next;
1910   kmp_adaptive_lock_info_t *p = lck->stats.prev;
1911 
1912   n->stats.prev = p;
1913   p->stats.next = n;
1914 }
1915 
1916 static void __kmp_zero_speculative_stats(kmp_adaptive_lock_info_t *lck) {
1917   memset((void *)&lck->stats, 0, sizeof(lck->stats));
1918   __kmp_remember_lock(lck);
1919 }
1920 
1921 static void __kmp_add_stats(kmp_adaptive_lock_statistics_t *t,
1922                             kmp_adaptive_lock_info_t *lck) {
1923   kmp_adaptive_lock_statistics_t volatile *s = &lck->stats;
1924 
1925   t->nonSpeculativeAcquireAttempts += lck->acquire_attempts;
1926   t->successfulSpeculations += s->successfulSpeculations;
1927   t->hardFailedSpeculations += s->hardFailedSpeculations;
1928   t->softFailedSpeculations += s->softFailedSpeculations;
1929   t->nonSpeculativeAcquires += s->nonSpeculativeAcquires;
1930   t->lemmingYields += s->lemmingYields;
1931 }
1932 
1933 static void __kmp_accumulate_speculative_stats(kmp_adaptive_lock_info_t *lck) {
1934   kmp_adaptive_lock_statistics_t *t = &destroyedStats;
1935 
1936   __kmp_acquire_bootstrap_lock(&chain_lock);
1937 
1938   __kmp_add_stats(&destroyedStats, lck);
1939   __kmp_forget_lock(lck);
1940 
1941   __kmp_release_bootstrap_lock(&chain_lock);
1942 }
1943 
1944 static float percent(kmp_uint32 count, kmp_uint32 total) {
1945   return (total == 0) ? 0.0 : (100.0 * count) / total;
1946 }
1947 
1948 static FILE *__kmp_open_stats_file() {
1949   if (strcmp(__kmp_speculative_statsfile, "-") == 0)
1950     return stdout;
1951 
1952   size_t buffLen = KMP_STRLEN(__kmp_speculative_statsfile) + 20;
1953   char buffer[buffLen];
1954   KMP_SNPRINTF(&buffer[0], buffLen, __kmp_speculative_statsfile,
1955                (kmp_int32)getpid());
1956   FILE *result = fopen(&buffer[0], "w");
1957 
1958   // Maybe we should issue a warning here...
1959   return result ? result : stdout;
1960 }
1961 
1962 void __kmp_print_speculative_stats() {
1963   if (__kmp_user_lock_kind != lk_adaptive)
1964     return;
1965 
1966   FILE *statsFile = __kmp_open_stats_file();
1967 
1968   kmp_adaptive_lock_statistics_t total = destroyedStats;
1969   kmp_adaptive_lock_info_t *lck;
1970 
1971   for (lck = liveLocks.stats.next; lck != &liveLocks; lck = lck->stats.next) {
1972     __kmp_add_stats(&total, lck);
1973   }
1974   kmp_adaptive_lock_statistics_t *t = &total;
1975   kmp_uint32 totalSections =
1976       t->nonSpeculativeAcquires + t->successfulSpeculations;
1977   kmp_uint32 totalSpeculations = t->successfulSpeculations +
1978                                  t->hardFailedSpeculations +
1979                                  t->softFailedSpeculations;
1980 
1981   fprintf(statsFile, "Speculative lock statistics (all approximate!)\n");
1982   fprintf(statsFile, " Lock parameters: \n"
1983                      "   max_soft_retries               : %10d\n"
1984                      "   max_badness                    : %10d\n",
1985           __kmp_adaptive_backoff_params.max_soft_retries,
1986           __kmp_adaptive_backoff_params.max_badness);
1987   fprintf(statsFile, " Non-speculative acquire attempts : %10d\n",
1988           t->nonSpeculativeAcquireAttempts);
1989   fprintf(statsFile, " Total critical sections          : %10d\n",
1990           totalSections);
1991   fprintf(statsFile, " Successful speculations          : %10d (%5.1f%%)\n",
1992           t->successfulSpeculations,
1993           percent(t->successfulSpeculations, totalSections));
1994   fprintf(statsFile, " Non-speculative acquires         : %10d (%5.1f%%)\n",
1995           t->nonSpeculativeAcquires,
1996           percent(t->nonSpeculativeAcquires, totalSections));
1997   fprintf(statsFile, " Lemming yields                   : %10d\n\n",
1998           t->lemmingYields);
1999 
2000   fprintf(statsFile, " Speculative acquire attempts     : %10d\n",
2001           totalSpeculations);
2002   fprintf(statsFile, " Successes                        : %10d (%5.1f%%)\n",
2003           t->successfulSpeculations,
2004           percent(t->successfulSpeculations, totalSpeculations));
2005   fprintf(statsFile, " Soft failures                    : %10d (%5.1f%%)\n",
2006           t->softFailedSpeculations,
2007           percent(t->softFailedSpeculations, totalSpeculations));
2008   fprintf(statsFile, " Hard failures                    : %10d (%5.1f%%)\n",
2009           t->hardFailedSpeculations,
2010           percent(t->hardFailedSpeculations, totalSpeculations));
2011 
2012   if (statsFile != stdout)
2013     fclose(statsFile);
2014 }
2015 
2016 #define KMP_INC_STAT(lck, stat) (lck->lk.adaptive.stats.stat++)
2017 #else
2018 #define KMP_INC_STAT(lck, stat)
2019 
2020 #endif // KMP_DEBUG_ADAPTIVE_LOCKS
2021 
2022 static inline bool __kmp_is_unlocked_queuing_lock(kmp_queuing_lock_t *lck) {
2023   // It is enough to check that the head_id is zero.
2024   // We don't also need to check the tail.
2025   bool res = lck->lk.head_id == 0;
2026 
2027 // We need a fence here, since we must ensure that no memory operations
2028 // from later in this thread float above that read.
2029 #if KMP_COMPILER_ICC
2030   _mm_mfence();
2031 #else
2032   __sync_synchronize();
2033 #endif
2034 
2035   return res;
2036 }
2037 
2038 // Functions for manipulating the badness
2039 static __inline void
2040 __kmp_update_badness_after_success(kmp_adaptive_lock_t *lck) {
2041   // Reset the badness to zero so we eagerly try to speculate again
2042   lck->lk.adaptive.badness = 0;
2043   KMP_INC_STAT(lck, successfulSpeculations);
2044 }
2045 
2046 // Create a bit mask with one more set bit.
2047 static __inline void __kmp_step_badness(kmp_adaptive_lock_t *lck) {
2048   kmp_uint32 newBadness = (lck->lk.adaptive.badness << 1) | 1;
2049   if (newBadness > lck->lk.adaptive.max_badness) {
2050     return;
2051   } else {
2052     lck->lk.adaptive.badness = newBadness;
2053   }
2054 }
2055 
2056 // Check whether speculation should be attempted.
2057 static __inline int __kmp_should_speculate(kmp_adaptive_lock_t *lck,
2058                                            kmp_int32 gtid) {
2059   kmp_uint32 badness = lck->lk.adaptive.badness;
2060   kmp_uint32 attempts = lck->lk.adaptive.acquire_attempts;
2061   int res = (attempts & badness) == 0;
2062   return res;
2063 }
2064 
2065 // Attempt to acquire only the speculative lock.
2066 // Does not back off to the non-speculative lock.
2067 static int __kmp_test_adaptive_lock_only(kmp_adaptive_lock_t *lck,
2068                                          kmp_int32 gtid) {
2069   int retries = lck->lk.adaptive.max_soft_retries;
2070 
2071   // We don't explicitly count the start of speculation, rather we record the
2072   // results (success, hard fail, soft fail). The sum of all of those is the
2073   // total number of times we started speculation since all speculations must
2074   // end one of those ways.
2075   do {
2076     kmp_uint32 status = _xbegin();
2077     // Switch this in to disable actual speculation but exercise at least some
2078     // of the rest of the code. Useful for debugging...
2079     // kmp_uint32 status = _XABORT_NESTED;
2080 
2081     if (status == _XBEGIN_STARTED) {
2082       /* We have successfully started speculation. Check that no-one acquired
2083          the lock for real between when we last looked and now. This also gets
2084          the lock cache line into our read-set, which we need so that we'll
2085          abort if anyone later claims it for real. */
2086       if (!__kmp_is_unlocked_queuing_lock(GET_QLK_PTR(lck))) {
2087         // Lock is now visibly acquired, so someone beat us to it. Abort the
2088         // transaction so we'll restart from _xbegin with the failure status.
2089         _xabort(0x01);
2090         KMP_ASSERT2(0, "should not get here");
2091       }
2092       return 1; // Lock has been acquired (speculatively)
2093     } else {
2094       // We have aborted, update the statistics
2095       if (status & SOFT_ABORT_MASK) {
2096         KMP_INC_STAT(lck, softFailedSpeculations);
2097         // and loop round to retry.
2098       } else {
2099         KMP_INC_STAT(lck, hardFailedSpeculations);
2100         // Give up if we had a hard failure.
2101         break;
2102       }
2103     }
2104   } while (retries--); // Loop while we have retries, and didn't fail hard.
2105 
2106   // Either we had a hard failure or we didn't succeed softly after
2107   // the full set of attempts, so back off the badness.
2108   __kmp_step_badness(lck);
2109   return 0;
2110 }
2111 
2112 // Attempt to acquire the speculative lock, or back off to the non-speculative
2113 // one if the speculative lock cannot be acquired.
2114 // We can succeed speculatively, non-speculatively, or fail.
2115 static int __kmp_test_adaptive_lock(kmp_adaptive_lock_t *lck, kmp_int32 gtid) {
2116   // First try to acquire the lock speculatively
2117   if (__kmp_should_speculate(lck, gtid) &&
2118       __kmp_test_adaptive_lock_only(lck, gtid))
2119     return 1;
2120 
2121   // Speculative acquisition failed, so try to acquire it non-speculatively.
2122   // Count the non-speculative acquire attempt
2123   lck->lk.adaptive.acquire_attempts++;
2124 
2125   // Use base, non-speculative lock.
2126   if (__kmp_test_queuing_lock(GET_QLK_PTR(lck), gtid)) {
2127     KMP_INC_STAT(lck, nonSpeculativeAcquires);
2128     return 1; // Lock is acquired (non-speculatively)
2129   } else {
2130     return 0; // Failed to acquire the lock, it's already visibly locked.
2131   }
2132 }
2133 
2134 static int __kmp_test_adaptive_lock_with_checks(kmp_adaptive_lock_t *lck,
2135                                                 kmp_int32 gtid) {
2136   char const *const func = "omp_test_lock";
2137   if (lck->lk.qlk.initialized != GET_QLK_PTR(lck)) {
2138     KMP_FATAL(LockIsUninitialized, func);
2139   }
2140 
2141   int retval = __kmp_test_adaptive_lock(lck, gtid);
2142 
2143   if (retval) {
2144     lck->lk.qlk.owner_id = gtid + 1;
2145   }
2146   return retval;
2147 }
2148 
2149 // Block until we can acquire a speculative, adaptive lock. We check whether we
2150 // should be trying to speculate. If we should be, we check the real lock to see
2151 // if it is free, and, if not, pause without attempting to acquire it until it
2152 // is. Then we try the speculative acquire. This means that although we suffer
2153 // from lemmings a little (because all we can't acquire the lock speculatively
2154 // until the queue of threads waiting has cleared), we don't get into a state
2155 // where we can never acquire the lock speculatively (because we force the queue
2156 // to clear by preventing new arrivals from entering the queue). This does mean
2157 // that when we're trying to break lemmings, the lock is no longer fair. However
2158 // OpenMP makes no guarantee that its locks are fair, so this isn't a real
2159 // problem.
2160 static void __kmp_acquire_adaptive_lock(kmp_adaptive_lock_t *lck,
2161                                         kmp_int32 gtid) {
2162   if (__kmp_should_speculate(lck, gtid)) {
2163     if (__kmp_is_unlocked_queuing_lock(GET_QLK_PTR(lck))) {
2164       if (__kmp_test_adaptive_lock_only(lck, gtid))
2165         return;
2166       // We tried speculation and failed, so give up.
2167     } else {
2168       // We can't try speculation until the lock is free, so we pause here
2169       // (without suspending on the queueing lock, to allow it to drain, then
2170       // try again. All other threads will also see the same result for
2171       // shouldSpeculate, so will be doing the same if they try to claim the
2172       // lock from now on.
2173       while (!__kmp_is_unlocked_queuing_lock(GET_QLK_PTR(lck))) {
2174         KMP_INC_STAT(lck, lemmingYields);
2175         __kmp_yield(TRUE);
2176       }
2177 
2178       if (__kmp_test_adaptive_lock_only(lck, gtid))
2179         return;
2180     }
2181   }
2182 
2183   // Speculative acquisition failed, so acquire it non-speculatively.
2184   // Count the non-speculative acquire attempt
2185   lck->lk.adaptive.acquire_attempts++;
2186 
2187   __kmp_acquire_queuing_lock_timed_template<FALSE>(GET_QLK_PTR(lck), gtid);
2188   // We have acquired the base lock, so count that.
2189   KMP_INC_STAT(lck, nonSpeculativeAcquires);
2190   ANNOTATE_QUEUING_ACQUIRED(lck);
2191 }
2192 
2193 static void __kmp_acquire_adaptive_lock_with_checks(kmp_adaptive_lock_t *lck,
2194                                                     kmp_int32 gtid) {
2195   char const *const func = "omp_set_lock";
2196   if (lck->lk.qlk.initialized != GET_QLK_PTR(lck)) {
2197     KMP_FATAL(LockIsUninitialized, func);
2198   }
2199   if (__kmp_get_queuing_lock_owner(GET_QLK_PTR(lck)) == gtid) {
2200     KMP_FATAL(LockIsAlreadyOwned, func);
2201   }
2202 
2203   __kmp_acquire_adaptive_lock(lck, gtid);
2204 
2205   lck->lk.qlk.owner_id = gtid + 1;
2206 }
2207 
2208 static int __kmp_release_adaptive_lock(kmp_adaptive_lock_t *lck,
2209                                        kmp_int32 gtid) {
2210   if (__kmp_is_unlocked_queuing_lock(GET_QLK_PTR(
2211           lck))) { // If the lock doesn't look claimed we must be speculating.
2212     // (Or the user's code is buggy and they're releasing without locking;
2213     // if we had XTEST we'd be able to check that case...)
2214     _xend(); // Exit speculation
2215     __kmp_update_badness_after_success(lck);
2216   } else { // Since the lock *is* visibly locked we're not speculating,
2217     // so should use the underlying lock's release scheme.
2218     __kmp_release_queuing_lock(GET_QLK_PTR(lck), gtid);
2219   }
2220   return KMP_LOCK_RELEASED;
2221 }
2222 
2223 static int __kmp_release_adaptive_lock_with_checks(kmp_adaptive_lock_t *lck,
2224                                                    kmp_int32 gtid) {
2225   char const *const func = "omp_unset_lock";
2226   KMP_MB(); /* in case another processor initialized lock */
2227   if (lck->lk.qlk.initialized != GET_QLK_PTR(lck)) {
2228     KMP_FATAL(LockIsUninitialized, func);
2229   }
2230   if (__kmp_get_queuing_lock_owner(GET_QLK_PTR(lck)) == -1) {
2231     KMP_FATAL(LockUnsettingFree, func);
2232   }
2233   if (__kmp_get_queuing_lock_owner(GET_QLK_PTR(lck)) != gtid) {
2234     KMP_FATAL(LockUnsettingSetByAnother, func);
2235   }
2236   lck->lk.qlk.owner_id = 0;
2237   __kmp_release_adaptive_lock(lck, gtid);
2238   return KMP_LOCK_RELEASED;
2239 }
2240 
2241 static void __kmp_init_adaptive_lock(kmp_adaptive_lock_t *lck) {
2242   __kmp_init_queuing_lock(GET_QLK_PTR(lck));
2243   lck->lk.adaptive.badness = 0;
2244   lck->lk.adaptive.acquire_attempts = 0; // nonSpeculativeAcquireAttempts = 0;
2245   lck->lk.adaptive.max_soft_retries =
2246       __kmp_adaptive_backoff_params.max_soft_retries;
2247   lck->lk.adaptive.max_badness = __kmp_adaptive_backoff_params.max_badness;
2248 #if KMP_DEBUG_ADAPTIVE_LOCKS
2249   __kmp_zero_speculative_stats(&lck->lk.adaptive);
2250 #endif
2251   KA_TRACE(1000, ("__kmp_init_adaptive_lock: lock %p initialized\n", lck));
2252 }
2253 
2254 static void __kmp_init_adaptive_lock_with_checks(kmp_adaptive_lock_t *lck) {
2255   __kmp_init_adaptive_lock(lck);
2256 }
2257 
2258 static void __kmp_destroy_adaptive_lock(kmp_adaptive_lock_t *lck) {
2259 #if KMP_DEBUG_ADAPTIVE_LOCKS
2260   __kmp_accumulate_speculative_stats(&lck->lk.adaptive);
2261 #endif
2262   __kmp_destroy_queuing_lock(GET_QLK_PTR(lck));
2263   // Nothing needed for the speculative part.
2264 }
2265 
2266 static void __kmp_destroy_adaptive_lock_with_checks(kmp_adaptive_lock_t *lck) {
2267   char const *const func = "omp_destroy_lock";
2268   if (lck->lk.qlk.initialized != GET_QLK_PTR(lck)) {
2269     KMP_FATAL(LockIsUninitialized, func);
2270   }
2271   if (__kmp_get_queuing_lock_owner(GET_QLK_PTR(lck)) != -1) {
2272     KMP_FATAL(LockStillOwned, func);
2273   }
2274   __kmp_destroy_adaptive_lock(lck);
2275 }
2276 
2277 #endif // KMP_USE_ADAPTIVE_LOCKS
2278 
2279 /* ------------------------------------------------------------------------ */
2280 /* DRDPA ticket locks                                                */
2281 /* "DRDPA" means Dynamically Reconfigurable Distributed Polling Area */
2282 
2283 static kmp_int32 __kmp_get_drdpa_lock_owner(kmp_drdpa_lock_t *lck) {
2284   return lck->lk.owner_id - 1;
2285 }
2286 
2287 static inline bool __kmp_is_drdpa_lock_nestable(kmp_drdpa_lock_t *lck) {
2288   return lck->lk.depth_locked != -1;
2289 }
2290 
2291 __forceinline static int
2292 __kmp_acquire_drdpa_lock_timed_template(kmp_drdpa_lock_t *lck, kmp_int32 gtid) {
2293   kmp_uint64 ticket = KMP_ATOMIC_INC(&lck->lk.next_ticket);
2294   kmp_uint64 mask = lck->lk.mask; // atomic load
2295   std::atomic<kmp_uint64> *polls = lck->lk.polls;
2296 
2297 #ifdef USE_LOCK_PROFILE
2298   if (polls[ticket & mask] != ticket)
2299     __kmp_printf("LOCK CONTENTION: %p\n", lck);
2300 /* else __kmp_printf( "." );*/
2301 #endif /* USE_LOCK_PROFILE */
2302 
2303   // Now spin-wait, but reload the polls pointer and mask, in case the
2304   // polling area has been reconfigured.  Unless it is reconfigured, the
2305   // reloads stay in L1 cache and are cheap.
2306   //
2307   // Keep this code in sync with KMP_WAIT_YIELD, in kmp_dispatch.cpp !!!
2308   //
2309   // The current implementation of KMP_WAIT_YIELD doesn't allow for mask
2310   // and poll to be re-read every spin iteration.
2311   kmp_uint32 spins;
2312 
2313   KMP_FSYNC_PREPARE(lck);
2314   KMP_INIT_YIELD(spins);
2315   while (polls[ticket & mask] < ticket) { // atomic load
2316     // If we are oversubscribed,
2317     // or have waited a bit (and KMP_LIBRARY=turnaround), then yield.
2318     // CPU Pause is in the macros for yield.
2319     //
2320     KMP_YIELD(TCR_4(__kmp_nth) >
2321               (__kmp_avail_proc ? __kmp_avail_proc : __kmp_xproc));
2322     KMP_YIELD_SPIN(spins);
2323 
2324     // Re-read the mask and the poll pointer from the lock structure.
2325     //
2326     // Make certain that "mask" is read before "polls" !!!
2327     //
2328     // If another thread picks reconfigures the polling area and updates their
2329     // values, and we get the new value of mask and the old polls pointer, we
2330     // could access memory beyond the end of the old polling area.
2331     mask = lck->lk.mask; // atomic load
2332     polls = lck->lk.polls; // atomic load
2333   }
2334 
2335   // Critical section starts here
2336   KMP_FSYNC_ACQUIRED(lck);
2337   KA_TRACE(1000, ("__kmp_acquire_drdpa_lock: ticket #%lld acquired lock %p\n",
2338                   ticket, lck));
2339   lck->lk.now_serving = ticket; // non-volatile store
2340 
2341   // Deallocate a garbage polling area if we know that we are the last
2342   // thread that could possibly access it.
2343   //
2344   // The >= check is in case __kmp_test_drdpa_lock() allocated the cleanup
2345   // ticket.
2346   if ((lck->lk.old_polls != NULL) && (ticket >= lck->lk.cleanup_ticket)) {
2347     __kmp_free(lck->lk.old_polls);
2348     lck->lk.old_polls = NULL;
2349     lck->lk.cleanup_ticket = 0;
2350   }
2351 
2352   // Check to see if we should reconfigure the polling area.
2353   // If there is still a garbage polling area to be deallocated from a
2354   // previous reconfiguration, let a later thread reconfigure it.
2355   if (lck->lk.old_polls == NULL) {
2356     bool reconfigure = false;
2357     std::atomic<kmp_uint64> *old_polls = polls;
2358     kmp_uint32 num_polls = TCR_4(lck->lk.num_polls);
2359 
2360     if (TCR_4(__kmp_nth) >
2361         (__kmp_avail_proc ? __kmp_avail_proc : __kmp_xproc)) {
2362       // We are in oversubscription mode.  Contract the polling area
2363       // down to a single location, if that hasn't been done already.
2364       if (num_polls > 1) {
2365         reconfigure = true;
2366         num_polls = TCR_4(lck->lk.num_polls);
2367         mask = 0;
2368         num_polls = 1;
2369         polls = (std::atomic<kmp_uint64> *)__kmp_allocate(num_polls *
2370                                                           sizeof(*polls));
2371         polls[0] = ticket;
2372       }
2373     } else {
2374       // We are in under/fully subscribed mode.  Check the number of
2375       // threads waiting on the lock.  The size of the polling area
2376       // should be at least the number of threads waiting.
2377       kmp_uint64 num_waiting = TCR_8(lck->lk.next_ticket) - ticket - 1;
2378       if (num_waiting > num_polls) {
2379         kmp_uint32 old_num_polls = num_polls;
2380         reconfigure = true;
2381         do {
2382           mask = (mask << 1) | 1;
2383           num_polls *= 2;
2384         } while (num_polls <= num_waiting);
2385 
2386         // Allocate the new polling area, and copy the relevant portion
2387         // of the old polling area to the new area.  __kmp_allocate()
2388         // zeroes the memory it allocates, and most of the old area is
2389         // just zero padding, so we only copy the release counters.
2390         polls = (std::atomic<kmp_uint64> *)__kmp_allocate(num_polls *
2391                                                           sizeof(*polls));
2392         kmp_uint32 i;
2393         for (i = 0; i < old_num_polls; i++) {
2394           polls[i].store(old_polls[i]);
2395         }
2396       }
2397     }
2398 
2399     if (reconfigure) {
2400       // Now write the updated fields back to the lock structure.
2401       //
2402       // Make certain that "polls" is written before "mask" !!!
2403       //
2404       // If another thread picks up the new value of mask and the old polls
2405       // pointer , it could access memory beyond the end of the old polling
2406       // area.
2407       //
2408       // On x86, we need memory fences.
2409       KA_TRACE(1000, ("__kmp_acquire_drdpa_lock: ticket #%lld reconfiguring "
2410                       "lock %p to %d polls\n",
2411                       ticket, lck, num_polls));
2412 
2413       lck->lk.old_polls = old_polls;
2414       lck->lk.polls = polls; // atomic store
2415 
2416       KMP_MB();
2417 
2418       lck->lk.num_polls = num_polls;
2419       lck->lk.mask = mask; // atomic store
2420 
2421       KMP_MB();
2422 
2423       // Only after the new polling area and mask have been flushed
2424       // to main memory can we update the cleanup ticket field.
2425       //
2426       // volatile load / non-volatile store
2427       lck->lk.cleanup_ticket = lck->lk.next_ticket;
2428     }
2429   }
2430   return KMP_LOCK_ACQUIRED_FIRST;
2431 }
2432 
2433 int __kmp_acquire_drdpa_lock(kmp_drdpa_lock_t *lck, kmp_int32 gtid) {
2434   int retval = __kmp_acquire_drdpa_lock_timed_template(lck, gtid);
2435   ANNOTATE_DRDPA_ACQUIRED(lck);
2436   return retval;
2437 }
2438 
2439 static int __kmp_acquire_drdpa_lock_with_checks(kmp_drdpa_lock_t *lck,
2440                                                 kmp_int32 gtid) {
2441   char const *const func = "omp_set_lock";
2442   if (lck->lk.initialized != lck) {
2443     KMP_FATAL(LockIsUninitialized, func);
2444   }
2445   if (__kmp_is_drdpa_lock_nestable(lck)) {
2446     KMP_FATAL(LockNestableUsedAsSimple, func);
2447   }
2448   if ((gtid >= 0) && (__kmp_get_drdpa_lock_owner(lck) == gtid)) {
2449     KMP_FATAL(LockIsAlreadyOwned, func);
2450   }
2451 
2452   __kmp_acquire_drdpa_lock(lck, gtid);
2453 
2454   lck->lk.owner_id = gtid + 1;
2455   return KMP_LOCK_ACQUIRED_FIRST;
2456 }
2457 
2458 int __kmp_test_drdpa_lock(kmp_drdpa_lock_t *lck, kmp_int32 gtid) {
2459   // First get a ticket, then read the polls pointer and the mask.
2460   // The polls pointer must be read before the mask!!! (See above)
2461   kmp_uint64 ticket = lck->lk.next_ticket; // atomic load
2462   std::atomic<kmp_uint64> *polls = lck->lk.polls;
2463   kmp_uint64 mask = lck->lk.mask; // atomic load
2464   if (polls[ticket & mask] == ticket) {
2465     kmp_uint64 next_ticket = ticket + 1;
2466     if (__kmp_atomic_compare_store_acq(&lck->lk.next_ticket, ticket,
2467                                        next_ticket)) {
2468       KMP_FSYNC_ACQUIRED(lck);
2469       KA_TRACE(1000, ("__kmp_test_drdpa_lock: ticket #%lld acquired lock %p\n",
2470                       ticket, lck));
2471       lck->lk.now_serving = ticket; // non-volatile store
2472 
2473       // Since no threads are waiting, there is no possibility that we would
2474       // want to reconfigure the polling area.  We might have the cleanup ticket
2475       // value (which says that it is now safe to deallocate old_polls), but
2476       // we'll let a later thread which calls __kmp_acquire_lock do that - this
2477       // routine isn't supposed to block, and we would risk blocks if we called
2478       // __kmp_free() to do the deallocation.
2479       return TRUE;
2480     }
2481   }
2482   return FALSE;
2483 }
2484 
2485 static int __kmp_test_drdpa_lock_with_checks(kmp_drdpa_lock_t *lck,
2486                                              kmp_int32 gtid) {
2487   char const *const func = "omp_test_lock";
2488   if (lck->lk.initialized != lck) {
2489     KMP_FATAL(LockIsUninitialized, func);
2490   }
2491   if (__kmp_is_drdpa_lock_nestable(lck)) {
2492     KMP_FATAL(LockNestableUsedAsSimple, func);
2493   }
2494 
2495   int retval = __kmp_test_drdpa_lock(lck, gtid);
2496 
2497   if (retval) {
2498     lck->lk.owner_id = gtid + 1;
2499   }
2500   return retval;
2501 }
2502 
2503 int __kmp_release_drdpa_lock(kmp_drdpa_lock_t *lck, kmp_int32 gtid) {
2504   // Read the ticket value from the lock data struct, then the polls pointer and
2505   // the mask.  The polls pointer must be read before the mask!!! (See above)
2506   kmp_uint64 ticket = lck->lk.now_serving + 1; // non-atomic load
2507   std::atomic<kmp_uint64> *polls = lck->lk.polls; // atomic load
2508   kmp_uint64 mask = lck->lk.mask; // atomic load
2509   KA_TRACE(1000, ("__kmp_release_drdpa_lock: ticket #%lld released lock %p\n",
2510                   ticket - 1, lck));
2511   KMP_FSYNC_RELEASING(lck);
2512   ANNOTATE_DRDPA_RELEASED(lck);
2513   polls[ticket & mask] = ticket; // atomic store
2514   return KMP_LOCK_RELEASED;
2515 }
2516 
2517 static int __kmp_release_drdpa_lock_with_checks(kmp_drdpa_lock_t *lck,
2518                                                 kmp_int32 gtid) {
2519   char const *const func = "omp_unset_lock";
2520   KMP_MB(); /* in case another processor initialized lock */
2521   if (lck->lk.initialized != lck) {
2522     KMP_FATAL(LockIsUninitialized, func);
2523   }
2524   if (__kmp_is_drdpa_lock_nestable(lck)) {
2525     KMP_FATAL(LockNestableUsedAsSimple, func);
2526   }
2527   if (__kmp_get_drdpa_lock_owner(lck) == -1) {
2528     KMP_FATAL(LockUnsettingFree, func);
2529   }
2530   if ((gtid >= 0) && (__kmp_get_drdpa_lock_owner(lck) >= 0) &&
2531       (__kmp_get_drdpa_lock_owner(lck) != gtid)) {
2532     KMP_FATAL(LockUnsettingSetByAnother, func);
2533   }
2534   lck->lk.owner_id = 0;
2535   return __kmp_release_drdpa_lock(lck, gtid);
2536 }
2537 
2538 void __kmp_init_drdpa_lock(kmp_drdpa_lock_t *lck) {
2539   lck->lk.location = NULL;
2540   lck->lk.mask = 0;
2541   lck->lk.num_polls = 1;
2542   lck->lk.polls = (std::atomic<kmp_uint64> *)__kmp_allocate(
2543       lck->lk.num_polls * sizeof(*(lck->lk.polls)));
2544   lck->lk.cleanup_ticket = 0;
2545   lck->lk.old_polls = NULL;
2546   lck->lk.next_ticket = 0;
2547   lck->lk.now_serving = 0;
2548   lck->lk.owner_id = 0; // no thread owns the lock.
2549   lck->lk.depth_locked = -1; // >= 0 for nestable locks, -1 for simple locks.
2550   lck->lk.initialized = lck;
2551 
2552   KA_TRACE(1000, ("__kmp_init_drdpa_lock: lock %p initialized\n", lck));
2553 }
2554 
2555 static void __kmp_init_drdpa_lock_with_checks(kmp_drdpa_lock_t *lck) {
2556   __kmp_init_drdpa_lock(lck);
2557 }
2558 
2559 void __kmp_destroy_drdpa_lock(kmp_drdpa_lock_t *lck) {
2560   lck->lk.initialized = NULL;
2561   lck->lk.location = NULL;
2562   if (lck->lk.polls.load() != NULL) {
2563     __kmp_free(lck->lk.polls.load());
2564     lck->lk.polls = NULL;
2565   }
2566   if (lck->lk.old_polls != NULL) {
2567     __kmp_free(lck->lk.old_polls);
2568     lck->lk.old_polls = NULL;
2569   }
2570   lck->lk.mask = 0;
2571   lck->lk.num_polls = 0;
2572   lck->lk.cleanup_ticket = 0;
2573   lck->lk.next_ticket = 0;
2574   lck->lk.now_serving = 0;
2575   lck->lk.owner_id = 0;
2576   lck->lk.depth_locked = -1;
2577 }
2578 
2579 static void __kmp_destroy_drdpa_lock_with_checks(kmp_drdpa_lock_t *lck) {
2580   char const *const func = "omp_destroy_lock";
2581   if (lck->lk.initialized != lck) {
2582     KMP_FATAL(LockIsUninitialized, func);
2583   }
2584   if (__kmp_is_drdpa_lock_nestable(lck)) {
2585     KMP_FATAL(LockNestableUsedAsSimple, func);
2586   }
2587   if (__kmp_get_drdpa_lock_owner(lck) != -1) {
2588     KMP_FATAL(LockStillOwned, func);
2589   }
2590   __kmp_destroy_drdpa_lock(lck);
2591 }
2592 
2593 // nested drdpa ticket locks
2594 
2595 int __kmp_acquire_nested_drdpa_lock(kmp_drdpa_lock_t *lck, kmp_int32 gtid) {
2596   KMP_DEBUG_ASSERT(gtid >= 0);
2597 
2598   if (__kmp_get_drdpa_lock_owner(lck) == gtid) {
2599     lck->lk.depth_locked += 1;
2600     return KMP_LOCK_ACQUIRED_NEXT;
2601   } else {
2602     __kmp_acquire_drdpa_lock_timed_template(lck, gtid);
2603     ANNOTATE_DRDPA_ACQUIRED(lck);
2604     KMP_MB();
2605     lck->lk.depth_locked = 1;
2606     KMP_MB();
2607     lck->lk.owner_id = gtid + 1;
2608     return KMP_LOCK_ACQUIRED_FIRST;
2609   }
2610 }
2611 
2612 static void __kmp_acquire_nested_drdpa_lock_with_checks(kmp_drdpa_lock_t *lck,
2613                                                         kmp_int32 gtid) {
2614   char const *const func = "omp_set_nest_lock";
2615   if (lck->lk.initialized != lck) {
2616     KMP_FATAL(LockIsUninitialized, func);
2617   }
2618   if (!__kmp_is_drdpa_lock_nestable(lck)) {
2619     KMP_FATAL(LockSimpleUsedAsNestable, func);
2620   }
2621   __kmp_acquire_nested_drdpa_lock(lck, gtid);
2622 }
2623 
2624 int __kmp_test_nested_drdpa_lock(kmp_drdpa_lock_t *lck, kmp_int32 gtid) {
2625   int retval;
2626 
2627   KMP_DEBUG_ASSERT(gtid >= 0);
2628 
2629   if (__kmp_get_drdpa_lock_owner(lck) == gtid) {
2630     retval = ++lck->lk.depth_locked;
2631   } else if (!__kmp_test_drdpa_lock(lck, gtid)) {
2632     retval = 0;
2633   } else {
2634     KMP_MB();
2635     retval = lck->lk.depth_locked = 1;
2636     KMP_MB();
2637     lck->lk.owner_id = gtid + 1;
2638   }
2639   return retval;
2640 }
2641 
2642 static int __kmp_test_nested_drdpa_lock_with_checks(kmp_drdpa_lock_t *lck,
2643                                                     kmp_int32 gtid) {
2644   char const *const func = "omp_test_nest_lock";
2645   if (lck->lk.initialized != lck) {
2646     KMP_FATAL(LockIsUninitialized, func);
2647   }
2648   if (!__kmp_is_drdpa_lock_nestable(lck)) {
2649     KMP_FATAL(LockSimpleUsedAsNestable, func);
2650   }
2651   return __kmp_test_nested_drdpa_lock(lck, gtid);
2652 }
2653 
2654 int __kmp_release_nested_drdpa_lock(kmp_drdpa_lock_t *lck, kmp_int32 gtid) {
2655   KMP_DEBUG_ASSERT(gtid >= 0);
2656 
2657   KMP_MB();
2658   if (--(lck->lk.depth_locked) == 0) {
2659     KMP_MB();
2660     lck->lk.owner_id = 0;
2661     __kmp_release_drdpa_lock(lck, gtid);
2662     return KMP_LOCK_RELEASED;
2663   }
2664   return KMP_LOCK_STILL_HELD;
2665 }
2666 
2667 static int __kmp_release_nested_drdpa_lock_with_checks(kmp_drdpa_lock_t *lck,
2668                                                        kmp_int32 gtid) {
2669   char const *const func = "omp_unset_nest_lock";
2670   KMP_MB(); /* in case another processor initialized lock */
2671   if (lck->lk.initialized != lck) {
2672     KMP_FATAL(LockIsUninitialized, func);
2673   }
2674   if (!__kmp_is_drdpa_lock_nestable(lck)) {
2675     KMP_FATAL(LockSimpleUsedAsNestable, func);
2676   }
2677   if (__kmp_get_drdpa_lock_owner(lck) == -1) {
2678     KMP_FATAL(LockUnsettingFree, func);
2679   }
2680   if (__kmp_get_drdpa_lock_owner(lck) != gtid) {
2681     KMP_FATAL(LockUnsettingSetByAnother, func);
2682   }
2683   return __kmp_release_nested_drdpa_lock(lck, gtid);
2684 }
2685 
2686 void __kmp_init_nested_drdpa_lock(kmp_drdpa_lock_t *lck) {
2687   __kmp_init_drdpa_lock(lck);
2688   lck->lk.depth_locked = 0; // >= 0 for nestable locks, -1 for simple locks
2689 }
2690 
2691 static void __kmp_init_nested_drdpa_lock_with_checks(kmp_drdpa_lock_t *lck) {
2692   __kmp_init_nested_drdpa_lock(lck);
2693 }
2694 
2695 void __kmp_destroy_nested_drdpa_lock(kmp_drdpa_lock_t *lck) {
2696   __kmp_destroy_drdpa_lock(lck);
2697   lck->lk.depth_locked = 0;
2698 }
2699 
2700 static void __kmp_destroy_nested_drdpa_lock_with_checks(kmp_drdpa_lock_t *lck) {
2701   char const *const func = "omp_destroy_nest_lock";
2702   if (lck->lk.initialized != lck) {
2703     KMP_FATAL(LockIsUninitialized, func);
2704   }
2705   if (!__kmp_is_drdpa_lock_nestable(lck)) {
2706     KMP_FATAL(LockSimpleUsedAsNestable, func);
2707   }
2708   if (__kmp_get_drdpa_lock_owner(lck) != -1) {
2709     KMP_FATAL(LockStillOwned, func);
2710   }
2711   __kmp_destroy_nested_drdpa_lock(lck);
2712 }
2713 
2714 // access functions to fields which don't exist for all lock kinds.
2715 
2716 static int __kmp_is_drdpa_lock_initialized(kmp_drdpa_lock_t *lck) {
2717   return lck == lck->lk.initialized;
2718 }
2719 
2720 static const ident_t *__kmp_get_drdpa_lock_location(kmp_drdpa_lock_t *lck) {
2721   return lck->lk.location;
2722 }
2723 
2724 static void __kmp_set_drdpa_lock_location(kmp_drdpa_lock_t *lck,
2725                                           const ident_t *loc) {
2726   lck->lk.location = loc;
2727 }
2728 
2729 static kmp_lock_flags_t __kmp_get_drdpa_lock_flags(kmp_drdpa_lock_t *lck) {
2730   return lck->lk.flags;
2731 }
2732 
2733 static void __kmp_set_drdpa_lock_flags(kmp_drdpa_lock_t *lck,
2734                                        kmp_lock_flags_t flags) {
2735   lck->lk.flags = flags;
2736 }
2737 
2738 // Time stamp counter
2739 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
2740 #define __kmp_tsc() __kmp_hardware_timestamp()
2741 // Runtime's default backoff parameters
2742 kmp_backoff_t __kmp_spin_backoff_params = {1, 4096, 100};
2743 #else
2744 // Use nanoseconds for other platforms
2745 extern kmp_uint64 __kmp_now_nsec();
2746 kmp_backoff_t __kmp_spin_backoff_params = {1, 256, 100};
2747 #define __kmp_tsc() __kmp_now_nsec()
2748 #endif
2749 
2750 // A useful predicate for dealing with timestamps that may wrap.
2751 // Is a before b? Since the timestamps may wrap, this is asking whether it's
2752 // shorter to go clockwise from a to b around the clock-face, or anti-clockwise.
2753 // Times where going clockwise is less distance than going anti-clockwise
2754 // are in the future, others are in the past. e.g. a = MAX-1, b = MAX+1 (=0),
2755 // then a > b (true) does not mean a reached b; whereas signed(a) = -2,
2756 // signed(b) = 0 captures the actual difference
2757 static inline bool before(kmp_uint64 a, kmp_uint64 b) {
2758   return ((kmp_int64)b - (kmp_int64)a) > 0;
2759 }
2760 
2761 // Truncated binary exponential backoff function
2762 void __kmp_spin_backoff(kmp_backoff_t *boff) {
2763   // We could flatten this loop, but making it a nested loop gives better result
2764   kmp_uint32 i;
2765   for (i = boff->step; i > 0; i--) {
2766     kmp_uint64 goal = __kmp_tsc() + boff->min_tick;
2767     do {
2768       KMP_CPU_PAUSE();
2769     } while (before(__kmp_tsc(), goal));
2770   }
2771   boff->step = (boff->step << 1 | 1) & (boff->max_backoff - 1);
2772 }
2773 
2774 #if KMP_USE_DYNAMIC_LOCK
2775 
2776 // Direct lock initializers. It simply writes a tag to the low 8 bits of the
2777 // lock word.
2778 static void __kmp_init_direct_lock(kmp_dyna_lock_t *lck,
2779                                    kmp_dyna_lockseq_t seq) {
2780   TCW_4(*lck, KMP_GET_D_TAG(seq));
2781   KA_TRACE(
2782       20,
2783       ("__kmp_init_direct_lock: initialized direct lock with type#%d\n", seq));
2784 }
2785 
2786 #if KMP_USE_TSX
2787 
2788 // HLE lock functions - imported from the testbed runtime.
2789 #define HLE_ACQUIRE ".byte 0xf2;"
2790 #define HLE_RELEASE ".byte 0xf3;"
2791 
2792 static inline kmp_uint32 swap4(kmp_uint32 volatile *p, kmp_uint32 v) {
2793   __asm__ volatile(HLE_ACQUIRE "xchg %1,%0" : "+r"(v), "+m"(*p) : : "memory");
2794   return v;
2795 }
2796 
2797 static void __kmp_destroy_hle_lock(kmp_dyna_lock_t *lck) { TCW_4(*lck, 0); }
2798 
2799 static void __kmp_acquire_hle_lock(kmp_dyna_lock_t *lck, kmp_int32 gtid) {
2800   // Use gtid for KMP_LOCK_BUSY if necessary
2801   if (swap4(lck, KMP_LOCK_BUSY(1, hle)) != KMP_LOCK_FREE(hle)) {
2802     int delay = 1;
2803     do {
2804       while (*(kmp_uint32 volatile *)lck != KMP_LOCK_FREE(hle)) {
2805         for (int i = delay; i != 0; --i)
2806           KMP_CPU_PAUSE();
2807         delay = ((delay << 1) | 1) & 7;
2808       }
2809     } while (swap4(lck, KMP_LOCK_BUSY(1, hle)) != KMP_LOCK_FREE(hle));
2810   }
2811 }
2812 
2813 static void __kmp_acquire_hle_lock_with_checks(kmp_dyna_lock_t *lck,
2814                                                kmp_int32 gtid) {
2815   __kmp_acquire_hle_lock(lck, gtid); // TODO: add checks
2816 }
2817 
2818 static int __kmp_release_hle_lock(kmp_dyna_lock_t *lck, kmp_int32 gtid) {
2819   __asm__ volatile(HLE_RELEASE "movl %1,%0"
2820                    : "=m"(*lck)
2821                    : "r"(KMP_LOCK_FREE(hle))
2822                    : "memory");
2823   return KMP_LOCK_RELEASED;
2824 }
2825 
2826 static int __kmp_release_hle_lock_with_checks(kmp_dyna_lock_t *lck,
2827                                               kmp_int32 gtid) {
2828   return __kmp_release_hle_lock(lck, gtid); // TODO: add checks
2829 }
2830 
2831 static int __kmp_test_hle_lock(kmp_dyna_lock_t *lck, kmp_int32 gtid) {
2832   return swap4(lck, KMP_LOCK_BUSY(1, hle)) == KMP_LOCK_FREE(hle);
2833 }
2834 
2835 static int __kmp_test_hle_lock_with_checks(kmp_dyna_lock_t *lck,
2836                                            kmp_int32 gtid) {
2837   return __kmp_test_hle_lock(lck, gtid); // TODO: add checks
2838 }
2839 
2840 static void __kmp_init_rtm_lock(kmp_queuing_lock_t *lck) {
2841   __kmp_init_queuing_lock(lck);
2842 }
2843 
2844 static void __kmp_destroy_rtm_lock(kmp_queuing_lock_t *lck) {
2845   __kmp_destroy_queuing_lock(lck);
2846 }
2847 
2848 static void __kmp_acquire_rtm_lock(kmp_queuing_lock_t *lck, kmp_int32 gtid) {
2849   unsigned retries = 3, status;
2850   do {
2851     status = _xbegin();
2852     if (status == _XBEGIN_STARTED) {
2853       if (__kmp_is_unlocked_queuing_lock(lck))
2854         return;
2855       _xabort(0xff);
2856     }
2857     if ((status & _XABORT_EXPLICIT) && _XABORT_CODE(status) == 0xff) {
2858       // Wait until lock becomes free
2859       while (!__kmp_is_unlocked_queuing_lock(lck))
2860         __kmp_yield(TRUE);
2861     } else if (!(status & _XABORT_RETRY))
2862       break;
2863   } while (retries--);
2864 
2865   // Fall-back non-speculative lock (xchg)
2866   __kmp_acquire_queuing_lock(lck, gtid);
2867 }
2868 
2869 static void __kmp_acquire_rtm_lock_with_checks(kmp_queuing_lock_t *lck,
2870                                                kmp_int32 gtid) {
2871   __kmp_acquire_rtm_lock(lck, gtid);
2872 }
2873 
2874 static int __kmp_release_rtm_lock(kmp_queuing_lock_t *lck, kmp_int32 gtid) {
2875   if (__kmp_is_unlocked_queuing_lock(lck)) {
2876     // Releasing from speculation
2877     _xend();
2878   } else {
2879     // Releasing from a real lock
2880     __kmp_release_queuing_lock(lck, gtid);
2881   }
2882   return KMP_LOCK_RELEASED;
2883 }
2884 
2885 static int __kmp_release_rtm_lock_with_checks(kmp_queuing_lock_t *lck,
2886                                               kmp_int32 gtid) {
2887   return __kmp_release_rtm_lock(lck, gtid);
2888 }
2889 
2890 static int __kmp_test_rtm_lock(kmp_queuing_lock_t *lck, kmp_int32 gtid) {
2891   unsigned retries = 3, status;
2892   do {
2893     status = _xbegin();
2894     if (status == _XBEGIN_STARTED && __kmp_is_unlocked_queuing_lock(lck)) {
2895       return 1;
2896     }
2897     if (!(status & _XABORT_RETRY))
2898       break;
2899   } while (retries--);
2900 
2901   return (__kmp_is_unlocked_queuing_lock(lck)) ? 1 : 0;
2902 }
2903 
2904 static int __kmp_test_rtm_lock_with_checks(kmp_queuing_lock_t *lck,
2905                                            kmp_int32 gtid) {
2906   return __kmp_test_rtm_lock(lck, gtid);
2907 }
2908 
2909 #endif // KMP_USE_TSX
2910 
2911 // Entry functions for indirect locks (first element of direct lock jump tables)
2912 static void __kmp_init_indirect_lock(kmp_dyna_lock_t *l,
2913                                      kmp_dyna_lockseq_t tag);
2914 static void __kmp_destroy_indirect_lock(kmp_dyna_lock_t *lock);
2915 static int __kmp_set_indirect_lock(kmp_dyna_lock_t *lock, kmp_int32);
2916 static int __kmp_unset_indirect_lock(kmp_dyna_lock_t *lock, kmp_int32);
2917 static int __kmp_test_indirect_lock(kmp_dyna_lock_t *lock, kmp_int32);
2918 static int __kmp_set_indirect_lock_with_checks(kmp_dyna_lock_t *lock,
2919                                                kmp_int32);
2920 static int __kmp_unset_indirect_lock_with_checks(kmp_dyna_lock_t *lock,
2921                                                  kmp_int32);
2922 static int __kmp_test_indirect_lock_with_checks(kmp_dyna_lock_t *lock,
2923                                                 kmp_int32);
2924 
2925 // Jump tables for the indirect lock functions
2926 // Only fill in the odd entries, that avoids the need to shift out the low bit
2927 
2928 // init functions
2929 #define expand(l, op) 0, __kmp_init_direct_lock,
2930 void (*__kmp_direct_init[])(kmp_dyna_lock_t *, kmp_dyna_lockseq_t) = {
2931     __kmp_init_indirect_lock, 0, KMP_FOREACH_D_LOCK(expand, init)};
2932 #undef expand
2933 
2934 // destroy functions
2935 #define expand(l, op) 0, (void (*)(kmp_dyna_lock_t *))__kmp_##op##_##l##_lock,
2936 void (*__kmp_direct_destroy[])(kmp_dyna_lock_t *) = {
2937     __kmp_destroy_indirect_lock, 0, KMP_FOREACH_D_LOCK(expand, destroy)};
2938 #undef expand
2939 
2940 // set/acquire functions
2941 #define expand(l, op)                                                          \
2942   0, (int (*)(kmp_dyna_lock_t *, kmp_int32))__kmp_##op##_##l##_lock,
2943 static int (*direct_set[])(kmp_dyna_lock_t *, kmp_int32) = {
2944     __kmp_set_indirect_lock, 0, KMP_FOREACH_D_LOCK(expand, acquire)};
2945 #undef expand
2946 #define expand(l, op)                                                          \
2947   0, (int (*)(kmp_dyna_lock_t *, kmp_int32))__kmp_##op##_##l##_lock_with_checks,
2948 static int (*direct_set_check[])(kmp_dyna_lock_t *, kmp_int32) = {
2949     __kmp_set_indirect_lock_with_checks, 0,
2950     KMP_FOREACH_D_LOCK(expand, acquire)};
2951 #undef expand
2952 
2953 // unset/release and test functions
2954 #define expand(l, op)                                                          \
2955   0, (int (*)(kmp_dyna_lock_t *, kmp_int32))__kmp_##op##_##l##_lock,
2956 static int (*direct_unset[])(kmp_dyna_lock_t *, kmp_int32) = {
2957     __kmp_unset_indirect_lock, 0, KMP_FOREACH_D_LOCK(expand, release)};
2958 static int (*direct_test[])(kmp_dyna_lock_t *, kmp_int32) = {
2959     __kmp_test_indirect_lock, 0, KMP_FOREACH_D_LOCK(expand, test)};
2960 #undef expand
2961 #define expand(l, op)                                                          \
2962   0, (int (*)(kmp_dyna_lock_t *, kmp_int32))__kmp_##op##_##l##_lock_with_checks,
2963 static int (*direct_unset_check[])(kmp_dyna_lock_t *, kmp_int32) = {
2964     __kmp_unset_indirect_lock_with_checks, 0,
2965     KMP_FOREACH_D_LOCK(expand, release)};
2966 static int (*direct_test_check[])(kmp_dyna_lock_t *, kmp_int32) = {
2967     __kmp_test_indirect_lock_with_checks, 0, KMP_FOREACH_D_LOCK(expand, test)};
2968 #undef expand
2969 
2970 // Exposes only one set of jump tables (*lock or *lock_with_checks).
2971 int (*(*__kmp_direct_set))(kmp_dyna_lock_t *, kmp_int32) = 0;
2972 int (*(*__kmp_direct_unset))(kmp_dyna_lock_t *, kmp_int32) = 0;
2973 int (*(*__kmp_direct_test))(kmp_dyna_lock_t *, kmp_int32) = 0;
2974 
2975 // Jump tables for the indirect lock functions
2976 #define expand(l, op) (void (*)(kmp_user_lock_p)) __kmp_##op##_##l##_##lock,
2977 void (*__kmp_indirect_init[])(kmp_user_lock_p) = {
2978     KMP_FOREACH_I_LOCK(expand, init)};
2979 void (*__kmp_indirect_destroy[])(kmp_user_lock_p) = {
2980     KMP_FOREACH_I_LOCK(expand, destroy)};
2981 #undef expand
2982 
2983 // set/acquire functions
2984 #define expand(l, op)                                                          \
2985   (int (*)(kmp_user_lock_p, kmp_int32)) __kmp_##op##_##l##_##lock,
2986 static int (*indirect_set[])(kmp_user_lock_p,
2987                              kmp_int32) = {KMP_FOREACH_I_LOCK(expand, acquire)};
2988 #undef expand
2989 #define expand(l, op)                                                          \
2990   (int (*)(kmp_user_lock_p, kmp_int32)) __kmp_##op##_##l##_##lock_with_checks,
2991 static int (*indirect_set_check[])(kmp_user_lock_p, kmp_int32) = {
2992     KMP_FOREACH_I_LOCK(expand, acquire)};
2993 #undef expand
2994 
2995 // unset/release and test functions
2996 #define expand(l, op)                                                          \
2997   (int (*)(kmp_user_lock_p, kmp_int32)) __kmp_##op##_##l##_##lock,
2998 static int (*indirect_unset[])(kmp_user_lock_p, kmp_int32) = {
2999     KMP_FOREACH_I_LOCK(expand, release)};
3000 static int (*indirect_test[])(kmp_user_lock_p,
3001                               kmp_int32) = {KMP_FOREACH_I_LOCK(expand, test)};
3002 #undef expand
3003 #define expand(l, op)                                                          \
3004   (int (*)(kmp_user_lock_p, kmp_int32)) __kmp_##op##_##l##_##lock_with_checks,
3005 static int (*indirect_unset_check[])(kmp_user_lock_p, kmp_int32) = {
3006     KMP_FOREACH_I_LOCK(expand, release)};
3007 static int (*indirect_test_check[])(kmp_user_lock_p, kmp_int32) = {
3008     KMP_FOREACH_I_LOCK(expand, test)};
3009 #undef expand
3010 
3011 // Exposes only one jump tables (*lock or *lock_with_checks).
3012 int (*(*__kmp_indirect_set))(kmp_user_lock_p, kmp_int32) = 0;
3013 int (*(*__kmp_indirect_unset))(kmp_user_lock_p, kmp_int32) = 0;
3014 int (*(*__kmp_indirect_test))(kmp_user_lock_p, kmp_int32) = 0;
3015 
3016 // Lock index table.
3017 kmp_indirect_lock_table_t __kmp_i_lock_table;
3018 
3019 // Size of indirect locks.
3020 static kmp_uint32 __kmp_indirect_lock_size[KMP_NUM_I_LOCKS] = {0};
3021 
3022 // Jump tables for lock accessor/modifier.
3023 void (*__kmp_indirect_set_location[KMP_NUM_I_LOCKS])(kmp_user_lock_p,
3024                                                      const ident_t *) = {0};
3025 void (*__kmp_indirect_set_flags[KMP_NUM_I_LOCKS])(kmp_user_lock_p,
3026                                                   kmp_lock_flags_t) = {0};
3027 const ident_t *(*__kmp_indirect_get_location[KMP_NUM_I_LOCKS])(
3028     kmp_user_lock_p) = {0};
3029 kmp_lock_flags_t (*__kmp_indirect_get_flags[KMP_NUM_I_LOCKS])(
3030     kmp_user_lock_p) = {0};
3031 
3032 // Use different lock pools for different lock types.
3033 static kmp_indirect_lock_t *__kmp_indirect_lock_pool[KMP_NUM_I_LOCKS] = {0};
3034 
3035 // User lock allocator for dynamically dispatched indirect locks. Every entry of
3036 // the indirect lock table holds the address and type of the allocated indrect
3037 // lock (kmp_indirect_lock_t), and the size of the table doubles when it is
3038 // full. A destroyed indirect lock object is returned to the reusable pool of
3039 // locks, unique to each lock type.
3040 kmp_indirect_lock_t *__kmp_allocate_indirect_lock(void **user_lock,
3041                                                   kmp_int32 gtid,
3042                                                   kmp_indirect_locktag_t tag) {
3043   kmp_indirect_lock_t *lck;
3044   kmp_lock_index_t idx;
3045 
3046   __kmp_acquire_lock(&__kmp_global_lock, gtid);
3047 
3048   if (__kmp_indirect_lock_pool[tag] != NULL) {
3049     // Reuse the allocated and destroyed lock object
3050     lck = __kmp_indirect_lock_pool[tag];
3051     if (OMP_LOCK_T_SIZE < sizeof(void *))
3052       idx = lck->lock->pool.index;
3053     __kmp_indirect_lock_pool[tag] = (kmp_indirect_lock_t *)lck->lock->pool.next;
3054     KA_TRACE(20, ("__kmp_allocate_indirect_lock: reusing an existing lock %p\n",
3055                   lck));
3056   } else {
3057     idx = __kmp_i_lock_table.next;
3058     // Check capacity and double the size if it is full
3059     if (idx == __kmp_i_lock_table.size) {
3060       // Double up the space for block pointers
3061       int row = __kmp_i_lock_table.size / KMP_I_LOCK_CHUNK;
3062       kmp_indirect_lock_t **new_table = (kmp_indirect_lock_t **)__kmp_allocate(
3063           2 * row * sizeof(kmp_indirect_lock_t *));
3064       KMP_MEMCPY(new_table, __kmp_i_lock_table.table,
3065                  row * sizeof(kmp_indirect_lock_t *));
3066       kmp_indirect_lock_t **old_table = __kmp_i_lock_table.table;
3067       __kmp_i_lock_table.table = new_table;
3068       __kmp_free(old_table);
3069       // Allocate new objects in the new blocks
3070       for (int i = row; i < 2 * row; ++i)
3071         *(__kmp_i_lock_table.table + i) = (kmp_indirect_lock_t *)__kmp_allocate(
3072             KMP_I_LOCK_CHUNK * sizeof(kmp_indirect_lock_t));
3073       __kmp_i_lock_table.size = 2 * idx;
3074     }
3075     __kmp_i_lock_table.next++;
3076     lck = KMP_GET_I_LOCK(idx);
3077     // Allocate a new base lock object
3078     lck->lock = (kmp_user_lock_p)__kmp_allocate(__kmp_indirect_lock_size[tag]);
3079     KA_TRACE(20,
3080              ("__kmp_allocate_indirect_lock: allocated a new lock %p\n", lck));
3081   }
3082 
3083   __kmp_release_lock(&__kmp_global_lock, gtid);
3084 
3085   lck->type = tag;
3086 
3087   if (OMP_LOCK_T_SIZE < sizeof(void *)) {
3088     *((kmp_lock_index_t *)user_lock) = idx
3089                                        << 1; // indirect lock word must be even
3090   } else {
3091     *((kmp_indirect_lock_t **)user_lock) = lck;
3092   }
3093 
3094   return lck;
3095 }
3096 
3097 // User lock lookup for dynamically dispatched locks.
3098 static __forceinline kmp_indirect_lock_t *
3099 __kmp_lookup_indirect_lock(void **user_lock, const char *func) {
3100   if (__kmp_env_consistency_check) {
3101     kmp_indirect_lock_t *lck = NULL;
3102     if (user_lock == NULL) {
3103       KMP_FATAL(LockIsUninitialized, func);
3104     }
3105     if (OMP_LOCK_T_SIZE < sizeof(void *)) {
3106       kmp_lock_index_t idx = KMP_EXTRACT_I_INDEX(user_lock);
3107       if (idx >= __kmp_i_lock_table.size) {
3108         KMP_FATAL(LockIsUninitialized, func);
3109       }
3110       lck = KMP_GET_I_LOCK(idx);
3111     } else {
3112       lck = *((kmp_indirect_lock_t **)user_lock);
3113     }
3114     if (lck == NULL) {
3115       KMP_FATAL(LockIsUninitialized, func);
3116     }
3117     return lck;
3118   } else {
3119     if (OMP_LOCK_T_SIZE < sizeof(void *)) {
3120       return KMP_GET_I_LOCK(KMP_EXTRACT_I_INDEX(user_lock));
3121     } else {
3122       return *((kmp_indirect_lock_t **)user_lock);
3123     }
3124   }
3125 }
3126 
3127 static void __kmp_init_indirect_lock(kmp_dyna_lock_t *lock,
3128                                      kmp_dyna_lockseq_t seq) {
3129 #if KMP_USE_ADAPTIVE_LOCKS
3130   if (seq == lockseq_adaptive && !__kmp_cpuinfo.rtm) {
3131     KMP_WARNING(AdaptiveNotSupported, "kmp_lockseq_t", "adaptive");
3132     seq = lockseq_queuing;
3133   }
3134 #endif
3135 #if KMP_USE_TSX
3136   if (seq == lockseq_rtm && !__kmp_cpuinfo.rtm) {
3137     seq = lockseq_queuing;
3138   }
3139 #endif
3140   kmp_indirect_locktag_t tag = KMP_GET_I_TAG(seq);
3141   kmp_indirect_lock_t *l =
3142       __kmp_allocate_indirect_lock((void **)lock, __kmp_entry_gtid(), tag);
3143   KMP_I_LOCK_FUNC(l, init)(l->lock);
3144   KA_TRACE(
3145       20, ("__kmp_init_indirect_lock: initialized indirect lock with type#%d\n",
3146            seq));
3147 }
3148 
3149 static void __kmp_destroy_indirect_lock(kmp_dyna_lock_t *lock) {
3150   kmp_uint32 gtid = __kmp_entry_gtid();
3151   kmp_indirect_lock_t *l =
3152       __kmp_lookup_indirect_lock((void **)lock, "omp_destroy_lock");
3153   KMP_I_LOCK_FUNC(l, destroy)(l->lock);
3154   kmp_indirect_locktag_t tag = l->type;
3155 
3156   __kmp_acquire_lock(&__kmp_global_lock, gtid);
3157 
3158   // Use the base lock's space to keep the pool chain.
3159   l->lock->pool.next = (kmp_user_lock_p)__kmp_indirect_lock_pool[tag];
3160   if (OMP_LOCK_T_SIZE < sizeof(void *)) {
3161     l->lock->pool.index = KMP_EXTRACT_I_INDEX(lock);
3162   }
3163   __kmp_indirect_lock_pool[tag] = l;
3164 
3165   __kmp_release_lock(&__kmp_global_lock, gtid);
3166 }
3167 
3168 static int __kmp_set_indirect_lock(kmp_dyna_lock_t *lock, kmp_int32 gtid) {
3169   kmp_indirect_lock_t *l = KMP_LOOKUP_I_LOCK(lock);
3170   return KMP_I_LOCK_FUNC(l, set)(l->lock, gtid);
3171 }
3172 
3173 static int __kmp_unset_indirect_lock(kmp_dyna_lock_t *lock, kmp_int32 gtid) {
3174   kmp_indirect_lock_t *l = KMP_LOOKUP_I_LOCK(lock);
3175   return KMP_I_LOCK_FUNC(l, unset)(l->lock, gtid);
3176 }
3177 
3178 static int __kmp_test_indirect_lock(kmp_dyna_lock_t *lock, kmp_int32 gtid) {
3179   kmp_indirect_lock_t *l = KMP_LOOKUP_I_LOCK(lock);
3180   return KMP_I_LOCK_FUNC(l, test)(l->lock, gtid);
3181 }
3182 
3183 static int __kmp_set_indirect_lock_with_checks(kmp_dyna_lock_t *lock,
3184                                                kmp_int32 gtid) {
3185   kmp_indirect_lock_t *l =
3186       __kmp_lookup_indirect_lock((void **)lock, "omp_set_lock");
3187   return KMP_I_LOCK_FUNC(l, set)(l->lock, gtid);
3188 }
3189 
3190 static int __kmp_unset_indirect_lock_with_checks(kmp_dyna_lock_t *lock,
3191                                                  kmp_int32 gtid) {
3192   kmp_indirect_lock_t *l =
3193       __kmp_lookup_indirect_lock((void **)lock, "omp_unset_lock");
3194   return KMP_I_LOCK_FUNC(l, unset)(l->lock, gtid);
3195 }
3196 
3197 static int __kmp_test_indirect_lock_with_checks(kmp_dyna_lock_t *lock,
3198                                                 kmp_int32 gtid) {
3199   kmp_indirect_lock_t *l =
3200       __kmp_lookup_indirect_lock((void **)lock, "omp_test_lock");
3201   return KMP_I_LOCK_FUNC(l, test)(l->lock, gtid);
3202 }
3203 
3204 kmp_dyna_lockseq_t __kmp_user_lock_seq = lockseq_queuing;
3205 
3206 // This is used only in kmp_error.cpp when consistency checking is on.
3207 kmp_int32 __kmp_get_user_lock_owner(kmp_user_lock_p lck, kmp_uint32 seq) {
3208   switch (seq) {
3209   case lockseq_tas:
3210   case lockseq_nested_tas:
3211     return __kmp_get_tas_lock_owner((kmp_tas_lock_t *)lck);
3212 #if KMP_USE_FUTEX
3213   case lockseq_futex:
3214   case lockseq_nested_futex:
3215     return __kmp_get_futex_lock_owner((kmp_futex_lock_t *)lck);
3216 #endif
3217   case lockseq_ticket:
3218   case lockseq_nested_ticket:
3219     return __kmp_get_ticket_lock_owner((kmp_ticket_lock_t *)lck);
3220   case lockseq_queuing:
3221   case lockseq_nested_queuing:
3222 #if KMP_USE_ADAPTIVE_LOCKS
3223   case lockseq_adaptive:
3224 #endif
3225     return __kmp_get_queuing_lock_owner((kmp_queuing_lock_t *)lck);
3226   case lockseq_drdpa:
3227   case lockseq_nested_drdpa:
3228     return __kmp_get_drdpa_lock_owner((kmp_drdpa_lock_t *)lck);
3229   default:
3230     return 0;
3231   }
3232 }
3233 
3234 // Initializes data for dynamic user locks.
3235 void __kmp_init_dynamic_user_locks() {
3236   // Initialize jump table for the lock functions
3237   if (__kmp_env_consistency_check) {
3238     __kmp_direct_set = direct_set_check;
3239     __kmp_direct_unset = direct_unset_check;
3240     __kmp_direct_test = direct_test_check;
3241     __kmp_indirect_set = indirect_set_check;
3242     __kmp_indirect_unset = indirect_unset_check;
3243     __kmp_indirect_test = indirect_test_check;
3244   } else {
3245     __kmp_direct_set = direct_set;
3246     __kmp_direct_unset = direct_unset;
3247     __kmp_direct_test = direct_test;
3248     __kmp_indirect_set = indirect_set;
3249     __kmp_indirect_unset = indirect_unset;
3250     __kmp_indirect_test = indirect_test;
3251   }
3252   // If the user locks have already been initialized, then return. Allow the
3253   // switch between different KMP_CONSISTENCY_CHECK values, but do not allocate
3254   // new lock tables if they have already been allocated.
3255   if (__kmp_init_user_locks)
3256     return;
3257 
3258   // Initialize lock index table
3259   __kmp_i_lock_table.size = KMP_I_LOCK_CHUNK;
3260   __kmp_i_lock_table.table =
3261       (kmp_indirect_lock_t **)__kmp_allocate(sizeof(kmp_indirect_lock_t *));
3262   *(__kmp_i_lock_table.table) = (kmp_indirect_lock_t *)__kmp_allocate(
3263       KMP_I_LOCK_CHUNK * sizeof(kmp_indirect_lock_t));
3264   __kmp_i_lock_table.next = 0;
3265 
3266   // Indirect lock size
3267   __kmp_indirect_lock_size[locktag_ticket] = sizeof(kmp_ticket_lock_t);
3268   __kmp_indirect_lock_size[locktag_queuing] = sizeof(kmp_queuing_lock_t);
3269 #if KMP_USE_ADAPTIVE_LOCKS
3270   __kmp_indirect_lock_size[locktag_adaptive] = sizeof(kmp_adaptive_lock_t);
3271 #endif
3272   __kmp_indirect_lock_size[locktag_drdpa] = sizeof(kmp_drdpa_lock_t);
3273 #if KMP_USE_TSX
3274   __kmp_indirect_lock_size[locktag_rtm] = sizeof(kmp_queuing_lock_t);
3275 #endif
3276   __kmp_indirect_lock_size[locktag_nested_tas] = sizeof(kmp_tas_lock_t);
3277 #if KMP_USE_FUTEX
3278   __kmp_indirect_lock_size[locktag_nested_futex] = sizeof(kmp_futex_lock_t);
3279 #endif
3280   __kmp_indirect_lock_size[locktag_nested_ticket] = sizeof(kmp_ticket_lock_t);
3281   __kmp_indirect_lock_size[locktag_nested_queuing] = sizeof(kmp_queuing_lock_t);
3282   __kmp_indirect_lock_size[locktag_nested_drdpa] = sizeof(kmp_drdpa_lock_t);
3283 
3284 // Initialize lock accessor/modifier
3285 #define fill_jumps(table, expand, sep)                                         \
3286   {                                                                            \
3287     table[locktag##sep##ticket] = expand(ticket);                              \
3288     table[locktag##sep##queuing] = expand(queuing);                            \
3289     table[locktag##sep##drdpa] = expand(drdpa);                                \
3290   }
3291 
3292 #if KMP_USE_ADAPTIVE_LOCKS
3293 #define fill_table(table, expand)                                              \
3294   {                                                                            \
3295     fill_jumps(table, expand, _);                                              \
3296     table[locktag_adaptive] = expand(queuing);                                 \
3297     fill_jumps(table, expand, _nested_);                                       \
3298   }
3299 #else
3300 #define fill_table(table, expand)                                              \
3301   {                                                                            \
3302     fill_jumps(table, expand, _);                                              \
3303     fill_jumps(table, expand, _nested_);                                       \
3304   }
3305 #endif // KMP_USE_ADAPTIVE_LOCKS
3306 
3307 #define expand(l)                                                              \
3308   (void (*)(kmp_user_lock_p, const ident_t *)) __kmp_set_##l##_lock_location
3309   fill_table(__kmp_indirect_set_location, expand);
3310 #undef expand
3311 #define expand(l)                                                              \
3312   (void (*)(kmp_user_lock_p, kmp_lock_flags_t)) __kmp_set_##l##_lock_flags
3313   fill_table(__kmp_indirect_set_flags, expand);
3314 #undef expand
3315 #define expand(l)                                                              \
3316   (const ident_t *(*)(kmp_user_lock_p)) __kmp_get_##l##_lock_location
3317   fill_table(__kmp_indirect_get_location, expand);
3318 #undef expand
3319 #define expand(l)                                                              \
3320   (kmp_lock_flags_t(*)(kmp_user_lock_p)) __kmp_get_##l##_lock_flags
3321   fill_table(__kmp_indirect_get_flags, expand);
3322 #undef expand
3323 
3324   __kmp_init_user_locks = TRUE;
3325 }
3326 
3327 // Clean up the lock table.
3328 void __kmp_cleanup_indirect_user_locks() {
3329   kmp_lock_index_t i;
3330   int k;
3331 
3332   // Clean up locks in the pools first (they were already destroyed before going
3333   // into the pools).
3334   for (k = 0; k < KMP_NUM_I_LOCKS; ++k) {
3335     kmp_indirect_lock_t *l = __kmp_indirect_lock_pool[k];
3336     while (l != NULL) {
3337       kmp_indirect_lock_t *ll = l;
3338       l = (kmp_indirect_lock_t *)l->lock->pool.next;
3339       KA_TRACE(20, ("__kmp_cleanup_indirect_user_locks: freeing %p from pool\n",
3340                     ll));
3341       __kmp_free(ll->lock);
3342       ll->lock = NULL;
3343     }
3344     __kmp_indirect_lock_pool[k] = NULL;
3345   }
3346   // Clean up the remaining undestroyed locks.
3347   for (i = 0; i < __kmp_i_lock_table.next; i++) {
3348     kmp_indirect_lock_t *l = KMP_GET_I_LOCK(i);
3349     if (l->lock != NULL) {
3350       // Locks not destroyed explicitly need to be destroyed here.
3351       KMP_I_LOCK_FUNC(l, destroy)(l->lock);
3352       KA_TRACE(
3353           20,
3354           ("__kmp_cleanup_indirect_user_locks: destroy/freeing %p from table\n",
3355            l));
3356       __kmp_free(l->lock);
3357     }
3358   }
3359   // Free the table
3360   for (i = 0; i < __kmp_i_lock_table.size / KMP_I_LOCK_CHUNK; i++)
3361     __kmp_free(__kmp_i_lock_table.table[i]);
3362   __kmp_free(__kmp_i_lock_table.table);
3363 
3364   __kmp_init_user_locks = FALSE;
3365 }
3366 
3367 enum kmp_lock_kind __kmp_user_lock_kind = lk_default;
3368 int __kmp_num_locks_in_block = 1; // FIXME - tune this value
3369 
3370 #else // KMP_USE_DYNAMIC_LOCK
3371 
3372 /* user locks
3373  * They are implemented as a table of function pointers which are set to the
3374  * lock functions of the appropriate kind, once that has been determined. */
3375 
3376 enum kmp_lock_kind __kmp_user_lock_kind = lk_default;
3377 
3378 size_t __kmp_base_user_lock_size = 0;
3379 size_t __kmp_user_lock_size = 0;
3380 
3381 kmp_int32 (*__kmp_get_user_lock_owner_)(kmp_user_lock_p lck) = NULL;
3382 int (*__kmp_acquire_user_lock_with_checks_)(kmp_user_lock_p lck,
3383                                             kmp_int32 gtid) = NULL;
3384 
3385 int (*__kmp_test_user_lock_with_checks_)(kmp_user_lock_p lck,
3386                                          kmp_int32 gtid) = NULL;
3387 int (*__kmp_release_user_lock_with_checks_)(kmp_user_lock_p lck,
3388                                             kmp_int32 gtid) = NULL;
3389 void (*__kmp_init_user_lock_with_checks_)(kmp_user_lock_p lck) = NULL;
3390 void (*__kmp_destroy_user_lock_)(kmp_user_lock_p lck) = NULL;
3391 void (*__kmp_destroy_user_lock_with_checks_)(kmp_user_lock_p lck) = NULL;
3392 int (*__kmp_acquire_nested_user_lock_with_checks_)(kmp_user_lock_p lck,
3393                                                    kmp_int32 gtid) = NULL;
3394 
3395 int (*__kmp_test_nested_user_lock_with_checks_)(kmp_user_lock_p lck,
3396                                                 kmp_int32 gtid) = NULL;
3397 int (*__kmp_release_nested_user_lock_with_checks_)(kmp_user_lock_p lck,
3398                                                    kmp_int32 gtid) = NULL;
3399 void (*__kmp_init_nested_user_lock_with_checks_)(kmp_user_lock_p lck) = NULL;
3400 void (*__kmp_destroy_nested_user_lock_with_checks_)(kmp_user_lock_p lck) = NULL;
3401 
3402 int (*__kmp_is_user_lock_initialized_)(kmp_user_lock_p lck) = NULL;
3403 const ident_t *(*__kmp_get_user_lock_location_)(kmp_user_lock_p lck) = NULL;
3404 void (*__kmp_set_user_lock_location_)(kmp_user_lock_p lck,
3405                                       const ident_t *loc) = NULL;
3406 kmp_lock_flags_t (*__kmp_get_user_lock_flags_)(kmp_user_lock_p lck) = NULL;
3407 void (*__kmp_set_user_lock_flags_)(kmp_user_lock_p lck,
3408                                    kmp_lock_flags_t flags) = NULL;
3409 
3410 void __kmp_set_user_lock_vptrs(kmp_lock_kind_t user_lock_kind) {
3411   switch (user_lock_kind) {
3412   case lk_default:
3413   default:
3414     KMP_ASSERT(0);
3415 
3416   case lk_tas: {
3417     __kmp_base_user_lock_size = sizeof(kmp_base_tas_lock_t);
3418     __kmp_user_lock_size = sizeof(kmp_tas_lock_t);
3419 
3420     __kmp_get_user_lock_owner_ =
3421         (kmp_int32(*)(kmp_user_lock_p))(&__kmp_get_tas_lock_owner);
3422 
3423     if (__kmp_env_consistency_check) {
3424       KMP_BIND_USER_LOCK_WITH_CHECKS(tas);
3425       KMP_BIND_NESTED_USER_LOCK_WITH_CHECKS(tas);
3426     } else {
3427       KMP_BIND_USER_LOCK(tas);
3428       KMP_BIND_NESTED_USER_LOCK(tas);
3429     }
3430 
3431     __kmp_destroy_user_lock_ =
3432         (void (*)(kmp_user_lock_p))(&__kmp_destroy_tas_lock);
3433 
3434     __kmp_is_user_lock_initialized_ = (int (*)(kmp_user_lock_p))NULL;
3435 
3436     __kmp_get_user_lock_location_ = (const ident_t *(*)(kmp_user_lock_p))NULL;
3437 
3438     __kmp_set_user_lock_location_ =
3439         (void (*)(kmp_user_lock_p, const ident_t *))NULL;
3440 
3441     __kmp_get_user_lock_flags_ = (kmp_lock_flags_t(*)(kmp_user_lock_p))NULL;
3442 
3443     __kmp_set_user_lock_flags_ =
3444         (void (*)(kmp_user_lock_p, kmp_lock_flags_t))NULL;
3445   } break;
3446 
3447 #if KMP_USE_FUTEX
3448 
3449   case lk_futex: {
3450     __kmp_base_user_lock_size = sizeof(kmp_base_futex_lock_t);
3451     __kmp_user_lock_size = sizeof(kmp_futex_lock_t);
3452 
3453     __kmp_get_user_lock_owner_ =
3454         (kmp_int32(*)(kmp_user_lock_p))(&__kmp_get_futex_lock_owner);
3455 
3456     if (__kmp_env_consistency_check) {
3457       KMP_BIND_USER_LOCK_WITH_CHECKS(futex);
3458       KMP_BIND_NESTED_USER_LOCK_WITH_CHECKS(futex);
3459     } else {
3460       KMP_BIND_USER_LOCK(futex);
3461       KMP_BIND_NESTED_USER_LOCK(futex);
3462     }
3463 
3464     __kmp_destroy_user_lock_ =
3465         (void (*)(kmp_user_lock_p))(&__kmp_destroy_futex_lock);
3466 
3467     __kmp_is_user_lock_initialized_ = (int (*)(kmp_user_lock_p))NULL;
3468 
3469     __kmp_get_user_lock_location_ = (const ident_t *(*)(kmp_user_lock_p))NULL;
3470 
3471     __kmp_set_user_lock_location_ =
3472         (void (*)(kmp_user_lock_p, const ident_t *))NULL;
3473 
3474     __kmp_get_user_lock_flags_ = (kmp_lock_flags_t(*)(kmp_user_lock_p))NULL;
3475 
3476     __kmp_set_user_lock_flags_ =
3477         (void (*)(kmp_user_lock_p, kmp_lock_flags_t))NULL;
3478   } break;
3479 
3480 #endif // KMP_USE_FUTEX
3481 
3482   case lk_ticket: {
3483     __kmp_base_user_lock_size = sizeof(kmp_base_ticket_lock_t);
3484     __kmp_user_lock_size = sizeof(kmp_ticket_lock_t);
3485 
3486     __kmp_get_user_lock_owner_ =
3487         (kmp_int32(*)(kmp_user_lock_p))(&__kmp_get_ticket_lock_owner);
3488 
3489     if (__kmp_env_consistency_check) {
3490       KMP_BIND_USER_LOCK_WITH_CHECKS(ticket);
3491       KMP_BIND_NESTED_USER_LOCK_WITH_CHECKS(ticket);
3492     } else {
3493       KMP_BIND_USER_LOCK(ticket);
3494       KMP_BIND_NESTED_USER_LOCK(ticket);
3495     }
3496 
3497     __kmp_destroy_user_lock_ =
3498         (void (*)(kmp_user_lock_p))(&__kmp_destroy_ticket_lock);
3499 
3500     __kmp_is_user_lock_initialized_ =
3501         (int (*)(kmp_user_lock_p))(&__kmp_is_ticket_lock_initialized);
3502 
3503     __kmp_get_user_lock_location_ =
3504         (const ident_t *(*)(kmp_user_lock_p))(&__kmp_get_ticket_lock_location);
3505 
3506     __kmp_set_user_lock_location_ = (void (*)(
3507         kmp_user_lock_p, const ident_t *))(&__kmp_set_ticket_lock_location);
3508 
3509     __kmp_get_user_lock_flags_ =
3510         (kmp_lock_flags_t(*)(kmp_user_lock_p))(&__kmp_get_ticket_lock_flags);
3511 
3512     __kmp_set_user_lock_flags_ = (void (*)(kmp_user_lock_p, kmp_lock_flags_t))(
3513         &__kmp_set_ticket_lock_flags);
3514   } break;
3515 
3516   case lk_queuing: {
3517     __kmp_base_user_lock_size = sizeof(kmp_base_queuing_lock_t);
3518     __kmp_user_lock_size = sizeof(kmp_queuing_lock_t);
3519 
3520     __kmp_get_user_lock_owner_ =
3521         (kmp_int32(*)(kmp_user_lock_p))(&__kmp_get_queuing_lock_owner);
3522 
3523     if (__kmp_env_consistency_check) {
3524       KMP_BIND_USER_LOCK_WITH_CHECKS(queuing);
3525       KMP_BIND_NESTED_USER_LOCK_WITH_CHECKS(queuing);
3526     } else {
3527       KMP_BIND_USER_LOCK(queuing);
3528       KMP_BIND_NESTED_USER_LOCK(queuing);
3529     }
3530 
3531     __kmp_destroy_user_lock_ =
3532         (void (*)(kmp_user_lock_p))(&__kmp_destroy_queuing_lock);
3533 
3534     __kmp_is_user_lock_initialized_ =
3535         (int (*)(kmp_user_lock_p))(&__kmp_is_queuing_lock_initialized);
3536 
3537     __kmp_get_user_lock_location_ =
3538         (const ident_t *(*)(kmp_user_lock_p))(&__kmp_get_queuing_lock_location);
3539 
3540     __kmp_set_user_lock_location_ = (void (*)(
3541         kmp_user_lock_p, const ident_t *))(&__kmp_set_queuing_lock_location);
3542 
3543     __kmp_get_user_lock_flags_ =
3544         (kmp_lock_flags_t(*)(kmp_user_lock_p))(&__kmp_get_queuing_lock_flags);
3545 
3546     __kmp_set_user_lock_flags_ = (void (*)(kmp_user_lock_p, kmp_lock_flags_t))(
3547         &__kmp_set_queuing_lock_flags);
3548   } break;
3549 
3550 #if KMP_USE_ADAPTIVE_LOCKS
3551   case lk_adaptive: {
3552     __kmp_base_user_lock_size = sizeof(kmp_base_adaptive_lock_t);
3553     __kmp_user_lock_size = sizeof(kmp_adaptive_lock_t);
3554 
3555     __kmp_get_user_lock_owner_ =
3556         (kmp_int32(*)(kmp_user_lock_p))(&__kmp_get_queuing_lock_owner);
3557 
3558     if (__kmp_env_consistency_check) {
3559       KMP_BIND_USER_LOCK_WITH_CHECKS(adaptive);
3560     } else {
3561       KMP_BIND_USER_LOCK(adaptive);
3562     }
3563 
3564     __kmp_destroy_user_lock_ =
3565         (void (*)(kmp_user_lock_p))(&__kmp_destroy_adaptive_lock);
3566 
3567     __kmp_is_user_lock_initialized_ =
3568         (int (*)(kmp_user_lock_p))(&__kmp_is_queuing_lock_initialized);
3569 
3570     __kmp_get_user_lock_location_ =
3571         (const ident_t *(*)(kmp_user_lock_p))(&__kmp_get_queuing_lock_location);
3572 
3573     __kmp_set_user_lock_location_ = (void (*)(
3574         kmp_user_lock_p, const ident_t *))(&__kmp_set_queuing_lock_location);
3575 
3576     __kmp_get_user_lock_flags_ =
3577         (kmp_lock_flags_t(*)(kmp_user_lock_p))(&__kmp_get_queuing_lock_flags);
3578 
3579     __kmp_set_user_lock_flags_ = (void (*)(kmp_user_lock_p, kmp_lock_flags_t))(
3580         &__kmp_set_queuing_lock_flags);
3581 
3582   } break;
3583 #endif // KMP_USE_ADAPTIVE_LOCKS
3584 
3585   case lk_drdpa: {
3586     __kmp_base_user_lock_size = sizeof(kmp_base_drdpa_lock_t);
3587     __kmp_user_lock_size = sizeof(kmp_drdpa_lock_t);
3588 
3589     __kmp_get_user_lock_owner_ =
3590         (kmp_int32(*)(kmp_user_lock_p))(&__kmp_get_drdpa_lock_owner);
3591 
3592     if (__kmp_env_consistency_check) {
3593       KMP_BIND_USER_LOCK_WITH_CHECKS(drdpa);
3594       KMP_BIND_NESTED_USER_LOCK_WITH_CHECKS(drdpa);
3595     } else {
3596       KMP_BIND_USER_LOCK(drdpa);
3597       KMP_BIND_NESTED_USER_LOCK(drdpa);
3598     }
3599 
3600     __kmp_destroy_user_lock_ =
3601         (void (*)(kmp_user_lock_p))(&__kmp_destroy_drdpa_lock);
3602 
3603     __kmp_is_user_lock_initialized_ =
3604         (int (*)(kmp_user_lock_p))(&__kmp_is_drdpa_lock_initialized);
3605 
3606     __kmp_get_user_lock_location_ =
3607         (const ident_t *(*)(kmp_user_lock_p))(&__kmp_get_drdpa_lock_location);
3608 
3609     __kmp_set_user_lock_location_ = (void (*)(
3610         kmp_user_lock_p, const ident_t *))(&__kmp_set_drdpa_lock_location);
3611 
3612     __kmp_get_user_lock_flags_ =
3613         (kmp_lock_flags_t(*)(kmp_user_lock_p))(&__kmp_get_drdpa_lock_flags);
3614 
3615     __kmp_set_user_lock_flags_ = (void (*)(kmp_user_lock_p, kmp_lock_flags_t))(
3616         &__kmp_set_drdpa_lock_flags);
3617   } break;
3618   }
3619 }
3620 
3621 // ----------------------------------------------------------------------------
3622 // User lock table & lock allocation
3623 
3624 kmp_lock_table_t __kmp_user_lock_table = {1, 0, NULL};
3625 kmp_user_lock_p __kmp_lock_pool = NULL;
3626 
3627 // Lock block-allocation support.
3628 kmp_block_of_locks *__kmp_lock_blocks = NULL;
3629 int __kmp_num_locks_in_block = 1; // FIXME - tune this value
3630 
3631 static kmp_lock_index_t __kmp_lock_table_insert(kmp_user_lock_p lck) {
3632   // Assume that kmp_global_lock is held upon entry/exit.
3633   kmp_lock_index_t index;
3634   if (__kmp_user_lock_table.used >= __kmp_user_lock_table.allocated) {
3635     kmp_lock_index_t size;
3636     kmp_user_lock_p *table;
3637     // Reallocate lock table.
3638     if (__kmp_user_lock_table.allocated == 0) {
3639       size = 1024;
3640     } else {
3641       size = __kmp_user_lock_table.allocated * 2;
3642     }
3643     table = (kmp_user_lock_p *)__kmp_allocate(sizeof(kmp_user_lock_p) * size);
3644     KMP_MEMCPY(table + 1, __kmp_user_lock_table.table + 1,
3645                sizeof(kmp_user_lock_p) * (__kmp_user_lock_table.used - 1));
3646     table[0] = (kmp_user_lock_p)__kmp_user_lock_table.table;
3647     // We cannot free the previous table now, since it may be in use by other
3648     // threads. So save the pointer to the previous table in in the first
3649     // element of the new table. All the tables will be organized into a list,
3650     // and could be freed when library shutting down.
3651     __kmp_user_lock_table.table = table;
3652     __kmp_user_lock_table.allocated = size;
3653   }
3654   KMP_DEBUG_ASSERT(__kmp_user_lock_table.used <
3655                    __kmp_user_lock_table.allocated);
3656   index = __kmp_user_lock_table.used;
3657   __kmp_user_lock_table.table[index] = lck;
3658   ++__kmp_user_lock_table.used;
3659   return index;
3660 }
3661 
3662 static kmp_user_lock_p __kmp_lock_block_allocate() {
3663   // Assume that kmp_global_lock is held upon entry/exit.
3664   static int last_index = 0;
3665   if ((last_index >= __kmp_num_locks_in_block) || (__kmp_lock_blocks == NULL)) {
3666     // Restart the index.
3667     last_index = 0;
3668     // Need to allocate a new block.
3669     KMP_DEBUG_ASSERT(__kmp_user_lock_size > 0);
3670     size_t space_for_locks = __kmp_user_lock_size * __kmp_num_locks_in_block;
3671     char *buffer =
3672         (char *)__kmp_allocate(space_for_locks + sizeof(kmp_block_of_locks));
3673     // Set up the new block.
3674     kmp_block_of_locks *new_block =
3675         (kmp_block_of_locks *)(&buffer[space_for_locks]);
3676     new_block->next_block = __kmp_lock_blocks;
3677     new_block->locks = (void *)buffer;
3678     // Publish the new block.
3679     KMP_MB();
3680     __kmp_lock_blocks = new_block;
3681   }
3682   kmp_user_lock_p ret = (kmp_user_lock_p)(&(
3683       ((char *)(__kmp_lock_blocks->locks))[last_index * __kmp_user_lock_size]));
3684   last_index++;
3685   return ret;
3686 }
3687 
3688 // Get memory for a lock. It may be freshly allocated memory or reused memory
3689 // from lock pool.
3690 kmp_user_lock_p __kmp_user_lock_allocate(void **user_lock, kmp_int32 gtid,
3691                                          kmp_lock_flags_t flags) {
3692   kmp_user_lock_p lck;
3693   kmp_lock_index_t index;
3694   KMP_DEBUG_ASSERT(user_lock);
3695 
3696   __kmp_acquire_lock(&__kmp_global_lock, gtid);
3697 
3698   if (__kmp_lock_pool == NULL) {
3699     // Lock pool is empty. Allocate new memory.
3700 
3701     // ANNOTATION: Found no good way to express the syncronisation
3702     // between allocation and usage, so ignore the allocation
3703     ANNOTATE_IGNORE_WRITES_BEGIN();
3704     if (__kmp_num_locks_in_block <= 1) { // Tune this cutoff point.
3705       lck = (kmp_user_lock_p)__kmp_allocate(__kmp_user_lock_size);
3706     } else {
3707       lck = __kmp_lock_block_allocate();
3708     }
3709     ANNOTATE_IGNORE_WRITES_END();
3710 
3711     // Insert lock in the table so that it can be freed in __kmp_cleanup,
3712     // and debugger has info on all allocated locks.
3713     index = __kmp_lock_table_insert(lck);
3714   } else {
3715     // Pick up lock from pool.
3716     lck = __kmp_lock_pool;
3717     index = __kmp_lock_pool->pool.index;
3718     __kmp_lock_pool = __kmp_lock_pool->pool.next;
3719   }
3720 
3721   // We could potentially differentiate between nested and regular locks
3722   // here, and do the lock table lookup for regular locks only.
3723   if (OMP_LOCK_T_SIZE < sizeof(void *)) {
3724     *((kmp_lock_index_t *)user_lock) = index;
3725   } else {
3726     *((kmp_user_lock_p *)user_lock) = lck;
3727   }
3728 
3729   // mark the lock if it is critical section lock.
3730   __kmp_set_user_lock_flags(lck, flags);
3731 
3732   __kmp_release_lock(&__kmp_global_lock, gtid); // AC: TODO move this line upper
3733 
3734   return lck;
3735 }
3736 
3737 // Put lock's memory to pool for reusing.
3738 void __kmp_user_lock_free(void **user_lock, kmp_int32 gtid,
3739                           kmp_user_lock_p lck) {
3740   KMP_DEBUG_ASSERT(user_lock != NULL);
3741   KMP_DEBUG_ASSERT(lck != NULL);
3742 
3743   __kmp_acquire_lock(&__kmp_global_lock, gtid);
3744 
3745   lck->pool.next = __kmp_lock_pool;
3746   __kmp_lock_pool = lck;
3747   if (OMP_LOCK_T_SIZE < sizeof(void *)) {
3748     kmp_lock_index_t index = *((kmp_lock_index_t *)user_lock);
3749     KMP_DEBUG_ASSERT(0 < index && index <= __kmp_user_lock_table.used);
3750     lck->pool.index = index;
3751   }
3752 
3753   __kmp_release_lock(&__kmp_global_lock, gtid);
3754 }
3755 
3756 kmp_user_lock_p __kmp_lookup_user_lock(void **user_lock, char const *func) {
3757   kmp_user_lock_p lck = NULL;
3758 
3759   if (__kmp_env_consistency_check) {
3760     if (user_lock == NULL) {
3761       KMP_FATAL(LockIsUninitialized, func);
3762     }
3763   }
3764 
3765   if (OMP_LOCK_T_SIZE < sizeof(void *)) {
3766     kmp_lock_index_t index = *((kmp_lock_index_t *)user_lock);
3767     if (__kmp_env_consistency_check) {
3768       if (!(0 < index && index < __kmp_user_lock_table.used)) {
3769         KMP_FATAL(LockIsUninitialized, func);
3770       }
3771     }
3772     KMP_DEBUG_ASSERT(0 < index && index < __kmp_user_lock_table.used);
3773     KMP_DEBUG_ASSERT(__kmp_user_lock_size > 0);
3774     lck = __kmp_user_lock_table.table[index];
3775   } else {
3776     lck = *((kmp_user_lock_p *)user_lock);
3777   }
3778 
3779   if (__kmp_env_consistency_check) {
3780     if (lck == NULL) {
3781       KMP_FATAL(LockIsUninitialized, func);
3782     }
3783   }
3784 
3785   return lck;
3786 }
3787 
3788 void __kmp_cleanup_user_locks(void) {
3789   // Reset lock pool. Don't worry about lock in the pool--we will free them when
3790   // iterating through lock table (it includes all the locks, dead or alive).
3791   __kmp_lock_pool = NULL;
3792 
3793 #define IS_CRITICAL(lck)                                                       \
3794   ((__kmp_get_user_lock_flags_ != NULL) &&                                     \
3795    ((*__kmp_get_user_lock_flags_)(lck)&kmp_lf_critical_section))
3796 
3797   // Loop through lock table, free all locks.
3798   // Do not free item [0], it is reserved for lock tables list.
3799   //
3800   // FIXME - we are iterating through a list of (pointers to) objects of type
3801   // union kmp_user_lock, but we have no way of knowing whether the base type is
3802   // currently "pool" or whatever the global user lock type is.
3803   //
3804   // We are relying on the fact that for all of the user lock types
3805   // (except "tas"), the first field in the lock struct is the "initialized"
3806   // field, which is set to the address of the lock object itself when
3807   // the lock is initialized.  When the union is of type "pool", the
3808   // first field is a pointer to the next object in the free list, which
3809   // will not be the same address as the object itself.
3810   //
3811   // This means that the check (*__kmp_is_user_lock_initialized_)(lck) will fail
3812   // for "pool" objects on the free list.  This must happen as the "location"
3813   // field of real user locks overlaps the "index" field of "pool" objects.
3814   //
3815   // It would be better to run through the free list, and remove all "pool"
3816   // objects from the lock table before executing this loop.  However,
3817   // "pool" objects do not always have their index field set (only on
3818   // lin_32e), and I don't want to search the lock table for the address
3819   // of every "pool" object on the free list.
3820   while (__kmp_user_lock_table.used > 1) {
3821     const ident *loc;
3822 
3823     // reduce __kmp_user_lock_table.used before freeing the lock,
3824     // so that state of locks is consistent
3825     kmp_user_lock_p lck =
3826         __kmp_user_lock_table.table[--__kmp_user_lock_table.used];
3827 
3828     if ((__kmp_is_user_lock_initialized_ != NULL) &&
3829         (*__kmp_is_user_lock_initialized_)(lck)) {
3830       // Issue a warning if: KMP_CONSISTENCY_CHECK AND lock is initialized AND
3831       // it is NOT a critical section (user is not responsible for destroying
3832       // criticals) AND we know source location to report.
3833       if (__kmp_env_consistency_check && (!IS_CRITICAL(lck)) &&
3834           ((loc = __kmp_get_user_lock_location(lck)) != NULL) &&
3835           (loc->psource != NULL)) {
3836         kmp_str_loc_t str_loc = __kmp_str_loc_init(loc->psource, 0);
3837         KMP_WARNING(CnsLockNotDestroyed, str_loc.file, str_loc.line);
3838         __kmp_str_loc_free(&str_loc);
3839       }
3840 
3841 #ifdef KMP_DEBUG
3842       if (IS_CRITICAL(lck)) {
3843         KA_TRACE(
3844             20,
3845             ("__kmp_cleanup_user_locks: free critical section lock %p (%p)\n",
3846              lck, *(void **)lck));
3847       } else {
3848         KA_TRACE(20, ("__kmp_cleanup_user_locks: free lock %p (%p)\n", lck,
3849                       *(void **)lck));
3850       }
3851 #endif // KMP_DEBUG
3852 
3853       // Cleanup internal lock dynamic resources (for drdpa locks particularly).
3854       __kmp_destroy_user_lock(lck);
3855     }
3856 
3857     // Free the lock if block allocation of locks is not used.
3858     if (__kmp_lock_blocks == NULL) {
3859       __kmp_free(lck);
3860     }
3861   }
3862 
3863 #undef IS_CRITICAL
3864 
3865   // delete lock table(s).
3866   kmp_user_lock_p *table_ptr = __kmp_user_lock_table.table;
3867   __kmp_user_lock_table.table = NULL;
3868   __kmp_user_lock_table.allocated = 0;
3869 
3870   while (table_ptr != NULL) {
3871     // In the first element we saved the pointer to the previous
3872     // (smaller) lock table.
3873     kmp_user_lock_p *next = (kmp_user_lock_p *)(table_ptr[0]);
3874     __kmp_free(table_ptr);
3875     table_ptr = next;
3876   }
3877 
3878   // Free buffers allocated for blocks of locks.
3879   kmp_block_of_locks_t *block_ptr = __kmp_lock_blocks;
3880   __kmp_lock_blocks = NULL;
3881 
3882   while (block_ptr != NULL) {
3883     kmp_block_of_locks_t *next = block_ptr->next_block;
3884     __kmp_free(block_ptr->locks);
3885     // *block_ptr itself was allocated at the end of the locks vector.
3886     block_ptr = next;
3887   }
3888 
3889   TCW_4(__kmp_init_user_locks, FALSE);
3890 }
3891 
3892 #endif // KMP_USE_DYNAMIC_LOCK
3893