xref: /freebsd/contrib/llvm-project/openmp/runtime/src/kmp_lock.cpp (revision 79ac3c12a714bcd3f2354c52d948aed9575c46d6)
1 /*
2  * kmp_lock.cpp -- lock-related functions
3  */
4 
5 //===----------------------------------------------------------------------===//
6 //
7 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
8 // See https://llvm.org/LICENSE.txt for license information.
9 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #include <stddef.h>
14 #include <atomic>
15 
16 #include "kmp.h"
17 #include "kmp_i18n.h"
18 #include "kmp_io.h"
19 #include "kmp_itt.h"
20 #include "kmp_lock.h"
21 #include "kmp_wait_release.h"
22 #include "kmp_wrapper_getpid.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   kmp_backoff_t backoff = __kmp_spin_backoff_params;
104   do {
105     __kmp_spin_backoff(&backoff);
106     KMP_YIELD_OVERSUB_ELSE_SPIN(spins);
107   } while (KMP_ATOMIC_LD_RLX(&lck->lk.poll) != tas_free ||
108            !__kmp_atomic_compare_store_acq(&lck->lk.poll, tas_free, tas_busy));
109   KMP_FSYNC_ACQUIRED(lck);
110   return KMP_LOCK_ACQUIRED_FIRST;
111 }
112 
113 int __kmp_acquire_tas_lock(kmp_tas_lock_t *lck, kmp_int32 gtid) {
114   int retval = __kmp_acquire_tas_lock_timed_template(lck, gtid);
115   ANNOTATE_TAS_ACQUIRED(lck);
116   return retval;
117 }
118 
119 static int __kmp_acquire_tas_lock_with_checks(kmp_tas_lock_t *lck,
120                                               kmp_int32 gtid) {
121   char const *const func = "omp_set_lock";
122   if ((sizeof(kmp_tas_lock_t) <= OMP_LOCK_T_SIZE) &&
123       __kmp_is_tas_lock_nestable(lck)) {
124     KMP_FATAL(LockNestableUsedAsSimple, func);
125   }
126   if ((gtid >= 0) && (__kmp_get_tas_lock_owner(lck) == gtid)) {
127     KMP_FATAL(LockIsAlreadyOwned, func);
128   }
129   return __kmp_acquire_tas_lock(lck, gtid);
130 }
131 
132 int __kmp_test_tas_lock(kmp_tas_lock_t *lck, kmp_int32 gtid) {
133   kmp_int32 tas_free = KMP_LOCK_FREE(tas);
134   kmp_int32 tas_busy = KMP_LOCK_BUSY(gtid + 1, tas);
135   if (KMP_ATOMIC_LD_RLX(&lck->lk.poll) == tas_free &&
136       __kmp_atomic_compare_store_acq(&lck->lk.poll, tas_free, tas_busy)) {
137     KMP_FSYNC_ACQUIRED(lck);
138     return TRUE;
139   }
140   return FALSE;
141 }
142 
143 static int __kmp_test_tas_lock_with_checks(kmp_tas_lock_t *lck,
144                                            kmp_int32 gtid) {
145   char const *const func = "omp_test_lock";
146   if ((sizeof(kmp_tas_lock_t) <= OMP_LOCK_T_SIZE) &&
147       __kmp_is_tas_lock_nestable(lck)) {
148     KMP_FATAL(LockNestableUsedAsSimple, func);
149   }
150   return __kmp_test_tas_lock(lck, gtid);
151 }
152 
153 int __kmp_release_tas_lock(kmp_tas_lock_t *lck, kmp_int32 gtid) {
154   KMP_MB(); /* Flush all pending memory write invalidates.  */
155 
156   KMP_FSYNC_RELEASING(lck);
157   ANNOTATE_TAS_RELEASED(lck);
158   KMP_ATOMIC_ST_REL(&lck->lk.poll, KMP_LOCK_FREE(tas));
159   KMP_MB(); /* Flush all pending memory write invalidates.  */
160 
161   KMP_YIELD_OVERSUB();
162   return KMP_LOCK_RELEASED;
163 }
164 
165 static int __kmp_release_tas_lock_with_checks(kmp_tas_lock_t *lck,
166                                               kmp_int32 gtid) {
167   char const *const func = "omp_unset_lock";
168   KMP_MB(); /* in case another processor initialized lock */
169   if ((sizeof(kmp_tas_lock_t) <= OMP_LOCK_T_SIZE) &&
170       __kmp_is_tas_lock_nestable(lck)) {
171     KMP_FATAL(LockNestableUsedAsSimple, func);
172   }
173   if (__kmp_get_tas_lock_owner(lck) == -1) {
174     KMP_FATAL(LockUnsettingFree, func);
175   }
176   if ((gtid >= 0) && (__kmp_get_tas_lock_owner(lck) >= 0) &&
177       (__kmp_get_tas_lock_owner(lck) != gtid)) {
178     KMP_FATAL(LockUnsettingSetByAnother, func);
179   }
180   return __kmp_release_tas_lock(lck, gtid);
181 }
182 
183 void __kmp_init_tas_lock(kmp_tas_lock_t *lck) {
184   lck->lk.poll = KMP_LOCK_FREE(tas);
185 }
186 
187 void __kmp_destroy_tas_lock(kmp_tas_lock_t *lck) { lck->lk.poll = 0; }
188 
189 static void __kmp_destroy_tas_lock_with_checks(kmp_tas_lock_t *lck) {
190   char const *const func = "omp_destroy_lock";
191   if ((sizeof(kmp_tas_lock_t) <= OMP_LOCK_T_SIZE) &&
192       __kmp_is_tas_lock_nestable(lck)) {
193     KMP_FATAL(LockNestableUsedAsSimple, func);
194   }
195   if (__kmp_get_tas_lock_owner(lck) != -1) {
196     KMP_FATAL(LockStillOwned, func);
197   }
198   __kmp_destroy_tas_lock(lck);
199 }
200 
201 // nested test and set locks
202 
203 int __kmp_acquire_nested_tas_lock(kmp_tas_lock_t *lck, kmp_int32 gtid) {
204   KMP_DEBUG_ASSERT(gtid >= 0);
205 
206   if (__kmp_get_tas_lock_owner(lck) == gtid) {
207     lck->lk.depth_locked += 1;
208     return KMP_LOCK_ACQUIRED_NEXT;
209   } else {
210     __kmp_acquire_tas_lock_timed_template(lck, gtid);
211     ANNOTATE_TAS_ACQUIRED(lck);
212     lck->lk.depth_locked = 1;
213     return KMP_LOCK_ACQUIRED_FIRST;
214   }
215 }
216 
217 static int __kmp_acquire_nested_tas_lock_with_checks(kmp_tas_lock_t *lck,
218                                                      kmp_int32 gtid) {
219   char const *const func = "omp_set_nest_lock";
220   if (!__kmp_is_tas_lock_nestable(lck)) {
221     KMP_FATAL(LockSimpleUsedAsNestable, func);
222   }
223   return __kmp_acquire_nested_tas_lock(lck, gtid);
224 }
225 
226 int __kmp_test_nested_tas_lock(kmp_tas_lock_t *lck, kmp_int32 gtid) {
227   int retval;
228 
229   KMP_DEBUG_ASSERT(gtid >= 0);
230 
231   if (__kmp_get_tas_lock_owner(lck) == gtid) {
232     retval = ++lck->lk.depth_locked;
233   } else if (!__kmp_test_tas_lock(lck, gtid)) {
234     retval = 0;
235   } else {
236     KMP_MB();
237     retval = lck->lk.depth_locked = 1;
238   }
239   return retval;
240 }
241 
242 static int __kmp_test_nested_tas_lock_with_checks(kmp_tas_lock_t *lck,
243                                                   kmp_int32 gtid) {
244   char const *const func = "omp_test_nest_lock";
245   if (!__kmp_is_tas_lock_nestable(lck)) {
246     KMP_FATAL(LockSimpleUsedAsNestable, func);
247   }
248   return __kmp_test_nested_tas_lock(lck, gtid);
249 }
250 
251 int __kmp_release_nested_tas_lock(kmp_tas_lock_t *lck, kmp_int32 gtid) {
252   KMP_DEBUG_ASSERT(gtid >= 0);
253 
254   KMP_MB();
255   if (--(lck->lk.depth_locked) == 0) {
256     __kmp_release_tas_lock(lck, gtid);
257     return KMP_LOCK_RELEASED;
258   }
259   return KMP_LOCK_STILL_HELD;
260 }
261 
262 static int __kmp_release_nested_tas_lock_with_checks(kmp_tas_lock_t *lck,
263                                                      kmp_int32 gtid) {
264   char const *const func = "omp_unset_nest_lock";
265   KMP_MB(); /* in case another processor initialized lock */
266   if (!__kmp_is_tas_lock_nestable(lck)) {
267     KMP_FATAL(LockSimpleUsedAsNestable, func);
268   }
269   if (__kmp_get_tas_lock_owner(lck) == -1) {
270     KMP_FATAL(LockUnsettingFree, func);
271   }
272   if (__kmp_get_tas_lock_owner(lck) != gtid) {
273     KMP_FATAL(LockUnsettingSetByAnother, func);
274   }
275   return __kmp_release_nested_tas_lock(lck, gtid);
276 }
277 
278 void __kmp_init_nested_tas_lock(kmp_tas_lock_t *lck) {
279   __kmp_init_tas_lock(lck);
280   lck->lk.depth_locked = 0; // >= 0 for nestable locks, -1 for simple locks
281 }
282 
283 void __kmp_destroy_nested_tas_lock(kmp_tas_lock_t *lck) {
284   __kmp_destroy_tas_lock(lck);
285   lck->lk.depth_locked = 0;
286 }
287 
288 static void __kmp_destroy_nested_tas_lock_with_checks(kmp_tas_lock_t *lck) {
289   char const *const func = "omp_destroy_nest_lock";
290   if (!__kmp_is_tas_lock_nestable(lck)) {
291     KMP_FATAL(LockSimpleUsedAsNestable, func);
292   }
293   if (__kmp_get_tas_lock_owner(lck) != -1) {
294     KMP_FATAL(LockStillOwned, func);
295   }
296   __kmp_destroy_nested_tas_lock(lck);
297 }
298 
299 #if KMP_USE_FUTEX
300 
301 /* ------------------------------------------------------------------------ */
302 /* futex locks */
303 
304 // futex locks are really just test and set locks, with a different method
305 // of handling contention.  They take the same amount of space as test and
306 // set locks, and are allocated the same way (i.e. use the area allocated by
307 // the compiler for non-nested locks / allocate nested locks on the heap).
308 
309 static kmp_int32 __kmp_get_futex_lock_owner(kmp_futex_lock_t *lck) {
310   return KMP_LOCK_STRIP((TCR_4(lck->lk.poll) >> 1)) - 1;
311 }
312 
313 static inline bool __kmp_is_futex_lock_nestable(kmp_futex_lock_t *lck) {
314   return lck->lk.depth_locked != -1;
315 }
316 
317 __forceinline static int
318 __kmp_acquire_futex_lock_timed_template(kmp_futex_lock_t *lck, kmp_int32 gtid) {
319   kmp_int32 gtid_code = (gtid + 1) << 1;
320 
321   KMP_MB();
322 
323 #ifdef USE_LOCK_PROFILE
324   kmp_uint32 curr = KMP_LOCK_STRIP(TCR_4(lck->lk.poll));
325   if ((curr != 0) && (curr != gtid_code))
326     __kmp_printf("LOCK CONTENTION: %p\n", lck);
327 /* else __kmp_printf( "." );*/
328 #endif /* USE_LOCK_PROFILE */
329 
330   KMP_FSYNC_PREPARE(lck);
331   KA_TRACE(1000, ("__kmp_acquire_futex_lock: lck:%p(0x%x), T#%d entering\n",
332                   lck, lck->lk.poll, gtid));
333 
334   kmp_int32 poll_val;
335 
336   while ((poll_val = KMP_COMPARE_AND_STORE_RET32(
337               &(lck->lk.poll), KMP_LOCK_FREE(futex),
338               KMP_LOCK_BUSY(gtid_code, futex))) != KMP_LOCK_FREE(futex)) {
339 
340     kmp_int32 cond = KMP_LOCK_STRIP(poll_val) & 1;
341     KA_TRACE(
342         1000,
343         ("__kmp_acquire_futex_lock: lck:%p, T#%d poll_val = 0x%x cond = 0x%x\n",
344          lck, gtid, poll_val, cond));
345 
346     // NOTE: if you try to use the following condition for this branch
347     //
348     // if ( poll_val & 1 == 0 )
349     //
350     // Then the 12.0 compiler has a bug where the following block will
351     // always be skipped, regardless of the value of the LSB of poll_val.
352     if (!cond) {
353       // Try to set the lsb in the poll to indicate to the owner
354       // thread that they need to wake this thread up.
355       if (!KMP_COMPARE_AND_STORE_REL32(&(lck->lk.poll), poll_val,
356                                        poll_val | KMP_LOCK_BUSY(1, futex))) {
357         KA_TRACE(
358             1000,
359             ("__kmp_acquire_futex_lock: lck:%p(0x%x), T#%d can't set bit 0\n",
360              lck, lck->lk.poll, gtid));
361         continue;
362       }
363       poll_val |= KMP_LOCK_BUSY(1, futex);
364 
365       KA_TRACE(1000,
366                ("__kmp_acquire_futex_lock: lck:%p(0x%x), T#%d bit 0 set\n", lck,
367                 lck->lk.poll, gtid));
368     }
369 
370     KA_TRACE(
371         1000,
372         ("__kmp_acquire_futex_lock: lck:%p, T#%d before futex_wait(0x%x)\n",
373          lck, gtid, poll_val));
374 
375     long rc;
376     if ((rc = syscall(__NR_futex, &(lck->lk.poll), FUTEX_WAIT, poll_val, NULL,
377                       NULL, 0)) != 0) {
378       KA_TRACE(1000, ("__kmp_acquire_futex_lock: lck:%p, T#%d futex_wait(0x%x) "
379                       "failed (rc=%ld errno=%d)\n",
380                       lck, gtid, poll_val, rc, errno));
381       continue;
382     }
383 
384     KA_TRACE(1000,
385              ("__kmp_acquire_futex_lock: lck:%p, T#%d after futex_wait(0x%x)\n",
386               lck, gtid, poll_val));
387     // This thread has now done a successful futex wait call and was entered on
388     // the OS futex queue.  We must now perform a futex wake call when releasing
389     // the lock, as we have no idea how many other threads are in the queue.
390     gtid_code |= 1;
391   }
392 
393   KMP_FSYNC_ACQUIRED(lck);
394   KA_TRACE(1000, ("__kmp_acquire_futex_lock: lck:%p(0x%x), T#%d exiting\n", lck,
395                   lck->lk.poll, gtid));
396   return KMP_LOCK_ACQUIRED_FIRST;
397 }
398 
399 int __kmp_acquire_futex_lock(kmp_futex_lock_t *lck, kmp_int32 gtid) {
400   int retval = __kmp_acquire_futex_lock_timed_template(lck, gtid);
401   ANNOTATE_FUTEX_ACQUIRED(lck);
402   return retval;
403 }
404 
405 static int __kmp_acquire_futex_lock_with_checks(kmp_futex_lock_t *lck,
406                                                 kmp_int32 gtid) {
407   char const *const func = "omp_set_lock";
408   if ((sizeof(kmp_futex_lock_t) <= OMP_LOCK_T_SIZE) &&
409       __kmp_is_futex_lock_nestable(lck)) {
410     KMP_FATAL(LockNestableUsedAsSimple, func);
411   }
412   if ((gtid >= 0) && (__kmp_get_futex_lock_owner(lck) == gtid)) {
413     KMP_FATAL(LockIsAlreadyOwned, func);
414   }
415   return __kmp_acquire_futex_lock(lck, gtid);
416 }
417 
418 int __kmp_test_futex_lock(kmp_futex_lock_t *lck, kmp_int32 gtid) {
419   if (KMP_COMPARE_AND_STORE_ACQ32(&(lck->lk.poll), KMP_LOCK_FREE(futex),
420                                   KMP_LOCK_BUSY((gtid + 1) << 1, futex))) {
421     KMP_FSYNC_ACQUIRED(lck);
422     return TRUE;
423   }
424   return FALSE;
425 }
426 
427 static int __kmp_test_futex_lock_with_checks(kmp_futex_lock_t *lck,
428                                              kmp_int32 gtid) {
429   char const *const func = "omp_test_lock";
430   if ((sizeof(kmp_futex_lock_t) <= OMP_LOCK_T_SIZE) &&
431       __kmp_is_futex_lock_nestable(lck)) {
432     KMP_FATAL(LockNestableUsedAsSimple, func);
433   }
434   return __kmp_test_futex_lock(lck, gtid);
435 }
436 
437 int __kmp_release_futex_lock(kmp_futex_lock_t *lck, kmp_int32 gtid) {
438   KMP_MB(); /* Flush all pending memory write invalidates.  */
439 
440   KA_TRACE(1000, ("__kmp_release_futex_lock: lck:%p(0x%x), T#%d entering\n",
441                   lck, lck->lk.poll, gtid));
442 
443   KMP_FSYNC_RELEASING(lck);
444   ANNOTATE_FUTEX_RELEASED(lck);
445 
446   kmp_int32 poll_val = KMP_XCHG_FIXED32(&(lck->lk.poll), KMP_LOCK_FREE(futex));
447 
448   KA_TRACE(1000,
449            ("__kmp_release_futex_lock: lck:%p, T#%d released poll_val = 0x%x\n",
450             lck, gtid, poll_val));
451 
452   if (KMP_LOCK_STRIP(poll_val) & 1) {
453     KA_TRACE(1000,
454              ("__kmp_release_futex_lock: lck:%p, T#%d futex_wake 1 thread\n",
455               lck, gtid));
456     syscall(__NR_futex, &(lck->lk.poll), FUTEX_WAKE, KMP_LOCK_BUSY(1, futex),
457             NULL, NULL, 0);
458   }
459 
460   KMP_MB(); /* Flush all pending memory write invalidates.  */
461 
462   KA_TRACE(1000, ("__kmp_release_futex_lock: lck:%p(0x%x), T#%d exiting\n", lck,
463                   lck->lk.poll, gtid));
464 
465   KMP_YIELD_OVERSUB();
466   return KMP_LOCK_RELEASED;
467 }
468 
469 static int __kmp_release_futex_lock_with_checks(kmp_futex_lock_t *lck,
470                                                 kmp_int32 gtid) {
471   char const *const func = "omp_unset_lock";
472   KMP_MB(); /* in case another processor initialized lock */
473   if ((sizeof(kmp_futex_lock_t) <= OMP_LOCK_T_SIZE) &&
474       __kmp_is_futex_lock_nestable(lck)) {
475     KMP_FATAL(LockNestableUsedAsSimple, func);
476   }
477   if (__kmp_get_futex_lock_owner(lck) == -1) {
478     KMP_FATAL(LockUnsettingFree, func);
479   }
480   if ((gtid >= 0) && (__kmp_get_futex_lock_owner(lck) >= 0) &&
481       (__kmp_get_futex_lock_owner(lck) != gtid)) {
482     KMP_FATAL(LockUnsettingSetByAnother, func);
483   }
484   return __kmp_release_futex_lock(lck, gtid);
485 }
486 
487 void __kmp_init_futex_lock(kmp_futex_lock_t *lck) {
488   TCW_4(lck->lk.poll, KMP_LOCK_FREE(futex));
489 }
490 
491 void __kmp_destroy_futex_lock(kmp_futex_lock_t *lck) { lck->lk.poll = 0; }
492 
493 static void __kmp_destroy_futex_lock_with_checks(kmp_futex_lock_t *lck) {
494   char const *const func = "omp_destroy_lock";
495   if ((sizeof(kmp_futex_lock_t) <= OMP_LOCK_T_SIZE) &&
496       __kmp_is_futex_lock_nestable(lck)) {
497     KMP_FATAL(LockNestableUsedAsSimple, func);
498   }
499   if (__kmp_get_futex_lock_owner(lck) != -1) {
500     KMP_FATAL(LockStillOwned, func);
501   }
502   __kmp_destroy_futex_lock(lck);
503 }
504 
505 // nested futex locks
506 
507 int __kmp_acquire_nested_futex_lock(kmp_futex_lock_t *lck, kmp_int32 gtid) {
508   KMP_DEBUG_ASSERT(gtid >= 0);
509 
510   if (__kmp_get_futex_lock_owner(lck) == gtid) {
511     lck->lk.depth_locked += 1;
512     return KMP_LOCK_ACQUIRED_NEXT;
513   } else {
514     __kmp_acquire_futex_lock_timed_template(lck, gtid);
515     ANNOTATE_FUTEX_ACQUIRED(lck);
516     lck->lk.depth_locked = 1;
517     return KMP_LOCK_ACQUIRED_FIRST;
518   }
519 }
520 
521 static int __kmp_acquire_nested_futex_lock_with_checks(kmp_futex_lock_t *lck,
522                                                        kmp_int32 gtid) {
523   char const *const func = "omp_set_nest_lock";
524   if (!__kmp_is_futex_lock_nestable(lck)) {
525     KMP_FATAL(LockSimpleUsedAsNestable, func);
526   }
527   return __kmp_acquire_nested_futex_lock(lck, gtid);
528 }
529 
530 int __kmp_test_nested_futex_lock(kmp_futex_lock_t *lck, kmp_int32 gtid) {
531   int retval;
532 
533   KMP_DEBUG_ASSERT(gtid >= 0);
534 
535   if (__kmp_get_futex_lock_owner(lck) == gtid) {
536     retval = ++lck->lk.depth_locked;
537   } else if (!__kmp_test_futex_lock(lck, gtid)) {
538     retval = 0;
539   } else {
540     KMP_MB();
541     retval = lck->lk.depth_locked = 1;
542   }
543   return retval;
544 }
545 
546 static int __kmp_test_nested_futex_lock_with_checks(kmp_futex_lock_t *lck,
547                                                     kmp_int32 gtid) {
548   char const *const func = "omp_test_nest_lock";
549   if (!__kmp_is_futex_lock_nestable(lck)) {
550     KMP_FATAL(LockSimpleUsedAsNestable, func);
551   }
552   return __kmp_test_nested_futex_lock(lck, gtid);
553 }
554 
555 int __kmp_release_nested_futex_lock(kmp_futex_lock_t *lck, kmp_int32 gtid) {
556   KMP_DEBUG_ASSERT(gtid >= 0);
557 
558   KMP_MB();
559   if (--(lck->lk.depth_locked) == 0) {
560     __kmp_release_futex_lock(lck, gtid);
561     return KMP_LOCK_RELEASED;
562   }
563   return KMP_LOCK_STILL_HELD;
564 }
565 
566 static int __kmp_release_nested_futex_lock_with_checks(kmp_futex_lock_t *lck,
567                                                        kmp_int32 gtid) {
568   char const *const func = "omp_unset_nest_lock";
569   KMP_MB(); /* in case another processor initialized lock */
570   if (!__kmp_is_futex_lock_nestable(lck)) {
571     KMP_FATAL(LockSimpleUsedAsNestable, func);
572   }
573   if (__kmp_get_futex_lock_owner(lck) == -1) {
574     KMP_FATAL(LockUnsettingFree, func);
575   }
576   if (__kmp_get_futex_lock_owner(lck) != gtid) {
577     KMP_FATAL(LockUnsettingSetByAnother, func);
578   }
579   return __kmp_release_nested_futex_lock(lck, gtid);
580 }
581 
582 void __kmp_init_nested_futex_lock(kmp_futex_lock_t *lck) {
583   __kmp_init_futex_lock(lck);
584   lck->lk.depth_locked = 0; // >= 0 for nestable locks, -1 for simple locks
585 }
586 
587 void __kmp_destroy_nested_futex_lock(kmp_futex_lock_t *lck) {
588   __kmp_destroy_futex_lock(lck);
589   lck->lk.depth_locked = 0;
590 }
591 
592 static void __kmp_destroy_nested_futex_lock_with_checks(kmp_futex_lock_t *lck) {
593   char const *const func = "omp_destroy_nest_lock";
594   if (!__kmp_is_futex_lock_nestable(lck)) {
595     KMP_FATAL(LockSimpleUsedAsNestable, func);
596   }
597   if (__kmp_get_futex_lock_owner(lck) != -1) {
598     KMP_FATAL(LockStillOwned, func);
599   }
600   __kmp_destroy_nested_futex_lock(lck);
601 }
602 
603 #endif // KMP_USE_FUTEX
604 
605 /* ------------------------------------------------------------------------ */
606 /* ticket (bakery) locks */
607 
608 static kmp_int32 __kmp_get_ticket_lock_owner(kmp_ticket_lock_t *lck) {
609   return std::atomic_load_explicit(&lck->lk.owner_id,
610                                    std::memory_order_relaxed) -
611          1;
612 }
613 
614 static inline bool __kmp_is_ticket_lock_nestable(kmp_ticket_lock_t *lck) {
615   return std::atomic_load_explicit(&lck->lk.depth_locked,
616                                    std::memory_order_relaxed) != -1;
617 }
618 
619 static kmp_uint32 __kmp_bakery_check(void *now_serving, kmp_uint32 my_ticket) {
620   return std::atomic_load_explicit((std::atomic<unsigned> *)now_serving,
621                                    std::memory_order_acquire) == my_ticket;
622 }
623 
624 __forceinline static int
625 __kmp_acquire_ticket_lock_timed_template(kmp_ticket_lock_t *lck,
626                                          kmp_int32 gtid) {
627   kmp_uint32 my_ticket = std::atomic_fetch_add_explicit(
628       &lck->lk.next_ticket, 1U, std::memory_order_relaxed);
629 
630 #ifdef USE_LOCK_PROFILE
631   if (std::atomic_load_explicit(&lck->lk.now_serving,
632                                 std::memory_order_relaxed) != my_ticket)
633     __kmp_printf("LOCK CONTENTION: %p\n", lck);
634 /* else __kmp_printf( "." );*/
635 #endif /* USE_LOCK_PROFILE */
636 
637   if (std::atomic_load_explicit(&lck->lk.now_serving,
638                                 std::memory_order_acquire) == my_ticket) {
639     return KMP_LOCK_ACQUIRED_FIRST;
640   }
641   KMP_WAIT_PTR(&lck->lk.now_serving, my_ticket, __kmp_bakery_check, lck);
642   return KMP_LOCK_ACQUIRED_FIRST;
643 }
644 
645 int __kmp_acquire_ticket_lock(kmp_ticket_lock_t *lck, kmp_int32 gtid) {
646   int retval = __kmp_acquire_ticket_lock_timed_template(lck, gtid);
647   ANNOTATE_TICKET_ACQUIRED(lck);
648   return retval;
649 }
650 
651 static int __kmp_acquire_ticket_lock_with_checks(kmp_ticket_lock_t *lck,
652                                                  kmp_int32 gtid) {
653   char const *const func = "omp_set_lock";
654 
655   if (!std::atomic_load_explicit(&lck->lk.initialized,
656                                  std::memory_order_relaxed)) {
657     KMP_FATAL(LockIsUninitialized, func);
658   }
659   if (lck->lk.self != lck) {
660     KMP_FATAL(LockIsUninitialized, func);
661   }
662   if (__kmp_is_ticket_lock_nestable(lck)) {
663     KMP_FATAL(LockNestableUsedAsSimple, func);
664   }
665   if ((gtid >= 0) && (__kmp_get_ticket_lock_owner(lck) == gtid)) {
666     KMP_FATAL(LockIsAlreadyOwned, func);
667   }
668 
669   __kmp_acquire_ticket_lock(lck, gtid);
670 
671   std::atomic_store_explicit(&lck->lk.owner_id, gtid + 1,
672                              std::memory_order_relaxed);
673   return KMP_LOCK_ACQUIRED_FIRST;
674 }
675 
676 int __kmp_test_ticket_lock(kmp_ticket_lock_t *lck, kmp_int32 gtid) {
677   kmp_uint32 my_ticket = std::atomic_load_explicit(&lck->lk.next_ticket,
678                                                    std::memory_order_relaxed);
679 
680   if (std::atomic_load_explicit(&lck->lk.now_serving,
681                                 std::memory_order_relaxed) == my_ticket) {
682     kmp_uint32 next_ticket = my_ticket + 1;
683     if (std::atomic_compare_exchange_strong_explicit(
684             &lck->lk.next_ticket, &my_ticket, next_ticket,
685             std::memory_order_acquire, std::memory_order_acquire)) {
686       return TRUE;
687     }
688   }
689   return FALSE;
690 }
691 
692 static int __kmp_test_ticket_lock_with_checks(kmp_ticket_lock_t *lck,
693                                               kmp_int32 gtid) {
694   char const *const func = "omp_test_lock";
695 
696   if (!std::atomic_load_explicit(&lck->lk.initialized,
697                                  std::memory_order_relaxed)) {
698     KMP_FATAL(LockIsUninitialized, func);
699   }
700   if (lck->lk.self != lck) {
701     KMP_FATAL(LockIsUninitialized, func);
702   }
703   if (__kmp_is_ticket_lock_nestable(lck)) {
704     KMP_FATAL(LockNestableUsedAsSimple, func);
705   }
706 
707   int retval = __kmp_test_ticket_lock(lck, gtid);
708 
709   if (retval) {
710     std::atomic_store_explicit(&lck->lk.owner_id, gtid + 1,
711                                std::memory_order_relaxed);
712   }
713   return retval;
714 }
715 
716 int __kmp_release_ticket_lock(kmp_ticket_lock_t *lck, kmp_int32 gtid) {
717   kmp_uint32 distance = std::atomic_load_explicit(&lck->lk.next_ticket,
718                                                   std::memory_order_relaxed) -
719                         std::atomic_load_explicit(&lck->lk.now_serving,
720                                                   std::memory_order_relaxed);
721 
722   ANNOTATE_TICKET_RELEASED(lck);
723   std::atomic_fetch_add_explicit(&lck->lk.now_serving, 1U,
724                                  std::memory_order_release);
725 
726   KMP_YIELD(distance >
727             (kmp_uint32)(__kmp_avail_proc ? __kmp_avail_proc : __kmp_xproc));
728   return KMP_LOCK_RELEASED;
729 }
730 
731 static int __kmp_release_ticket_lock_with_checks(kmp_ticket_lock_t *lck,
732                                                  kmp_int32 gtid) {
733   char const *const func = "omp_unset_lock";
734 
735   if (!std::atomic_load_explicit(&lck->lk.initialized,
736                                  std::memory_order_relaxed)) {
737     KMP_FATAL(LockIsUninitialized, func);
738   }
739   if (lck->lk.self != lck) {
740     KMP_FATAL(LockIsUninitialized, func);
741   }
742   if (__kmp_is_ticket_lock_nestable(lck)) {
743     KMP_FATAL(LockNestableUsedAsSimple, func);
744   }
745   if (__kmp_get_ticket_lock_owner(lck) == -1) {
746     KMP_FATAL(LockUnsettingFree, func);
747   }
748   if ((gtid >= 0) && (__kmp_get_ticket_lock_owner(lck) >= 0) &&
749       (__kmp_get_ticket_lock_owner(lck) != gtid)) {
750     KMP_FATAL(LockUnsettingSetByAnother, func);
751   }
752   std::atomic_store_explicit(&lck->lk.owner_id, 0, std::memory_order_relaxed);
753   return __kmp_release_ticket_lock(lck, gtid);
754 }
755 
756 void __kmp_init_ticket_lock(kmp_ticket_lock_t *lck) {
757   lck->lk.location = NULL;
758   lck->lk.self = lck;
759   std::atomic_store_explicit(&lck->lk.next_ticket, 0U,
760                              std::memory_order_relaxed);
761   std::atomic_store_explicit(&lck->lk.now_serving, 0U,
762                              std::memory_order_relaxed);
763   std::atomic_store_explicit(
764       &lck->lk.owner_id, 0,
765       std::memory_order_relaxed); // no thread owns the lock.
766   std::atomic_store_explicit(
767       &lck->lk.depth_locked, -1,
768       std::memory_order_relaxed); // -1 => not a nested lock.
769   std::atomic_store_explicit(&lck->lk.initialized, true,
770                              std::memory_order_release);
771 }
772 
773 void __kmp_destroy_ticket_lock(kmp_ticket_lock_t *lck) {
774   std::atomic_store_explicit(&lck->lk.initialized, false,
775                              std::memory_order_release);
776   lck->lk.self = NULL;
777   lck->lk.location = NULL;
778   std::atomic_store_explicit(&lck->lk.next_ticket, 0U,
779                              std::memory_order_relaxed);
780   std::atomic_store_explicit(&lck->lk.now_serving, 0U,
781                              std::memory_order_relaxed);
782   std::atomic_store_explicit(&lck->lk.owner_id, 0, std::memory_order_relaxed);
783   std::atomic_store_explicit(&lck->lk.depth_locked, -1,
784                              std::memory_order_relaxed);
785 }
786 
787 static void __kmp_destroy_ticket_lock_with_checks(kmp_ticket_lock_t *lck) {
788   char const *const func = "omp_destroy_lock";
789 
790   if (!std::atomic_load_explicit(&lck->lk.initialized,
791                                  std::memory_order_relaxed)) {
792     KMP_FATAL(LockIsUninitialized, func);
793   }
794   if (lck->lk.self != lck) {
795     KMP_FATAL(LockIsUninitialized, func);
796   }
797   if (__kmp_is_ticket_lock_nestable(lck)) {
798     KMP_FATAL(LockNestableUsedAsSimple, func);
799   }
800   if (__kmp_get_ticket_lock_owner(lck) != -1) {
801     KMP_FATAL(LockStillOwned, func);
802   }
803   __kmp_destroy_ticket_lock(lck);
804 }
805 
806 // nested ticket locks
807 
808 int __kmp_acquire_nested_ticket_lock(kmp_ticket_lock_t *lck, kmp_int32 gtid) {
809   KMP_DEBUG_ASSERT(gtid >= 0);
810 
811   if (__kmp_get_ticket_lock_owner(lck) == gtid) {
812     std::atomic_fetch_add_explicit(&lck->lk.depth_locked, 1,
813                                    std::memory_order_relaxed);
814     return KMP_LOCK_ACQUIRED_NEXT;
815   } else {
816     __kmp_acquire_ticket_lock_timed_template(lck, gtid);
817     ANNOTATE_TICKET_ACQUIRED(lck);
818     std::atomic_store_explicit(&lck->lk.depth_locked, 1,
819                                std::memory_order_relaxed);
820     std::atomic_store_explicit(&lck->lk.owner_id, gtid + 1,
821                                std::memory_order_relaxed);
822     return KMP_LOCK_ACQUIRED_FIRST;
823   }
824 }
825 
826 static int __kmp_acquire_nested_ticket_lock_with_checks(kmp_ticket_lock_t *lck,
827                                                         kmp_int32 gtid) {
828   char const *const func = "omp_set_nest_lock";
829 
830   if (!std::atomic_load_explicit(&lck->lk.initialized,
831                                  std::memory_order_relaxed)) {
832     KMP_FATAL(LockIsUninitialized, func);
833   }
834   if (lck->lk.self != lck) {
835     KMP_FATAL(LockIsUninitialized, func);
836   }
837   if (!__kmp_is_ticket_lock_nestable(lck)) {
838     KMP_FATAL(LockSimpleUsedAsNestable, func);
839   }
840   return __kmp_acquire_nested_ticket_lock(lck, gtid);
841 }
842 
843 int __kmp_test_nested_ticket_lock(kmp_ticket_lock_t *lck, kmp_int32 gtid) {
844   int retval;
845 
846   KMP_DEBUG_ASSERT(gtid >= 0);
847 
848   if (__kmp_get_ticket_lock_owner(lck) == gtid) {
849     retval = std::atomic_fetch_add_explicit(&lck->lk.depth_locked, 1,
850                                             std::memory_order_relaxed) +
851              1;
852   } else if (!__kmp_test_ticket_lock(lck, gtid)) {
853     retval = 0;
854   } else {
855     std::atomic_store_explicit(&lck->lk.depth_locked, 1,
856                                std::memory_order_relaxed);
857     std::atomic_store_explicit(&lck->lk.owner_id, gtid + 1,
858                                std::memory_order_relaxed);
859     retval = 1;
860   }
861   return retval;
862 }
863 
864 static int __kmp_test_nested_ticket_lock_with_checks(kmp_ticket_lock_t *lck,
865                                                      kmp_int32 gtid) {
866   char const *const func = "omp_test_nest_lock";
867 
868   if (!std::atomic_load_explicit(&lck->lk.initialized,
869                                  std::memory_order_relaxed)) {
870     KMP_FATAL(LockIsUninitialized, func);
871   }
872   if (lck->lk.self != lck) {
873     KMP_FATAL(LockIsUninitialized, func);
874   }
875   if (!__kmp_is_ticket_lock_nestable(lck)) {
876     KMP_FATAL(LockSimpleUsedAsNestable, func);
877   }
878   return __kmp_test_nested_ticket_lock(lck, gtid);
879 }
880 
881 int __kmp_release_nested_ticket_lock(kmp_ticket_lock_t *lck, kmp_int32 gtid) {
882   KMP_DEBUG_ASSERT(gtid >= 0);
883 
884   if ((std::atomic_fetch_add_explicit(&lck->lk.depth_locked, -1,
885                                       std::memory_order_relaxed) -
886        1) == 0) {
887     std::atomic_store_explicit(&lck->lk.owner_id, 0, std::memory_order_relaxed);
888     __kmp_release_ticket_lock(lck, gtid);
889     return KMP_LOCK_RELEASED;
890   }
891   return KMP_LOCK_STILL_HELD;
892 }
893 
894 static int __kmp_release_nested_ticket_lock_with_checks(kmp_ticket_lock_t *lck,
895                                                         kmp_int32 gtid) {
896   char const *const func = "omp_unset_nest_lock";
897 
898   if (!std::atomic_load_explicit(&lck->lk.initialized,
899                                  std::memory_order_relaxed)) {
900     KMP_FATAL(LockIsUninitialized, func);
901   }
902   if (lck->lk.self != lck) {
903     KMP_FATAL(LockIsUninitialized, func);
904   }
905   if (!__kmp_is_ticket_lock_nestable(lck)) {
906     KMP_FATAL(LockSimpleUsedAsNestable, func);
907   }
908   if (__kmp_get_ticket_lock_owner(lck) == -1) {
909     KMP_FATAL(LockUnsettingFree, func);
910   }
911   if (__kmp_get_ticket_lock_owner(lck) != gtid) {
912     KMP_FATAL(LockUnsettingSetByAnother, func);
913   }
914   return __kmp_release_nested_ticket_lock(lck, gtid);
915 }
916 
917 void __kmp_init_nested_ticket_lock(kmp_ticket_lock_t *lck) {
918   __kmp_init_ticket_lock(lck);
919   std::atomic_store_explicit(&lck->lk.depth_locked, 0,
920                              std::memory_order_relaxed);
921   // >= 0 for nestable locks, -1 for simple locks
922 }
923 
924 void __kmp_destroy_nested_ticket_lock(kmp_ticket_lock_t *lck) {
925   __kmp_destroy_ticket_lock(lck);
926   std::atomic_store_explicit(&lck->lk.depth_locked, 0,
927                              std::memory_order_relaxed);
928 }
929 
930 static void
931 __kmp_destroy_nested_ticket_lock_with_checks(kmp_ticket_lock_t *lck) {
932   char const *const func = "omp_destroy_nest_lock";
933 
934   if (!std::atomic_load_explicit(&lck->lk.initialized,
935                                  std::memory_order_relaxed)) {
936     KMP_FATAL(LockIsUninitialized, func);
937   }
938   if (lck->lk.self != lck) {
939     KMP_FATAL(LockIsUninitialized, func);
940   }
941   if (!__kmp_is_ticket_lock_nestable(lck)) {
942     KMP_FATAL(LockSimpleUsedAsNestable, func);
943   }
944   if (__kmp_get_ticket_lock_owner(lck) != -1) {
945     KMP_FATAL(LockStillOwned, func);
946   }
947   __kmp_destroy_nested_ticket_lock(lck);
948 }
949 
950 // access functions to fields which don't exist for all lock kinds.
951 
952 static const ident_t *__kmp_get_ticket_lock_location(kmp_ticket_lock_t *lck) {
953   return lck->lk.location;
954 }
955 
956 static void __kmp_set_ticket_lock_location(kmp_ticket_lock_t *lck,
957                                            const ident_t *loc) {
958   lck->lk.location = loc;
959 }
960 
961 static kmp_lock_flags_t __kmp_get_ticket_lock_flags(kmp_ticket_lock_t *lck) {
962   return lck->lk.flags;
963 }
964 
965 static void __kmp_set_ticket_lock_flags(kmp_ticket_lock_t *lck,
966                                         kmp_lock_flags_t flags) {
967   lck->lk.flags = flags;
968 }
969 
970 /* ------------------------------------------------------------------------ */
971 /* queuing locks */
972 
973 /* First the states
974    (head,tail) =              0, 0  means lock is unheld, nobody on queue
975                  UINT_MAX or -1, 0  means lock is held, nobody on queue
976                               h, h  means lock held or about to transition,
977                                     1 element on queue
978                               h, t  h <> t, means lock is held or about to
979                                     transition, >1 elements on queue
980 
981    Now the transitions
982       Acquire(0,0)  = -1 ,0
983       Release(0,0)  = Error
984       Acquire(-1,0) =  h ,h    h > 0
985       Release(-1,0) =  0 ,0
986       Acquire(h,h)  =  h ,t    h > 0, t > 0, h <> t
987       Release(h,h)  = -1 ,0    h > 0
988       Acquire(h,t)  =  h ,t'   h > 0, t > 0, t' > 0, h <> t, h <> t', t <> t'
989       Release(h,t)  =  h',t    h > 0, t > 0, h <> t, h <> h', h' maybe = t
990 
991    And pictorially
992 
993            +-----+
994            | 0, 0|------- release -------> Error
995            +-----+
996              |  ^
997       acquire|  |release
998              |  |
999              |  |
1000              v  |
1001            +-----+
1002            |-1, 0|
1003            +-----+
1004              |  ^
1005       acquire|  |release
1006              |  |
1007              |  |
1008              v  |
1009            +-----+
1010            | h, h|
1011            +-----+
1012              |  ^
1013       acquire|  |release
1014              |  |
1015              |  |
1016              v  |
1017            +-----+
1018            | h, t|----- acquire, release loopback ---+
1019            +-----+                                   |
1020                 ^                                    |
1021                 |                                    |
1022                 +------------------------------------+
1023  */
1024 
1025 #ifdef DEBUG_QUEUING_LOCKS
1026 
1027 /* Stuff for circular trace buffer */
1028 #define TRACE_BUF_ELE 1024
1029 static char traces[TRACE_BUF_ELE][128] = {0};
1030 static int tc = 0;
1031 #define TRACE_LOCK(X, Y)                                                       \
1032   KMP_SNPRINTF(traces[tc++ % TRACE_BUF_ELE], 128, "t%d at %s\n", X, Y);
1033 #define TRACE_LOCK_T(X, Y, Z)                                                  \
1034   KMP_SNPRINTF(traces[tc++ % TRACE_BUF_ELE], 128, "t%d at %s%d\n", X, Y, Z);
1035 #define TRACE_LOCK_HT(X, Y, Z, Q)                                              \
1036   KMP_SNPRINTF(traces[tc++ % TRACE_BUF_ELE], 128, "t%d at %s %d,%d\n", X, Y,   \
1037                Z, Q);
1038 
1039 static void __kmp_dump_queuing_lock(kmp_info_t *this_thr, kmp_int32 gtid,
1040                                     kmp_queuing_lock_t *lck, kmp_int32 head_id,
1041                                     kmp_int32 tail_id) {
1042   kmp_int32 t, i;
1043 
1044   __kmp_printf_no_lock("\n__kmp_dump_queuing_lock: TRACE BEGINS HERE! \n");
1045 
1046   i = tc % TRACE_BUF_ELE;
1047   __kmp_printf_no_lock("%s\n", traces[i]);
1048   i = (i + 1) % TRACE_BUF_ELE;
1049   while (i != (tc % TRACE_BUF_ELE)) {
1050     __kmp_printf_no_lock("%s", traces[i]);
1051     i = (i + 1) % TRACE_BUF_ELE;
1052   }
1053   __kmp_printf_no_lock("\n");
1054 
1055   __kmp_printf_no_lock("\n__kmp_dump_queuing_lock: gtid+1:%d, spin_here:%d, "
1056                        "next_wait:%d, head_id:%d, tail_id:%d\n",
1057                        gtid + 1, this_thr->th.th_spin_here,
1058                        this_thr->th.th_next_waiting, head_id, tail_id);
1059 
1060   __kmp_printf_no_lock("\t\thead: %d ", lck->lk.head_id);
1061 
1062   if (lck->lk.head_id >= 1) {
1063     t = __kmp_threads[lck->lk.head_id - 1]->th.th_next_waiting;
1064     while (t > 0) {
1065       __kmp_printf_no_lock("-> %d ", t);
1066       t = __kmp_threads[t - 1]->th.th_next_waiting;
1067     }
1068   }
1069   __kmp_printf_no_lock(";  tail: %d ", lck->lk.tail_id);
1070   __kmp_printf_no_lock("\n\n");
1071 }
1072 
1073 #endif /* DEBUG_QUEUING_LOCKS */
1074 
1075 static kmp_int32 __kmp_get_queuing_lock_owner(kmp_queuing_lock_t *lck) {
1076   return TCR_4(lck->lk.owner_id) - 1;
1077 }
1078 
1079 static inline bool __kmp_is_queuing_lock_nestable(kmp_queuing_lock_t *lck) {
1080   return lck->lk.depth_locked != -1;
1081 }
1082 
1083 /* Acquire a lock using a the queuing lock implementation */
1084 template <bool takeTime>
1085 /* [TLW] The unused template above is left behind because of what BEB believes
1086    is a potential compiler problem with __forceinline. */
1087 __forceinline static int
1088 __kmp_acquire_queuing_lock_timed_template(kmp_queuing_lock_t *lck,
1089                                           kmp_int32 gtid) {
1090   kmp_info_t *this_thr = __kmp_thread_from_gtid(gtid);
1091   volatile kmp_int32 *head_id_p = &lck->lk.head_id;
1092   volatile kmp_int32 *tail_id_p = &lck->lk.tail_id;
1093   volatile kmp_uint32 *spin_here_p;
1094   kmp_int32 need_mf = 1;
1095 
1096 #if OMPT_SUPPORT
1097   ompt_state_t prev_state = ompt_state_undefined;
1098 #endif
1099 
1100   KA_TRACE(1000,
1101            ("__kmp_acquire_queuing_lock: lck:%p, T#%d entering\n", lck, gtid));
1102 
1103   KMP_FSYNC_PREPARE(lck);
1104   KMP_DEBUG_ASSERT(this_thr != NULL);
1105   spin_here_p = &this_thr->th.th_spin_here;
1106 
1107 #ifdef DEBUG_QUEUING_LOCKS
1108   TRACE_LOCK(gtid + 1, "acq ent");
1109   if (*spin_here_p)
1110     __kmp_dump_queuing_lock(this_thr, gtid, lck, *head_id_p, *tail_id_p);
1111   if (this_thr->th.th_next_waiting != 0)
1112     __kmp_dump_queuing_lock(this_thr, gtid, lck, *head_id_p, *tail_id_p);
1113 #endif
1114   KMP_DEBUG_ASSERT(!*spin_here_p);
1115   KMP_DEBUG_ASSERT(this_thr->th.th_next_waiting == 0);
1116 
1117   /* The following st.rel to spin_here_p needs to precede the cmpxchg.acq to
1118      head_id_p that may follow, not just in execution order, but also in
1119      visibility order. This way, when a releasing thread observes the changes to
1120      the queue by this thread, it can rightly assume that spin_here_p has
1121      already been set to TRUE, so that when it sets spin_here_p to FALSE, it is
1122      not premature.  If the releasing thread sets spin_here_p to FALSE before
1123      this thread sets it to TRUE, this thread will hang. */
1124   *spin_here_p = TRUE; /* before enqueuing to prevent race */
1125 
1126   while (1) {
1127     kmp_int32 enqueued;
1128     kmp_int32 head;
1129     kmp_int32 tail;
1130 
1131     head = *head_id_p;
1132 
1133     switch (head) {
1134 
1135     case -1: {
1136 #ifdef DEBUG_QUEUING_LOCKS
1137       tail = *tail_id_p;
1138       TRACE_LOCK_HT(gtid + 1, "acq read: ", head, tail);
1139 #endif
1140       tail = 0; /* to make sure next link asynchronously read is not set
1141                 accidentally; this assignment prevents us from entering the
1142                 if ( t > 0 ) condition in the enqueued case below, which is not
1143                 necessary for this state transition */
1144 
1145       need_mf = 0;
1146       /* try (-1,0)->(tid,tid) */
1147       enqueued = KMP_COMPARE_AND_STORE_ACQ64((volatile kmp_int64 *)tail_id_p,
1148                                              KMP_PACK_64(-1, 0),
1149                                              KMP_PACK_64(gtid + 1, gtid + 1));
1150 #ifdef DEBUG_QUEUING_LOCKS
1151       if (enqueued)
1152         TRACE_LOCK(gtid + 1, "acq enq: (-1,0)->(tid,tid)");
1153 #endif
1154     } break;
1155 
1156     default: {
1157       tail = *tail_id_p;
1158       KMP_DEBUG_ASSERT(tail != gtid + 1);
1159 
1160 #ifdef DEBUG_QUEUING_LOCKS
1161       TRACE_LOCK_HT(gtid + 1, "acq read: ", head, tail);
1162 #endif
1163 
1164       if (tail == 0) {
1165         enqueued = FALSE;
1166       } else {
1167         need_mf = 0;
1168         /* try (h,t) or (h,h)->(h,tid) */
1169         enqueued = KMP_COMPARE_AND_STORE_ACQ32(tail_id_p, tail, gtid + 1);
1170 
1171 #ifdef DEBUG_QUEUING_LOCKS
1172         if (enqueued)
1173           TRACE_LOCK(gtid + 1, "acq enq: (h,t)->(h,tid)");
1174 #endif
1175       }
1176     } break;
1177 
1178     case 0: /* empty queue */
1179     {
1180       kmp_int32 grabbed_lock;
1181 
1182 #ifdef DEBUG_QUEUING_LOCKS
1183       tail = *tail_id_p;
1184       TRACE_LOCK_HT(gtid + 1, "acq read: ", head, tail);
1185 #endif
1186       /* try (0,0)->(-1,0) */
1187 
1188       /* only legal transition out of head = 0 is head = -1 with no change to
1189        * tail */
1190       grabbed_lock = KMP_COMPARE_AND_STORE_ACQ32(head_id_p, 0, -1);
1191 
1192       if (grabbed_lock) {
1193 
1194         *spin_here_p = FALSE;
1195 
1196         KA_TRACE(
1197             1000,
1198             ("__kmp_acquire_queuing_lock: lck:%p, T#%d exiting: no queuing\n",
1199              lck, gtid));
1200 #ifdef DEBUG_QUEUING_LOCKS
1201         TRACE_LOCK_HT(gtid + 1, "acq exit: ", head, 0);
1202 #endif
1203 
1204 #if OMPT_SUPPORT
1205         if (ompt_enabled.enabled && prev_state != ompt_state_undefined) {
1206           /* change the state before clearing wait_id */
1207           this_thr->th.ompt_thread_info.state = prev_state;
1208           this_thr->th.ompt_thread_info.wait_id = 0;
1209         }
1210 #endif
1211 
1212         KMP_FSYNC_ACQUIRED(lck);
1213         return KMP_LOCK_ACQUIRED_FIRST; /* lock holder cannot be on queue */
1214       }
1215       enqueued = FALSE;
1216     } break;
1217     }
1218 
1219 #if OMPT_SUPPORT
1220     if (ompt_enabled.enabled && prev_state == ompt_state_undefined) {
1221       /* this thread will spin; set wait_id before entering wait state */
1222       prev_state = this_thr->th.ompt_thread_info.state;
1223       this_thr->th.ompt_thread_info.wait_id = (uint64_t)lck;
1224       this_thr->th.ompt_thread_info.state = ompt_state_wait_lock;
1225     }
1226 #endif
1227 
1228     if (enqueued) {
1229       if (tail > 0) {
1230         kmp_info_t *tail_thr = __kmp_thread_from_gtid(tail - 1);
1231         KMP_ASSERT(tail_thr != NULL);
1232         tail_thr->th.th_next_waiting = gtid + 1;
1233         /* corresponding wait for this write in release code */
1234       }
1235       KA_TRACE(1000,
1236                ("__kmp_acquire_queuing_lock: lck:%p, T#%d waiting for lock\n",
1237                 lck, gtid));
1238 
1239       KMP_MB();
1240       // ToDo: Use __kmp_wait_sleep or similar when blocktime != inf
1241       KMP_WAIT(spin_here_p, FALSE, KMP_EQ, lck);
1242       // Synchronize writes to both runtime thread structures
1243       // and writes in user code.
1244       KMP_MB();
1245 
1246 #ifdef DEBUG_QUEUING_LOCKS
1247       TRACE_LOCK(gtid + 1, "acq spin");
1248 
1249       if (this_thr->th.th_next_waiting != 0)
1250         __kmp_dump_queuing_lock(this_thr, gtid, lck, *head_id_p, *tail_id_p);
1251 #endif
1252       KMP_DEBUG_ASSERT(this_thr->th.th_next_waiting == 0);
1253       KA_TRACE(1000, ("__kmp_acquire_queuing_lock: lck:%p, T#%d exiting: after "
1254                       "waiting on queue\n",
1255                       lck, gtid));
1256 
1257 #ifdef DEBUG_QUEUING_LOCKS
1258       TRACE_LOCK(gtid + 1, "acq exit 2");
1259 #endif
1260 
1261 #if OMPT_SUPPORT
1262       /* change the state before clearing wait_id */
1263       this_thr->th.ompt_thread_info.state = prev_state;
1264       this_thr->th.ompt_thread_info.wait_id = 0;
1265 #endif
1266 
1267       /* got lock, we were dequeued by the thread that released lock */
1268       return KMP_LOCK_ACQUIRED_FIRST;
1269     }
1270 
1271     /* Yield if number of threads > number of logical processors */
1272     /* ToDo: Not sure why this should only be in oversubscription case,
1273        maybe should be traditional YIELD_INIT/YIELD_WHEN loop */
1274     KMP_YIELD_OVERSUB();
1275 
1276 #ifdef DEBUG_QUEUING_LOCKS
1277     TRACE_LOCK(gtid + 1, "acq retry");
1278 #endif
1279   }
1280   KMP_ASSERT2(0, "should not get here");
1281   return KMP_LOCK_ACQUIRED_FIRST;
1282 }
1283 
1284 int __kmp_acquire_queuing_lock(kmp_queuing_lock_t *lck, kmp_int32 gtid) {
1285   KMP_DEBUG_ASSERT(gtid >= 0);
1286 
1287   int retval = __kmp_acquire_queuing_lock_timed_template<false>(lck, gtid);
1288   ANNOTATE_QUEUING_ACQUIRED(lck);
1289   return retval;
1290 }
1291 
1292 static int __kmp_acquire_queuing_lock_with_checks(kmp_queuing_lock_t *lck,
1293                                                   kmp_int32 gtid) {
1294   char const *const func = "omp_set_lock";
1295   if (lck->lk.initialized != lck) {
1296     KMP_FATAL(LockIsUninitialized, func);
1297   }
1298   if (__kmp_is_queuing_lock_nestable(lck)) {
1299     KMP_FATAL(LockNestableUsedAsSimple, func);
1300   }
1301   if (__kmp_get_queuing_lock_owner(lck) == gtid) {
1302     KMP_FATAL(LockIsAlreadyOwned, func);
1303   }
1304 
1305   __kmp_acquire_queuing_lock(lck, gtid);
1306 
1307   lck->lk.owner_id = gtid + 1;
1308   return KMP_LOCK_ACQUIRED_FIRST;
1309 }
1310 
1311 int __kmp_test_queuing_lock(kmp_queuing_lock_t *lck, kmp_int32 gtid) {
1312   volatile kmp_int32 *head_id_p = &lck->lk.head_id;
1313   kmp_int32 head;
1314 #ifdef KMP_DEBUG
1315   kmp_info_t *this_thr;
1316 #endif
1317 
1318   KA_TRACE(1000, ("__kmp_test_queuing_lock: T#%d entering\n", gtid));
1319   KMP_DEBUG_ASSERT(gtid >= 0);
1320 #ifdef KMP_DEBUG
1321   this_thr = __kmp_thread_from_gtid(gtid);
1322   KMP_DEBUG_ASSERT(this_thr != NULL);
1323   KMP_DEBUG_ASSERT(!this_thr->th.th_spin_here);
1324 #endif
1325 
1326   head = *head_id_p;
1327 
1328   if (head == 0) { /* nobody on queue, nobody holding */
1329     /* try (0,0)->(-1,0) */
1330     if (KMP_COMPARE_AND_STORE_ACQ32(head_id_p, 0, -1)) {
1331       KA_TRACE(1000,
1332                ("__kmp_test_queuing_lock: T#%d exiting: holding lock\n", gtid));
1333       KMP_FSYNC_ACQUIRED(lck);
1334       ANNOTATE_QUEUING_ACQUIRED(lck);
1335       return TRUE;
1336     }
1337   }
1338 
1339   KA_TRACE(1000,
1340            ("__kmp_test_queuing_lock: T#%d exiting: without lock\n", gtid));
1341   return FALSE;
1342 }
1343 
1344 static int __kmp_test_queuing_lock_with_checks(kmp_queuing_lock_t *lck,
1345                                                kmp_int32 gtid) {
1346   char const *const func = "omp_test_lock";
1347   if (lck->lk.initialized != lck) {
1348     KMP_FATAL(LockIsUninitialized, func);
1349   }
1350   if (__kmp_is_queuing_lock_nestable(lck)) {
1351     KMP_FATAL(LockNestableUsedAsSimple, func);
1352   }
1353 
1354   int retval = __kmp_test_queuing_lock(lck, gtid);
1355 
1356   if (retval) {
1357     lck->lk.owner_id = gtid + 1;
1358   }
1359   return retval;
1360 }
1361 
1362 int __kmp_release_queuing_lock(kmp_queuing_lock_t *lck, kmp_int32 gtid) {
1363   kmp_info_t *this_thr;
1364   volatile kmp_int32 *head_id_p = &lck->lk.head_id;
1365   volatile kmp_int32 *tail_id_p = &lck->lk.tail_id;
1366 
1367   KA_TRACE(1000,
1368            ("__kmp_release_queuing_lock: lck:%p, T#%d entering\n", lck, gtid));
1369   KMP_DEBUG_ASSERT(gtid >= 0);
1370   this_thr = __kmp_thread_from_gtid(gtid);
1371   KMP_DEBUG_ASSERT(this_thr != NULL);
1372 #ifdef DEBUG_QUEUING_LOCKS
1373   TRACE_LOCK(gtid + 1, "rel ent");
1374 
1375   if (this_thr->th.th_spin_here)
1376     __kmp_dump_queuing_lock(this_thr, gtid, lck, *head_id_p, *tail_id_p);
1377   if (this_thr->th.th_next_waiting != 0)
1378     __kmp_dump_queuing_lock(this_thr, gtid, lck, *head_id_p, *tail_id_p);
1379 #endif
1380   KMP_DEBUG_ASSERT(!this_thr->th.th_spin_here);
1381   KMP_DEBUG_ASSERT(this_thr->th.th_next_waiting == 0);
1382 
1383   KMP_FSYNC_RELEASING(lck);
1384   ANNOTATE_QUEUING_RELEASED(lck);
1385 
1386   while (1) {
1387     kmp_int32 dequeued;
1388     kmp_int32 head;
1389     kmp_int32 tail;
1390 
1391     head = *head_id_p;
1392 
1393 #ifdef DEBUG_QUEUING_LOCKS
1394     tail = *tail_id_p;
1395     TRACE_LOCK_HT(gtid + 1, "rel read: ", head, tail);
1396     if (head == 0)
1397       __kmp_dump_queuing_lock(this_thr, gtid, lck, head, tail);
1398 #endif
1399     KMP_DEBUG_ASSERT(head !=
1400                      0); /* holding the lock, head must be -1 or queue head */
1401 
1402     if (head == -1) { /* nobody on queue */
1403       /* try (-1,0)->(0,0) */
1404       if (KMP_COMPARE_AND_STORE_REL32(head_id_p, -1, 0)) {
1405         KA_TRACE(
1406             1000,
1407             ("__kmp_release_queuing_lock: lck:%p, T#%d exiting: queue empty\n",
1408              lck, gtid));
1409 #ifdef DEBUG_QUEUING_LOCKS
1410         TRACE_LOCK_HT(gtid + 1, "rel exit: ", 0, 0);
1411 #endif
1412 
1413 #if OMPT_SUPPORT
1414 /* nothing to do - no other thread is trying to shift blame */
1415 #endif
1416         return KMP_LOCK_RELEASED;
1417       }
1418       dequeued = FALSE;
1419     } else {
1420       KMP_MB();
1421       tail = *tail_id_p;
1422       if (head == tail) { /* only one thread on the queue */
1423 #ifdef DEBUG_QUEUING_LOCKS
1424         if (head <= 0)
1425           __kmp_dump_queuing_lock(this_thr, gtid, lck, head, tail);
1426 #endif
1427         KMP_DEBUG_ASSERT(head > 0);
1428 
1429         /* try (h,h)->(-1,0) */
1430         dequeued = KMP_COMPARE_AND_STORE_REL64(
1431             RCAST(volatile kmp_int64 *, tail_id_p), KMP_PACK_64(head, head),
1432             KMP_PACK_64(-1, 0));
1433 #ifdef DEBUG_QUEUING_LOCKS
1434         TRACE_LOCK(gtid + 1, "rel deq: (h,h)->(-1,0)");
1435 #endif
1436 
1437       } else {
1438         volatile kmp_int32 *waiting_id_p;
1439         kmp_info_t *head_thr = __kmp_thread_from_gtid(head - 1);
1440         KMP_DEBUG_ASSERT(head_thr != NULL);
1441         waiting_id_p = &head_thr->th.th_next_waiting;
1442 
1443 /* Does this require synchronous reads? */
1444 #ifdef DEBUG_QUEUING_LOCKS
1445         if (head <= 0 || tail <= 0)
1446           __kmp_dump_queuing_lock(this_thr, gtid, lck, head, tail);
1447 #endif
1448         KMP_DEBUG_ASSERT(head > 0 && tail > 0);
1449 
1450         /* try (h,t)->(h',t) or (t,t) */
1451         KMP_MB();
1452         /* make sure enqueuing thread has time to update next waiting thread
1453          * field */
1454         *head_id_p =
1455             KMP_WAIT((volatile kmp_uint32 *)waiting_id_p, 0, KMP_NEQ, NULL);
1456 #ifdef DEBUG_QUEUING_LOCKS
1457         TRACE_LOCK(gtid + 1, "rel deq: (h,t)->(h',t)");
1458 #endif
1459         dequeued = TRUE;
1460       }
1461     }
1462 
1463     if (dequeued) {
1464       kmp_info_t *head_thr = __kmp_thread_from_gtid(head - 1);
1465       KMP_DEBUG_ASSERT(head_thr != NULL);
1466 
1467 /* Does this require synchronous reads? */
1468 #ifdef DEBUG_QUEUING_LOCKS
1469       if (head <= 0 || tail <= 0)
1470         __kmp_dump_queuing_lock(this_thr, gtid, lck, head, tail);
1471 #endif
1472       KMP_DEBUG_ASSERT(head > 0 && tail > 0);
1473 
1474       /* For clean code only. Thread not released until next statement prevents
1475          race with acquire code. */
1476       head_thr->th.th_next_waiting = 0;
1477 #ifdef DEBUG_QUEUING_LOCKS
1478       TRACE_LOCK_T(gtid + 1, "rel nw=0 for t=", head);
1479 #endif
1480 
1481       KMP_MB();
1482       /* reset spin value */
1483       head_thr->th.th_spin_here = FALSE;
1484 
1485       KA_TRACE(1000, ("__kmp_release_queuing_lock: lck:%p, T#%d exiting: after "
1486                       "dequeuing\n",
1487                       lck, gtid));
1488 #ifdef DEBUG_QUEUING_LOCKS
1489       TRACE_LOCK(gtid + 1, "rel exit 2");
1490 #endif
1491       return KMP_LOCK_RELEASED;
1492     }
1493 /* KMP_CPU_PAUSE(); don't want to make releasing thread hold up acquiring
1494    threads */
1495 
1496 #ifdef DEBUG_QUEUING_LOCKS
1497     TRACE_LOCK(gtid + 1, "rel retry");
1498 #endif
1499 
1500   } /* while */
1501   KMP_ASSERT2(0, "should not get here");
1502   return KMP_LOCK_RELEASED;
1503 }
1504 
1505 static int __kmp_release_queuing_lock_with_checks(kmp_queuing_lock_t *lck,
1506                                                   kmp_int32 gtid) {
1507   char const *const func = "omp_unset_lock";
1508   KMP_MB(); /* in case another processor initialized lock */
1509   if (lck->lk.initialized != lck) {
1510     KMP_FATAL(LockIsUninitialized, func);
1511   }
1512   if (__kmp_is_queuing_lock_nestable(lck)) {
1513     KMP_FATAL(LockNestableUsedAsSimple, func);
1514   }
1515   if (__kmp_get_queuing_lock_owner(lck) == -1) {
1516     KMP_FATAL(LockUnsettingFree, func);
1517   }
1518   if (__kmp_get_queuing_lock_owner(lck) != gtid) {
1519     KMP_FATAL(LockUnsettingSetByAnother, func);
1520   }
1521   lck->lk.owner_id = 0;
1522   return __kmp_release_queuing_lock(lck, gtid);
1523 }
1524 
1525 void __kmp_init_queuing_lock(kmp_queuing_lock_t *lck) {
1526   lck->lk.location = NULL;
1527   lck->lk.head_id = 0;
1528   lck->lk.tail_id = 0;
1529   lck->lk.next_ticket = 0;
1530   lck->lk.now_serving = 0;
1531   lck->lk.owner_id = 0; // no thread owns the lock.
1532   lck->lk.depth_locked = -1; // >= 0 for nestable locks, -1 for simple locks.
1533   lck->lk.initialized = lck;
1534 
1535   KA_TRACE(1000, ("__kmp_init_queuing_lock: lock %p initialized\n", lck));
1536 }
1537 
1538 void __kmp_destroy_queuing_lock(kmp_queuing_lock_t *lck) {
1539   lck->lk.initialized = NULL;
1540   lck->lk.location = NULL;
1541   lck->lk.head_id = 0;
1542   lck->lk.tail_id = 0;
1543   lck->lk.next_ticket = 0;
1544   lck->lk.now_serving = 0;
1545   lck->lk.owner_id = 0;
1546   lck->lk.depth_locked = -1;
1547 }
1548 
1549 static void __kmp_destroy_queuing_lock_with_checks(kmp_queuing_lock_t *lck) {
1550   char const *const func = "omp_destroy_lock";
1551   if (lck->lk.initialized != lck) {
1552     KMP_FATAL(LockIsUninitialized, func);
1553   }
1554   if (__kmp_is_queuing_lock_nestable(lck)) {
1555     KMP_FATAL(LockNestableUsedAsSimple, func);
1556   }
1557   if (__kmp_get_queuing_lock_owner(lck) != -1) {
1558     KMP_FATAL(LockStillOwned, func);
1559   }
1560   __kmp_destroy_queuing_lock(lck);
1561 }
1562 
1563 // nested queuing locks
1564 
1565 int __kmp_acquire_nested_queuing_lock(kmp_queuing_lock_t *lck, kmp_int32 gtid) {
1566   KMP_DEBUG_ASSERT(gtid >= 0);
1567 
1568   if (__kmp_get_queuing_lock_owner(lck) == gtid) {
1569     lck->lk.depth_locked += 1;
1570     return KMP_LOCK_ACQUIRED_NEXT;
1571   } else {
1572     __kmp_acquire_queuing_lock_timed_template<false>(lck, gtid);
1573     ANNOTATE_QUEUING_ACQUIRED(lck);
1574     KMP_MB();
1575     lck->lk.depth_locked = 1;
1576     KMP_MB();
1577     lck->lk.owner_id = gtid + 1;
1578     return KMP_LOCK_ACQUIRED_FIRST;
1579   }
1580 }
1581 
1582 static int
1583 __kmp_acquire_nested_queuing_lock_with_checks(kmp_queuing_lock_t *lck,
1584                                               kmp_int32 gtid) {
1585   char const *const func = "omp_set_nest_lock";
1586   if (lck->lk.initialized != lck) {
1587     KMP_FATAL(LockIsUninitialized, func);
1588   }
1589   if (!__kmp_is_queuing_lock_nestable(lck)) {
1590     KMP_FATAL(LockSimpleUsedAsNestable, func);
1591   }
1592   return __kmp_acquire_nested_queuing_lock(lck, gtid);
1593 }
1594 
1595 int __kmp_test_nested_queuing_lock(kmp_queuing_lock_t *lck, kmp_int32 gtid) {
1596   int retval;
1597 
1598   KMP_DEBUG_ASSERT(gtid >= 0);
1599 
1600   if (__kmp_get_queuing_lock_owner(lck) == gtid) {
1601     retval = ++lck->lk.depth_locked;
1602   } else if (!__kmp_test_queuing_lock(lck, gtid)) {
1603     retval = 0;
1604   } else {
1605     KMP_MB();
1606     retval = lck->lk.depth_locked = 1;
1607     KMP_MB();
1608     lck->lk.owner_id = gtid + 1;
1609   }
1610   return retval;
1611 }
1612 
1613 static int __kmp_test_nested_queuing_lock_with_checks(kmp_queuing_lock_t *lck,
1614                                                       kmp_int32 gtid) {
1615   char const *const func = "omp_test_nest_lock";
1616   if (lck->lk.initialized != lck) {
1617     KMP_FATAL(LockIsUninitialized, func);
1618   }
1619   if (!__kmp_is_queuing_lock_nestable(lck)) {
1620     KMP_FATAL(LockSimpleUsedAsNestable, func);
1621   }
1622   return __kmp_test_nested_queuing_lock(lck, gtid);
1623 }
1624 
1625 int __kmp_release_nested_queuing_lock(kmp_queuing_lock_t *lck, kmp_int32 gtid) {
1626   KMP_DEBUG_ASSERT(gtid >= 0);
1627 
1628   KMP_MB();
1629   if (--(lck->lk.depth_locked) == 0) {
1630     KMP_MB();
1631     lck->lk.owner_id = 0;
1632     __kmp_release_queuing_lock(lck, gtid);
1633     return KMP_LOCK_RELEASED;
1634   }
1635   return KMP_LOCK_STILL_HELD;
1636 }
1637 
1638 static int
1639 __kmp_release_nested_queuing_lock_with_checks(kmp_queuing_lock_t *lck,
1640                                               kmp_int32 gtid) {
1641   char const *const func = "omp_unset_nest_lock";
1642   KMP_MB(); /* in case another processor initialized lock */
1643   if (lck->lk.initialized != lck) {
1644     KMP_FATAL(LockIsUninitialized, func);
1645   }
1646   if (!__kmp_is_queuing_lock_nestable(lck)) {
1647     KMP_FATAL(LockSimpleUsedAsNestable, func);
1648   }
1649   if (__kmp_get_queuing_lock_owner(lck) == -1) {
1650     KMP_FATAL(LockUnsettingFree, func);
1651   }
1652   if (__kmp_get_queuing_lock_owner(lck) != gtid) {
1653     KMP_FATAL(LockUnsettingSetByAnother, func);
1654   }
1655   return __kmp_release_nested_queuing_lock(lck, gtid);
1656 }
1657 
1658 void __kmp_init_nested_queuing_lock(kmp_queuing_lock_t *lck) {
1659   __kmp_init_queuing_lock(lck);
1660   lck->lk.depth_locked = 0; // >= 0 for nestable locks, -1 for simple locks
1661 }
1662 
1663 void __kmp_destroy_nested_queuing_lock(kmp_queuing_lock_t *lck) {
1664   __kmp_destroy_queuing_lock(lck);
1665   lck->lk.depth_locked = 0;
1666 }
1667 
1668 static void
1669 __kmp_destroy_nested_queuing_lock_with_checks(kmp_queuing_lock_t *lck) {
1670   char const *const func = "omp_destroy_nest_lock";
1671   if (lck->lk.initialized != lck) {
1672     KMP_FATAL(LockIsUninitialized, func);
1673   }
1674   if (!__kmp_is_queuing_lock_nestable(lck)) {
1675     KMP_FATAL(LockSimpleUsedAsNestable, func);
1676   }
1677   if (__kmp_get_queuing_lock_owner(lck) != -1) {
1678     KMP_FATAL(LockStillOwned, func);
1679   }
1680   __kmp_destroy_nested_queuing_lock(lck);
1681 }
1682 
1683 // access functions to fields which don't exist for all lock kinds.
1684 
1685 static const ident_t *__kmp_get_queuing_lock_location(kmp_queuing_lock_t *lck) {
1686   return lck->lk.location;
1687 }
1688 
1689 static void __kmp_set_queuing_lock_location(kmp_queuing_lock_t *lck,
1690                                             const ident_t *loc) {
1691   lck->lk.location = loc;
1692 }
1693 
1694 static kmp_lock_flags_t __kmp_get_queuing_lock_flags(kmp_queuing_lock_t *lck) {
1695   return lck->lk.flags;
1696 }
1697 
1698 static void __kmp_set_queuing_lock_flags(kmp_queuing_lock_t *lck,
1699                                          kmp_lock_flags_t flags) {
1700   lck->lk.flags = flags;
1701 }
1702 
1703 #if KMP_USE_ADAPTIVE_LOCKS
1704 
1705 /* RTM Adaptive locks */
1706 
1707 #if KMP_HAVE_RTM_INTRINSICS
1708 #include <immintrin.h>
1709 #define SOFT_ABORT_MASK (_XABORT_RETRY | _XABORT_CONFLICT | _XABORT_EXPLICIT)
1710 
1711 #else
1712 
1713 // Values from the status register after failed speculation.
1714 #define _XBEGIN_STARTED (~0u)
1715 #define _XABORT_EXPLICIT (1 << 0)
1716 #define _XABORT_RETRY (1 << 1)
1717 #define _XABORT_CONFLICT (1 << 2)
1718 #define _XABORT_CAPACITY (1 << 3)
1719 #define _XABORT_DEBUG (1 << 4)
1720 #define _XABORT_NESTED (1 << 5)
1721 #define _XABORT_CODE(x) ((unsigned char)(((x) >> 24) & 0xFF))
1722 
1723 // Aborts for which it's worth trying again immediately
1724 #define SOFT_ABORT_MASK (_XABORT_RETRY | _XABORT_CONFLICT | _XABORT_EXPLICIT)
1725 
1726 #define STRINGIZE_INTERNAL(arg) #arg
1727 #define STRINGIZE(arg) STRINGIZE_INTERNAL(arg)
1728 
1729 // Access to RTM instructions
1730 /*A version of XBegin which returns -1 on speculation, and the value of EAX on
1731   an abort. This is the same definition as the compiler intrinsic that will be
1732   supported at some point. */
1733 static __inline int _xbegin() {
1734   int res = -1;
1735 
1736 #if KMP_OS_WINDOWS
1737 #if KMP_ARCH_X86_64
1738   _asm {
1739         _emit 0xC7
1740         _emit 0xF8
1741         _emit 2
1742         _emit 0
1743         _emit 0
1744         _emit 0
1745         jmp   L2
1746         mov   res, eax
1747     L2:
1748   }
1749 #else /* IA32 */
1750   _asm {
1751         _emit 0xC7
1752         _emit 0xF8
1753         _emit 2
1754         _emit 0
1755         _emit 0
1756         _emit 0
1757         jmp   L2
1758         mov   res, eax
1759     L2:
1760   }
1761 #endif // KMP_ARCH_X86_64
1762 #else
1763   /* Note that %eax must be noted as killed (clobbered), because the XSR is
1764      returned in %eax(%rax) on abort.  Other register values are restored, so
1765      don't need to be killed.
1766 
1767      We must also mark 'res' as an input and an output, since otherwise
1768      'res=-1' may be dropped as being dead, whereas we do need the assignment on
1769      the successful (i.e., non-abort) path. */
1770   __asm__ volatile("1: .byte  0xC7; .byte 0xF8;\n"
1771                    "   .long  1f-1b-6\n"
1772                    "    jmp   2f\n"
1773                    "1:  movl  %%eax,%0\n"
1774                    "2:"
1775                    : "+r"(res)::"memory", "%eax");
1776 #endif // KMP_OS_WINDOWS
1777   return res;
1778 }
1779 
1780 /* Transaction end */
1781 static __inline void _xend() {
1782 #if KMP_OS_WINDOWS
1783   __asm {
1784         _emit 0x0f
1785         _emit 0x01
1786         _emit 0xd5
1787   }
1788 #else
1789   __asm__ volatile(".byte 0x0f; .byte 0x01; .byte 0xd5" ::: "memory");
1790 #endif
1791 }
1792 
1793 /* This is a macro, the argument must be a single byte constant which can be
1794    evaluated by the inline assembler, since it is emitted as a byte into the
1795    assembly code. */
1796 // clang-format off
1797 #if KMP_OS_WINDOWS
1798 #define _xabort(ARG) _asm _emit 0xc6 _asm _emit 0xf8 _asm _emit ARG
1799 #else
1800 #define _xabort(ARG)                                                           \
1801   __asm__ volatile(".byte 0xC6; .byte 0xF8; .byte " STRINGIZE(ARG):::"memory");
1802 #endif
1803 // clang-format on
1804 #endif // KMP_COMPILER_ICC && __INTEL_COMPILER >= 1300
1805 
1806 // Statistics is collected for testing purpose
1807 #if KMP_DEBUG_ADAPTIVE_LOCKS
1808 
1809 // We accumulate speculative lock statistics when the lock is destroyed. We
1810 // keep locks that haven't been destroyed in the liveLocks list so that we can
1811 // grab their statistics too.
1812 static kmp_adaptive_lock_statistics_t destroyedStats;
1813 
1814 // To hold the list of live locks.
1815 static kmp_adaptive_lock_info_t liveLocks;
1816 
1817 // A lock so we can safely update the list of locks.
1818 static kmp_bootstrap_lock_t chain_lock =
1819     KMP_BOOTSTRAP_LOCK_INITIALIZER(chain_lock);
1820 
1821 // Initialize the list of stats.
1822 void __kmp_init_speculative_stats() {
1823   kmp_adaptive_lock_info_t *lck = &liveLocks;
1824 
1825   memset(CCAST(kmp_adaptive_lock_statistics_t *, &(lck->stats)), 0,
1826          sizeof(lck->stats));
1827   lck->stats.next = lck;
1828   lck->stats.prev = lck;
1829 
1830   KMP_ASSERT(lck->stats.next->stats.prev == lck);
1831   KMP_ASSERT(lck->stats.prev->stats.next == lck);
1832 
1833   __kmp_init_bootstrap_lock(&chain_lock);
1834 }
1835 
1836 // Insert the lock into the circular list
1837 static void __kmp_remember_lock(kmp_adaptive_lock_info_t *lck) {
1838   __kmp_acquire_bootstrap_lock(&chain_lock);
1839 
1840   lck->stats.next = liveLocks.stats.next;
1841   lck->stats.prev = &liveLocks;
1842 
1843   liveLocks.stats.next = lck;
1844   lck->stats.next->stats.prev = lck;
1845 
1846   KMP_ASSERT(lck->stats.next->stats.prev == lck);
1847   KMP_ASSERT(lck->stats.prev->stats.next == lck);
1848 
1849   __kmp_release_bootstrap_lock(&chain_lock);
1850 }
1851 
1852 static void __kmp_forget_lock(kmp_adaptive_lock_info_t *lck) {
1853   KMP_ASSERT(lck->stats.next->stats.prev == lck);
1854   KMP_ASSERT(lck->stats.prev->stats.next == lck);
1855 
1856   kmp_adaptive_lock_info_t *n = lck->stats.next;
1857   kmp_adaptive_lock_info_t *p = lck->stats.prev;
1858 
1859   n->stats.prev = p;
1860   p->stats.next = n;
1861 }
1862 
1863 static void __kmp_zero_speculative_stats(kmp_adaptive_lock_info_t *lck) {
1864   memset(CCAST(kmp_adaptive_lock_statistics_t *, &lck->stats), 0,
1865          sizeof(lck->stats));
1866   __kmp_remember_lock(lck);
1867 }
1868 
1869 static void __kmp_add_stats(kmp_adaptive_lock_statistics_t *t,
1870                             kmp_adaptive_lock_info_t *lck) {
1871   kmp_adaptive_lock_statistics_t volatile *s = &lck->stats;
1872 
1873   t->nonSpeculativeAcquireAttempts += lck->acquire_attempts;
1874   t->successfulSpeculations += s->successfulSpeculations;
1875   t->hardFailedSpeculations += s->hardFailedSpeculations;
1876   t->softFailedSpeculations += s->softFailedSpeculations;
1877   t->nonSpeculativeAcquires += s->nonSpeculativeAcquires;
1878   t->lemmingYields += s->lemmingYields;
1879 }
1880 
1881 static void __kmp_accumulate_speculative_stats(kmp_adaptive_lock_info_t *lck) {
1882   __kmp_acquire_bootstrap_lock(&chain_lock);
1883 
1884   __kmp_add_stats(&destroyedStats, lck);
1885   __kmp_forget_lock(lck);
1886 
1887   __kmp_release_bootstrap_lock(&chain_lock);
1888 }
1889 
1890 static float percent(kmp_uint32 count, kmp_uint32 total) {
1891   return (total == 0) ? 0.0 : (100.0 * count) / total;
1892 }
1893 
1894 void __kmp_print_speculative_stats() {
1895   kmp_adaptive_lock_statistics_t total = destroyedStats;
1896   kmp_adaptive_lock_info_t *lck;
1897 
1898   for (lck = liveLocks.stats.next; lck != &liveLocks; lck = lck->stats.next) {
1899     __kmp_add_stats(&total, lck);
1900   }
1901   kmp_adaptive_lock_statistics_t *t = &total;
1902   kmp_uint32 totalSections =
1903       t->nonSpeculativeAcquires + t->successfulSpeculations;
1904   kmp_uint32 totalSpeculations = t->successfulSpeculations +
1905                                  t->hardFailedSpeculations +
1906                                  t->softFailedSpeculations;
1907   if (totalSections <= 0)
1908     return;
1909 
1910   kmp_safe_raii_file_t statsFile;
1911   if (strcmp(__kmp_speculative_statsfile, "-") == 0) {
1912     statsFile.set_stdout();
1913   } else {
1914     size_t buffLen = KMP_STRLEN(__kmp_speculative_statsfile) + 20;
1915     char buffer[buffLen];
1916     KMP_SNPRINTF(&buffer[0], buffLen, __kmp_speculative_statsfile,
1917                  (kmp_int32)getpid());
1918     statsFile.open(buffer, "w");
1919   }
1920 
1921   fprintf(statsFile, "Speculative lock statistics (all approximate!)\n");
1922   fprintf(statsFile, " Lock parameters: \n"
1923                      "   max_soft_retries               : %10d\n"
1924                      "   max_badness                    : %10d\n",
1925           __kmp_adaptive_backoff_params.max_soft_retries,
1926           __kmp_adaptive_backoff_params.max_badness);
1927   fprintf(statsFile, " Non-speculative acquire attempts : %10d\n",
1928           t->nonSpeculativeAcquireAttempts);
1929   fprintf(statsFile, " Total critical sections          : %10d\n",
1930           totalSections);
1931   fprintf(statsFile, " Successful speculations          : %10d (%5.1f%%)\n",
1932           t->successfulSpeculations,
1933           percent(t->successfulSpeculations, totalSections));
1934   fprintf(statsFile, " Non-speculative acquires         : %10d (%5.1f%%)\n",
1935           t->nonSpeculativeAcquires,
1936           percent(t->nonSpeculativeAcquires, totalSections));
1937   fprintf(statsFile, " Lemming yields                   : %10d\n\n",
1938           t->lemmingYields);
1939 
1940   fprintf(statsFile, " Speculative acquire attempts     : %10d\n",
1941           totalSpeculations);
1942   fprintf(statsFile, " Successes                        : %10d (%5.1f%%)\n",
1943           t->successfulSpeculations,
1944           percent(t->successfulSpeculations, totalSpeculations));
1945   fprintf(statsFile, " Soft failures                    : %10d (%5.1f%%)\n",
1946           t->softFailedSpeculations,
1947           percent(t->softFailedSpeculations, totalSpeculations));
1948   fprintf(statsFile, " Hard failures                    : %10d (%5.1f%%)\n",
1949           t->hardFailedSpeculations,
1950           percent(t->hardFailedSpeculations, totalSpeculations));
1951 }
1952 
1953 #define KMP_INC_STAT(lck, stat) (lck->lk.adaptive.stats.stat++)
1954 #else
1955 #define KMP_INC_STAT(lck, stat)
1956 
1957 #endif // KMP_DEBUG_ADAPTIVE_LOCKS
1958 
1959 static inline bool __kmp_is_unlocked_queuing_lock(kmp_queuing_lock_t *lck) {
1960   // It is enough to check that the head_id is zero.
1961   // We don't also need to check the tail.
1962   bool res = lck->lk.head_id == 0;
1963 
1964 // We need a fence here, since we must ensure that no memory operations
1965 // from later in this thread float above that read.
1966 #if KMP_COMPILER_ICC
1967   _mm_mfence();
1968 #else
1969   __sync_synchronize();
1970 #endif
1971 
1972   return res;
1973 }
1974 
1975 // Functions for manipulating the badness
1976 static __inline void
1977 __kmp_update_badness_after_success(kmp_adaptive_lock_t *lck) {
1978   // Reset the badness to zero so we eagerly try to speculate again
1979   lck->lk.adaptive.badness = 0;
1980   KMP_INC_STAT(lck, successfulSpeculations);
1981 }
1982 
1983 // Create a bit mask with one more set bit.
1984 static __inline void __kmp_step_badness(kmp_adaptive_lock_t *lck) {
1985   kmp_uint32 newBadness = (lck->lk.adaptive.badness << 1) | 1;
1986   if (newBadness > lck->lk.adaptive.max_badness) {
1987     return;
1988   } else {
1989     lck->lk.adaptive.badness = newBadness;
1990   }
1991 }
1992 
1993 // Check whether speculation should be attempted.
1994 KMP_ATTRIBUTE_TARGET_RTM
1995 static __inline int __kmp_should_speculate(kmp_adaptive_lock_t *lck,
1996                                            kmp_int32 gtid) {
1997   kmp_uint32 badness = lck->lk.adaptive.badness;
1998   kmp_uint32 attempts = lck->lk.adaptive.acquire_attempts;
1999   int res = (attempts & badness) == 0;
2000   return res;
2001 }
2002 
2003 // Attempt to acquire only the speculative lock.
2004 // Does not back off to the non-speculative lock.
2005 KMP_ATTRIBUTE_TARGET_RTM
2006 static int __kmp_test_adaptive_lock_only(kmp_adaptive_lock_t *lck,
2007                                          kmp_int32 gtid) {
2008   int retries = lck->lk.adaptive.max_soft_retries;
2009 
2010   // We don't explicitly count the start of speculation, rather we record the
2011   // results (success, hard fail, soft fail). The sum of all of those is the
2012   // total number of times we started speculation since all speculations must
2013   // end one of those ways.
2014   do {
2015     kmp_uint32 status = _xbegin();
2016     // Switch this in to disable actual speculation but exercise at least some
2017     // of the rest of the code. Useful for debugging...
2018     // kmp_uint32 status = _XABORT_NESTED;
2019 
2020     if (status == _XBEGIN_STARTED) {
2021       /* We have successfully started speculation. Check that no-one acquired
2022          the lock for real between when we last looked and now. This also gets
2023          the lock cache line into our read-set, which we need so that we'll
2024          abort if anyone later claims it for real. */
2025       if (!__kmp_is_unlocked_queuing_lock(GET_QLK_PTR(lck))) {
2026         // Lock is now visibly acquired, so someone beat us to it. Abort the
2027         // transaction so we'll restart from _xbegin with the failure status.
2028         _xabort(0x01);
2029         KMP_ASSERT2(0, "should not get here");
2030       }
2031       return 1; // Lock has been acquired (speculatively)
2032     } else {
2033       // We have aborted, update the statistics
2034       if (status & SOFT_ABORT_MASK) {
2035         KMP_INC_STAT(lck, softFailedSpeculations);
2036         // and loop round to retry.
2037       } else {
2038         KMP_INC_STAT(lck, hardFailedSpeculations);
2039         // Give up if we had a hard failure.
2040         break;
2041       }
2042     }
2043   } while (retries--); // Loop while we have retries, and didn't fail hard.
2044 
2045   // Either we had a hard failure or we didn't succeed softly after
2046   // the full set of attempts, so back off the badness.
2047   __kmp_step_badness(lck);
2048   return 0;
2049 }
2050 
2051 // Attempt to acquire the speculative lock, or back off to the non-speculative
2052 // one if the speculative lock cannot be acquired.
2053 // We can succeed speculatively, non-speculatively, or fail.
2054 static int __kmp_test_adaptive_lock(kmp_adaptive_lock_t *lck, kmp_int32 gtid) {
2055   // First try to acquire the lock speculatively
2056   if (__kmp_should_speculate(lck, gtid) &&
2057       __kmp_test_adaptive_lock_only(lck, gtid))
2058     return 1;
2059 
2060   // Speculative acquisition failed, so try to acquire it non-speculatively.
2061   // Count the non-speculative acquire attempt
2062   lck->lk.adaptive.acquire_attempts++;
2063 
2064   // Use base, non-speculative lock.
2065   if (__kmp_test_queuing_lock(GET_QLK_PTR(lck), gtid)) {
2066     KMP_INC_STAT(lck, nonSpeculativeAcquires);
2067     return 1; // Lock is acquired (non-speculatively)
2068   } else {
2069     return 0; // Failed to acquire the lock, it's already visibly locked.
2070   }
2071 }
2072 
2073 static int __kmp_test_adaptive_lock_with_checks(kmp_adaptive_lock_t *lck,
2074                                                 kmp_int32 gtid) {
2075   char const *const func = "omp_test_lock";
2076   if (lck->lk.qlk.initialized != GET_QLK_PTR(lck)) {
2077     KMP_FATAL(LockIsUninitialized, func);
2078   }
2079 
2080   int retval = __kmp_test_adaptive_lock(lck, gtid);
2081 
2082   if (retval) {
2083     lck->lk.qlk.owner_id = gtid + 1;
2084   }
2085   return retval;
2086 }
2087 
2088 // Block until we can acquire a speculative, adaptive lock. We check whether we
2089 // should be trying to speculate. If we should be, we check the real lock to see
2090 // if it is free, and, if not, pause without attempting to acquire it until it
2091 // is. Then we try the speculative acquire. This means that although we suffer
2092 // from lemmings a little (because all we can't acquire the lock speculatively
2093 // until the queue of threads waiting has cleared), we don't get into a state
2094 // where we can never acquire the lock speculatively (because we force the queue
2095 // to clear by preventing new arrivals from entering the queue). This does mean
2096 // that when we're trying to break lemmings, the lock is no longer fair. However
2097 // OpenMP makes no guarantee that its locks are fair, so this isn't a real
2098 // problem.
2099 static void __kmp_acquire_adaptive_lock(kmp_adaptive_lock_t *lck,
2100                                         kmp_int32 gtid) {
2101   if (__kmp_should_speculate(lck, gtid)) {
2102     if (__kmp_is_unlocked_queuing_lock(GET_QLK_PTR(lck))) {
2103       if (__kmp_test_adaptive_lock_only(lck, gtid))
2104         return;
2105       // We tried speculation and failed, so give up.
2106     } else {
2107       // We can't try speculation until the lock is free, so we pause here
2108       // (without suspending on the queueing lock, to allow it to drain, then
2109       // try again. All other threads will also see the same result for
2110       // shouldSpeculate, so will be doing the same if they try to claim the
2111       // lock from now on.
2112       while (!__kmp_is_unlocked_queuing_lock(GET_QLK_PTR(lck))) {
2113         KMP_INC_STAT(lck, lemmingYields);
2114         KMP_YIELD(TRUE);
2115       }
2116 
2117       if (__kmp_test_adaptive_lock_only(lck, gtid))
2118         return;
2119     }
2120   }
2121 
2122   // Speculative acquisition failed, so acquire it non-speculatively.
2123   // Count the non-speculative acquire attempt
2124   lck->lk.adaptive.acquire_attempts++;
2125 
2126   __kmp_acquire_queuing_lock_timed_template<FALSE>(GET_QLK_PTR(lck), gtid);
2127   // We have acquired the base lock, so count that.
2128   KMP_INC_STAT(lck, nonSpeculativeAcquires);
2129   ANNOTATE_QUEUING_ACQUIRED(lck);
2130 }
2131 
2132 static void __kmp_acquire_adaptive_lock_with_checks(kmp_adaptive_lock_t *lck,
2133                                                     kmp_int32 gtid) {
2134   char const *const func = "omp_set_lock";
2135   if (lck->lk.qlk.initialized != GET_QLK_PTR(lck)) {
2136     KMP_FATAL(LockIsUninitialized, func);
2137   }
2138   if (__kmp_get_queuing_lock_owner(GET_QLK_PTR(lck)) == gtid) {
2139     KMP_FATAL(LockIsAlreadyOwned, func);
2140   }
2141 
2142   __kmp_acquire_adaptive_lock(lck, gtid);
2143 
2144   lck->lk.qlk.owner_id = gtid + 1;
2145 }
2146 
2147 KMP_ATTRIBUTE_TARGET_RTM
2148 static int __kmp_release_adaptive_lock(kmp_adaptive_lock_t *lck,
2149                                        kmp_int32 gtid) {
2150   if (__kmp_is_unlocked_queuing_lock(GET_QLK_PTR(
2151           lck))) { // If the lock doesn't look claimed we must be speculating.
2152     // (Or the user's code is buggy and they're releasing without locking;
2153     // if we had XTEST we'd be able to check that case...)
2154     _xend(); // Exit speculation
2155     __kmp_update_badness_after_success(lck);
2156   } else { // Since the lock *is* visibly locked we're not speculating,
2157     // so should use the underlying lock's release scheme.
2158     __kmp_release_queuing_lock(GET_QLK_PTR(lck), gtid);
2159   }
2160   return KMP_LOCK_RELEASED;
2161 }
2162 
2163 static int __kmp_release_adaptive_lock_with_checks(kmp_adaptive_lock_t *lck,
2164                                                    kmp_int32 gtid) {
2165   char const *const func = "omp_unset_lock";
2166   KMP_MB(); /* in case another processor initialized lock */
2167   if (lck->lk.qlk.initialized != GET_QLK_PTR(lck)) {
2168     KMP_FATAL(LockIsUninitialized, func);
2169   }
2170   if (__kmp_get_queuing_lock_owner(GET_QLK_PTR(lck)) == -1) {
2171     KMP_FATAL(LockUnsettingFree, func);
2172   }
2173   if (__kmp_get_queuing_lock_owner(GET_QLK_PTR(lck)) != gtid) {
2174     KMP_FATAL(LockUnsettingSetByAnother, func);
2175   }
2176   lck->lk.qlk.owner_id = 0;
2177   __kmp_release_adaptive_lock(lck, gtid);
2178   return KMP_LOCK_RELEASED;
2179 }
2180 
2181 static void __kmp_init_adaptive_lock(kmp_adaptive_lock_t *lck) {
2182   __kmp_init_queuing_lock(GET_QLK_PTR(lck));
2183   lck->lk.adaptive.badness = 0;
2184   lck->lk.adaptive.acquire_attempts = 0; // nonSpeculativeAcquireAttempts = 0;
2185   lck->lk.adaptive.max_soft_retries =
2186       __kmp_adaptive_backoff_params.max_soft_retries;
2187   lck->lk.adaptive.max_badness = __kmp_adaptive_backoff_params.max_badness;
2188 #if KMP_DEBUG_ADAPTIVE_LOCKS
2189   __kmp_zero_speculative_stats(&lck->lk.adaptive);
2190 #endif
2191   KA_TRACE(1000, ("__kmp_init_adaptive_lock: lock %p initialized\n", lck));
2192 }
2193 
2194 static void __kmp_destroy_adaptive_lock(kmp_adaptive_lock_t *lck) {
2195 #if KMP_DEBUG_ADAPTIVE_LOCKS
2196   __kmp_accumulate_speculative_stats(&lck->lk.adaptive);
2197 #endif
2198   __kmp_destroy_queuing_lock(GET_QLK_PTR(lck));
2199   // Nothing needed for the speculative part.
2200 }
2201 
2202 static void __kmp_destroy_adaptive_lock_with_checks(kmp_adaptive_lock_t *lck) {
2203   char const *const func = "omp_destroy_lock";
2204   if (lck->lk.qlk.initialized != GET_QLK_PTR(lck)) {
2205     KMP_FATAL(LockIsUninitialized, func);
2206   }
2207   if (__kmp_get_queuing_lock_owner(GET_QLK_PTR(lck)) != -1) {
2208     KMP_FATAL(LockStillOwned, func);
2209   }
2210   __kmp_destroy_adaptive_lock(lck);
2211 }
2212 
2213 #endif // KMP_USE_ADAPTIVE_LOCKS
2214 
2215 /* ------------------------------------------------------------------------ */
2216 /* DRDPA ticket locks                                                */
2217 /* "DRDPA" means Dynamically Reconfigurable Distributed Polling Area */
2218 
2219 static kmp_int32 __kmp_get_drdpa_lock_owner(kmp_drdpa_lock_t *lck) {
2220   return lck->lk.owner_id - 1;
2221 }
2222 
2223 static inline bool __kmp_is_drdpa_lock_nestable(kmp_drdpa_lock_t *lck) {
2224   return lck->lk.depth_locked != -1;
2225 }
2226 
2227 __forceinline static int
2228 __kmp_acquire_drdpa_lock_timed_template(kmp_drdpa_lock_t *lck, kmp_int32 gtid) {
2229   kmp_uint64 ticket = KMP_ATOMIC_INC(&lck->lk.next_ticket);
2230   kmp_uint64 mask = lck->lk.mask; // atomic load
2231   std::atomic<kmp_uint64> *polls = lck->lk.polls;
2232 
2233 #ifdef USE_LOCK_PROFILE
2234   if (polls[ticket & mask] != ticket)
2235     __kmp_printf("LOCK CONTENTION: %p\n", lck);
2236 /* else __kmp_printf( "." );*/
2237 #endif /* USE_LOCK_PROFILE */
2238 
2239   // Now spin-wait, but reload the polls pointer and mask, in case the
2240   // polling area has been reconfigured.  Unless it is reconfigured, the
2241   // reloads stay in L1 cache and are cheap.
2242   //
2243   // Keep this code in sync with KMP_WAIT, in kmp_dispatch.cpp !!!
2244   // The current implementation of KMP_WAIT doesn't allow for mask
2245   // and poll to be re-read every spin iteration.
2246   kmp_uint32 spins;
2247   KMP_FSYNC_PREPARE(lck);
2248   KMP_INIT_YIELD(spins);
2249   while (polls[ticket & mask] < ticket) { // atomic load
2250     KMP_YIELD_OVERSUB_ELSE_SPIN(spins);
2251     // Re-read the mask and the poll pointer from the lock structure.
2252     //
2253     // Make certain that "mask" is read before "polls" !!!
2254     //
2255     // If another thread picks reconfigures the polling area and updates their
2256     // values, and we get the new value of mask and the old polls pointer, we
2257     // could access memory beyond the end of the old polling area.
2258     mask = lck->lk.mask; // atomic load
2259     polls = lck->lk.polls; // atomic load
2260   }
2261 
2262   // Critical section starts here
2263   KMP_FSYNC_ACQUIRED(lck);
2264   KA_TRACE(1000, ("__kmp_acquire_drdpa_lock: ticket #%lld acquired lock %p\n",
2265                   ticket, lck));
2266   lck->lk.now_serving = ticket; // non-volatile store
2267 
2268   // Deallocate a garbage polling area if we know that we are the last
2269   // thread that could possibly access it.
2270   //
2271   // The >= check is in case __kmp_test_drdpa_lock() allocated the cleanup
2272   // ticket.
2273   if ((lck->lk.old_polls != NULL) && (ticket >= lck->lk.cleanup_ticket)) {
2274     __kmp_free(lck->lk.old_polls);
2275     lck->lk.old_polls = NULL;
2276     lck->lk.cleanup_ticket = 0;
2277   }
2278 
2279   // Check to see if we should reconfigure the polling area.
2280   // If there is still a garbage polling area to be deallocated from a
2281   // previous reconfiguration, let a later thread reconfigure it.
2282   if (lck->lk.old_polls == NULL) {
2283     bool reconfigure = false;
2284     std::atomic<kmp_uint64> *old_polls = polls;
2285     kmp_uint32 num_polls = TCR_4(lck->lk.num_polls);
2286 
2287     if (TCR_4(__kmp_nth) >
2288         (__kmp_avail_proc ? __kmp_avail_proc : __kmp_xproc)) {
2289       // We are in oversubscription mode.  Contract the polling area
2290       // down to a single location, if that hasn't been done already.
2291       if (num_polls > 1) {
2292         reconfigure = true;
2293         num_polls = TCR_4(lck->lk.num_polls);
2294         mask = 0;
2295         num_polls = 1;
2296         polls = (std::atomic<kmp_uint64> *)__kmp_allocate(num_polls *
2297                                                           sizeof(*polls));
2298         polls[0] = ticket;
2299       }
2300     } else {
2301       // We are in under/fully subscribed mode.  Check the number of
2302       // threads waiting on the lock.  The size of the polling area
2303       // should be at least the number of threads waiting.
2304       kmp_uint64 num_waiting = TCR_8(lck->lk.next_ticket) - ticket - 1;
2305       if (num_waiting > num_polls) {
2306         kmp_uint32 old_num_polls = num_polls;
2307         reconfigure = true;
2308         do {
2309           mask = (mask << 1) | 1;
2310           num_polls *= 2;
2311         } while (num_polls <= num_waiting);
2312 
2313         // Allocate the new polling area, and copy the relevant portion
2314         // of the old polling area to the new area.  __kmp_allocate()
2315         // zeroes the memory it allocates, and most of the old area is
2316         // just zero padding, so we only copy the release counters.
2317         polls = (std::atomic<kmp_uint64> *)__kmp_allocate(num_polls *
2318                                                           sizeof(*polls));
2319         kmp_uint32 i;
2320         for (i = 0; i < old_num_polls; i++) {
2321           polls[i].store(old_polls[i]);
2322         }
2323       }
2324     }
2325 
2326     if (reconfigure) {
2327       // Now write the updated fields back to the lock structure.
2328       //
2329       // Make certain that "polls" is written before "mask" !!!
2330       //
2331       // If another thread picks up the new value of mask and the old polls
2332       // pointer , it could access memory beyond the end of the old polling
2333       // area.
2334       //
2335       // On x86, we need memory fences.
2336       KA_TRACE(1000, ("__kmp_acquire_drdpa_lock: ticket #%lld reconfiguring "
2337                       "lock %p to %d polls\n",
2338                       ticket, lck, num_polls));
2339 
2340       lck->lk.old_polls = old_polls;
2341       lck->lk.polls = polls; // atomic store
2342 
2343       KMP_MB();
2344 
2345       lck->lk.num_polls = num_polls;
2346       lck->lk.mask = mask; // atomic store
2347 
2348       KMP_MB();
2349 
2350       // Only after the new polling area and mask have been flushed
2351       // to main memory can we update the cleanup ticket field.
2352       //
2353       // volatile load / non-volatile store
2354       lck->lk.cleanup_ticket = lck->lk.next_ticket;
2355     }
2356   }
2357   return KMP_LOCK_ACQUIRED_FIRST;
2358 }
2359 
2360 int __kmp_acquire_drdpa_lock(kmp_drdpa_lock_t *lck, kmp_int32 gtid) {
2361   int retval = __kmp_acquire_drdpa_lock_timed_template(lck, gtid);
2362   ANNOTATE_DRDPA_ACQUIRED(lck);
2363   return retval;
2364 }
2365 
2366 static int __kmp_acquire_drdpa_lock_with_checks(kmp_drdpa_lock_t *lck,
2367                                                 kmp_int32 gtid) {
2368   char const *const func = "omp_set_lock";
2369   if (lck->lk.initialized != lck) {
2370     KMP_FATAL(LockIsUninitialized, func);
2371   }
2372   if (__kmp_is_drdpa_lock_nestable(lck)) {
2373     KMP_FATAL(LockNestableUsedAsSimple, func);
2374   }
2375   if ((gtid >= 0) && (__kmp_get_drdpa_lock_owner(lck) == gtid)) {
2376     KMP_FATAL(LockIsAlreadyOwned, func);
2377   }
2378 
2379   __kmp_acquire_drdpa_lock(lck, gtid);
2380 
2381   lck->lk.owner_id = gtid + 1;
2382   return KMP_LOCK_ACQUIRED_FIRST;
2383 }
2384 
2385 int __kmp_test_drdpa_lock(kmp_drdpa_lock_t *lck, kmp_int32 gtid) {
2386   // First get a ticket, then read the polls pointer and the mask.
2387   // The polls pointer must be read before the mask!!! (See above)
2388   kmp_uint64 ticket = lck->lk.next_ticket; // atomic load
2389   std::atomic<kmp_uint64> *polls = lck->lk.polls;
2390   kmp_uint64 mask = lck->lk.mask; // atomic load
2391   if (polls[ticket & mask] == ticket) {
2392     kmp_uint64 next_ticket = ticket + 1;
2393     if (__kmp_atomic_compare_store_acq(&lck->lk.next_ticket, ticket,
2394                                        next_ticket)) {
2395       KMP_FSYNC_ACQUIRED(lck);
2396       KA_TRACE(1000, ("__kmp_test_drdpa_lock: ticket #%lld acquired lock %p\n",
2397                       ticket, lck));
2398       lck->lk.now_serving = ticket; // non-volatile store
2399 
2400       // Since no threads are waiting, there is no possibility that we would
2401       // want to reconfigure the polling area.  We might have the cleanup ticket
2402       // value (which says that it is now safe to deallocate old_polls), but
2403       // we'll let a later thread which calls __kmp_acquire_lock do that - this
2404       // routine isn't supposed to block, and we would risk blocks if we called
2405       // __kmp_free() to do the deallocation.
2406       return TRUE;
2407     }
2408   }
2409   return FALSE;
2410 }
2411 
2412 static int __kmp_test_drdpa_lock_with_checks(kmp_drdpa_lock_t *lck,
2413                                              kmp_int32 gtid) {
2414   char const *const func = "omp_test_lock";
2415   if (lck->lk.initialized != lck) {
2416     KMP_FATAL(LockIsUninitialized, func);
2417   }
2418   if (__kmp_is_drdpa_lock_nestable(lck)) {
2419     KMP_FATAL(LockNestableUsedAsSimple, func);
2420   }
2421 
2422   int retval = __kmp_test_drdpa_lock(lck, gtid);
2423 
2424   if (retval) {
2425     lck->lk.owner_id = gtid + 1;
2426   }
2427   return retval;
2428 }
2429 
2430 int __kmp_release_drdpa_lock(kmp_drdpa_lock_t *lck, kmp_int32 gtid) {
2431   // Read the ticket value from the lock data struct, then the polls pointer and
2432   // the mask.  The polls pointer must be read before the mask!!! (See above)
2433   kmp_uint64 ticket = lck->lk.now_serving + 1; // non-atomic load
2434   std::atomic<kmp_uint64> *polls = lck->lk.polls; // atomic load
2435   kmp_uint64 mask = lck->lk.mask; // atomic load
2436   KA_TRACE(1000, ("__kmp_release_drdpa_lock: ticket #%lld released lock %p\n",
2437                   ticket - 1, lck));
2438   KMP_FSYNC_RELEASING(lck);
2439   ANNOTATE_DRDPA_RELEASED(lck);
2440   polls[ticket & mask] = ticket; // atomic store
2441   return KMP_LOCK_RELEASED;
2442 }
2443 
2444 static int __kmp_release_drdpa_lock_with_checks(kmp_drdpa_lock_t *lck,
2445                                                 kmp_int32 gtid) {
2446   char const *const func = "omp_unset_lock";
2447   KMP_MB(); /* in case another processor initialized lock */
2448   if (lck->lk.initialized != lck) {
2449     KMP_FATAL(LockIsUninitialized, func);
2450   }
2451   if (__kmp_is_drdpa_lock_nestable(lck)) {
2452     KMP_FATAL(LockNestableUsedAsSimple, func);
2453   }
2454   if (__kmp_get_drdpa_lock_owner(lck) == -1) {
2455     KMP_FATAL(LockUnsettingFree, func);
2456   }
2457   if ((gtid >= 0) && (__kmp_get_drdpa_lock_owner(lck) >= 0) &&
2458       (__kmp_get_drdpa_lock_owner(lck) != gtid)) {
2459     KMP_FATAL(LockUnsettingSetByAnother, func);
2460   }
2461   lck->lk.owner_id = 0;
2462   return __kmp_release_drdpa_lock(lck, gtid);
2463 }
2464 
2465 void __kmp_init_drdpa_lock(kmp_drdpa_lock_t *lck) {
2466   lck->lk.location = NULL;
2467   lck->lk.mask = 0;
2468   lck->lk.num_polls = 1;
2469   lck->lk.polls = (std::atomic<kmp_uint64> *)__kmp_allocate(
2470       lck->lk.num_polls * sizeof(*(lck->lk.polls)));
2471   lck->lk.cleanup_ticket = 0;
2472   lck->lk.old_polls = NULL;
2473   lck->lk.next_ticket = 0;
2474   lck->lk.now_serving = 0;
2475   lck->lk.owner_id = 0; // no thread owns the lock.
2476   lck->lk.depth_locked = -1; // >= 0 for nestable locks, -1 for simple locks.
2477   lck->lk.initialized = lck;
2478 
2479   KA_TRACE(1000, ("__kmp_init_drdpa_lock: lock %p initialized\n", lck));
2480 }
2481 
2482 void __kmp_destroy_drdpa_lock(kmp_drdpa_lock_t *lck) {
2483   lck->lk.initialized = NULL;
2484   lck->lk.location = NULL;
2485   if (lck->lk.polls.load() != NULL) {
2486     __kmp_free(lck->lk.polls.load());
2487     lck->lk.polls = NULL;
2488   }
2489   if (lck->lk.old_polls != NULL) {
2490     __kmp_free(lck->lk.old_polls);
2491     lck->lk.old_polls = NULL;
2492   }
2493   lck->lk.mask = 0;
2494   lck->lk.num_polls = 0;
2495   lck->lk.cleanup_ticket = 0;
2496   lck->lk.next_ticket = 0;
2497   lck->lk.now_serving = 0;
2498   lck->lk.owner_id = 0;
2499   lck->lk.depth_locked = -1;
2500 }
2501 
2502 static void __kmp_destroy_drdpa_lock_with_checks(kmp_drdpa_lock_t *lck) {
2503   char const *const func = "omp_destroy_lock";
2504   if (lck->lk.initialized != lck) {
2505     KMP_FATAL(LockIsUninitialized, func);
2506   }
2507   if (__kmp_is_drdpa_lock_nestable(lck)) {
2508     KMP_FATAL(LockNestableUsedAsSimple, func);
2509   }
2510   if (__kmp_get_drdpa_lock_owner(lck) != -1) {
2511     KMP_FATAL(LockStillOwned, func);
2512   }
2513   __kmp_destroy_drdpa_lock(lck);
2514 }
2515 
2516 // nested drdpa ticket locks
2517 
2518 int __kmp_acquire_nested_drdpa_lock(kmp_drdpa_lock_t *lck, kmp_int32 gtid) {
2519   KMP_DEBUG_ASSERT(gtid >= 0);
2520 
2521   if (__kmp_get_drdpa_lock_owner(lck) == gtid) {
2522     lck->lk.depth_locked += 1;
2523     return KMP_LOCK_ACQUIRED_NEXT;
2524   } else {
2525     __kmp_acquire_drdpa_lock_timed_template(lck, gtid);
2526     ANNOTATE_DRDPA_ACQUIRED(lck);
2527     KMP_MB();
2528     lck->lk.depth_locked = 1;
2529     KMP_MB();
2530     lck->lk.owner_id = gtid + 1;
2531     return KMP_LOCK_ACQUIRED_FIRST;
2532   }
2533 }
2534 
2535 static void __kmp_acquire_nested_drdpa_lock_with_checks(kmp_drdpa_lock_t *lck,
2536                                                         kmp_int32 gtid) {
2537   char const *const func = "omp_set_nest_lock";
2538   if (lck->lk.initialized != lck) {
2539     KMP_FATAL(LockIsUninitialized, func);
2540   }
2541   if (!__kmp_is_drdpa_lock_nestable(lck)) {
2542     KMP_FATAL(LockSimpleUsedAsNestable, func);
2543   }
2544   __kmp_acquire_nested_drdpa_lock(lck, gtid);
2545 }
2546 
2547 int __kmp_test_nested_drdpa_lock(kmp_drdpa_lock_t *lck, kmp_int32 gtid) {
2548   int retval;
2549 
2550   KMP_DEBUG_ASSERT(gtid >= 0);
2551 
2552   if (__kmp_get_drdpa_lock_owner(lck) == gtid) {
2553     retval = ++lck->lk.depth_locked;
2554   } else if (!__kmp_test_drdpa_lock(lck, gtid)) {
2555     retval = 0;
2556   } else {
2557     KMP_MB();
2558     retval = lck->lk.depth_locked = 1;
2559     KMP_MB();
2560     lck->lk.owner_id = gtid + 1;
2561   }
2562   return retval;
2563 }
2564 
2565 static int __kmp_test_nested_drdpa_lock_with_checks(kmp_drdpa_lock_t *lck,
2566                                                     kmp_int32 gtid) {
2567   char const *const func = "omp_test_nest_lock";
2568   if (lck->lk.initialized != lck) {
2569     KMP_FATAL(LockIsUninitialized, func);
2570   }
2571   if (!__kmp_is_drdpa_lock_nestable(lck)) {
2572     KMP_FATAL(LockSimpleUsedAsNestable, func);
2573   }
2574   return __kmp_test_nested_drdpa_lock(lck, gtid);
2575 }
2576 
2577 int __kmp_release_nested_drdpa_lock(kmp_drdpa_lock_t *lck, kmp_int32 gtid) {
2578   KMP_DEBUG_ASSERT(gtid >= 0);
2579 
2580   KMP_MB();
2581   if (--(lck->lk.depth_locked) == 0) {
2582     KMP_MB();
2583     lck->lk.owner_id = 0;
2584     __kmp_release_drdpa_lock(lck, gtid);
2585     return KMP_LOCK_RELEASED;
2586   }
2587   return KMP_LOCK_STILL_HELD;
2588 }
2589 
2590 static int __kmp_release_nested_drdpa_lock_with_checks(kmp_drdpa_lock_t *lck,
2591                                                        kmp_int32 gtid) {
2592   char const *const func = "omp_unset_nest_lock";
2593   KMP_MB(); /* in case another processor initialized lock */
2594   if (lck->lk.initialized != lck) {
2595     KMP_FATAL(LockIsUninitialized, func);
2596   }
2597   if (!__kmp_is_drdpa_lock_nestable(lck)) {
2598     KMP_FATAL(LockSimpleUsedAsNestable, func);
2599   }
2600   if (__kmp_get_drdpa_lock_owner(lck) == -1) {
2601     KMP_FATAL(LockUnsettingFree, func);
2602   }
2603   if (__kmp_get_drdpa_lock_owner(lck) != gtid) {
2604     KMP_FATAL(LockUnsettingSetByAnother, func);
2605   }
2606   return __kmp_release_nested_drdpa_lock(lck, gtid);
2607 }
2608 
2609 void __kmp_init_nested_drdpa_lock(kmp_drdpa_lock_t *lck) {
2610   __kmp_init_drdpa_lock(lck);
2611   lck->lk.depth_locked = 0; // >= 0 for nestable locks, -1 for simple locks
2612 }
2613 
2614 void __kmp_destroy_nested_drdpa_lock(kmp_drdpa_lock_t *lck) {
2615   __kmp_destroy_drdpa_lock(lck);
2616   lck->lk.depth_locked = 0;
2617 }
2618 
2619 static void __kmp_destroy_nested_drdpa_lock_with_checks(kmp_drdpa_lock_t *lck) {
2620   char const *const func = "omp_destroy_nest_lock";
2621   if (lck->lk.initialized != lck) {
2622     KMP_FATAL(LockIsUninitialized, func);
2623   }
2624   if (!__kmp_is_drdpa_lock_nestable(lck)) {
2625     KMP_FATAL(LockSimpleUsedAsNestable, func);
2626   }
2627   if (__kmp_get_drdpa_lock_owner(lck) != -1) {
2628     KMP_FATAL(LockStillOwned, func);
2629   }
2630   __kmp_destroy_nested_drdpa_lock(lck);
2631 }
2632 
2633 // access functions to fields which don't exist for all lock kinds.
2634 
2635 static const ident_t *__kmp_get_drdpa_lock_location(kmp_drdpa_lock_t *lck) {
2636   return lck->lk.location;
2637 }
2638 
2639 static void __kmp_set_drdpa_lock_location(kmp_drdpa_lock_t *lck,
2640                                           const ident_t *loc) {
2641   lck->lk.location = loc;
2642 }
2643 
2644 static kmp_lock_flags_t __kmp_get_drdpa_lock_flags(kmp_drdpa_lock_t *lck) {
2645   return lck->lk.flags;
2646 }
2647 
2648 static void __kmp_set_drdpa_lock_flags(kmp_drdpa_lock_t *lck,
2649                                        kmp_lock_flags_t flags) {
2650   lck->lk.flags = flags;
2651 }
2652 
2653 // Time stamp counter
2654 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
2655 #define __kmp_tsc() __kmp_hardware_timestamp()
2656 // Runtime's default backoff parameters
2657 kmp_backoff_t __kmp_spin_backoff_params = {1, 4096, 100};
2658 #else
2659 // Use nanoseconds for other platforms
2660 extern kmp_uint64 __kmp_now_nsec();
2661 kmp_backoff_t __kmp_spin_backoff_params = {1, 256, 100};
2662 #define __kmp_tsc() __kmp_now_nsec()
2663 #endif
2664 
2665 // A useful predicate for dealing with timestamps that may wrap.
2666 // Is a before b? Since the timestamps may wrap, this is asking whether it's
2667 // shorter to go clockwise from a to b around the clock-face, or anti-clockwise.
2668 // Times where going clockwise is less distance than going anti-clockwise
2669 // are in the future, others are in the past. e.g. a = MAX-1, b = MAX+1 (=0),
2670 // then a > b (true) does not mean a reached b; whereas signed(a) = -2,
2671 // signed(b) = 0 captures the actual difference
2672 static inline bool before(kmp_uint64 a, kmp_uint64 b) {
2673   return ((kmp_int64)b - (kmp_int64)a) > 0;
2674 }
2675 
2676 // Truncated binary exponential backoff function
2677 void __kmp_spin_backoff(kmp_backoff_t *boff) {
2678   // We could flatten this loop, but making it a nested loop gives better result
2679   kmp_uint32 i;
2680   for (i = boff->step; i > 0; i--) {
2681     kmp_uint64 goal = __kmp_tsc() + boff->min_tick;
2682     do {
2683       KMP_CPU_PAUSE();
2684     } while (before(__kmp_tsc(), goal));
2685   }
2686   boff->step = (boff->step << 1 | 1) & (boff->max_backoff - 1);
2687 }
2688 
2689 #if KMP_USE_DYNAMIC_LOCK
2690 
2691 // Direct lock initializers. It simply writes a tag to the low 8 bits of the
2692 // lock word.
2693 static void __kmp_init_direct_lock(kmp_dyna_lock_t *lck,
2694                                    kmp_dyna_lockseq_t seq) {
2695   TCW_4(*lck, KMP_GET_D_TAG(seq));
2696   KA_TRACE(
2697       20,
2698       ("__kmp_init_direct_lock: initialized direct lock with type#%d\n", seq));
2699 }
2700 
2701 #if KMP_USE_TSX
2702 
2703 // HLE lock functions - imported from the testbed runtime.
2704 #define HLE_ACQUIRE ".byte 0xf2;"
2705 #define HLE_RELEASE ".byte 0xf3;"
2706 
2707 static inline kmp_uint32 swap4(kmp_uint32 volatile *p, kmp_uint32 v) {
2708   __asm__ volatile(HLE_ACQUIRE "xchg %1,%0" : "+r"(v), "+m"(*p) : : "memory");
2709   return v;
2710 }
2711 
2712 static void __kmp_destroy_hle_lock(kmp_dyna_lock_t *lck) { TCW_4(*lck, 0); }
2713 
2714 static void __kmp_destroy_hle_lock_with_checks(kmp_dyna_lock_t *lck) {
2715   TCW_4(*lck, 0);
2716 }
2717 
2718 static void __kmp_acquire_hle_lock(kmp_dyna_lock_t *lck, kmp_int32 gtid) {
2719   // Use gtid for KMP_LOCK_BUSY if necessary
2720   if (swap4(lck, KMP_LOCK_BUSY(1, hle)) != KMP_LOCK_FREE(hle)) {
2721     int delay = 1;
2722     do {
2723       while (*(kmp_uint32 volatile *)lck != KMP_LOCK_FREE(hle)) {
2724         for (int i = delay; i != 0; --i)
2725           KMP_CPU_PAUSE();
2726         delay = ((delay << 1) | 1) & 7;
2727       }
2728     } while (swap4(lck, KMP_LOCK_BUSY(1, hle)) != KMP_LOCK_FREE(hle));
2729   }
2730 }
2731 
2732 static void __kmp_acquire_hle_lock_with_checks(kmp_dyna_lock_t *lck,
2733                                                kmp_int32 gtid) {
2734   __kmp_acquire_hle_lock(lck, gtid); // TODO: add checks
2735 }
2736 
2737 static int __kmp_release_hle_lock(kmp_dyna_lock_t *lck, kmp_int32 gtid) {
2738   __asm__ volatile(HLE_RELEASE "movl %1,%0"
2739                    : "=m"(*lck)
2740                    : "r"(KMP_LOCK_FREE(hle))
2741                    : "memory");
2742   return KMP_LOCK_RELEASED;
2743 }
2744 
2745 static int __kmp_release_hle_lock_with_checks(kmp_dyna_lock_t *lck,
2746                                               kmp_int32 gtid) {
2747   return __kmp_release_hle_lock(lck, gtid); // TODO: add checks
2748 }
2749 
2750 static int __kmp_test_hle_lock(kmp_dyna_lock_t *lck, kmp_int32 gtid) {
2751   return swap4(lck, KMP_LOCK_BUSY(1, hle)) == KMP_LOCK_FREE(hle);
2752 }
2753 
2754 static int __kmp_test_hle_lock_with_checks(kmp_dyna_lock_t *lck,
2755                                            kmp_int32 gtid) {
2756   return __kmp_test_hle_lock(lck, gtid); // TODO: add checks
2757 }
2758 
2759 static void __kmp_init_rtm_queuing_lock(kmp_queuing_lock_t *lck) {
2760   __kmp_init_queuing_lock(lck);
2761 }
2762 
2763 static void __kmp_destroy_rtm_queuing_lock(kmp_queuing_lock_t *lck) {
2764   __kmp_destroy_queuing_lock(lck);
2765 }
2766 
2767 static void
2768 __kmp_destroy_rtm_queuing_lock_with_checks(kmp_queuing_lock_t *lck) {
2769   __kmp_destroy_queuing_lock_with_checks(lck);
2770 }
2771 
2772 KMP_ATTRIBUTE_TARGET_RTM
2773 static void __kmp_acquire_rtm_queuing_lock(kmp_queuing_lock_t *lck,
2774                                            kmp_int32 gtid) {
2775   unsigned retries = 3, status;
2776   do {
2777     status = _xbegin();
2778     if (status == _XBEGIN_STARTED) {
2779       if (__kmp_is_unlocked_queuing_lock(lck))
2780         return;
2781       _xabort(0xff);
2782     }
2783     if ((status & _XABORT_EXPLICIT) && _XABORT_CODE(status) == 0xff) {
2784       // Wait until lock becomes free
2785       while (!__kmp_is_unlocked_queuing_lock(lck)) {
2786         KMP_YIELD(TRUE);
2787       }
2788     } else if (!(status & _XABORT_RETRY))
2789       break;
2790   } while (retries--);
2791 
2792   // Fall-back non-speculative lock (xchg)
2793   __kmp_acquire_queuing_lock(lck, gtid);
2794 }
2795 
2796 static void __kmp_acquire_rtm_queuing_lock_with_checks(kmp_queuing_lock_t *lck,
2797                                                        kmp_int32 gtid) {
2798   __kmp_acquire_rtm_queuing_lock(lck, gtid);
2799 }
2800 
2801 KMP_ATTRIBUTE_TARGET_RTM
2802 static int __kmp_release_rtm_queuing_lock(kmp_queuing_lock_t *lck,
2803                                           kmp_int32 gtid) {
2804   if (__kmp_is_unlocked_queuing_lock(lck)) {
2805     // Releasing from speculation
2806     _xend();
2807   } else {
2808     // Releasing from a real lock
2809     __kmp_release_queuing_lock(lck, gtid);
2810   }
2811   return KMP_LOCK_RELEASED;
2812 }
2813 
2814 static int __kmp_release_rtm_queuing_lock_with_checks(kmp_queuing_lock_t *lck,
2815                                                       kmp_int32 gtid) {
2816   return __kmp_release_rtm_queuing_lock(lck, gtid);
2817 }
2818 
2819 KMP_ATTRIBUTE_TARGET_RTM
2820 static int __kmp_test_rtm_queuing_lock(kmp_queuing_lock_t *lck,
2821                                        kmp_int32 gtid) {
2822   unsigned retries = 3, status;
2823   do {
2824     status = _xbegin();
2825     if (status == _XBEGIN_STARTED && __kmp_is_unlocked_queuing_lock(lck)) {
2826       return 1;
2827     }
2828     if (!(status & _XABORT_RETRY))
2829       break;
2830   } while (retries--);
2831 
2832   return __kmp_test_queuing_lock(lck, gtid);
2833 }
2834 
2835 static int __kmp_test_rtm_queuing_lock_with_checks(kmp_queuing_lock_t *lck,
2836                                                    kmp_int32 gtid) {
2837   return __kmp_test_rtm_queuing_lock(lck, gtid);
2838 }
2839 
2840 // Reuse kmp_tas_lock_t for TSX lock which use RTM with fall-back spin lock.
2841 typedef kmp_tas_lock_t kmp_rtm_spin_lock_t;
2842 
2843 static void __kmp_destroy_rtm_spin_lock(kmp_rtm_spin_lock_t *lck) {
2844   KMP_ATOMIC_ST_REL(&lck->lk.poll, 0);
2845 }
2846 
2847 static void __kmp_destroy_rtm_spin_lock_with_checks(kmp_rtm_spin_lock_t *lck) {
2848   __kmp_destroy_rtm_spin_lock(lck);
2849 }
2850 
2851 KMP_ATTRIBUTE_TARGET_RTM
2852 static int __kmp_acquire_rtm_spin_lock(kmp_rtm_spin_lock_t *lck,
2853                                        kmp_int32 gtid) {
2854   unsigned retries = 3, status;
2855   kmp_int32 lock_free = KMP_LOCK_FREE(rtm_spin);
2856   kmp_int32 lock_busy = KMP_LOCK_BUSY(1, rtm_spin);
2857   do {
2858     status = _xbegin();
2859     if (status == _XBEGIN_STARTED) {
2860       if (KMP_ATOMIC_LD_RLX(&lck->lk.poll) == lock_free)
2861         return KMP_LOCK_ACQUIRED_FIRST;
2862       _xabort(0xff);
2863     }
2864     if ((status & _XABORT_EXPLICIT) && _XABORT_CODE(status) == 0xff) {
2865       // Wait until lock becomes free
2866       while (KMP_ATOMIC_LD_RLX(&lck->lk.poll) != lock_free) {
2867         KMP_YIELD(TRUE);
2868       }
2869     } else if (!(status & _XABORT_RETRY))
2870       break;
2871   } while (retries--);
2872 
2873   // Fall-back spin lock
2874   KMP_FSYNC_PREPARE(lck);
2875   kmp_backoff_t backoff = __kmp_spin_backoff_params;
2876   while (KMP_ATOMIC_LD_RLX(&lck->lk.poll) != lock_free ||
2877          !__kmp_atomic_compare_store_acq(&lck->lk.poll, lock_free, lock_busy)) {
2878     __kmp_spin_backoff(&backoff);
2879   }
2880   KMP_FSYNC_ACQUIRED(lck);
2881   return KMP_LOCK_ACQUIRED_FIRST;
2882 }
2883 
2884 static int __kmp_acquire_rtm_spin_lock_with_checks(kmp_rtm_spin_lock_t *lck,
2885                                                    kmp_int32 gtid) {
2886   return __kmp_acquire_rtm_spin_lock(lck, gtid);
2887 }
2888 
2889 KMP_ATTRIBUTE_TARGET_RTM
2890 static int __kmp_release_rtm_spin_lock(kmp_rtm_spin_lock_t *lck,
2891                                        kmp_int32 gtid) {
2892   if (KMP_ATOMIC_LD_RLX(&lck->lk.poll) == KMP_LOCK_FREE(rtm_spin)) {
2893     // Releasing from speculation
2894     _xend();
2895   } else {
2896     // Releasing from a real lock
2897     KMP_FSYNC_RELEASING(lck);
2898     KMP_ATOMIC_ST_REL(&lck->lk.poll, KMP_LOCK_FREE(rtm_spin));
2899   }
2900   return KMP_LOCK_RELEASED;
2901 }
2902 
2903 static int __kmp_release_rtm_spin_lock_with_checks(kmp_rtm_spin_lock_t *lck,
2904                                                    kmp_int32 gtid) {
2905   return __kmp_release_rtm_spin_lock(lck, gtid);
2906 }
2907 
2908 KMP_ATTRIBUTE_TARGET_RTM
2909 static int __kmp_test_rtm_spin_lock(kmp_rtm_spin_lock_t *lck, kmp_int32 gtid) {
2910   unsigned retries = 3, status;
2911   kmp_int32 lock_free = KMP_LOCK_FREE(rtm_spin);
2912   kmp_int32 lock_busy = KMP_LOCK_BUSY(1, rtm_spin);
2913   do {
2914     status = _xbegin();
2915     if (status == _XBEGIN_STARTED &&
2916         KMP_ATOMIC_LD_RLX(&lck->lk.poll) == lock_free) {
2917       return TRUE;
2918     }
2919     if (!(status & _XABORT_RETRY))
2920       break;
2921   } while (retries--);
2922 
2923   if (KMP_ATOMIC_LD_RLX(&lck->lk.poll) == lock_free &&
2924       __kmp_atomic_compare_store_acq(&lck->lk.poll, lock_free, lock_busy)) {
2925     KMP_FSYNC_ACQUIRED(lck);
2926     return TRUE;
2927   }
2928   return FALSE;
2929 }
2930 
2931 static int __kmp_test_rtm_spin_lock_with_checks(kmp_rtm_spin_lock_t *lck,
2932                                                 kmp_int32 gtid) {
2933   return __kmp_test_rtm_spin_lock(lck, gtid);
2934 }
2935 
2936 #endif // KMP_USE_TSX
2937 
2938 // Entry functions for indirect locks (first element of direct lock jump tables)
2939 static void __kmp_init_indirect_lock(kmp_dyna_lock_t *l,
2940                                      kmp_dyna_lockseq_t tag);
2941 static void __kmp_destroy_indirect_lock(kmp_dyna_lock_t *lock);
2942 static int __kmp_set_indirect_lock(kmp_dyna_lock_t *lock, kmp_int32);
2943 static int __kmp_unset_indirect_lock(kmp_dyna_lock_t *lock, kmp_int32);
2944 static int __kmp_test_indirect_lock(kmp_dyna_lock_t *lock, kmp_int32);
2945 static int __kmp_set_indirect_lock_with_checks(kmp_dyna_lock_t *lock,
2946                                                kmp_int32);
2947 static int __kmp_unset_indirect_lock_with_checks(kmp_dyna_lock_t *lock,
2948                                                  kmp_int32);
2949 static int __kmp_test_indirect_lock_with_checks(kmp_dyna_lock_t *lock,
2950                                                 kmp_int32);
2951 
2952 // Lock function definitions for the union parameter type
2953 #define KMP_FOREACH_LOCK_KIND(m, a) m(ticket, a) m(queuing, a) m(drdpa, a)
2954 
2955 #define expand1(lk, op)                                                        \
2956   static void __kmp_##op##_##lk##_##lock(kmp_user_lock_p lock) {               \
2957     __kmp_##op##_##lk##_##lock(&lock->lk);                                     \
2958   }
2959 #define expand2(lk, op)                                                        \
2960   static int __kmp_##op##_##lk##_##lock(kmp_user_lock_p lock,                  \
2961                                         kmp_int32 gtid) {                      \
2962     return __kmp_##op##_##lk##_##lock(&lock->lk, gtid);                        \
2963   }
2964 #define expand3(lk, op)                                                        \
2965   static void __kmp_set_##lk##_##lock_flags(kmp_user_lock_p lock,              \
2966                                             kmp_lock_flags_t flags) {          \
2967     __kmp_set_##lk##_lock_flags(&lock->lk, flags);                             \
2968   }
2969 #define expand4(lk, op)                                                        \
2970   static void __kmp_set_##lk##_##lock_location(kmp_user_lock_p lock,           \
2971                                                const ident_t *loc) {           \
2972     __kmp_set_##lk##_lock_location(&lock->lk, loc);                            \
2973   }
2974 
2975 KMP_FOREACH_LOCK_KIND(expand1, init)
2976 KMP_FOREACH_LOCK_KIND(expand1, init_nested)
2977 KMP_FOREACH_LOCK_KIND(expand1, destroy)
2978 KMP_FOREACH_LOCK_KIND(expand1, destroy_nested)
2979 KMP_FOREACH_LOCK_KIND(expand2, acquire)
2980 KMP_FOREACH_LOCK_KIND(expand2, acquire_nested)
2981 KMP_FOREACH_LOCK_KIND(expand2, release)
2982 KMP_FOREACH_LOCK_KIND(expand2, release_nested)
2983 KMP_FOREACH_LOCK_KIND(expand2, test)
2984 KMP_FOREACH_LOCK_KIND(expand2, test_nested)
2985 KMP_FOREACH_LOCK_KIND(expand3, )
2986 KMP_FOREACH_LOCK_KIND(expand4, )
2987 
2988 #undef expand1
2989 #undef expand2
2990 #undef expand3
2991 #undef expand4
2992 
2993 // Jump tables for the indirect lock functions
2994 // Only fill in the odd entries, that avoids the need to shift out the low bit
2995 
2996 // init functions
2997 #define expand(l, op) 0, __kmp_init_direct_lock,
2998 void (*__kmp_direct_init[])(kmp_dyna_lock_t *, kmp_dyna_lockseq_t) = {
2999     __kmp_init_indirect_lock, 0, KMP_FOREACH_D_LOCK(expand, init)};
3000 #undef expand
3001 
3002 // destroy functions
3003 #define expand(l, op) 0, (void (*)(kmp_dyna_lock_t *))__kmp_##op##_##l##_lock,
3004 static void (*direct_destroy[])(kmp_dyna_lock_t *) = {
3005     __kmp_destroy_indirect_lock, 0, KMP_FOREACH_D_LOCK(expand, destroy)};
3006 #undef expand
3007 #define expand(l, op)                                                          \
3008   0, (void (*)(kmp_dyna_lock_t *))__kmp_destroy_##l##_lock_with_checks,
3009 static void (*direct_destroy_check[])(kmp_dyna_lock_t *) = {
3010     __kmp_destroy_indirect_lock, 0, KMP_FOREACH_D_LOCK(expand, destroy)};
3011 #undef expand
3012 
3013 // set/acquire functions
3014 #define expand(l, op)                                                          \
3015   0, (int (*)(kmp_dyna_lock_t *, kmp_int32))__kmp_##op##_##l##_lock,
3016 static int (*direct_set[])(kmp_dyna_lock_t *, kmp_int32) = {
3017     __kmp_set_indirect_lock, 0, KMP_FOREACH_D_LOCK(expand, acquire)};
3018 #undef expand
3019 #define expand(l, op)                                                          \
3020   0, (int (*)(kmp_dyna_lock_t *, kmp_int32))__kmp_##op##_##l##_lock_with_checks,
3021 static int (*direct_set_check[])(kmp_dyna_lock_t *, kmp_int32) = {
3022     __kmp_set_indirect_lock_with_checks, 0,
3023     KMP_FOREACH_D_LOCK(expand, acquire)};
3024 #undef expand
3025 
3026 // unset/release and test functions
3027 #define expand(l, op)                                                          \
3028   0, (int (*)(kmp_dyna_lock_t *, kmp_int32))__kmp_##op##_##l##_lock,
3029 static int (*direct_unset[])(kmp_dyna_lock_t *, kmp_int32) = {
3030     __kmp_unset_indirect_lock, 0, KMP_FOREACH_D_LOCK(expand, release)};
3031 static int (*direct_test[])(kmp_dyna_lock_t *, kmp_int32) = {
3032     __kmp_test_indirect_lock, 0, KMP_FOREACH_D_LOCK(expand, test)};
3033 #undef expand
3034 #define expand(l, op)                                                          \
3035   0, (int (*)(kmp_dyna_lock_t *, kmp_int32))__kmp_##op##_##l##_lock_with_checks,
3036 static int (*direct_unset_check[])(kmp_dyna_lock_t *, kmp_int32) = {
3037     __kmp_unset_indirect_lock_with_checks, 0,
3038     KMP_FOREACH_D_LOCK(expand, release)};
3039 static int (*direct_test_check[])(kmp_dyna_lock_t *, kmp_int32) = {
3040     __kmp_test_indirect_lock_with_checks, 0, KMP_FOREACH_D_LOCK(expand, test)};
3041 #undef expand
3042 
3043 // Exposes only one set of jump tables (*lock or *lock_with_checks).
3044 void (**__kmp_direct_destroy)(kmp_dyna_lock_t *) = 0;
3045 int (**__kmp_direct_set)(kmp_dyna_lock_t *, kmp_int32) = 0;
3046 int (**__kmp_direct_unset)(kmp_dyna_lock_t *, kmp_int32) = 0;
3047 int (**__kmp_direct_test)(kmp_dyna_lock_t *, kmp_int32) = 0;
3048 
3049 // Jump tables for the indirect lock functions
3050 #define expand(l, op) (void (*)(kmp_user_lock_p)) __kmp_##op##_##l##_##lock,
3051 void (*__kmp_indirect_init[])(kmp_user_lock_p) = {
3052     KMP_FOREACH_I_LOCK(expand, init)};
3053 #undef expand
3054 
3055 #define expand(l, op) (void (*)(kmp_user_lock_p)) __kmp_##op##_##l##_##lock,
3056 static void (*indirect_destroy[])(kmp_user_lock_p) = {
3057     KMP_FOREACH_I_LOCK(expand, destroy)};
3058 #undef expand
3059 #define expand(l, op)                                                          \
3060   (void (*)(kmp_user_lock_p)) __kmp_##op##_##l##_##lock_with_checks,
3061 static void (*indirect_destroy_check[])(kmp_user_lock_p) = {
3062     KMP_FOREACH_I_LOCK(expand, destroy)};
3063 #undef expand
3064 
3065 // set/acquire functions
3066 #define expand(l, op)                                                          \
3067   (int (*)(kmp_user_lock_p, kmp_int32)) __kmp_##op##_##l##_##lock,
3068 static int (*indirect_set[])(kmp_user_lock_p,
3069                              kmp_int32) = {KMP_FOREACH_I_LOCK(expand, acquire)};
3070 #undef expand
3071 #define expand(l, op)                                                          \
3072   (int (*)(kmp_user_lock_p, kmp_int32)) __kmp_##op##_##l##_##lock_with_checks,
3073 static int (*indirect_set_check[])(kmp_user_lock_p, kmp_int32) = {
3074     KMP_FOREACH_I_LOCK(expand, acquire)};
3075 #undef expand
3076 
3077 // unset/release and test functions
3078 #define expand(l, op)                                                          \
3079   (int (*)(kmp_user_lock_p, kmp_int32)) __kmp_##op##_##l##_##lock,
3080 static int (*indirect_unset[])(kmp_user_lock_p, kmp_int32) = {
3081     KMP_FOREACH_I_LOCK(expand, release)};
3082 static int (*indirect_test[])(kmp_user_lock_p,
3083                               kmp_int32) = {KMP_FOREACH_I_LOCK(expand, test)};
3084 #undef expand
3085 #define expand(l, op)                                                          \
3086   (int (*)(kmp_user_lock_p, kmp_int32)) __kmp_##op##_##l##_##lock_with_checks,
3087 static int (*indirect_unset_check[])(kmp_user_lock_p, kmp_int32) = {
3088     KMP_FOREACH_I_LOCK(expand, release)};
3089 static int (*indirect_test_check[])(kmp_user_lock_p, kmp_int32) = {
3090     KMP_FOREACH_I_LOCK(expand, test)};
3091 #undef expand
3092 
3093 // Exposes only one jump tables (*lock or *lock_with_checks).
3094 void (**__kmp_indirect_destroy)(kmp_user_lock_p) = 0;
3095 int (**__kmp_indirect_set)(kmp_user_lock_p, kmp_int32) = 0;
3096 int (**__kmp_indirect_unset)(kmp_user_lock_p, kmp_int32) = 0;
3097 int (**__kmp_indirect_test)(kmp_user_lock_p, kmp_int32) = 0;
3098 
3099 // Lock index table.
3100 kmp_indirect_lock_table_t __kmp_i_lock_table;
3101 
3102 // Size of indirect locks.
3103 static kmp_uint32 __kmp_indirect_lock_size[KMP_NUM_I_LOCKS] = {0};
3104 
3105 // Jump tables for lock accessor/modifier.
3106 void (*__kmp_indirect_set_location[KMP_NUM_I_LOCKS])(kmp_user_lock_p,
3107                                                      const ident_t *) = {0};
3108 void (*__kmp_indirect_set_flags[KMP_NUM_I_LOCKS])(kmp_user_lock_p,
3109                                                   kmp_lock_flags_t) = {0};
3110 const ident_t *(*__kmp_indirect_get_location[KMP_NUM_I_LOCKS])(
3111     kmp_user_lock_p) = {0};
3112 kmp_lock_flags_t (*__kmp_indirect_get_flags[KMP_NUM_I_LOCKS])(
3113     kmp_user_lock_p) = {0};
3114 
3115 // Use different lock pools for different lock types.
3116 static kmp_indirect_lock_t *__kmp_indirect_lock_pool[KMP_NUM_I_LOCKS] = {0};
3117 
3118 // User lock allocator for dynamically dispatched indirect locks. Every entry of
3119 // the indirect lock table holds the address and type of the allocated indirect
3120 // lock (kmp_indirect_lock_t), and the size of the table doubles when it is
3121 // full. A destroyed indirect lock object is returned to the reusable pool of
3122 // locks, unique to each lock type.
3123 kmp_indirect_lock_t *__kmp_allocate_indirect_lock(void **user_lock,
3124                                                   kmp_int32 gtid,
3125                                                   kmp_indirect_locktag_t tag) {
3126   kmp_indirect_lock_t *lck;
3127   kmp_lock_index_t idx;
3128 
3129   __kmp_acquire_lock(&__kmp_global_lock, gtid);
3130 
3131   if (__kmp_indirect_lock_pool[tag] != NULL) {
3132     // Reuse the allocated and destroyed lock object
3133     lck = __kmp_indirect_lock_pool[tag];
3134     if (OMP_LOCK_T_SIZE < sizeof(void *))
3135       idx = lck->lock->pool.index;
3136     __kmp_indirect_lock_pool[tag] = (kmp_indirect_lock_t *)lck->lock->pool.next;
3137     KA_TRACE(20, ("__kmp_allocate_indirect_lock: reusing an existing lock %p\n",
3138                   lck));
3139   } else {
3140     idx = __kmp_i_lock_table.next;
3141     // Check capacity and double the size if it is full
3142     if (idx == __kmp_i_lock_table.size) {
3143       // Double up the space for block pointers
3144       int row = __kmp_i_lock_table.size / KMP_I_LOCK_CHUNK;
3145       kmp_indirect_lock_t **new_table = (kmp_indirect_lock_t **)__kmp_allocate(
3146           2 * row * sizeof(kmp_indirect_lock_t *));
3147       KMP_MEMCPY(new_table, __kmp_i_lock_table.table,
3148                  row * sizeof(kmp_indirect_lock_t *));
3149       kmp_indirect_lock_t **old_table = __kmp_i_lock_table.table;
3150       __kmp_i_lock_table.table = new_table;
3151       __kmp_free(old_table);
3152       // Allocate new objects in the new blocks
3153       for (int i = row; i < 2 * row; ++i)
3154         *(__kmp_i_lock_table.table + i) = (kmp_indirect_lock_t *)__kmp_allocate(
3155             KMP_I_LOCK_CHUNK * sizeof(kmp_indirect_lock_t));
3156       __kmp_i_lock_table.size = 2 * idx;
3157     }
3158     __kmp_i_lock_table.next++;
3159     lck = KMP_GET_I_LOCK(idx);
3160     // Allocate a new base lock object
3161     lck->lock = (kmp_user_lock_p)__kmp_allocate(__kmp_indirect_lock_size[tag]);
3162     KA_TRACE(20,
3163              ("__kmp_allocate_indirect_lock: allocated a new lock %p\n", lck));
3164   }
3165 
3166   __kmp_release_lock(&__kmp_global_lock, gtid);
3167 
3168   lck->type = tag;
3169 
3170   if (OMP_LOCK_T_SIZE < sizeof(void *)) {
3171     *((kmp_lock_index_t *)user_lock) = idx
3172                                        << 1; // indirect lock word must be even
3173   } else {
3174     *((kmp_indirect_lock_t **)user_lock) = lck;
3175   }
3176 
3177   return lck;
3178 }
3179 
3180 // User lock lookup for dynamically dispatched locks.
3181 static __forceinline kmp_indirect_lock_t *
3182 __kmp_lookup_indirect_lock(void **user_lock, const char *func) {
3183   if (__kmp_env_consistency_check) {
3184     kmp_indirect_lock_t *lck = NULL;
3185     if (user_lock == NULL) {
3186       KMP_FATAL(LockIsUninitialized, func);
3187     }
3188     if (OMP_LOCK_T_SIZE < sizeof(void *)) {
3189       kmp_lock_index_t idx = KMP_EXTRACT_I_INDEX(user_lock);
3190       if (idx >= __kmp_i_lock_table.size) {
3191         KMP_FATAL(LockIsUninitialized, func);
3192       }
3193       lck = KMP_GET_I_LOCK(idx);
3194     } else {
3195       lck = *((kmp_indirect_lock_t **)user_lock);
3196     }
3197     if (lck == NULL) {
3198       KMP_FATAL(LockIsUninitialized, func);
3199     }
3200     return lck;
3201   } else {
3202     if (OMP_LOCK_T_SIZE < sizeof(void *)) {
3203       return KMP_GET_I_LOCK(KMP_EXTRACT_I_INDEX(user_lock));
3204     } else {
3205       return *((kmp_indirect_lock_t **)user_lock);
3206     }
3207   }
3208 }
3209 
3210 static void __kmp_init_indirect_lock(kmp_dyna_lock_t *lock,
3211                                      kmp_dyna_lockseq_t seq) {
3212 #if KMP_USE_ADAPTIVE_LOCKS
3213   if (seq == lockseq_adaptive && !__kmp_cpuinfo.rtm) {
3214     KMP_WARNING(AdaptiveNotSupported, "kmp_lockseq_t", "adaptive");
3215     seq = lockseq_queuing;
3216   }
3217 #endif
3218 #if KMP_USE_TSX
3219   if (seq == lockseq_rtm_queuing && !__kmp_cpuinfo.rtm) {
3220     seq = lockseq_queuing;
3221   }
3222 #endif
3223   kmp_indirect_locktag_t tag = KMP_GET_I_TAG(seq);
3224   kmp_indirect_lock_t *l =
3225       __kmp_allocate_indirect_lock((void **)lock, __kmp_entry_gtid(), tag);
3226   KMP_I_LOCK_FUNC(l, init)(l->lock);
3227   KA_TRACE(
3228       20, ("__kmp_init_indirect_lock: initialized indirect lock with type#%d\n",
3229            seq));
3230 }
3231 
3232 static void __kmp_destroy_indirect_lock(kmp_dyna_lock_t *lock) {
3233   kmp_uint32 gtid = __kmp_entry_gtid();
3234   kmp_indirect_lock_t *l =
3235       __kmp_lookup_indirect_lock((void **)lock, "omp_destroy_lock");
3236   KMP_I_LOCK_FUNC(l, destroy)(l->lock);
3237   kmp_indirect_locktag_t tag = l->type;
3238 
3239   __kmp_acquire_lock(&__kmp_global_lock, gtid);
3240 
3241   // Use the base lock's space to keep the pool chain.
3242   l->lock->pool.next = (kmp_user_lock_p)__kmp_indirect_lock_pool[tag];
3243   if (OMP_LOCK_T_SIZE < sizeof(void *)) {
3244     l->lock->pool.index = KMP_EXTRACT_I_INDEX(lock);
3245   }
3246   __kmp_indirect_lock_pool[tag] = l;
3247 
3248   __kmp_release_lock(&__kmp_global_lock, gtid);
3249 }
3250 
3251 static int __kmp_set_indirect_lock(kmp_dyna_lock_t *lock, kmp_int32 gtid) {
3252   kmp_indirect_lock_t *l = KMP_LOOKUP_I_LOCK(lock);
3253   return KMP_I_LOCK_FUNC(l, set)(l->lock, gtid);
3254 }
3255 
3256 static int __kmp_unset_indirect_lock(kmp_dyna_lock_t *lock, kmp_int32 gtid) {
3257   kmp_indirect_lock_t *l = KMP_LOOKUP_I_LOCK(lock);
3258   return KMP_I_LOCK_FUNC(l, unset)(l->lock, gtid);
3259 }
3260 
3261 static int __kmp_test_indirect_lock(kmp_dyna_lock_t *lock, kmp_int32 gtid) {
3262   kmp_indirect_lock_t *l = KMP_LOOKUP_I_LOCK(lock);
3263   return KMP_I_LOCK_FUNC(l, test)(l->lock, gtid);
3264 }
3265 
3266 static int __kmp_set_indirect_lock_with_checks(kmp_dyna_lock_t *lock,
3267                                                kmp_int32 gtid) {
3268   kmp_indirect_lock_t *l =
3269       __kmp_lookup_indirect_lock((void **)lock, "omp_set_lock");
3270   return KMP_I_LOCK_FUNC(l, set)(l->lock, gtid);
3271 }
3272 
3273 static int __kmp_unset_indirect_lock_with_checks(kmp_dyna_lock_t *lock,
3274                                                  kmp_int32 gtid) {
3275   kmp_indirect_lock_t *l =
3276       __kmp_lookup_indirect_lock((void **)lock, "omp_unset_lock");
3277   return KMP_I_LOCK_FUNC(l, unset)(l->lock, gtid);
3278 }
3279 
3280 static int __kmp_test_indirect_lock_with_checks(kmp_dyna_lock_t *lock,
3281                                                 kmp_int32 gtid) {
3282   kmp_indirect_lock_t *l =
3283       __kmp_lookup_indirect_lock((void **)lock, "omp_test_lock");
3284   return KMP_I_LOCK_FUNC(l, test)(l->lock, gtid);
3285 }
3286 
3287 kmp_dyna_lockseq_t __kmp_user_lock_seq = lockseq_queuing;
3288 
3289 // This is used only in kmp_error.cpp when consistency checking is on.
3290 kmp_int32 __kmp_get_user_lock_owner(kmp_user_lock_p lck, kmp_uint32 seq) {
3291   switch (seq) {
3292   case lockseq_tas:
3293   case lockseq_nested_tas:
3294     return __kmp_get_tas_lock_owner((kmp_tas_lock_t *)lck);
3295 #if KMP_USE_FUTEX
3296   case lockseq_futex:
3297   case lockseq_nested_futex:
3298     return __kmp_get_futex_lock_owner((kmp_futex_lock_t *)lck);
3299 #endif
3300   case lockseq_ticket:
3301   case lockseq_nested_ticket:
3302     return __kmp_get_ticket_lock_owner((kmp_ticket_lock_t *)lck);
3303   case lockseq_queuing:
3304   case lockseq_nested_queuing:
3305 #if KMP_USE_ADAPTIVE_LOCKS
3306   case lockseq_adaptive:
3307 #endif
3308     return __kmp_get_queuing_lock_owner((kmp_queuing_lock_t *)lck);
3309   case lockseq_drdpa:
3310   case lockseq_nested_drdpa:
3311     return __kmp_get_drdpa_lock_owner((kmp_drdpa_lock_t *)lck);
3312   default:
3313     return 0;
3314   }
3315 }
3316 
3317 // Initializes data for dynamic user locks.
3318 void __kmp_init_dynamic_user_locks() {
3319   // Initialize jump table for the lock functions
3320   if (__kmp_env_consistency_check) {
3321     __kmp_direct_set = direct_set_check;
3322     __kmp_direct_unset = direct_unset_check;
3323     __kmp_direct_test = direct_test_check;
3324     __kmp_direct_destroy = direct_destroy_check;
3325     __kmp_indirect_set = indirect_set_check;
3326     __kmp_indirect_unset = indirect_unset_check;
3327     __kmp_indirect_test = indirect_test_check;
3328     __kmp_indirect_destroy = indirect_destroy_check;
3329   } else {
3330     __kmp_direct_set = direct_set;
3331     __kmp_direct_unset = direct_unset;
3332     __kmp_direct_test = direct_test;
3333     __kmp_direct_destroy = direct_destroy;
3334     __kmp_indirect_set = indirect_set;
3335     __kmp_indirect_unset = indirect_unset;
3336     __kmp_indirect_test = indirect_test;
3337     __kmp_indirect_destroy = indirect_destroy;
3338   }
3339   // If the user locks have already been initialized, then return. Allow the
3340   // switch between different KMP_CONSISTENCY_CHECK values, but do not allocate
3341   // new lock tables if they have already been allocated.
3342   if (__kmp_init_user_locks)
3343     return;
3344 
3345   // Initialize lock index table
3346   __kmp_i_lock_table.size = KMP_I_LOCK_CHUNK;
3347   __kmp_i_lock_table.table =
3348       (kmp_indirect_lock_t **)__kmp_allocate(sizeof(kmp_indirect_lock_t *));
3349   *(__kmp_i_lock_table.table) = (kmp_indirect_lock_t *)__kmp_allocate(
3350       KMP_I_LOCK_CHUNK * sizeof(kmp_indirect_lock_t));
3351   __kmp_i_lock_table.next = 0;
3352 
3353   // Indirect lock size
3354   __kmp_indirect_lock_size[locktag_ticket] = sizeof(kmp_ticket_lock_t);
3355   __kmp_indirect_lock_size[locktag_queuing] = sizeof(kmp_queuing_lock_t);
3356 #if KMP_USE_ADAPTIVE_LOCKS
3357   __kmp_indirect_lock_size[locktag_adaptive] = sizeof(kmp_adaptive_lock_t);
3358 #endif
3359   __kmp_indirect_lock_size[locktag_drdpa] = sizeof(kmp_drdpa_lock_t);
3360 #if KMP_USE_TSX
3361   __kmp_indirect_lock_size[locktag_rtm_queuing] = sizeof(kmp_queuing_lock_t);
3362 #endif
3363   __kmp_indirect_lock_size[locktag_nested_tas] = sizeof(kmp_tas_lock_t);
3364 #if KMP_USE_FUTEX
3365   __kmp_indirect_lock_size[locktag_nested_futex] = sizeof(kmp_futex_lock_t);
3366 #endif
3367   __kmp_indirect_lock_size[locktag_nested_ticket] = sizeof(kmp_ticket_lock_t);
3368   __kmp_indirect_lock_size[locktag_nested_queuing] = sizeof(kmp_queuing_lock_t);
3369   __kmp_indirect_lock_size[locktag_nested_drdpa] = sizeof(kmp_drdpa_lock_t);
3370 
3371 // Initialize lock accessor/modifier
3372 #define fill_jumps(table, expand, sep)                                         \
3373   {                                                                            \
3374     table[locktag##sep##ticket] = expand(ticket);                              \
3375     table[locktag##sep##queuing] = expand(queuing);                            \
3376     table[locktag##sep##drdpa] = expand(drdpa);                                \
3377   }
3378 
3379 #if KMP_USE_ADAPTIVE_LOCKS
3380 #define fill_table(table, expand)                                              \
3381   {                                                                            \
3382     fill_jumps(table, expand, _);                                              \
3383     table[locktag_adaptive] = expand(queuing);                                 \
3384     fill_jumps(table, expand, _nested_);                                       \
3385   }
3386 #else
3387 #define fill_table(table, expand)                                              \
3388   {                                                                            \
3389     fill_jumps(table, expand, _);                                              \
3390     fill_jumps(table, expand, _nested_);                                       \
3391   }
3392 #endif // KMP_USE_ADAPTIVE_LOCKS
3393 
3394 #define expand(l)                                                              \
3395   (void (*)(kmp_user_lock_p, const ident_t *)) __kmp_set_##l##_lock_location
3396   fill_table(__kmp_indirect_set_location, expand);
3397 #undef expand
3398 #define expand(l)                                                              \
3399   (void (*)(kmp_user_lock_p, kmp_lock_flags_t)) __kmp_set_##l##_lock_flags
3400   fill_table(__kmp_indirect_set_flags, expand);
3401 #undef expand
3402 #define expand(l)                                                              \
3403   (const ident_t *(*)(kmp_user_lock_p)) __kmp_get_##l##_lock_location
3404   fill_table(__kmp_indirect_get_location, expand);
3405 #undef expand
3406 #define expand(l)                                                              \
3407   (kmp_lock_flags_t(*)(kmp_user_lock_p)) __kmp_get_##l##_lock_flags
3408   fill_table(__kmp_indirect_get_flags, expand);
3409 #undef expand
3410 
3411   __kmp_init_user_locks = TRUE;
3412 }
3413 
3414 // Clean up the lock table.
3415 void __kmp_cleanup_indirect_user_locks() {
3416   kmp_lock_index_t i;
3417   int k;
3418 
3419   // Clean up locks in the pools first (they were already destroyed before going
3420   // into the pools).
3421   for (k = 0; k < KMP_NUM_I_LOCKS; ++k) {
3422     kmp_indirect_lock_t *l = __kmp_indirect_lock_pool[k];
3423     while (l != NULL) {
3424       kmp_indirect_lock_t *ll = l;
3425       l = (kmp_indirect_lock_t *)l->lock->pool.next;
3426       KA_TRACE(20, ("__kmp_cleanup_indirect_user_locks: freeing %p from pool\n",
3427                     ll));
3428       __kmp_free(ll->lock);
3429       ll->lock = NULL;
3430     }
3431     __kmp_indirect_lock_pool[k] = NULL;
3432   }
3433   // Clean up the remaining undestroyed locks.
3434   for (i = 0; i < __kmp_i_lock_table.next; i++) {
3435     kmp_indirect_lock_t *l = KMP_GET_I_LOCK(i);
3436     if (l->lock != NULL) {
3437       // Locks not destroyed explicitly need to be destroyed here.
3438       KMP_I_LOCK_FUNC(l, destroy)(l->lock);
3439       KA_TRACE(
3440           20,
3441           ("__kmp_cleanup_indirect_user_locks: destroy/freeing %p from table\n",
3442            l));
3443       __kmp_free(l->lock);
3444     }
3445   }
3446   // Free the table
3447   for (i = 0; i < __kmp_i_lock_table.size / KMP_I_LOCK_CHUNK; i++)
3448     __kmp_free(__kmp_i_lock_table.table[i]);
3449   __kmp_free(__kmp_i_lock_table.table);
3450 
3451   __kmp_init_user_locks = FALSE;
3452 }
3453 
3454 enum kmp_lock_kind __kmp_user_lock_kind = lk_default;
3455 int __kmp_num_locks_in_block = 1; // FIXME - tune this value
3456 
3457 #else // KMP_USE_DYNAMIC_LOCK
3458 
3459 static void __kmp_init_tas_lock_with_checks(kmp_tas_lock_t *lck) {
3460   __kmp_init_tas_lock(lck);
3461 }
3462 
3463 static void __kmp_init_nested_tas_lock_with_checks(kmp_tas_lock_t *lck) {
3464   __kmp_init_nested_tas_lock(lck);
3465 }
3466 
3467 #if KMP_USE_FUTEX
3468 static void __kmp_init_futex_lock_with_checks(kmp_futex_lock_t *lck) {
3469   __kmp_init_futex_lock(lck);
3470 }
3471 
3472 static void __kmp_init_nested_futex_lock_with_checks(kmp_futex_lock_t *lck) {
3473   __kmp_init_nested_futex_lock(lck);
3474 }
3475 #endif
3476 
3477 static int __kmp_is_ticket_lock_initialized(kmp_ticket_lock_t *lck) {
3478   return lck == lck->lk.self;
3479 }
3480 
3481 static void __kmp_init_ticket_lock_with_checks(kmp_ticket_lock_t *lck) {
3482   __kmp_init_ticket_lock(lck);
3483 }
3484 
3485 static void __kmp_init_nested_ticket_lock_with_checks(kmp_ticket_lock_t *lck) {
3486   __kmp_init_nested_ticket_lock(lck);
3487 }
3488 
3489 static int __kmp_is_queuing_lock_initialized(kmp_queuing_lock_t *lck) {
3490   return lck == lck->lk.initialized;
3491 }
3492 
3493 static void __kmp_init_queuing_lock_with_checks(kmp_queuing_lock_t *lck) {
3494   __kmp_init_queuing_lock(lck);
3495 }
3496 
3497 static void
3498 __kmp_init_nested_queuing_lock_with_checks(kmp_queuing_lock_t *lck) {
3499   __kmp_init_nested_queuing_lock(lck);
3500 }
3501 
3502 #if KMP_USE_ADAPTIVE_LOCKS
3503 static void __kmp_init_adaptive_lock_with_checks(kmp_adaptive_lock_t *lck) {
3504   __kmp_init_adaptive_lock(lck);
3505 }
3506 #endif
3507 
3508 static int __kmp_is_drdpa_lock_initialized(kmp_drdpa_lock_t *lck) {
3509   return lck == lck->lk.initialized;
3510 }
3511 
3512 static void __kmp_init_drdpa_lock_with_checks(kmp_drdpa_lock_t *lck) {
3513   __kmp_init_drdpa_lock(lck);
3514 }
3515 
3516 static void __kmp_init_nested_drdpa_lock_with_checks(kmp_drdpa_lock_t *lck) {
3517   __kmp_init_nested_drdpa_lock(lck);
3518 }
3519 
3520 /* user locks
3521  * They are implemented as a table of function pointers which are set to the
3522  * lock functions of the appropriate kind, once that has been determined. */
3523 
3524 enum kmp_lock_kind __kmp_user_lock_kind = lk_default;
3525 
3526 size_t __kmp_base_user_lock_size = 0;
3527 size_t __kmp_user_lock_size = 0;
3528 
3529 kmp_int32 (*__kmp_get_user_lock_owner_)(kmp_user_lock_p lck) = NULL;
3530 int (*__kmp_acquire_user_lock_with_checks_)(kmp_user_lock_p lck,
3531                                             kmp_int32 gtid) = NULL;
3532 
3533 int (*__kmp_test_user_lock_with_checks_)(kmp_user_lock_p lck,
3534                                          kmp_int32 gtid) = NULL;
3535 int (*__kmp_release_user_lock_with_checks_)(kmp_user_lock_p lck,
3536                                             kmp_int32 gtid) = NULL;
3537 void (*__kmp_init_user_lock_with_checks_)(kmp_user_lock_p lck) = NULL;
3538 void (*__kmp_destroy_user_lock_)(kmp_user_lock_p lck) = NULL;
3539 void (*__kmp_destroy_user_lock_with_checks_)(kmp_user_lock_p lck) = NULL;
3540 int (*__kmp_acquire_nested_user_lock_with_checks_)(kmp_user_lock_p lck,
3541                                                    kmp_int32 gtid) = NULL;
3542 
3543 int (*__kmp_test_nested_user_lock_with_checks_)(kmp_user_lock_p lck,
3544                                                 kmp_int32 gtid) = NULL;
3545 int (*__kmp_release_nested_user_lock_with_checks_)(kmp_user_lock_p lck,
3546                                                    kmp_int32 gtid) = NULL;
3547 void (*__kmp_init_nested_user_lock_with_checks_)(kmp_user_lock_p lck) = NULL;
3548 void (*__kmp_destroy_nested_user_lock_with_checks_)(kmp_user_lock_p lck) = NULL;
3549 
3550 int (*__kmp_is_user_lock_initialized_)(kmp_user_lock_p lck) = NULL;
3551 const ident_t *(*__kmp_get_user_lock_location_)(kmp_user_lock_p lck) = NULL;
3552 void (*__kmp_set_user_lock_location_)(kmp_user_lock_p lck,
3553                                       const ident_t *loc) = NULL;
3554 kmp_lock_flags_t (*__kmp_get_user_lock_flags_)(kmp_user_lock_p lck) = NULL;
3555 void (*__kmp_set_user_lock_flags_)(kmp_user_lock_p lck,
3556                                    kmp_lock_flags_t flags) = NULL;
3557 
3558 void __kmp_set_user_lock_vptrs(kmp_lock_kind_t user_lock_kind) {
3559   switch (user_lock_kind) {
3560   case lk_default:
3561   default:
3562     KMP_ASSERT(0);
3563 
3564   case lk_tas: {
3565     __kmp_base_user_lock_size = sizeof(kmp_base_tas_lock_t);
3566     __kmp_user_lock_size = sizeof(kmp_tas_lock_t);
3567 
3568     __kmp_get_user_lock_owner_ =
3569         (kmp_int32(*)(kmp_user_lock_p))(&__kmp_get_tas_lock_owner);
3570 
3571     if (__kmp_env_consistency_check) {
3572       KMP_BIND_USER_LOCK_WITH_CHECKS(tas);
3573       KMP_BIND_NESTED_USER_LOCK_WITH_CHECKS(tas);
3574     } else {
3575       KMP_BIND_USER_LOCK(tas);
3576       KMP_BIND_NESTED_USER_LOCK(tas);
3577     }
3578 
3579     __kmp_destroy_user_lock_ =
3580         (void (*)(kmp_user_lock_p))(&__kmp_destroy_tas_lock);
3581 
3582     __kmp_is_user_lock_initialized_ = (int (*)(kmp_user_lock_p))NULL;
3583 
3584     __kmp_get_user_lock_location_ = (const ident_t *(*)(kmp_user_lock_p))NULL;
3585 
3586     __kmp_set_user_lock_location_ =
3587         (void (*)(kmp_user_lock_p, const ident_t *))NULL;
3588 
3589     __kmp_get_user_lock_flags_ = (kmp_lock_flags_t(*)(kmp_user_lock_p))NULL;
3590 
3591     __kmp_set_user_lock_flags_ =
3592         (void (*)(kmp_user_lock_p, kmp_lock_flags_t))NULL;
3593   } break;
3594 
3595 #if KMP_USE_FUTEX
3596 
3597   case lk_futex: {
3598     __kmp_base_user_lock_size = sizeof(kmp_base_futex_lock_t);
3599     __kmp_user_lock_size = sizeof(kmp_futex_lock_t);
3600 
3601     __kmp_get_user_lock_owner_ =
3602         (kmp_int32(*)(kmp_user_lock_p))(&__kmp_get_futex_lock_owner);
3603 
3604     if (__kmp_env_consistency_check) {
3605       KMP_BIND_USER_LOCK_WITH_CHECKS(futex);
3606       KMP_BIND_NESTED_USER_LOCK_WITH_CHECKS(futex);
3607     } else {
3608       KMP_BIND_USER_LOCK(futex);
3609       KMP_BIND_NESTED_USER_LOCK(futex);
3610     }
3611 
3612     __kmp_destroy_user_lock_ =
3613         (void (*)(kmp_user_lock_p))(&__kmp_destroy_futex_lock);
3614 
3615     __kmp_is_user_lock_initialized_ = (int (*)(kmp_user_lock_p))NULL;
3616 
3617     __kmp_get_user_lock_location_ = (const ident_t *(*)(kmp_user_lock_p))NULL;
3618 
3619     __kmp_set_user_lock_location_ =
3620         (void (*)(kmp_user_lock_p, const ident_t *))NULL;
3621 
3622     __kmp_get_user_lock_flags_ = (kmp_lock_flags_t(*)(kmp_user_lock_p))NULL;
3623 
3624     __kmp_set_user_lock_flags_ =
3625         (void (*)(kmp_user_lock_p, kmp_lock_flags_t))NULL;
3626   } break;
3627 
3628 #endif // KMP_USE_FUTEX
3629 
3630   case lk_ticket: {
3631     __kmp_base_user_lock_size = sizeof(kmp_base_ticket_lock_t);
3632     __kmp_user_lock_size = sizeof(kmp_ticket_lock_t);
3633 
3634     __kmp_get_user_lock_owner_ =
3635         (kmp_int32(*)(kmp_user_lock_p))(&__kmp_get_ticket_lock_owner);
3636 
3637     if (__kmp_env_consistency_check) {
3638       KMP_BIND_USER_LOCK_WITH_CHECKS(ticket);
3639       KMP_BIND_NESTED_USER_LOCK_WITH_CHECKS(ticket);
3640     } else {
3641       KMP_BIND_USER_LOCK(ticket);
3642       KMP_BIND_NESTED_USER_LOCK(ticket);
3643     }
3644 
3645     __kmp_destroy_user_lock_ =
3646         (void (*)(kmp_user_lock_p))(&__kmp_destroy_ticket_lock);
3647 
3648     __kmp_is_user_lock_initialized_ =
3649         (int (*)(kmp_user_lock_p))(&__kmp_is_ticket_lock_initialized);
3650 
3651     __kmp_get_user_lock_location_ =
3652         (const ident_t *(*)(kmp_user_lock_p))(&__kmp_get_ticket_lock_location);
3653 
3654     __kmp_set_user_lock_location_ = (void (*)(
3655         kmp_user_lock_p, const ident_t *))(&__kmp_set_ticket_lock_location);
3656 
3657     __kmp_get_user_lock_flags_ =
3658         (kmp_lock_flags_t(*)(kmp_user_lock_p))(&__kmp_get_ticket_lock_flags);
3659 
3660     __kmp_set_user_lock_flags_ = (void (*)(kmp_user_lock_p, kmp_lock_flags_t))(
3661         &__kmp_set_ticket_lock_flags);
3662   } break;
3663 
3664   case lk_queuing: {
3665     __kmp_base_user_lock_size = sizeof(kmp_base_queuing_lock_t);
3666     __kmp_user_lock_size = sizeof(kmp_queuing_lock_t);
3667 
3668     __kmp_get_user_lock_owner_ =
3669         (kmp_int32(*)(kmp_user_lock_p))(&__kmp_get_queuing_lock_owner);
3670 
3671     if (__kmp_env_consistency_check) {
3672       KMP_BIND_USER_LOCK_WITH_CHECKS(queuing);
3673       KMP_BIND_NESTED_USER_LOCK_WITH_CHECKS(queuing);
3674     } else {
3675       KMP_BIND_USER_LOCK(queuing);
3676       KMP_BIND_NESTED_USER_LOCK(queuing);
3677     }
3678 
3679     __kmp_destroy_user_lock_ =
3680         (void (*)(kmp_user_lock_p))(&__kmp_destroy_queuing_lock);
3681 
3682     __kmp_is_user_lock_initialized_ =
3683         (int (*)(kmp_user_lock_p))(&__kmp_is_queuing_lock_initialized);
3684 
3685     __kmp_get_user_lock_location_ =
3686         (const ident_t *(*)(kmp_user_lock_p))(&__kmp_get_queuing_lock_location);
3687 
3688     __kmp_set_user_lock_location_ = (void (*)(
3689         kmp_user_lock_p, const ident_t *))(&__kmp_set_queuing_lock_location);
3690 
3691     __kmp_get_user_lock_flags_ =
3692         (kmp_lock_flags_t(*)(kmp_user_lock_p))(&__kmp_get_queuing_lock_flags);
3693 
3694     __kmp_set_user_lock_flags_ = (void (*)(kmp_user_lock_p, kmp_lock_flags_t))(
3695         &__kmp_set_queuing_lock_flags);
3696   } break;
3697 
3698 #if KMP_USE_ADAPTIVE_LOCKS
3699   case lk_adaptive: {
3700     __kmp_base_user_lock_size = sizeof(kmp_base_adaptive_lock_t);
3701     __kmp_user_lock_size = sizeof(kmp_adaptive_lock_t);
3702 
3703     __kmp_get_user_lock_owner_ =
3704         (kmp_int32(*)(kmp_user_lock_p))(&__kmp_get_queuing_lock_owner);
3705 
3706     if (__kmp_env_consistency_check) {
3707       KMP_BIND_USER_LOCK_WITH_CHECKS(adaptive);
3708     } else {
3709       KMP_BIND_USER_LOCK(adaptive);
3710     }
3711 
3712     __kmp_destroy_user_lock_ =
3713         (void (*)(kmp_user_lock_p))(&__kmp_destroy_adaptive_lock);
3714 
3715     __kmp_is_user_lock_initialized_ =
3716         (int (*)(kmp_user_lock_p))(&__kmp_is_queuing_lock_initialized);
3717 
3718     __kmp_get_user_lock_location_ =
3719         (const ident_t *(*)(kmp_user_lock_p))(&__kmp_get_queuing_lock_location);
3720 
3721     __kmp_set_user_lock_location_ = (void (*)(
3722         kmp_user_lock_p, const ident_t *))(&__kmp_set_queuing_lock_location);
3723 
3724     __kmp_get_user_lock_flags_ =
3725         (kmp_lock_flags_t(*)(kmp_user_lock_p))(&__kmp_get_queuing_lock_flags);
3726 
3727     __kmp_set_user_lock_flags_ = (void (*)(kmp_user_lock_p, kmp_lock_flags_t))(
3728         &__kmp_set_queuing_lock_flags);
3729 
3730   } break;
3731 #endif // KMP_USE_ADAPTIVE_LOCKS
3732 
3733   case lk_drdpa: {
3734     __kmp_base_user_lock_size = sizeof(kmp_base_drdpa_lock_t);
3735     __kmp_user_lock_size = sizeof(kmp_drdpa_lock_t);
3736 
3737     __kmp_get_user_lock_owner_ =
3738         (kmp_int32(*)(kmp_user_lock_p))(&__kmp_get_drdpa_lock_owner);
3739 
3740     if (__kmp_env_consistency_check) {
3741       KMP_BIND_USER_LOCK_WITH_CHECKS(drdpa);
3742       KMP_BIND_NESTED_USER_LOCK_WITH_CHECKS(drdpa);
3743     } else {
3744       KMP_BIND_USER_LOCK(drdpa);
3745       KMP_BIND_NESTED_USER_LOCK(drdpa);
3746     }
3747 
3748     __kmp_destroy_user_lock_ =
3749         (void (*)(kmp_user_lock_p))(&__kmp_destroy_drdpa_lock);
3750 
3751     __kmp_is_user_lock_initialized_ =
3752         (int (*)(kmp_user_lock_p))(&__kmp_is_drdpa_lock_initialized);
3753 
3754     __kmp_get_user_lock_location_ =
3755         (const ident_t *(*)(kmp_user_lock_p))(&__kmp_get_drdpa_lock_location);
3756 
3757     __kmp_set_user_lock_location_ = (void (*)(
3758         kmp_user_lock_p, const ident_t *))(&__kmp_set_drdpa_lock_location);
3759 
3760     __kmp_get_user_lock_flags_ =
3761         (kmp_lock_flags_t(*)(kmp_user_lock_p))(&__kmp_get_drdpa_lock_flags);
3762 
3763     __kmp_set_user_lock_flags_ = (void (*)(kmp_user_lock_p, kmp_lock_flags_t))(
3764         &__kmp_set_drdpa_lock_flags);
3765   } break;
3766   }
3767 }
3768 
3769 // ----------------------------------------------------------------------------
3770 // User lock table & lock allocation
3771 
3772 kmp_lock_table_t __kmp_user_lock_table = {1, 0, NULL};
3773 kmp_user_lock_p __kmp_lock_pool = NULL;
3774 
3775 // Lock block-allocation support.
3776 kmp_block_of_locks *__kmp_lock_blocks = NULL;
3777 int __kmp_num_locks_in_block = 1; // FIXME - tune this value
3778 
3779 static kmp_lock_index_t __kmp_lock_table_insert(kmp_user_lock_p lck) {
3780   // Assume that kmp_global_lock is held upon entry/exit.
3781   kmp_lock_index_t index;
3782   if (__kmp_user_lock_table.used >= __kmp_user_lock_table.allocated) {
3783     kmp_lock_index_t size;
3784     kmp_user_lock_p *table;
3785     // Reallocate lock table.
3786     if (__kmp_user_lock_table.allocated == 0) {
3787       size = 1024;
3788     } else {
3789       size = __kmp_user_lock_table.allocated * 2;
3790     }
3791     table = (kmp_user_lock_p *)__kmp_allocate(sizeof(kmp_user_lock_p) * size);
3792     KMP_MEMCPY(table + 1, __kmp_user_lock_table.table + 1,
3793                sizeof(kmp_user_lock_p) * (__kmp_user_lock_table.used - 1));
3794     table[0] = (kmp_user_lock_p)__kmp_user_lock_table.table;
3795     // We cannot free the previous table now, since it may be in use by other
3796     // threads. So save the pointer to the previous table in in the first
3797     // element of the new table. All the tables will be organized into a list,
3798     // and could be freed when library shutting down.
3799     __kmp_user_lock_table.table = table;
3800     __kmp_user_lock_table.allocated = size;
3801   }
3802   KMP_DEBUG_ASSERT(__kmp_user_lock_table.used <
3803                    __kmp_user_lock_table.allocated);
3804   index = __kmp_user_lock_table.used;
3805   __kmp_user_lock_table.table[index] = lck;
3806   ++__kmp_user_lock_table.used;
3807   return index;
3808 }
3809 
3810 static kmp_user_lock_p __kmp_lock_block_allocate() {
3811   // Assume that kmp_global_lock is held upon entry/exit.
3812   static int last_index = 0;
3813   if ((last_index >= __kmp_num_locks_in_block) || (__kmp_lock_blocks == NULL)) {
3814     // Restart the index.
3815     last_index = 0;
3816     // Need to allocate a new block.
3817     KMP_DEBUG_ASSERT(__kmp_user_lock_size > 0);
3818     size_t space_for_locks = __kmp_user_lock_size * __kmp_num_locks_in_block;
3819     char *buffer =
3820         (char *)__kmp_allocate(space_for_locks + sizeof(kmp_block_of_locks));
3821     // Set up the new block.
3822     kmp_block_of_locks *new_block =
3823         (kmp_block_of_locks *)(&buffer[space_for_locks]);
3824     new_block->next_block = __kmp_lock_blocks;
3825     new_block->locks = (void *)buffer;
3826     // Publish the new block.
3827     KMP_MB();
3828     __kmp_lock_blocks = new_block;
3829   }
3830   kmp_user_lock_p ret = (kmp_user_lock_p)(&(
3831       ((char *)(__kmp_lock_blocks->locks))[last_index * __kmp_user_lock_size]));
3832   last_index++;
3833   return ret;
3834 }
3835 
3836 // Get memory for a lock. It may be freshly allocated memory or reused memory
3837 // from lock pool.
3838 kmp_user_lock_p __kmp_user_lock_allocate(void **user_lock, kmp_int32 gtid,
3839                                          kmp_lock_flags_t flags) {
3840   kmp_user_lock_p lck;
3841   kmp_lock_index_t index;
3842   KMP_DEBUG_ASSERT(user_lock);
3843 
3844   __kmp_acquire_lock(&__kmp_global_lock, gtid);
3845 
3846   if (__kmp_lock_pool == NULL) {
3847     // Lock pool is empty. Allocate new memory.
3848 
3849     // ANNOTATION: Found no good way to express the syncronisation
3850     // between allocation and usage, so ignore the allocation
3851     ANNOTATE_IGNORE_WRITES_BEGIN();
3852     if (__kmp_num_locks_in_block <= 1) { // Tune this cutoff point.
3853       lck = (kmp_user_lock_p)__kmp_allocate(__kmp_user_lock_size);
3854     } else {
3855       lck = __kmp_lock_block_allocate();
3856     }
3857     ANNOTATE_IGNORE_WRITES_END();
3858 
3859     // Insert lock in the table so that it can be freed in __kmp_cleanup,
3860     // and debugger has info on all allocated locks.
3861     index = __kmp_lock_table_insert(lck);
3862   } else {
3863     // Pick up lock from pool.
3864     lck = __kmp_lock_pool;
3865     index = __kmp_lock_pool->pool.index;
3866     __kmp_lock_pool = __kmp_lock_pool->pool.next;
3867   }
3868 
3869   // We could potentially differentiate between nested and regular locks
3870   // here, and do the lock table lookup for regular locks only.
3871   if (OMP_LOCK_T_SIZE < sizeof(void *)) {
3872     *((kmp_lock_index_t *)user_lock) = index;
3873   } else {
3874     *((kmp_user_lock_p *)user_lock) = lck;
3875   }
3876 
3877   // mark the lock if it is critical section lock.
3878   __kmp_set_user_lock_flags(lck, flags);
3879 
3880   __kmp_release_lock(&__kmp_global_lock, gtid); // AC: TODO move this line upper
3881 
3882   return lck;
3883 }
3884 
3885 // Put lock's memory to pool for reusing.
3886 void __kmp_user_lock_free(void **user_lock, kmp_int32 gtid,
3887                           kmp_user_lock_p lck) {
3888   KMP_DEBUG_ASSERT(user_lock != NULL);
3889   KMP_DEBUG_ASSERT(lck != NULL);
3890 
3891   __kmp_acquire_lock(&__kmp_global_lock, gtid);
3892 
3893   lck->pool.next = __kmp_lock_pool;
3894   __kmp_lock_pool = lck;
3895   if (OMP_LOCK_T_SIZE < sizeof(void *)) {
3896     kmp_lock_index_t index = *((kmp_lock_index_t *)user_lock);
3897     KMP_DEBUG_ASSERT(0 < index && index <= __kmp_user_lock_table.used);
3898     lck->pool.index = index;
3899   }
3900 
3901   __kmp_release_lock(&__kmp_global_lock, gtid);
3902 }
3903 
3904 kmp_user_lock_p __kmp_lookup_user_lock(void **user_lock, char const *func) {
3905   kmp_user_lock_p lck = NULL;
3906 
3907   if (__kmp_env_consistency_check) {
3908     if (user_lock == NULL) {
3909       KMP_FATAL(LockIsUninitialized, func);
3910     }
3911   }
3912 
3913   if (OMP_LOCK_T_SIZE < sizeof(void *)) {
3914     kmp_lock_index_t index = *((kmp_lock_index_t *)user_lock);
3915     if (__kmp_env_consistency_check) {
3916       if (!(0 < index && index < __kmp_user_lock_table.used)) {
3917         KMP_FATAL(LockIsUninitialized, func);
3918       }
3919     }
3920     KMP_DEBUG_ASSERT(0 < index && index < __kmp_user_lock_table.used);
3921     KMP_DEBUG_ASSERT(__kmp_user_lock_size > 0);
3922     lck = __kmp_user_lock_table.table[index];
3923   } else {
3924     lck = *((kmp_user_lock_p *)user_lock);
3925   }
3926 
3927   if (__kmp_env_consistency_check) {
3928     if (lck == NULL) {
3929       KMP_FATAL(LockIsUninitialized, func);
3930     }
3931   }
3932 
3933   return lck;
3934 }
3935 
3936 void __kmp_cleanup_user_locks(void) {
3937   // Reset lock pool. Don't worry about lock in the pool--we will free them when
3938   // iterating through lock table (it includes all the locks, dead or alive).
3939   __kmp_lock_pool = NULL;
3940 
3941 #define IS_CRITICAL(lck)                                                       \
3942   ((__kmp_get_user_lock_flags_ != NULL) &&                                     \
3943    ((*__kmp_get_user_lock_flags_)(lck)&kmp_lf_critical_section))
3944 
3945   // Loop through lock table, free all locks.
3946   // Do not free item [0], it is reserved for lock tables list.
3947   //
3948   // FIXME - we are iterating through a list of (pointers to) objects of type
3949   // union kmp_user_lock, but we have no way of knowing whether the base type is
3950   // currently "pool" or whatever the global user lock type is.
3951   //
3952   // We are relying on the fact that for all of the user lock types
3953   // (except "tas"), the first field in the lock struct is the "initialized"
3954   // field, which is set to the address of the lock object itself when
3955   // the lock is initialized.  When the union is of type "pool", the
3956   // first field is a pointer to the next object in the free list, which
3957   // will not be the same address as the object itself.
3958   //
3959   // This means that the check (*__kmp_is_user_lock_initialized_)(lck) will fail
3960   // for "pool" objects on the free list.  This must happen as the "location"
3961   // field of real user locks overlaps the "index" field of "pool" objects.
3962   //
3963   // It would be better to run through the free list, and remove all "pool"
3964   // objects from the lock table before executing this loop.  However,
3965   // "pool" objects do not always have their index field set (only on
3966   // lin_32e), and I don't want to search the lock table for the address
3967   // of every "pool" object on the free list.
3968   while (__kmp_user_lock_table.used > 1) {
3969     const ident *loc;
3970 
3971     // reduce __kmp_user_lock_table.used before freeing the lock,
3972     // so that state of locks is consistent
3973     kmp_user_lock_p lck =
3974         __kmp_user_lock_table.table[--__kmp_user_lock_table.used];
3975 
3976     if ((__kmp_is_user_lock_initialized_ != NULL) &&
3977         (*__kmp_is_user_lock_initialized_)(lck)) {
3978       // Issue a warning if: KMP_CONSISTENCY_CHECK AND lock is initialized AND
3979       // it is NOT a critical section (user is not responsible for destroying
3980       // criticals) AND we know source location to report.
3981       if (__kmp_env_consistency_check && (!IS_CRITICAL(lck)) &&
3982           ((loc = __kmp_get_user_lock_location(lck)) != NULL) &&
3983           (loc->psource != NULL)) {
3984         kmp_str_loc_t str_loc = __kmp_str_loc_init(loc->psource, false);
3985         KMP_WARNING(CnsLockNotDestroyed, str_loc.file, str_loc.line);
3986         __kmp_str_loc_free(&str_loc);
3987       }
3988 
3989 #ifdef KMP_DEBUG
3990       if (IS_CRITICAL(lck)) {
3991         KA_TRACE(
3992             20,
3993             ("__kmp_cleanup_user_locks: free critical section lock %p (%p)\n",
3994              lck, *(void **)lck));
3995       } else {
3996         KA_TRACE(20, ("__kmp_cleanup_user_locks: free lock %p (%p)\n", lck,
3997                       *(void **)lck));
3998       }
3999 #endif // KMP_DEBUG
4000 
4001       // Cleanup internal lock dynamic resources (for drdpa locks particularly).
4002       __kmp_destroy_user_lock(lck);
4003     }
4004 
4005     // Free the lock if block allocation of locks is not used.
4006     if (__kmp_lock_blocks == NULL) {
4007       __kmp_free(lck);
4008     }
4009   }
4010 
4011 #undef IS_CRITICAL
4012 
4013   // delete lock table(s).
4014   kmp_user_lock_p *table_ptr = __kmp_user_lock_table.table;
4015   __kmp_user_lock_table.table = NULL;
4016   __kmp_user_lock_table.allocated = 0;
4017 
4018   while (table_ptr != NULL) {
4019     // In the first element we saved the pointer to the previous
4020     // (smaller) lock table.
4021     kmp_user_lock_p *next = (kmp_user_lock_p *)(table_ptr[0]);
4022     __kmp_free(table_ptr);
4023     table_ptr = next;
4024   }
4025 
4026   // Free buffers allocated for blocks of locks.
4027   kmp_block_of_locks_t *block_ptr = __kmp_lock_blocks;
4028   __kmp_lock_blocks = NULL;
4029 
4030   while (block_ptr != NULL) {
4031     kmp_block_of_locks_t *next = block_ptr->next_block;
4032     __kmp_free(block_ptr->locks);
4033     // *block_ptr itself was allocated at the end of the locks vector.
4034     block_ptr = next;
4035   }
4036 
4037   TCW_4(__kmp_init_user_locks, FALSE);
4038 }
4039 
4040 #endif // KMP_USE_DYNAMIC_LOCK
4041