xref: /freebsd-src/contrib/llvm-project/openmp/runtime/src/kmp_lock.cpp (revision 48aaf27bca484047c923eb06644c7666d5d89c24)
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     kmp_int32 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=%d 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_COMPILER_ICC && __INTEL_COMPILER >= 1300) ||                          \
1708     (KMP_COMPILER_MSVC && _MSC_VER >= 1700) ||                                 \
1709     (KMP_COMPILER_CLANG && KMP_MSVC_COMPAT)
1710 
1711 #include <immintrin.h>
1712 #define SOFT_ABORT_MASK (_XABORT_RETRY | _XABORT_CONFLICT | _XABORT_EXPLICIT)
1713 
1714 #else
1715 
1716 // Values from the status register after failed speculation.
1717 #define _XBEGIN_STARTED (~0u)
1718 #define _XABORT_EXPLICIT (1 << 0)
1719 #define _XABORT_RETRY (1 << 1)
1720 #define _XABORT_CONFLICT (1 << 2)
1721 #define _XABORT_CAPACITY (1 << 3)
1722 #define _XABORT_DEBUG (1 << 4)
1723 #define _XABORT_NESTED (1 << 5)
1724 #define _XABORT_CODE(x) ((unsigned char)(((x) >> 24) & 0xFF))
1725 
1726 // Aborts for which it's worth trying again immediately
1727 #define SOFT_ABORT_MASK (_XABORT_RETRY | _XABORT_CONFLICT | _XABORT_EXPLICIT)
1728 
1729 #define STRINGIZE_INTERNAL(arg) #arg
1730 #define STRINGIZE(arg) STRINGIZE_INTERNAL(arg)
1731 
1732 // Access to RTM instructions
1733 /*A version of XBegin which returns -1 on speculation, and the value of EAX on
1734   an abort. This is the same definition as the compiler intrinsic that will be
1735   supported at some point. */
1736 static __inline int _xbegin() {
1737   int res = -1;
1738 
1739 #if KMP_OS_WINDOWS
1740 #if KMP_ARCH_X86_64
1741   _asm {
1742         _emit 0xC7
1743         _emit 0xF8
1744         _emit 2
1745         _emit 0
1746         _emit 0
1747         _emit 0
1748         jmp   L2
1749         mov   res, eax
1750     L2:
1751   }
1752 #else /* IA32 */
1753   _asm {
1754         _emit 0xC7
1755         _emit 0xF8
1756         _emit 2
1757         _emit 0
1758         _emit 0
1759         _emit 0
1760         jmp   L2
1761         mov   res, eax
1762     L2:
1763   }
1764 #endif // KMP_ARCH_X86_64
1765 #else
1766   /* Note that %eax must be noted as killed (clobbered), because the XSR is
1767      returned in %eax(%rax) on abort.  Other register values are restored, so
1768      don't need to be killed.
1769 
1770      We must also mark 'res' as an input and an output, since otherwise
1771      'res=-1' may be dropped as being dead, whereas we do need the assignment on
1772      the successful (i.e., non-abort) path. */
1773   __asm__ volatile("1: .byte  0xC7; .byte 0xF8;\n"
1774                    "   .long  1f-1b-6\n"
1775                    "    jmp   2f\n"
1776                    "1:  movl  %%eax,%0\n"
1777                    "2:"
1778                    : "+r"(res)::"memory", "%eax");
1779 #endif // KMP_OS_WINDOWS
1780   return res;
1781 }
1782 
1783 /* Transaction end */
1784 static __inline void _xend() {
1785 #if KMP_OS_WINDOWS
1786   __asm {
1787         _emit 0x0f
1788         _emit 0x01
1789         _emit 0xd5
1790   }
1791 #else
1792   __asm__ volatile(".byte 0x0f; .byte 0x01; .byte 0xd5" ::: "memory");
1793 #endif
1794 }
1795 
1796 /* This is a macro, the argument must be a single byte constant which can be
1797    evaluated by the inline assembler, since it is emitted as a byte into the
1798    assembly code. */
1799 // clang-format off
1800 #if KMP_OS_WINDOWS
1801 #define _xabort(ARG) _asm _emit 0xc6 _asm _emit 0xf8 _asm _emit ARG
1802 #else
1803 #define _xabort(ARG)                                                           \
1804   __asm__ volatile(".byte 0xC6; .byte 0xF8; .byte " STRINGIZE(ARG):::"memory");
1805 #endif
1806 // clang-format on
1807 #endif // KMP_COMPILER_ICC && __INTEL_COMPILER >= 1300
1808 
1809 // Statistics is collected for testing purpose
1810 #if KMP_DEBUG_ADAPTIVE_LOCKS
1811 
1812 // We accumulate speculative lock statistics when the lock is destroyed. We
1813 // keep locks that haven't been destroyed in the liveLocks list so that we can
1814 // grab their statistics too.
1815 static kmp_adaptive_lock_statistics_t destroyedStats;
1816 
1817 // To hold the list of live locks.
1818 static kmp_adaptive_lock_info_t liveLocks;
1819 
1820 // A lock so we can safely update the list of locks.
1821 static kmp_bootstrap_lock_t chain_lock =
1822     KMP_BOOTSTRAP_LOCK_INITIALIZER(chain_lock);
1823 
1824 // Initialize the list of stats.
1825 void __kmp_init_speculative_stats() {
1826   kmp_adaptive_lock_info_t *lck = &liveLocks;
1827 
1828   memset(CCAST(kmp_adaptive_lock_statistics_t *, &(lck->stats)), 0,
1829          sizeof(lck->stats));
1830   lck->stats.next = lck;
1831   lck->stats.prev = lck;
1832 
1833   KMP_ASSERT(lck->stats.next->stats.prev == lck);
1834   KMP_ASSERT(lck->stats.prev->stats.next == lck);
1835 
1836   __kmp_init_bootstrap_lock(&chain_lock);
1837 }
1838 
1839 // Insert the lock into the circular list
1840 static void __kmp_remember_lock(kmp_adaptive_lock_info_t *lck) {
1841   __kmp_acquire_bootstrap_lock(&chain_lock);
1842 
1843   lck->stats.next = liveLocks.stats.next;
1844   lck->stats.prev = &liveLocks;
1845 
1846   liveLocks.stats.next = lck;
1847   lck->stats.next->stats.prev = lck;
1848 
1849   KMP_ASSERT(lck->stats.next->stats.prev == lck);
1850   KMP_ASSERT(lck->stats.prev->stats.next == lck);
1851 
1852   __kmp_release_bootstrap_lock(&chain_lock);
1853 }
1854 
1855 static void __kmp_forget_lock(kmp_adaptive_lock_info_t *lck) {
1856   KMP_ASSERT(lck->stats.next->stats.prev == lck);
1857   KMP_ASSERT(lck->stats.prev->stats.next == lck);
1858 
1859   kmp_adaptive_lock_info_t *n = lck->stats.next;
1860   kmp_adaptive_lock_info_t *p = lck->stats.prev;
1861 
1862   n->stats.prev = p;
1863   p->stats.next = n;
1864 }
1865 
1866 static void __kmp_zero_speculative_stats(kmp_adaptive_lock_info_t *lck) {
1867   memset(CCAST(kmp_adaptive_lock_statistics_t *, &lck->stats), 0,
1868          sizeof(lck->stats));
1869   __kmp_remember_lock(lck);
1870 }
1871 
1872 static void __kmp_add_stats(kmp_adaptive_lock_statistics_t *t,
1873                             kmp_adaptive_lock_info_t *lck) {
1874   kmp_adaptive_lock_statistics_t volatile *s = &lck->stats;
1875 
1876   t->nonSpeculativeAcquireAttempts += lck->acquire_attempts;
1877   t->successfulSpeculations += s->successfulSpeculations;
1878   t->hardFailedSpeculations += s->hardFailedSpeculations;
1879   t->softFailedSpeculations += s->softFailedSpeculations;
1880   t->nonSpeculativeAcquires += s->nonSpeculativeAcquires;
1881   t->lemmingYields += s->lemmingYields;
1882 }
1883 
1884 static void __kmp_accumulate_speculative_stats(kmp_adaptive_lock_info_t *lck) {
1885   __kmp_acquire_bootstrap_lock(&chain_lock);
1886 
1887   __kmp_add_stats(&destroyedStats, lck);
1888   __kmp_forget_lock(lck);
1889 
1890   __kmp_release_bootstrap_lock(&chain_lock);
1891 }
1892 
1893 static float percent(kmp_uint32 count, kmp_uint32 total) {
1894   return (total == 0) ? 0.0 : (100.0 * count) / total;
1895 }
1896 
1897 static FILE *__kmp_open_stats_file() {
1898   if (strcmp(__kmp_speculative_statsfile, "-") == 0)
1899     return stdout;
1900 
1901   size_t buffLen = KMP_STRLEN(__kmp_speculative_statsfile) + 20;
1902   char buffer[buffLen];
1903   KMP_SNPRINTF(&buffer[0], buffLen, __kmp_speculative_statsfile,
1904                (kmp_int32)getpid());
1905   FILE *result = fopen(&buffer[0], "w");
1906 
1907   // Maybe we should issue a warning here...
1908   return result ? result : stdout;
1909 }
1910 
1911 void __kmp_print_speculative_stats() {
1912   kmp_adaptive_lock_statistics_t total = destroyedStats;
1913   kmp_adaptive_lock_info_t *lck;
1914 
1915   for (lck = liveLocks.stats.next; lck != &liveLocks; lck = lck->stats.next) {
1916     __kmp_add_stats(&total, lck);
1917   }
1918   kmp_adaptive_lock_statistics_t *t = &total;
1919   kmp_uint32 totalSections =
1920       t->nonSpeculativeAcquires + t->successfulSpeculations;
1921   kmp_uint32 totalSpeculations = t->successfulSpeculations +
1922                                  t->hardFailedSpeculations +
1923                                  t->softFailedSpeculations;
1924   if (totalSections <= 0)
1925     return;
1926 
1927   FILE *statsFile = __kmp_open_stats_file();
1928 
1929   fprintf(statsFile, "Speculative lock statistics (all approximate!)\n");
1930   fprintf(statsFile, " Lock parameters: \n"
1931                      "   max_soft_retries               : %10d\n"
1932                      "   max_badness                    : %10d\n",
1933           __kmp_adaptive_backoff_params.max_soft_retries,
1934           __kmp_adaptive_backoff_params.max_badness);
1935   fprintf(statsFile, " Non-speculative acquire attempts : %10d\n",
1936           t->nonSpeculativeAcquireAttempts);
1937   fprintf(statsFile, " Total critical sections          : %10d\n",
1938           totalSections);
1939   fprintf(statsFile, " Successful speculations          : %10d (%5.1f%%)\n",
1940           t->successfulSpeculations,
1941           percent(t->successfulSpeculations, totalSections));
1942   fprintf(statsFile, " Non-speculative acquires         : %10d (%5.1f%%)\n",
1943           t->nonSpeculativeAcquires,
1944           percent(t->nonSpeculativeAcquires, totalSections));
1945   fprintf(statsFile, " Lemming yields                   : %10d\n\n",
1946           t->lemmingYields);
1947 
1948   fprintf(statsFile, " Speculative acquire attempts     : %10d\n",
1949           totalSpeculations);
1950   fprintf(statsFile, " Successes                        : %10d (%5.1f%%)\n",
1951           t->successfulSpeculations,
1952           percent(t->successfulSpeculations, totalSpeculations));
1953   fprintf(statsFile, " Soft failures                    : %10d (%5.1f%%)\n",
1954           t->softFailedSpeculations,
1955           percent(t->softFailedSpeculations, totalSpeculations));
1956   fprintf(statsFile, " Hard failures                    : %10d (%5.1f%%)\n",
1957           t->hardFailedSpeculations,
1958           percent(t->hardFailedSpeculations, totalSpeculations));
1959 
1960   if (statsFile != stdout)
1961     fclose(statsFile);
1962 }
1963 
1964 #define KMP_INC_STAT(lck, stat) (lck->lk.adaptive.stats.stat++)
1965 #else
1966 #define KMP_INC_STAT(lck, stat)
1967 
1968 #endif // KMP_DEBUG_ADAPTIVE_LOCKS
1969 
1970 static inline bool __kmp_is_unlocked_queuing_lock(kmp_queuing_lock_t *lck) {
1971   // It is enough to check that the head_id is zero.
1972   // We don't also need to check the tail.
1973   bool res = lck->lk.head_id == 0;
1974 
1975 // We need a fence here, since we must ensure that no memory operations
1976 // from later in this thread float above that read.
1977 #if KMP_COMPILER_ICC
1978   _mm_mfence();
1979 #else
1980   __sync_synchronize();
1981 #endif
1982 
1983   return res;
1984 }
1985 
1986 // Functions for manipulating the badness
1987 static __inline void
1988 __kmp_update_badness_after_success(kmp_adaptive_lock_t *lck) {
1989   // Reset the badness to zero so we eagerly try to speculate again
1990   lck->lk.adaptive.badness = 0;
1991   KMP_INC_STAT(lck, successfulSpeculations);
1992 }
1993 
1994 // Create a bit mask with one more set bit.
1995 static __inline void __kmp_step_badness(kmp_adaptive_lock_t *lck) {
1996   kmp_uint32 newBadness = (lck->lk.adaptive.badness << 1) | 1;
1997   if (newBadness > lck->lk.adaptive.max_badness) {
1998     return;
1999   } else {
2000     lck->lk.adaptive.badness = newBadness;
2001   }
2002 }
2003 
2004 // Check whether speculation should be attempted.
2005 static __inline int __kmp_should_speculate(kmp_adaptive_lock_t *lck,
2006                                            kmp_int32 gtid) {
2007   kmp_uint32 badness = lck->lk.adaptive.badness;
2008   kmp_uint32 attempts = lck->lk.adaptive.acquire_attempts;
2009   int res = (attempts & badness) == 0;
2010   return res;
2011 }
2012 
2013 // Attempt to acquire only the speculative lock.
2014 // Does not back off to the non-speculative lock.
2015 static int __kmp_test_adaptive_lock_only(kmp_adaptive_lock_t *lck,
2016                                          kmp_int32 gtid) {
2017   int retries = lck->lk.adaptive.max_soft_retries;
2018 
2019   // We don't explicitly count the start of speculation, rather we record the
2020   // results (success, hard fail, soft fail). The sum of all of those is the
2021   // total number of times we started speculation since all speculations must
2022   // end one of those ways.
2023   do {
2024     kmp_uint32 status = _xbegin();
2025     // Switch this in to disable actual speculation but exercise at least some
2026     // of the rest of the code. Useful for debugging...
2027     // kmp_uint32 status = _XABORT_NESTED;
2028 
2029     if (status == _XBEGIN_STARTED) {
2030       /* We have successfully started speculation. Check that no-one acquired
2031          the lock for real between when we last looked and now. This also gets
2032          the lock cache line into our read-set, which we need so that we'll
2033          abort if anyone later claims it for real. */
2034       if (!__kmp_is_unlocked_queuing_lock(GET_QLK_PTR(lck))) {
2035         // Lock is now visibly acquired, so someone beat us to it. Abort the
2036         // transaction so we'll restart from _xbegin with the failure status.
2037         _xabort(0x01);
2038         KMP_ASSERT2(0, "should not get here");
2039       }
2040       return 1; // Lock has been acquired (speculatively)
2041     } else {
2042       // We have aborted, update the statistics
2043       if (status & SOFT_ABORT_MASK) {
2044         KMP_INC_STAT(lck, softFailedSpeculations);
2045         // and loop round to retry.
2046       } else {
2047         KMP_INC_STAT(lck, hardFailedSpeculations);
2048         // Give up if we had a hard failure.
2049         break;
2050       }
2051     }
2052   } while (retries--); // Loop while we have retries, and didn't fail hard.
2053 
2054   // Either we had a hard failure or we didn't succeed softly after
2055   // the full set of attempts, so back off the badness.
2056   __kmp_step_badness(lck);
2057   return 0;
2058 }
2059 
2060 // Attempt to acquire the speculative lock, or back off to the non-speculative
2061 // one if the speculative lock cannot be acquired.
2062 // We can succeed speculatively, non-speculatively, or fail.
2063 static int __kmp_test_adaptive_lock(kmp_adaptive_lock_t *lck, kmp_int32 gtid) {
2064   // First try to acquire the lock speculatively
2065   if (__kmp_should_speculate(lck, gtid) &&
2066       __kmp_test_adaptive_lock_only(lck, gtid))
2067     return 1;
2068 
2069   // Speculative acquisition failed, so try to acquire it non-speculatively.
2070   // Count the non-speculative acquire attempt
2071   lck->lk.adaptive.acquire_attempts++;
2072 
2073   // Use base, non-speculative lock.
2074   if (__kmp_test_queuing_lock(GET_QLK_PTR(lck), gtid)) {
2075     KMP_INC_STAT(lck, nonSpeculativeAcquires);
2076     return 1; // Lock is acquired (non-speculatively)
2077   } else {
2078     return 0; // Failed to acquire the lock, it's already visibly locked.
2079   }
2080 }
2081 
2082 static int __kmp_test_adaptive_lock_with_checks(kmp_adaptive_lock_t *lck,
2083                                                 kmp_int32 gtid) {
2084   char const *const func = "omp_test_lock";
2085   if (lck->lk.qlk.initialized != GET_QLK_PTR(lck)) {
2086     KMP_FATAL(LockIsUninitialized, func);
2087   }
2088 
2089   int retval = __kmp_test_adaptive_lock(lck, gtid);
2090 
2091   if (retval) {
2092     lck->lk.qlk.owner_id = gtid + 1;
2093   }
2094   return retval;
2095 }
2096 
2097 // Block until we can acquire a speculative, adaptive lock. We check whether we
2098 // should be trying to speculate. If we should be, we check the real lock to see
2099 // if it is free, and, if not, pause without attempting to acquire it until it
2100 // is. Then we try the speculative acquire. This means that although we suffer
2101 // from lemmings a little (because all we can't acquire the lock speculatively
2102 // until the queue of threads waiting has cleared), we don't get into a state
2103 // where we can never acquire the lock speculatively (because we force the queue
2104 // to clear by preventing new arrivals from entering the queue). This does mean
2105 // that when we're trying to break lemmings, the lock is no longer fair. However
2106 // OpenMP makes no guarantee that its locks are fair, so this isn't a real
2107 // problem.
2108 static void __kmp_acquire_adaptive_lock(kmp_adaptive_lock_t *lck,
2109                                         kmp_int32 gtid) {
2110   if (__kmp_should_speculate(lck, gtid)) {
2111     if (__kmp_is_unlocked_queuing_lock(GET_QLK_PTR(lck))) {
2112       if (__kmp_test_adaptive_lock_only(lck, gtid))
2113         return;
2114       // We tried speculation and failed, so give up.
2115     } else {
2116       // We can't try speculation until the lock is free, so we pause here
2117       // (without suspending on the queueing lock, to allow it to drain, then
2118       // try again. All other threads will also see the same result for
2119       // shouldSpeculate, so will be doing the same if they try to claim the
2120       // lock from now on.
2121       while (!__kmp_is_unlocked_queuing_lock(GET_QLK_PTR(lck))) {
2122         KMP_INC_STAT(lck, lemmingYields);
2123         KMP_YIELD(TRUE);
2124       }
2125 
2126       if (__kmp_test_adaptive_lock_only(lck, gtid))
2127         return;
2128     }
2129   }
2130 
2131   // Speculative acquisition failed, so acquire it non-speculatively.
2132   // Count the non-speculative acquire attempt
2133   lck->lk.adaptive.acquire_attempts++;
2134 
2135   __kmp_acquire_queuing_lock_timed_template<FALSE>(GET_QLK_PTR(lck), gtid);
2136   // We have acquired the base lock, so count that.
2137   KMP_INC_STAT(lck, nonSpeculativeAcquires);
2138   ANNOTATE_QUEUING_ACQUIRED(lck);
2139 }
2140 
2141 static void __kmp_acquire_adaptive_lock_with_checks(kmp_adaptive_lock_t *lck,
2142                                                     kmp_int32 gtid) {
2143   char const *const func = "omp_set_lock";
2144   if (lck->lk.qlk.initialized != GET_QLK_PTR(lck)) {
2145     KMP_FATAL(LockIsUninitialized, func);
2146   }
2147   if (__kmp_get_queuing_lock_owner(GET_QLK_PTR(lck)) == gtid) {
2148     KMP_FATAL(LockIsAlreadyOwned, func);
2149   }
2150 
2151   __kmp_acquire_adaptive_lock(lck, gtid);
2152 
2153   lck->lk.qlk.owner_id = gtid + 1;
2154 }
2155 
2156 static int __kmp_release_adaptive_lock(kmp_adaptive_lock_t *lck,
2157                                        kmp_int32 gtid) {
2158   if (__kmp_is_unlocked_queuing_lock(GET_QLK_PTR(
2159           lck))) { // If the lock doesn't look claimed we must be speculating.
2160     // (Or the user's code is buggy and they're releasing without locking;
2161     // if we had XTEST we'd be able to check that case...)
2162     _xend(); // Exit speculation
2163     __kmp_update_badness_after_success(lck);
2164   } else { // Since the lock *is* visibly locked we're not speculating,
2165     // so should use the underlying lock's release scheme.
2166     __kmp_release_queuing_lock(GET_QLK_PTR(lck), gtid);
2167   }
2168   return KMP_LOCK_RELEASED;
2169 }
2170 
2171 static int __kmp_release_adaptive_lock_with_checks(kmp_adaptive_lock_t *lck,
2172                                                    kmp_int32 gtid) {
2173   char const *const func = "omp_unset_lock";
2174   KMP_MB(); /* in case another processor initialized lock */
2175   if (lck->lk.qlk.initialized != GET_QLK_PTR(lck)) {
2176     KMP_FATAL(LockIsUninitialized, func);
2177   }
2178   if (__kmp_get_queuing_lock_owner(GET_QLK_PTR(lck)) == -1) {
2179     KMP_FATAL(LockUnsettingFree, func);
2180   }
2181   if (__kmp_get_queuing_lock_owner(GET_QLK_PTR(lck)) != gtid) {
2182     KMP_FATAL(LockUnsettingSetByAnother, func);
2183   }
2184   lck->lk.qlk.owner_id = 0;
2185   __kmp_release_adaptive_lock(lck, gtid);
2186   return KMP_LOCK_RELEASED;
2187 }
2188 
2189 static void __kmp_init_adaptive_lock(kmp_adaptive_lock_t *lck) {
2190   __kmp_init_queuing_lock(GET_QLK_PTR(lck));
2191   lck->lk.adaptive.badness = 0;
2192   lck->lk.adaptive.acquire_attempts = 0; // nonSpeculativeAcquireAttempts = 0;
2193   lck->lk.adaptive.max_soft_retries =
2194       __kmp_adaptive_backoff_params.max_soft_retries;
2195   lck->lk.adaptive.max_badness = __kmp_adaptive_backoff_params.max_badness;
2196 #if KMP_DEBUG_ADAPTIVE_LOCKS
2197   __kmp_zero_speculative_stats(&lck->lk.adaptive);
2198 #endif
2199   KA_TRACE(1000, ("__kmp_init_adaptive_lock: lock %p initialized\n", lck));
2200 }
2201 
2202 static void __kmp_destroy_adaptive_lock(kmp_adaptive_lock_t *lck) {
2203 #if KMP_DEBUG_ADAPTIVE_LOCKS
2204   __kmp_accumulate_speculative_stats(&lck->lk.adaptive);
2205 #endif
2206   __kmp_destroy_queuing_lock(GET_QLK_PTR(lck));
2207   // Nothing needed for the speculative part.
2208 }
2209 
2210 static void __kmp_destroy_adaptive_lock_with_checks(kmp_adaptive_lock_t *lck) {
2211   char const *const func = "omp_destroy_lock";
2212   if (lck->lk.qlk.initialized != GET_QLK_PTR(lck)) {
2213     KMP_FATAL(LockIsUninitialized, func);
2214   }
2215   if (__kmp_get_queuing_lock_owner(GET_QLK_PTR(lck)) != -1) {
2216     KMP_FATAL(LockStillOwned, func);
2217   }
2218   __kmp_destroy_adaptive_lock(lck);
2219 }
2220 
2221 #endif // KMP_USE_ADAPTIVE_LOCKS
2222 
2223 /* ------------------------------------------------------------------------ */
2224 /* DRDPA ticket locks                                                */
2225 /* "DRDPA" means Dynamically Reconfigurable Distributed Polling Area */
2226 
2227 static kmp_int32 __kmp_get_drdpa_lock_owner(kmp_drdpa_lock_t *lck) {
2228   return lck->lk.owner_id - 1;
2229 }
2230 
2231 static inline bool __kmp_is_drdpa_lock_nestable(kmp_drdpa_lock_t *lck) {
2232   return lck->lk.depth_locked != -1;
2233 }
2234 
2235 __forceinline static int
2236 __kmp_acquire_drdpa_lock_timed_template(kmp_drdpa_lock_t *lck, kmp_int32 gtid) {
2237   kmp_uint64 ticket = KMP_ATOMIC_INC(&lck->lk.next_ticket);
2238   kmp_uint64 mask = lck->lk.mask; // atomic load
2239   std::atomic<kmp_uint64> *polls = lck->lk.polls;
2240 
2241 #ifdef USE_LOCK_PROFILE
2242   if (polls[ticket & mask] != ticket)
2243     __kmp_printf("LOCK CONTENTION: %p\n", lck);
2244 /* else __kmp_printf( "." );*/
2245 #endif /* USE_LOCK_PROFILE */
2246 
2247   // Now spin-wait, but reload the polls pointer and mask, in case the
2248   // polling area has been reconfigured.  Unless it is reconfigured, the
2249   // reloads stay in L1 cache and are cheap.
2250   //
2251   // Keep this code in sync with KMP_WAIT, in kmp_dispatch.cpp !!!
2252   // The current implementation of KMP_WAIT doesn't allow for mask
2253   // and poll to be re-read every spin iteration.
2254   kmp_uint32 spins;
2255   KMP_FSYNC_PREPARE(lck);
2256   KMP_INIT_YIELD(spins);
2257   while (polls[ticket & mask] < ticket) { // atomic load
2258     KMP_YIELD_OVERSUB_ELSE_SPIN(spins);
2259     // Re-read the mask and the poll pointer from the lock structure.
2260     //
2261     // Make certain that "mask" is read before "polls" !!!
2262     //
2263     // If another thread picks reconfigures the polling area and updates their
2264     // values, and we get the new value of mask and the old polls pointer, we
2265     // could access memory beyond the end of the old polling area.
2266     mask = lck->lk.mask; // atomic load
2267     polls = lck->lk.polls; // atomic load
2268   }
2269 
2270   // Critical section starts here
2271   KMP_FSYNC_ACQUIRED(lck);
2272   KA_TRACE(1000, ("__kmp_acquire_drdpa_lock: ticket #%lld acquired lock %p\n",
2273                   ticket, lck));
2274   lck->lk.now_serving = ticket; // non-volatile store
2275 
2276   // Deallocate a garbage polling area if we know that we are the last
2277   // thread that could possibly access it.
2278   //
2279   // The >= check is in case __kmp_test_drdpa_lock() allocated the cleanup
2280   // ticket.
2281   if ((lck->lk.old_polls != NULL) && (ticket >= lck->lk.cleanup_ticket)) {
2282     __kmp_free(lck->lk.old_polls);
2283     lck->lk.old_polls = NULL;
2284     lck->lk.cleanup_ticket = 0;
2285   }
2286 
2287   // Check to see if we should reconfigure the polling area.
2288   // If there is still a garbage polling area to be deallocated from a
2289   // previous reconfiguration, let a later thread reconfigure it.
2290   if (lck->lk.old_polls == NULL) {
2291     bool reconfigure = false;
2292     std::atomic<kmp_uint64> *old_polls = polls;
2293     kmp_uint32 num_polls = TCR_4(lck->lk.num_polls);
2294 
2295     if (TCR_4(__kmp_nth) >
2296         (__kmp_avail_proc ? __kmp_avail_proc : __kmp_xproc)) {
2297       // We are in oversubscription mode.  Contract the polling area
2298       // down to a single location, if that hasn't been done already.
2299       if (num_polls > 1) {
2300         reconfigure = true;
2301         num_polls = TCR_4(lck->lk.num_polls);
2302         mask = 0;
2303         num_polls = 1;
2304         polls = (std::atomic<kmp_uint64> *)__kmp_allocate(num_polls *
2305                                                           sizeof(*polls));
2306         polls[0] = ticket;
2307       }
2308     } else {
2309       // We are in under/fully subscribed mode.  Check the number of
2310       // threads waiting on the lock.  The size of the polling area
2311       // should be at least the number of threads waiting.
2312       kmp_uint64 num_waiting = TCR_8(lck->lk.next_ticket) - ticket - 1;
2313       if (num_waiting > num_polls) {
2314         kmp_uint32 old_num_polls = num_polls;
2315         reconfigure = true;
2316         do {
2317           mask = (mask << 1) | 1;
2318           num_polls *= 2;
2319         } while (num_polls <= num_waiting);
2320 
2321         // Allocate the new polling area, and copy the relevant portion
2322         // of the old polling area to the new area.  __kmp_allocate()
2323         // zeroes the memory it allocates, and most of the old area is
2324         // just zero padding, so we only copy the release counters.
2325         polls = (std::atomic<kmp_uint64> *)__kmp_allocate(num_polls *
2326                                                           sizeof(*polls));
2327         kmp_uint32 i;
2328         for (i = 0; i < old_num_polls; i++) {
2329           polls[i].store(old_polls[i]);
2330         }
2331       }
2332     }
2333 
2334     if (reconfigure) {
2335       // Now write the updated fields back to the lock structure.
2336       //
2337       // Make certain that "polls" is written before "mask" !!!
2338       //
2339       // If another thread picks up the new value of mask and the old polls
2340       // pointer , it could access memory beyond the end of the old polling
2341       // area.
2342       //
2343       // On x86, we need memory fences.
2344       KA_TRACE(1000, ("__kmp_acquire_drdpa_lock: ticket #%lld reconfiguring "
2345                       "lock %p to %d polls\n",
2346                       ticket, lck, num_polls));
2347 
2348       lck->lk.old_polls = old_polls;
2349       lck->lk.polls = polls; // atomic store
2350 
2351       KMP_MB();
2352 
2353       lck->lk.num_polls = num_polls;
2354       lck->lk.mask = mask; // atomic store
2355 
2356       KMP_MB();
2357 
2358       // Only after the new polling area and mask have been flushed
2359       // to main memory can we update the cleanup ticket field.
2360       //
2361       // volatile load / non-volatile store
2362       lck->lk.cleanup_ticket = lck->lk.next_ticket;
2363     }
2364   }
2365   return KMP_LOCK_ACQUIRED_FIRST;
2366 }
2367 
2368 int __kmp_acquire_drdpa_lock(kmp_drdpa_lock_t *lck, kmp_int32 gtid) {
2369   int retval = __kmp_acquire_drdpa_lock_timed_template(lck, gtid);
2370   ANNOTATE_DRDPA_ACQUIRED(lck);
2371   return retval;
2372 }
2373 
2374 static int __kmp_acquire_drdpa_lock_with_checks(kmp_drdpa_lock_t *lck,
2375                                                 kmp_int32 gtid) {
2376   char const *const func = "omp_set_lock";
2377   if (lck->lk.initialized != lck) {
2378     KMP_FATAL(LockIsUninitialized, func);
2379   }
2380   if (__kmp_is_drdpa_lock_nestable(lck)) {
2381     KMP_FATAL(LockNestableUsedAsSimple, func);
2382   }
2383   if ((gtid >= 0) && (__kmp_get_drdpa_lock_owner(lck) == gtid)) {
2384     KMP_FATAL(LockIsAlreadyOwned, func);
2385   }
2386 
2387   __kmp_acquire_drdpa_lock(lck, gtid);
2388 
2389   lck->lk.owner_id = gtid + 1;
2390   return KMP_LOCK_ACQUIRED_FIRST;
2391 }
2392 
2393 int __kmp_test_drdpa_lock(kmp_drdpa_lock_t *lck, kmp_int32 gtid) {
2394   // First get a ticket, then read the polls pointer and the mask.
2395   // The polls pointer must be read before the mask!!! (See above)
2396   kmp_uint64 ticket = lck->lk.next_ticket; // atomic load
2397   std::atomic<kmp_uint64> *polls = lck->lk.polls;
2398   kmp_uint64 mask = lck->lk.mask; // atomic load
2399   if (polls[ticket & mask] == ticket) {
2400     kmp_uint64 next_ticket = ticket + 1;
2401     if (__kmp_atomic_compare_store_acq(&lck->lk.next_ticket, ticket,
2402                                        next_ticket)) {
2403       KMP_FSYNC_ACQUIRED(lck);
2404       KA_TRACE(1000, ("__kmp_test_drdpa_lock: ticket #%lld acquired lock %p\n",
2405                       ticket, lck));
2406       lck->lk.now_serving = ticket; // non-volatile store
2407 
2408       // Since no threads are waiting, there is no possibility that we would
2409       // want to reconfigure the polling area.  We might have the cleanup ticket
2410       // value (which says that it is now safe to deallocate old_polls), but
2411       // we'll let a later thread which calls __kmp_acquire_lock do that - this
2412       // routine isn't supposed to block, and we would risk blocks if we called
2413       // __kmp_free() to do the deallocation.
2414       return TRUE;
2415     }
2416   }
2417   return FALSE;
2418 }
2419 
2420 static int __kmp_test_drdpa_lock_with_checks(kmp_drdpa_lock_t *lck,
2421                                              kmp_int32 gtid) {
2422   char const *const func = "omp_test_lock";
2423   if (lck->lk.initialized != lck) {
2424     KMP_FATAL(LockIsUninitialized, func);
2425   }
2426   if (__kmp_is_drdpa_lock_nestable(lck)) {
2427     KMP_FATAL(LockNestableUsedAsSimple, func);
2428   }
2429 
2430   int retval = __kmp_test_drdpa_lock(lck, gtid);
2431 
2432   if (retval) {
2433     lck->lk.owner_id = gtid + 1;
2434   }
2435   return retval;
2436 }
2437 
2438 int __kmp_release_drdpa_lock(kmp_drdpa_lock_t *lck, kmp_int32 gtid) {
2439   // Read the ticket value from the lock data struct, then the polls pointer and
2440   // the mask.  The polls pointer must be read before the mask!!! (See above)
2441   kmp_uint64 ticket = lck->lk.now_serving + 1; // non-atomic load
2442   std::atomic<kmp_uint64> *polls = lck->lk.polls; // atomic load
2443   kmp_uint64 mask = lck->lk.mask; // atomic load
2444   KA_TRACE(1000, ("__kmp_release_drdpa_lock: ticket #%lld released lock %p\n",
2445                   ticket - 1, lck));
2446   KMP_FSYNC_RELEASING(lck);
2447   ANNOTATE_DRDPA_RELEASED(lck);
2448   polls[ticket & mask] = ticket; // atomic store
2449   return KMP_LOCK_RELEASED;
2450 }
2451 
2452 static int __kmp_release_drdpa_lock_with_checks(kmp_drdpa_lock_t *lck,
2453                                                 kmp_int32 gtid) {
2454   char const *const func = "omp_unset_lock";
2455   KMP_MB(); /* in case another processor initialized lock */
2456   if (lck->lk.initialized != lck) {
2457     KMP_FATAL(LockIsUninitialized, func);
2458   }
2459   if (__kmp_is_drdpa_lock_nestable(lck)) {
2460     KMP_FATAL(LockNestableUsedAsSimple, func);
2461   }
2462   if (__kmp_get_drdpa_lock_owner(lck) == -1) {
2463     KMP_FATAL(LockUnsettingFree, func);
2464   }
2465   if ((gtid >= 0) && (__kmp_get_drdpa_lock_owner(lck) >= 0) &&
2466       (__kmp_get_drdpa_lock_owner(lck) != gtid)) {
2467     KMP_FATAL(LockUnsettingSetByAnother, func);
2468   }
2469   lck->lk.owner_id = 0;
2470   return __kmp_release_drdpa_lock(lck, gtid);
2471 }
2472 
2473 void __kmp_init_drdpa_lock(kmp_drdpa_lock_t *lck) {
2474   lck->lk.location = NULL;
2475   lck->lk.mask = 0;
2476   lck->lk.num_polls = 1;
2477   lck->lk.polls = (std::atomic<kmp_uint64> *)__kmp_allocate(
2478       lck->lk.num_polls * sizeof(*(lck->lk.polls)));
2479   lck->lk.cleanup_ticket = 0;
2480   lck->lk.old_polls = NULL;
2481   lck->lk.next_ticket = 0;
2482   lck->lk.now_serving = 0;
2483   lck->lk.owner_id = 0; // no thread owns the lock.
2484   lck->lk.depth_locked = -1; // >= 0 for nestable locks, -1 for simple locks.
2485   lck->lk.initialized = lck;
2486 
2487   KA_TRACE(1000, ("__kmp_init_drdpa_lock: lock %p initialized\n", lck));
2488 }
2489 
2490 void __kmp_destroy_drdpa_lock(kmp_drdpa_lock_t *lck) {
2491   lck->lk.initialized = NULL;
2492   lck->lk.location = NULL;
2493   if (lck->lk.polls.load() != NULL) {
2494     __kmp_free(lck->lk.polls.load());
2495     lck->lk.polls = NULL;
2496   }
2497   if (lck->lk.old_polls != NULL) {
2498     __kmp_free(lck->lk.old_polls);
2499     lck->lk.old_polls = NULL;
2500   }
2501   lck->lk.mask = 0;
2502   lck->lk.num_polls = 0;
2503   lck->lk.cleanup_ticket = 0;
2504   lck->lk.next_ticket = 0;
2505   lck->lk.now_serving = 0;
2506   lck->lk.owner_id = 0;
2507   lck->lk.depth_locked = -1;
2508 }
2509 
2510 static void __kmp_destroy_drdpa_lock_with_checks(kmp_drdpa_lock_t *lck) {
2511   char const *const func = "omp_destroy_lock";
2512   if (lck->lk.initialized != lck) {
2513     KMP_FATAL(LockIsUninitialized, func);
2514   }
2515   if (__kmp_is_drdpa_lock_nestable(lck)) {
2516     KMP_FATAL(LockNestableUsedAsSimple, func);
2517   }
2518   if (__kmp_get_drdpa_lock_owner(lck) != -1) {
2519     KMP_FATAL(LockStillOwned, func);
2520   }
2521   __kmp_destroy_drdpa_lock(lck);
2522 }
2523 
2524 // nested drdpa ticket locks
2525 
2526 int __kmp_acquire_nested_drdpa_lock(kmp_drdpa_lock_t *lck, kmp_int32 gtid) {
2527   KMP_DEBUG_ASSERT(gtid >= 0);
2528 
2529   if (__kmp_get_drdpa_lock_owner(lck) == gtid) {
2530     lck->lk.depth_locked += 1;
2531     return KMP_LOCK_ACQUIRED_NEXT;
2532   } else {
2533     __kmp_acquire_drdpa_lock_timed_template(lck, gtid);
2534     ANNOTATE_DRDPA_ACQUIRED(lck);
2535     KMP_MB();
2536     lck->lk.depth_locked = 1;
2537     KMP_MB();
2538     lck->lk.owner_id = gtid + 1;
2539     return KMP_LOCK_ACQUIRED_FIRST;
2540   }
2541 }
2542 
2543 static void __kmp_acquire_nested_drdpa_lock_with_checks(kmp_drdpa_lock_t *lck,
2544                                                         kmp_int32 gtid) {
2545   char const *const func = "omp_set_nest_lock";
2546   if (lck->lk.initialized != lck) {
2547     KMP_FATAL(LockIsUninitialized, func);
2548   }
2549   if (!__kmp_is_drdpa_lock_nestable(lck)) {
2550     KMP_FATAL(LockSimpleUsedAsNestable, func);
2551   }
2552   __kmp_acquire_nested_drdpa_lock(lck, gtid);
2553 }
2554 
2555 int __kmp_test_nested_drdpa_lock(kmp_drdpa_lock_t *lck, kmp_int32 gtid) {
2556   int retval;
2557 
2558   KMP_DEBUG_ASSERT(gtid >= 0);
2559 
2560   if (__kmp_get_drdpa_lock_owner(lck) == gtid) {
2561     retval = ++lck->lk.depth_locked;
2562   } else if (!__kmp_test_drdpa_lock(lck, gtid)) {
2563     retval = 0;
2564   } else {
2565     KMP_MB();
2566     retval = lck->lk.depth_locked = 1;
2567     KMP_MB();
2568     lck->lk.owner_id = gtid + 1;
2569   }
2570   return retval;
2571 }
2572 
2573 static int __kmp_test_nested_drdpa_lock_with_checks(kmp_drdpa_lock_t *lck,
2574                                                     kmp_int32 gtid) {
2575   char const *const func = "omp_test_nest_lock";
2576   if (lck->lk.initialized != lck) {
2577     KMP_FATAL(LockIsUninitialized, func);
2578   }
2579   if (!__kmp_is_drdpa_lock_nestable(lck)) {
2580     KMP_FATAL(LockSimpleUsedAsNestable, func);
2581   }
2582   return __kmp_test_nested_drdpa_lock(lck, gtid);
2583 }
2584 
2585 int __kmp_release_nested_drdpa_lock(kmp_drdpa_lock_t *lck, kmp_int32 gtid) {
2586   KMP_DEBUG_ASSERT(gtid >= 0);
2587 
2588   KMP_MB();
2589   if (--(lck->lk.depth_locked) == 0) {
2590     KMP_MB();
2591     lck->lk.owner_id = 0;
2592     __kmp_release_drdpa_lock(lck, gtid);
2593     return KMP_LOCK_RELEASED;
2594   }
2595   return KMP_LOCK_STILL_HELD;
2596 }
2597 
2598 static int __kmp_release_nested_drdpa_lock_with_checks(kmp_drdpa_lock_t *lck,
2599                                                        kmp_int32 gtid) {
2600   char const *const func = "omp_unset_nest_lock";
2601   KMP_MB(); /* in case another processor initialized lock */
2602   if (lck->lk.initialized != lck) {
2603     KMP_FATAL(LockIsUninitialized, func);
2604   }
2605   if (!__kmp_is_drdpa_lock_nestable(lck)) {
2606     KMP_FATAL(LockSimpleUsedAsNestable, func);
2607   }
2608   if (__kmp_get_drdpa_lock_owner(lck) == -1) {
2609     KMP_FATAL(LockUnsettingFree, func);
2610   }
2611   if (__kmp_get_drdpa_lock_owner(lck) != gtid) {
2612     KMP_FATAL(LockUnsettingSetByAnother, func);
2613   }
2614   return __kmp_release_nested_drdpa_lock(lck, gtid);
2615 }
2616 
2617 void __kmp_init_nested_drdpa_lock(kmp_drdpa_lock_t *lck) {
2618   __kmp_init_drdpa_lock(lck);
2619   lck->lk.depth_locked = 0; // >= 0 for nestable locks, -1 for simple locks
2620 }
2621 
2622 void __kmp_destroy_nested_drdpa_lock(kmp_drdpa_lock_t *lck) {
2623   __kmp_destroy_drdpa_lock(lck);
2624   lck->lk.depth_locked = 0;
2625 }
2626 
2627 static void __kmp_destroy_nested_drdpa_lock_with_checks(kmp_drdpa_lock_t *lck) {
2628   char const *const func = "omp_destroy_nest_lock";
2629   if (lck->lk.initialized != lck) {
2630     KMP_FATAL(LockIsUninitialized, func);
2631   }
2632   if (!__kmp_is_drdpa_lock_nestable(lck)) {
2633     KMP_FATAL(LockSimpleUsedAsNestable, func);
2634   }
2635   if (__kmp_get_drdpa_lock_owner(lck) != -1) {
2636     KMP_FATAL(LockStillOwned, func);
2637   }
2638   __kmp_destroy_nested_drdpa_lock(lck);
2639 }
2640 
2641 // access functions to fields which don't exist for all lock kinds.
2642 
2643 static const ident_t *__kmp_get_drdpa_lock_location(kmp_drdpa_lock_t *lck) {
2644   return lck->lk.location;
2645 }
2646 
2647 static void __kmp_set_drdpa_lock_location(kmp_drdpa_lock_t *lck,
2648                                           const ident_t *loc) {
2649   lck->lk.location = loc;
2650 }
2651 
2652 static kmp_lock_flags_t __kmp_get_drdpa_lock_flags(kmp_drdpa_lock_t *lck) {
2653   return lck->lk.flags;
2654 }
2655 
2656 static void __kmp_set_drdpa_lock_flags(kmp_drdpa_lock_t *lck,
2657                                        kmp_lock_flags_t flags) {
2658   lck->lk.flags = flags;
2659 }
2660 
2661 // Time stamp counter
2662 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
2663 #define __kmp_tsc() __kmp_hardware_timestamp()
2664 // Runtime's default backoff parameters
2665 kmp_backoff_t __kmp_spin_backoff_params = {1, 4096, 100};
2666 #else
2667 // Use nanoseconds for other platforms
2668 extern kmp_uint64 __kmp_now_nsec();
2669 kmp_backoff_t __kmp_spin_backoff_params = {1, 256, 100};
2670 #define __kmp_tsc() __kmp_now_nsec()
2671 #endif
2672 
2673 // A useful predicate for dealing with timestamps that may wrap.
2674 // Is a before b? Since the timestamps may wrap, this is asking whether it's
2675 // shorter to go clockwise from a to b around the clock-face, or anti-clockwise.
2676 // Times where going clockwise is less distance than going anti-clockwise
2677 // are in the future, others are in the past. e.g. a = MAX-1, b = MAX+1 (=0),
2678 // then a > b (true) does not mean a reached b; whereas signed(a) = -2,
2679 // signed(b) = 0 captures the actual difference
2680 static inline bool before(kmp_uint64 a, kmp_uint64 b) {
2681   return ((kmp_int64)b - (kmp_int64)a) > 0;
2682 }
2683 
2684 // Truncated binary exponential backoff function
2685 void __kmp_spin_backoff(kmp_backoff_t *boff) {
2686   // We could flatten this loop, but making it a nested loop gives better result
2687   kmp_uint32 i;
2688   for (i = boff->step; i > 0; i--) {
2689     kmp_uint64 goal = __kmp_tsc() + boff->min_tick;
2690     do {
2691       KMP_CPU_PAUSE();
2692     } while (before(__kmp_tsc(), goal));
2693   }
2694   boff->step = (boff->step << 1 | 1) & (boff->max_backoff - 1);
2695 }
2696 
2697 #if KMP_USE_DYNAMIC_LOCK
2698 
2699 // Direct lock initializers. It simply writes a tag to the low 8 bits of the
2700 // lock word.
2701 static void __kmp_init_direct_lock(kmp_dyna_lock_t *lck,
2702                                    kmp_dyna_lockseq_t seq) {
2703   TCW_4(*lck, KMP_GET_D_TAG(seq));
2704   KA_TRACE(
2705       20,
2706       ("__kmp_init_direct_lock: initialized direct lock with type#%d\n", seq));
2707 }
2708 
2709 #if KMP_USE_TSX
2710 
2711 // HLE lock functions - imported from the testbed runtime.
2712 #define HLE_ACQUIRE ".byte 0xf2;"
2713 #define HLE_RELEASE ".byte 0xf3;"
2714 
2715 static inline kmp_uint32 swap4(kmp_uint32 volatile *p, kmp_uint32 v) {
2716   __asm__ volatile(HLE_ACQUIRE "xchg %1,%0" : "+r"(v), "+m"(*p) : : "memory");
2717   return v;
2718 }
2719 
2720 static void __kmp_destroy_hle_lock(kmp_dyna_lock_t *lck) { TCW_4(*lck, 0); }
2721 
2722 static void __kmp_destroy_hle_lock_with_checks(kmp_dyna_lock_t *lck) {
2723   TCW_4(*lck, 0);
2724 }
2725 
2726 static void __kmp_acquire_hle_lock(kmp_dyna_lock_t *lck, kmp_int32 gtid) {
2727   // Use gtid for KMP_LOCK_BUSY if necessary
2728   if (swap4(lck, KMP_LOCK_BUSY(1, hle)) != KMP_LOCK_FREE(hle)) {
2729     int delay = 1;
2730     do {
2731       while (*(kmp_uint32 volatile *)lck != KMP_LOCK_FREE(hle)) {
2732         for (int i = delay; i != 0; --i)
2733           KMP_CPU_PAUSE();
2734         delay = ((delay << 1) | 1) & 7;
2735       }
2736     } while (swap4(lck, KMP_LOCK_BUSY(1, hle)) != KMP_LOCK_FREE(hle));
2737   }
2738 }
2739 
2740 static void __kmp_acquire_hle_lock_with_checks(kmp_dyna_lock_t *lck,
2741                                                kmp_int32 gtid) {
2742   __kmp_acquire_hle_lock(lck, gtid); // TODO: add checks
2743 }
2744 
2745 static int __kmp_release_hle_lock(kmp_dyna_lock_t *lck, kmp_int32 gtid) {
2746   __asm__ volatile(HLE_RELEASE "movl %1,%0"
2747                    : "=m"(*lck)
2748                    : "r"(KMP_LOCK_FREE(hle))
2749                    : "memory");
2750   return KMP_LOCK_RELEASED;
2751 }
2752 
2753 static int __kmp_release_hle_lock_with_checks(kmp_dyna_lock_t *lck,
2754                                               kmp_int32 gtid) {
2755   return __kmp_release_hle_lock(lck, gtid); // TODO: add checks
2756 }
2757 
2758 static int __kmp_test_hle_lock(kmp_dyna_lock_t *lck, kmp_int32 gtid) {
2759   return swap4(lck, KMP_LOCK_BUSY(1, hle)) == KMP_LOCK_FREE(hle);
2760 }
2761 
2762 static int __kmp_test_hle_lock_with_checks(kmp_dyna_lock_t *lck,
2763                                            kmp_int32 gtid) {
2764   return __kmp_test_hle_lock(lck, gtid); // TODO: add checks
2765 }
2766 
2767 static void __kmp_init_rtm_lock(kmp_queuing_lock_t *lck) {
2768   __kmp_init_queuing_lock(lck);
2769 }
2770 
2771 static void __kmp_destroy_rtm_lock(kmp_queuing_lock_t *lck) {
2772   __kmp_destroy_queuing_lock(lck);
2773 }
2774 
2775 static void __kmp_destroy_rtm_lock_with_checks(kmp_queuing_lock_t *lck) {
2776   __kmp_destroy_queuing_lock_with_checks(lck);
2777 }
2778 
2779 static void __kmp_acquire_rtm_lock(kmp_queuing_lock_t *lck, kmp_int32 gtid) {
2780   unsigned retries = 3, status;
2781   do {
2782     status = _xbegin();
2783     if (status == _XBEGIN_STARTED) {
2784       if (__kmp_is_unlocked_queuing_lock(lck))
2785         return;
2786       _xabort(0xff);
2787     }
2788     if ((status & _XABORT_EXPLICIT) && _XABORT_CODE(status) == 0xff) {
2789       // Wait until lock becomes free
2790       while (!__kmp_is_unlocked_queuing_lock(lck)) {
2791         KMP_YIELD(TRUE);
2792       }
2793     } else if (!(status & _XABORT_RETRY))
2794       break;
2795   } while (retries--);
2796 
2797   // Fall-back non-speculative lock (xchg)
2798   __kmp_acquire_queuing_lock(lck, gtid);
2799 }
2800 
2801 static void __kmp_acquire_rtm_lock_with_checks(kmp_queuing_lock_t *lck,
2802                                                kmp_int32 gtid) {
2803   __kmp_acquire_rtm_lock(lck, gtid);
2804 }
2805 
2806 static int __kmp_release_rtm_lock(kmp_queuing_lock_t *lck, kmp_int32 gtid) {
2807   if (__kmp_is_unlocked_queuing_lock(lck)) {
2808     // Releasing from speculation
2809     _xend();
2810   } else {
2811     // Releasing from a real lock
2812     __kmp_release_queuing_lock(lck, gtid);
2813   }
2814   return KMP_LOCK_RELEASED;
2815 }
2816 
2817 static int __kmp_release_rtm_lock_with_checks(kmp_queuing_lock_t *lck,
2818                                               kmp_int32 gtid) {
2819   return __kmp_release_rtm_lock(lck, gtid);
2820 }
2821 
2822 static int __kmp_test_rtm_lock(kmp_queuing_lock_t *lck, kmp_int32 gtid) {
2823   unsigned retries = 3, status;
2824   do {
2825     status = _xbegin();
2826     if (status == _XBEGIN_STARTED && __kmp_is_unlocked_queuing_lock(lck)) {
2827       return 1;
2828     }
2829     if (!(status & _XABORT_RETRY))
2830       break;
2831   } while (retries--);
2832 
2833   return (__kmp_is_unlocked_queuing_lock(lck)) ? 1 : 0;
2834 }
2835 
2836 static int __kmp_test_rtm_lock_with_checks(kmp_queuing_lock_t *lck,
2837                                            kmp_int32 gtid) {
2838   return __kmp_test_rtm_lock(lck, gtid);
2839 }
2840 
2841 #endif // KMP_USE_TSX
2842 
2843 // Entry functions for indirect locks (first element of direct lock jump tables)
2844 static void __kmp_init_indirect_lock(kmp_dyna_lock_t *l,
2845                                      kmp_dyna_lockseq_t tag);
2846 static void __kmp_destroy_indirect_lock(kmp_dyna_lock_t *lock);
2847 static int __kmp_set_indirect_lock(kmp_dyna_lock_t *lock, kmp_int32);
2848 static int __kmp_unset_indirect_lock(kmp_dyna_lock_t *lock, kmp_int32);
2849 static int __kmp_test_indirect_lock(kmp_dyna_lock_t *lock, kmp_int32);
2850 static int __kmp_set_indirect_lock_with_checks(kmp_dyna_lock_t *lock,
2851                                                kmp_int32);
2852 static int __kmp_unset_indirect_lock_with_checks(kmp_dyna_lock_t *lock,
2853                                                  kmp_int32);
2854 static int __kmp_test_indirect_lock_with_checks(kmp_dyna_lock_t *lock,
2855                                                 kmp_int32);
2856 
2857 // Lock function definitions for the union parameter type
2858 #define KMP_FOREACH_LOCK_KIND(m, a) m(ticket, a) m(queuing, a) m(drdpa, a)
2859 
2860 #define expand1(lk, op)                                                        \
2861   static void __kmp_##op##_##lk##_##lock(kmp_user_lock_p lock) {               \
2862     __kmp_##op##_##lk##_##lock(&lock->lk);                                     \
2863   }
2864 #define expand2(lk, op)                                                        \
2865   static int __kmp_##op##_##lk##_##lock(kmp_user_lock_p lock,                  \
2866                                         kmp_int32 gtid) {                      \
2867     return __kmp_##op##_##lk##_##lock(&lock->lk, gtid);                        \
2868   }
2869 #define expand3(lk, op)                                                        \
2870   static void __kmp_set_##lk##_##lock_flags(kmp_user_lock_p lock,              \
2871                                             kmp_lock_flags_t flags) {          \
2872     __kmp_set_##lk##_lock_flags(&lock->lk, flags);                             \
2873   }
2874 #define expand4(lk, op)                                                        \
2875   static void __kmp_set_##lk##_##lock_location(kmp_user_lock_p lock,           \
2876                                                const ident_t *loc) {           \
2877     __kmp_set_##lk##_lock_location(&lock->lk, loc);                            \
2878   }
2879 
2880 KMP_FOREACH_LOCK_KIND(expand1, init)
2881 KMP_FOREACH_LOCK_KIND(expand1, init_nested)
2882 KMP_FOREACH_LOCK_KIND(expand1, destroy)
2883 KMP_FOREACH_LOCK_KIND(expand1, destroy_nested)
2884 KMP_FOREACH_LOCK_KIND(expand2, acquire)
2885 KMP_FOREACH_LOCK_KIND(expand2, acquire_nested)
2886 KMP_FOREACH_LOCK_KIND(expand2, release)
2887 KMP_FOREACH_LOCK_KIND(expand2, release_nested)
2888 KMP_FOREACH_LOCK_KIND(expand2, test)
2889 KMP_FOREACH_LOCK_KIND(expand2, test_nested)
2890 KMP_FOREACH_LOCK_KIND(expand3, )
2891 KMP_FOREACH_LOCK_KIND(expand4, )
2892 
2893 #undef expand1
2894 #undef expand2
2895 #undef expand3
2896 #undef expand4
2897 
2898 // Jump tables for the indirect lock functions
2899 // Only fill in the odd entries, that avoids the need to shift out the low bit
2900 
2901 // init functions
2902 #define expand(l, op) 0, __kmp_init_direct_lock,
2903 void (*__kmp_direct_init[])(kmp_dyna_lock_t *, kmp_dyna_lockseq_t) = {
2904     __kmp_init_indirect_lock, 0, KMP_FOREACH_D_LOCK(expand, init)};
2905 #undef expand
2906 
2907 // destroy functions
2908 #define expand(l, op) 0, (void (*)(kmp_dyna_lock_t *))__kmp_##op##_##l##_lock,
2909 static void (*direct_destroy[])(kmp_dyna_lock_t *) = {
2910     __kmp_destroy_indirect_lock, 0, KMP_FOREACH_D_LOCK(expand, destroy)};
2911 #undef expand
2912 #define expand(l, op)                                                          \
2913   0, (void (*)(kmp_dyna_lock_t *))__kmp_destroy_##l##_lock_with_checks,
2914 static void (*direct_destroy_check[])(kmp_dyna_lock_t *) = {
2915     __kmp_destroy_indirect_lock, 0, KMP_FOREACH_D_LOCK(expand, destroy)};
2916 #undef expand
2917 
2918 // set/acquire functions
2919 #define expand(l, op)                                                          \
2920   0, (int (*)(kmp_dyna_lock_t *, kmp_int32))__kmp_##op##_##l##_lock,
2921 static int (*direct_set[])(kmp_dyna_lock_t *, kmp_int32) = {
2922     __kmp_set_indirect_lock, 0, KMP_FOREACH_D_LOCK(expand, acquire)};
2923 #undef expand
2924 #define expand(l, op)                                                          \
2925   0, (int (*)(kmp_dyna_lock_t *, kmp_int32))__kmp_##op##_##l##_lock_with_checks,
2926 static int (*direct_set_check[])(kmp_dyna_lock_t *, kmp_int32) = {
2927     __kmp_set_indirect_lock_with_checks, 0,
2928     KMP_FOREACH_D_LOCK(expand, acquire)};
2929 #undef expand
2930 
2931 // unset/release and test functions
2932 #define expand(l, op)                                                          \
2933   0, (int (*)(kmp_dyna_lock_t *, kmp_int32))__kmp_##op##_##l##_lock,
2934 static int (*direct_unset[])(kmp_dyna_lock_t *, kmp_int32) = {
2935     __kmp_unset_indirect_lock, 0, KMP_FOREACH_D_LOCK(expand, release)};
2936 static int (*direct_test[])(kmp_dyna_lock_t *, kmp_int32) = {
2937     __kmp_test_indirect_lock, 0, KMP_FOREACH_D_LOCK(expand, test)};
2938 #undef expand
2939 #define expand(l, op)                                                          \
2940   0, (int (*)(kmp_dyna_lock_t *, kmp_int32))__kmp_##op##_##l##_lock_with_checks,
2941 static int (*direct_unset_check[])(kmp_dyna_lock_t *, kmp_int32) = {
2942     __kmp_unset_indirect_lock_with_checks, 0,
2943     KMP_FOREACH_D_LOCK(expand, release)};
2944 static int (*direct_test_check[])(kmp_dyna_lock_t *, kmp_int32) = {
2945     __kmp_test_indirect_lock_with_checks, 0, KMP_FOREACH_D_LOCK(expand, test)};
2946 #undef expand
2947 
2948 // Exposes only one set of jump tables (*lock or *lock_with_checks).
2949 void (**__kmp_direct_destroy)(kmp_dyna_lock_t *) = 0;
2950 int (**__kmp_direct_set)(kmp_dyna_lock_t *, kmp_int32) = 0;
2951 int (**__kmp_direct_unset)(kmp_dyna_lock_t *, kmp_int32) = 0;
2952 int (**__kmp_direct_test)(kmp_dyna_lock_t *, kmp_int32) = 0;
2953 
2954 // Jump tables for the indirect lock functions
2955 #define expand(l, op) (void (*)(kmp_user_lock_p)) __kmp_##op##_##l##_##lock,
2956 void (*__kmp_indirect_init[])(kmp_user_lock_p) = {
2957     KMP_FOREACH_I_LOCK(expand, init)};
2958 #undef expand
2959 
2960 #define expand(l, op) (void (*)(kmp_user_lock_p)) __kmp_##op##_##l##_##lock,
2961 static void (*indirect_destroy[])(kmp_user_lock_p) = {
2962     KMP_FOREACH_I_LOCK(expand, destroy)};
2963 #undef expand
2964 #define expand(l, op)                                                          \
2965   (void (*)(kmp_user_lock_p)) __kmp_##op##_##l##_##lock_with_checks,
2966 static void (*indirect_destroy_check[])(kmp_user_lock_p) = {
2967     KMP_FOREACH_I_LOCK(expand, destroy)};
2968 #undef expand
2969 
2970 // set/acquire functions
2971 #define expand(l, op)                                                          \
2972   (int (*)(kmp_user_lock_p, kmp_int32)) __kmp_##op##_##l##_##lock,
2973 static int (*indirect_set[])(kmp_user_lock_p,
2974                              kmp_int32) = {KMP_FOREACH_I_LOCK(expand, acquire)};
2975 #undef expand
2976 #define expand(l, op)                                                          \
2977   (int (*)(kmp_user_lock_p, kmp_int32)) __kmp_##op##_##l##_##lock_with_checks,
2978 static int (*indirect_set_check[])(kmp_user_lock_p, kmp_int32) = {
2979     KMP_FOREACH_I_LOCK(expand, acquire)};
2980 #undef expand
2981 
2982 // unset/release and test functions
2983 #define expand(l, op)                                                          \
2984   (int (*)(kmp_user_lock_p, kmp_int32)) __kmp_##op##_##l##_##lock,
2985 static int (*indirect_unset[])(kmp_user_lock_p, kmp_int32) = {
2986     KMP_FOREACH_I_LOCK(expand, release)};
2987 static int (*indirect_test[])(kmp_user_lock_p,
2988                               kmp_int32) = {KMP_FOREACH_I_LOCK(expand, test)};
2989 #undef expand
2990 #define expand(l, op)                                                          \
2991   (int (*)(kmp_user_lock_p, kmp_int32)) __kmp_##op##_##l##_##lock_with_checks,
2992 static int (*indirect_unset_check[])(kmp_user_lock_p, kmp_int32) = {
2993     KMP_FOREACH_I_LOCK(expand, release)};
2994 static int (*indirect_test_check[])(kmp_user_lock_p, kmp_int32) = {
2995     KMP_FOREACH_I_LOCK(expand, test)};
2996 #undef expand
2997 
2998 // Exposes only one jump tables (*lock or *lock_with_checks).
2999 void (**__kmp_indirect_destroy)(kmp_user_lock_p) = 0;
3000 int (**__kmp_indirect_set)(kmp_user_lock_p, kmp_int32) = 0;
3001 int (**__kmp_indirect_unset)(kmp_user_lock_p, kmp_int32) = 0;
3002 int (**__kmp_indirect_test)(kmp_user_lock_p, kmp_int32) = 0;
3003 
3004 // Lock index table.
3005 kmp_indirect_lock_table_t __kmp_i_lock_table;
3006 
3007 // Size of indirect locks.
3008 static kmp_uint32 __kmp_indirect_lock_size[KMP_NUM_I_LOCKS] = {0};
3009 
3010 // Jump tables for lock accessor/modifier.
3011 void (*__kmp_indirect_set_location[KMP_NUM_I_LOCKS])(kmp_user_lock_p,
3012                                                      const ident_t *) = {0};
3013 void (*__kmp_indirect_set_flags[KMP_NUM_I_LOCKS])(kmp_user_lock_p,
3014                                                   kmp_lock_flags_t) = {0};
3015 const ident_t *(*__kmp_indirect_get_location[KMP_NUM_I_LOCKS])(
3016     kmp_user_lock_p) = {0};
3017 kmp_lock_flags_t (*__kmp_indirect_get_flags[KMP_NUM_I_LOCKS])(
3018     kmp_user_lock_p) = {0};
3019 
3020 // Use different lock pools for different lock types.
3021 static kmp_indirect_lock_t *__kmp_indirect_lock_pool[KMP_NUM_I_LOCKS] = {0};
3022 
3023 // User lock allocator for dynamically dispatched indirect locks. Every entry of
3024 // the indirect lock table holds the address and type of the allocated indirect
3025 // lock (kmp_indirect_lock_t), and the size of the table doubles when it is
3026 // full. A destroyed indirect lock object is returned to the reusable pool of
3027 // locks, unique to each lock type.
3028 kmp_indirect_lock_t *__kmp_allocate_indirect_lock(void **user_lock,
3029                                                   kmp_int32 gtid,
3030                                                   kmp_indirect_locktag_t tag) {
3031   kmp_indirect_lock_t *lck;
3032   kmp_lock_index_t idx;
3033 
3034   __kmp_acquire_lock(&__kmp_global_lock, gtid);
3035 
3036   if (__kmp_indirect_lock_pool[tag] != NULL) {
3037     // Reuse the allocated and destroyed lock object
3038     lck = __kmp_indirect_lock_pool[tag];
3039     if (OMP_LOCK_T_SIZE < sizeof(void *))
3040       idx = lck->lock->pool.index;
3041     __kmp_indirect_lock_pool[tag] = (kmp_indirect_lock_t *)lck->lock->pool.next;
3042     KA_TRACE(20, ("__kmp_allocate_indirect_lock: reusing an existing lock %p\n",
3043                   lck));
3044   } else {
3045     idx = __kmp_i_lock_table.next;
3046     // Check capacity and double the size if it is full
3047     if (idx == __kmp_i_lock_table.size) {
3048       // Double up the space for block pointers
3049       int row = __kmp_i_lock_table.size / KMP_I_LOCK_CHUNK;
3050       kmp_indirect_lock_t **new_table = (kmp_indirect_lock_t **)__kmp_allocate(
3051           2 * row * sizeof(kmp_indirect_lock_t *));
3052       KMP_MEMCPY(new_table, __kmp_i_lock_table.table,
3053                  row * sizeof(kmp_indirect_lock_t *));
3054       kmp_indirect_lock_t **old_table = __kmp_i_lock_table.table;
3055       __kmp_i_lock_table.table = new_table;
3056       __kmp_free(old_table);
3057       // Allocate new objects in the new blocks
3058       for (int i = row; i < 2 * row; ++i)
3059         *(__kmp_i_lock_table.table + i) = (kmp_indirect_lock_t *)__kmp_allocate(
3060             KMP_I_LOCK_CHUNK * sizeof(kmp_indirect_lock_t));
3061       __kmp_i_lock_table.size = 2 * idx;
3062     }
3063     __kmp_i_lock_table.next++;
3064     lck = KMP_GET_I_LOCK(idx);
3065     // Allocate a new base lock object
3066     lck->lock = (kmp_user_lock_p)__kmp_allocate(__kmp_indirect_lock_size[tag]);
3067     KA_TRACE(20,
3068              ("__kmp_allocate_indirect_lock: allocated a new lock %p\n", lck));
3069   }
3070 
3071   __kmp_release_lock(&__kmp_global_lock, gtid);
3072 
3073   lck->type = tag;
3074 
3075   if (OMP_LOCK_T_SIZE < sizeof(void *)) {
3076     *((kmp_lock_index_t *)user_lock) = idx
3077                                        << 1; // indirect lock word must be even
3078   } else {
3079     *((kmp_indirect_lock_t **)user_lock) = lck;
3080   }
3081 
3082   return lck;
3083 }
3084 
3085 // User lock lookup for dynamically dispatched locks.
3086 static __forceinline kmp_indirect_lock_t *
3087 __kmp_lookup_indirect_lock(void **user_lock, const char *func) {
3088   if (__kmp_env_consistency_check) {
3089     kmp_indirect_lock_t *lck = NULL;
3090     if (user_lock == NULL) {
3091       KMP_FATAL(LockIsUninitialized, func);
3092     }
3093     if (OMP_LOCK_T_SIZE < sizeof(void *)) {
3094       kmp_lock_index_t idx = KMP_EXTRACT_I_INDEX(user_lock);
3095       if (idx >= __kmp_i_lock_table.size) {
3096         KMP_FATAL(LockIsUninitialized, func);
3097       }
3098       lck = KMP_GET_I_LOCK(idx);
3099     } else {
3100       lck = *((kmp_indirect_lock_t **)user_lock);
3101     }
3102     if (lck == NULL) {
3103       KMP_FATAL(LockIsUninitialized, func);
3104     }
3105     return lck;
3106   } else {
3107     if (OMP_LOCK_T_SIZE < sizeof(void *)) {
3108       return KMP_GET_I_LOCK(KMP_EXTRACT_I_INDEX(user_lock));
3109     } else {
3110       return *((kmp_indirect_lock_t **)user_lock);
3111     }
3112   }
3113 }
3114 
3115 static void __kmp_init_indirect_lock(kmp_dyna_lock_t *lock,
3116                                      kmp_dyna_lockseq_t seq) {
3117 #if KMP_USE_ADAPTIVE_LOCKS
3118   if (seq == lockseq_adaptive && !__kmp_cpuinfo.rtm) {
3119     KMP_WARNING(AdaptiveNotSupported, "kmp_lockseq_t", "adaptive");
3120     seq = lockseq_queuing;
3121   }
3122 #endif
3123 #if KMP_USE_TSX
3124   if (seq == lockseq_rtm && !__kmp_cpuinfo.rtm) {
3125     seq = lockseq_queuing;
3126   }
3127 #endif
3128   kmp_indirect_locktag_t tag = KMP_GET_I_TAG(seq);
3129   kmp_indirect_lock_t *l =
3130       __kmp_allocate_indirect_lock((void **)lock, __kmp_entry_gtid(), tag);
3131   KMP_I_LOCK_FUNC(l, init)(l->lock);
3132   KA_TRACE(
3133       20, ("__kmp_init_indirect_lock: initialized indirect lock with type#%d\n",
3134            seq));
3135 }
3136 
3137 static void __kmp_destroy_indirect_lock(kmp_dyna_lock_t *lock) {
3138   kmp_uint32 gtid = __kmp_entry_gtid();
3139   kmp_indirect_lock_t *l =
3140       __kmp_lookup_indirect_lock((void **)lock, "omp_destroy_lock");
3141   KMP_I_LOCK_FUNC(l, destroy)(l->lock);
3142   kmp_indirect_locktag_t tag = l->type;
3143 
3144   __kmp_acquire_lock(&__kmp_global_lock, gtid);
3145 
3146   // Use the base lock's space to keep the pool chain.
3147   l->lock->pool.next = (kmp_user_lock_p)__kmp_indirect_lock_pool[tag];
3148   if (OMP_LOCK_T_SIZE < sizeof(void *)) {
3149     l->lock->pool.index = KMP_EXTRACT_I_INDEX(lock);
3150   }
3151   __kmp_indirect_lock_pool[tag] = l;
3152 
3153   __kmp_release_lock(&__kmp_global_lock, gtid);
3154 }
3155 
3156 static int __kmp_set_indirect_lock(kmp_dyna_lock_t *lock, kmp_int32 gtid) {
3157   kmp_indirect_lock_t *l = KMP_LOOKUP_I_LOCK(lock);
3158   return KMP_I_LOCK_FUNC(l, set)(l->lock, gtid);
3159 }
3160 
3161 static int __kmp_unset_indirect_lock(kmp_dyna_lock_t *lock, kmp_int32 gtid) {
3162   kmp_indirect_lock_t *l = KMP_LOOKUP_I_LOCK(lock);
3163   return KMP_I_LOCK_FUNC(l, unset)(l->lock, gtid);
3164 }
3165 
3166 static int __kmp_test_indirect_lock(kmp_dyna_lock_t *lock, kmp_int32 gtid) {
3167   kmp_indirect_lock_t *l = KMP_LOOKUP_I_LOCK(lock);
3168   return KMP_I_LOCK_FUNC(l, test)(l->lock, gtid);
3169 }
3170 
3171 static int __kmp_set_indirect_lock_with_checks(kmp_dyna_lock_t *lock,
3172                                                kmp_int32 gtid) {
3173   kmp_indirect_lock_t *l =
3174       __kmp_lookup_indirect_lock((void **)lock, "omp_set_lock");
3175   return KMP_I_LOCK_FUNC(l, set)(l->lock, gtid);
3176 }
3177 
3178 static int __kmp_unset_indirect_lock_with_checks(kmp_dyna_lock_t *lock,
3179                                                  kmp_int32 gtid) {
3180   kmp_indirect_lock_t *l =
3181       __kmp_lookup_indirect_lock((void **)lock, "omp_unset_lock");
3182   return KMP_I_LOCK_FUNC(l, unset)(l->lock, gtid);
3183 }
3184 
3185 static int __kmp_test_indirect_lock_with_checks(kmp_dyna_lock_t *lock,
3186                                                 kmp_int32 gtid) {
3187   kmp_indirect_lock_t *l =
3188       __kmp_lookup_indirect_lock((void **)lock, "omp_test_lock");
3189   return KMP_I_LOCK_FUNC(l, test)(l->lock, gtid);
3190 }
3191 
3192 kmp_dyna_lockseq_t __kmp_user_lock_seq = lockseq_queuing;
3193 
3194 // This is used only in kmp_error.cpp when consistency checking is on.
3195 kmp_int32 __kmp_get_user_lock_owner(kmp_user_lock_p lck, kmp_uint32 seq) {
3196   switch (seq) {
3197   case lockseq_tas:
3198   case lockseq_nested_tas:
3199     return __kmp_get_tas_lock_owner((kmp_tas_lock_t *)lck);
3200 #if KMP_USE_FUTEX
3201   case lockseq_futex:
3202   case lockseq_nested_futex:
3203     return __kmp_get_futex_lock_owner((kmp_futex_lock_t *)lck);
3204 #endif
3205   case lockseq_ticket:
3206   case lockseq_nested_ticket:
3207     return __kmp_get_ticket_lock_owner((kmp_ticket_lock_t *)lck);
3208   case lockseq_queuing:
3209   case lockseq_nested_queuing:
3210 #if KMP_USE_ADAPTIVE_LOCKS
3211   case lockseq_adaptive:
3212 #endif
3213     return __kmp_get_queuing_lock_owner((kmp_queuing_lock_t *)lck);
3214   case lockseq_drdpa:
3215   case lockseq_nested_drdpa:
3216     return __kmp_get_drdpa_lock_owner((kmp_drdpa_lock_t *)lck);
3217   default:
3218     return 0;
3219   }
3220 }
3221 
3222 // Initializes data for dynamic user locks.
3223 void __kmp_init_dynamic_user_locks() {
3224   // Initialize jump table for the lock functions
3225   if (__kmp_env_consistency_check) {
3226     __kmp_direct_set = direct_set_check;
3227     __kmp_direct_unset = direct_unset_check;
3228     __kmp_direct_test = direct_test_check;
3229     __kmp_direct_destroy = direct_destroy_check;
3230     __kmp_indirect_set = indirect_set_check;
3231     __kmp_indirect_unset = indirect_unset_check;
3232     __kmp_indirect_test = indirect_test_check;
3233     __kmp_indirect_destroy = indirect_destroy_check;
3234   } else {
3235     __kmp_direct_set = direct_set;
3236     __kmp_direct_unset = direct_unset;
3237     __kmp_direct_test = direct_test;
3238     __kmp_direct_destroy = direct_destroy;
3239     __kmp_indirect_set = indirect_set;
3240     __kmp_indirect_unset = indirect_unset;
3241     __kmp_indirect_test = indirect_test;
3242     __kmp_indirect_destroy = indirect_destroy;
3243   }
3244   // If the user locks have already been initialized, then return. Allow the
3245   // switch between different KMP_CONSISTENCY_CHECK values, but do not allocate
3246   // new lock tables if they have already been allocated.
3247   if (__kmp_init_user_locks)
3248     return;
3249 
3250   // Initialize lock index table
3251   __kmp_i_lock_table.size = KMP_I_LOCK_CHUNK;
3252   __kmp_i_lock_table.table =
3253       (kmp_indirect_lock_t **)__kmp_allocate(sizeof(kmp_indirect_lock_t *));
3254   *(__kmp_i_lock_table.table) = (kmp_indirect_lock_t *)__kmp_allocate(
3255       KMP_I_LOCK_CHUNK * sizeof(kmp_indirect_lock_t));
3256   __kmp_i_lock_table.next = 0;
3257 
3258   // Indirect lock size
3259   __kmp_indirect_lock_size[locktag_ticket] = sizeof(kmp_ticket_lock_t);
3260   __kmp_indirect_lock_size[locktag_queuing] = sizeof(kmp_queuing_lock_t);
3261 #if KMP_USE_ADAPTIVE_LOCKS
3262   __kmp_indirect_lock_size[locktag_adaptive] = sizeof(kmp_adaptive_lock_t);
3263 #endif
3264   __kmp_indirect_lock_size[locktag_drdpa] = sizeof(kmp_drdpa_lock_t);
3265 #if KMP_USE_TSX
3266   __kmp_indirect_lock_size[locktag_rtm] = sizeof(kmp_queuing_lock_t);
3267 #endif
3268   __kmp_indirect_lock_size[locktag_nested_tas] = sizeof(kmp_tas_lock_t);
3269 #if KMP_USE_FUTEX
3270   __kmp_indirect_lock_size[locktag_nested_futex] = sizeof(kmp_futex_lock_t);
3271 #endif
3272   __kmp_indirect_lock_size[locktag_nested_ticket] = sizeof(kmp_ticket_lock_t);
3273   __kmp_indirect_lock_size[locktag_nested_queuing] = sizeof(kmp_queuing_lock_t);
3274   __kmp_indirect_lock_size[locktag_nested_drdpa] = sizeof(kmp_drdpa_lock_t);
3275 
3276 // Initialize lock accessor/modifier
3277 #define fill_jumps(table, expand, sep)                                         \
3278   {                                                                            \
3279     table[locktag##sep##ticket] = expand(ticket);                              \
3280     table[locktag##sep##queuing] = expand(queuing);                            \
3281     table[locktag##sep##drdpa] = expand(drdpa);                                \
3282   }
3283 
3284 #if KMP_USE_ADAPTIVE_LOCKS
3285 #define fill_table(table, expand)                                              \
3286   {                                                                            \
3287     fill_jumps(table, expand, _);                                              \
3288     table[locktag_adaptive] = expand(queuing);                                 \
3289     fill_jumps(table, expand, _nested_);                                       \
3290   }
3291 #else
3292 #define fill_table(table, expand)                                              \
3293   {                                                                            \
3294     fill_jumps(table, expand, _);                                              \
3295     fill_jumps(table, expand, _nested_);                                       \
3296   }
3297 #endif // KMP_USE_ADAPTIVE_LOCKS
3298 
3299 #define expand(l)                                                              \
3300   (void (*)(kmp_user_lock_p, const ident_t *)) __kmp_set_##l##_lock_location
3301   fill_table(__kmp_indirect_set_location, expand);
3302 #undef expand
3303 #define expand(l)                                                              \
3304   (void (*)(kmp_user_lock_p, kmp_lock_flags_t)) __kmp_set_##l##_lock_flags
3305   fill_table(__kmp_indirect_set_flags, expand);
3306 #undef expand
3307 #define expand(l)                                                              \
3308   (const ident_t *(*)(kmp_user_lock_p)) __kmp_get_##l##_lock_location
3309   fill_table(__kmp_indirect_get_location, expand);
3310 #undef expand
3311 #define expand(l)                                                              \
3312   (kmp_lock_flags_t(*)(kmp_user_lock_p)) __kmp_get_##l##_lock_flags
3313   fill_table(__kmp_indirect_get_flags, expand);
3314 #undef expand
3315 
3316   __kmp_init_user_locks = TRUE;
3317 }
3318 
3319 // Clean up the lock table.
3320 void __kmp_cleanup_indirect_user_locks() {
3321   kmp_lock_index_t i;
3322   int k;
3323 
3324   // Clean up locks in the pools first (they were already destroyed before going
3325   // into the pools).
3326   for (k = 0; k < KMP_NUM_I_LOCKS; ++k) {
3327     kmp_indirect_lock_t *l = __kmp_indirect_lock_pool[k];
3328     while (l != NULL) {
3329       kmp_indirect_lock_t *ll = l;
3330       l = (kmp_indirect_lock_t *)l->lock->pool.next;
3331       KA_TRACE(20, ("__kmp_cleanup_indirect_user_locks: freeing %p from pool\n",
3332                     ll));
3333       __kmp_free(ll->lock);
3334       ll->lock = NULL;
3335     }
3336     __kmp_indirect_lock_pool[k] = NULL;
3337   }
3338   // Clean up the remaining undestroyed locks.
3339   for (i = 0; i < __kmp_i_lock_table.next; i++) {
3340     kmp_indirect_lock_t *l = KMP_GET_I_LOCK(i);
3341     if (l->lock != NULL) {
3342       // Locks not destroyed explicitly need to be destroyed here.
3343       KMP_I_LOCK_FUNC(l, destroy)(l->lock);
3344       KA_TRACE(
3345           20,
3346           ("__kmp_cleanup_indirect_user_locks: destroy/freeing %p from table\n",
3347            l));
3348       __kmp_free(l->lock);
3349     }
3350   }
3351   // Free the table
3352   for (i = 0; i < __kmp_i_lock_table.size / KMP_I_LOCK_CHUNK; i++)
3353     __kmp_free(__kmp_i_lock_table.table[i]);
3354   __kmp_free(__kmp_i_lock_table.table);
3355 
3356   __kmp_init_user_locks = FALSE;
3357 }
3358 
3359 enum kmp_lock_kind __kmp_user_lock_kind = lk_default;
3360 int __kmp_num_locks_in_block = 1; // FIXME - tune this value
3361 
3362 #else // KMP_USE_DYNAMIC_LOCK
3363 
3364 static void __kmp_init_tas_lock_with_checks(kmp_tas_lock_t *lck) {
3365   __kmp_init_tas_lock(lck);
3366 }
3367 
3368 static void __kmp_init_nested_tas_lock_with_checks(kmp_tas_lock_t *lck) {
3369   __kmp_init_nested_tas_lock(lck);
3370 }
3371 
3372 #if KMP_USE_FUTEX
3373 static void __kmp_init_futex_lock_with_checks(kmp_futex_lock_t *lck) {
3374   __kmp_init_futex_lock(lck);
3375 }
3376 
3377 static void __kmp_init_nested_futex_lock_with_checks(kmp_futex_lock_t *lck) {
3378   __kmp_init_nested_futex_lock(lck);
3379 }
3380 #endif
3381 
3382 static int __kmp_is_ticket_lock_initialized(kmp_ticket_lock_t *lck) {
3383   return lck == lck->lk.self;
3384 }
3385 
3386 static void __kmp_init_ticket_lock_with_checks(kmp_ticket_lock_t *lck) {
3387   __kmp_init_ticket_lock(lck);
3388 }
3389 
3390 static void __kmp_init_nested_ticket_lock_with_checks(kmp_ticket_lock_t *lck) {
3391   __kmp_init_nested_ticket_lock(lck);
3392 }
3393 
3394 static int __kmp_is_queuing_lock_initialized(kmp_queuing_lock_t *lck) {
3395   return lck == lck->lk.initialized;
3396 }
3397 
3398 static void __kmp_init_queuing_lock_with_checks(kmp_queuing_lock_t *lck) {
3399   __kmp_init_queuing_lock(lck);
3400 }
3401 
3402 static void
3403 __kmp_init_nested_queuing_lock_with_checks(kmp_queuing_lock_t *lck) {
3404   __kmp_init_nested_queuing_lock(lck);
3405 }
3406 
3407 #if KMP_USE_ADAPTIVE_LOCKS
3408 static void __kmp_init_adaptive_lock_with_checks(kmp_adaptive_lock_t *lck) {
3409   __kmp_init_adaptive_lock(lck);
3410 }
3411 #endif
3412 
3413 static int __kmp_is_drdpa_lock_initialized(kmp_drdpa_lock_t *lck) {
3414   return lck == lck->lk.initialized;
3415 }
3416 
3417 static void __kmp_init_drdpa_lock_with_checks(kmp_drdpa_lock_t *lck) {
3418   __kmp_init_drdpa_lock(lck);
3419 }
3420 
3421 static void __kmp_init_nested_drdpa_lock_with_checks(kmp_drdpa_lock_t *lck) {
3422   __kmp_init_nested_drdpa_lock(lck);
3423 }
3424 
3425 /* user locks
3426  * They are implemented as a table of function pointers which are set to the
3427  * lock functions of the appropriate kind, once that has been determined. */
3428 
3429 enum kmp_lock_kind __kmp_user_lock_kind = lk_default;
3430 
3431 size_t __kmp_base_user_lock_size = 0;
3432 size_t __kmp_user_lock_size = 0;
3433 
3434 kmp_int32 (*__kmp_get_user_lock_owner_)(kmp_user_lock_p lck) = NULL;
3435 int (*__kmp_acquire_user_lock_with_checks_)(kmp_user_lock_p lck,
3436                                             kmp_int32 gtid) = NULL;
3437 
3438 int (*__kmp_test_user_lock_with_checks_)(kmp_user_lock_p lck,
3439                                          kmp_int32 gtid) = NULL;
3440 int (*__kmp_release_user_lock_with_checks_)(kmp_user_lock_p lck,
3441                                             kmp_int32 gtid) = NULL;
3442 void (*__kmp_init_user_lock_with_checks_)(kmp_user_lock_p lck) = NULL;
3443 void (*__kmp_destroy_user_lock_)(kmp_user_lock_p lck) = NULL;
3444 void (*__kmp_destroy_user_lock_with_checks_)(kmp_user_lock_p lck) = NULL;
3445 int (*__kmp_acquire_nested_user_lock_with_checks_)(kmp_user_lock_p lck,
3446                                                    kmp_int32 gtid) = NULL;
3447 
3448 int (*__kmp_test_nested_user_lock_with_checks_)(kmp_user_lock_p lck,
3449                                                 kmp_int32 gtid) = NULL;
3450 int (*__kmp_release_nested_user_lock_with_checks_)(kmp_user_lock_p lck,
3451                                                    kmp_int32 gtid) = NULL;
3452 void (*__kmp_init_nested_user_lock_with_checks_)(kmp_user_lock_p lck) = NULL;
3453 void (*__kmp_destroy_nested_user_lock_with_checks_)(kmp_user_lock_p lck) = NULL;
3454 
3455 int (*__kmp_is_user_lock_initialized_)(kmp_user_lock_p lck) = NULL;
3456 const ident_t *(*__kmp_get_user_lock_location_)(kmp_user_lock_p lck) = NULL;
3457 void (*__kmp_set_user_lock_location_)(kmp_user_lock_p lck,
3458                                       const ident_t *loc) = NULL;
3459 kmp_lock_flags_t (*__kmp_get_user_lock_flags_)(kmp_user_lock_p lck) = NULL;
3460 void (*__kmp_set_user_lock_flags_)(kmp_user_lock_p lck,
3461                                    kmp_lock_flags_t flags) = NULL;
3462 
3463 void __kmp_set_user_lock_vptrs(kmp_lock_kind_t user_lock_kind) {
3464   switch (user_lock_kind) {
3465   case lk_default:
3466   default:
3467     KMP_ASSERT(0);
3468 
3469   case lk_tas: {
3470     __kmp_base_user_lock_size = sizeof(kmp_base_tas_lock_t);
3471     __kmp_user_lock_size = sizeof(kmp_tas_lock_t);
3472 
3473     __kmp_get_user_lock_owner_ =
3474         (kmp_int32(*)(kmp_user_lock_p))(&__kmp_get_tas_lock_owner);
3475 
3476     if (__kmp_env_consistency_check) {
3477       KMP_BIND_USER_LOCK_WITH_CHECKS(tas);
3478       KMP_BIND_NESTED_USER_LOCK_WITH_CHECKS(tas);
3479     } else {
3480       KMP_BIND_USER_LOCK(tas);
3481       KMP_BIND_NESTED_USER_LOCK(tas);
3482     }
3483 
3484     __kmp_destroy_user_lock_ =
3485         (void (*)(kmp_user_lock_p))(&__kmp_destroy_tas_lock);
3486 
3487     __kmp_is_user_lock_initialized_ = (int (*)(kmp_user_lock_p))NULL;
3488 
3489     __kmp_get_user_lock_location_ = (const ident_t *(*)(kmp_user_lock_p))NULL;
3490 
3491     __kmp_set_user_lock_location_ =
3492         (void (*)(kmp_user_lock_p, const ident_t *))NULL;
3493 
3494     __kmp_get_user_lock_flags_ = (kmp_lock_flags_t(*)(kmp_user_lock_p))NULL;
3495 
3496     __kmp_set_user_lock_flags_ =
3497         (void (*)(kmp_user_lock_p, kmp_lock_flags_t))NULL;
3498   } break;
3499 
3500 #if KMP_USE_FUTEX
3501 
3502   case lk_futex: {
3503     __kmp_base_user_lock_size = sizeof(kmp_base_futex_lock_t);
3504     __kmp_user_lock_size = sizeof(kmp_futex_lock_t);
3505 
3506     __kmp_get_user_lock_owner_ =
3507         (kmp_int32(*)(kmp_user_lock_p))(&__kmp_get_futex_lock_owner);
3508 
3509     if (__kmp_env_consistency_check) {
3510       KMP_BIND_USER_LOCK_WITH_CHECKS(futex);
3511       KMP_BIND_NESTED_USER_LOCK_WITH_CHECKS(futex);
3512     } else {
3513       KMP_BIND_USER_LOCK(futex);
3514       KMP_BIND_NESTED_USER_LOCK(futex);
3515     }
3516 
3517     __kmp_destroy_user_lock_ =
3518         (void (*)(kmp_user_lock_p))(&__kmp_destroy_futex_lock);
3519 
3520     __kmp_is_user_lock_initialized_ = (int (*)(kmp_user_lock_p))NULL;
3521 
3522     __kmp_get_user_lock_location_ = (const ident_t *(*)(kmp_user_lock_p))NULL;
3523 
3524     __kmp_set_user_lock_location_ =
3525         (void (*)(kmp_user_lock_p, const ident_t *))NULL;
3526 
3527     __kmp_get_user_lock_flags_ = (kmp_lock_flags_t(*)(kmp_user_lock_p))NULL;
3528 
3529     __kmp_set_user_lock_flags_ =
3530         (void (*)(kmp_user_lock_p, kmp_lock_flags_t))NULL;
3531   } break;
3532 
3533 #endif // KMP_USE_FUTEX
3534 
3535   case lk_ticket: {
3536     __kmp_base_user_lock_size = sizeof(kmp_base_ticket_lock_t);
3537     __kmp_user_lock_size = sizeof(kmp_ticket_lock_t);
3538 
3539     __kmp_get_user_lock_owner_ =
3540         (kmp_int32(*)(kmp_user_lock_p))(&__kmp_get_ticket_lock_owner);
3541 
3542     if (__kmp_env_consistency_check) {
3543       KMP_BIND_USER_LOCK_WITH_CHECKS(ticket);
3544       KMP_BIND_NESTED_USER_LOCK_WITH_CHECKS(ticket);
3545     } else {
3546       KMP_BIND_USER_LOCK(ticket);
3547       KMP_BIND_NESTED_USER_LOCK(ticket);
3548     }
3549 
3550     __kmp_destroy_user_lock_ =
3551         (void (*)(kmp_user_lock_p))(&__kmp_destroy_ticket_lock);
3552 
3553     __kmp_is_user_lock_initialized_ =
3554         (int (*)(kmp_user_lock_p))(&__kmp_is_ticket_lock_initialized);
3555 
3556     __kmp_get_user_lock_location_ =
3557         (const ident_t *(*)(kmp_user_lock_p))(&__kmp_get_ticket_lock_location);
3558 
3559     __kmp_set_user_lock_location_ = (void (*)(
3560         kmp_user_lock_p, const ident_t *))(&__kmp_set_ticket_lock_location);
3561 
3562     __kmp_get_user_lock_flags_ =
3563         (kmp_lock_flags_t(*)(kmp_user_lock_p))(&__kmp_get_ticket_lock_flags);
3564 
3565     __kmp_set_user_lock_flags_ = (void (*)(kmp_user_lock_p, kmp_lock_flags_t))(
3566         &__kmp_set_ticket_lock_flags);
3567   } break;
3568 
3569   case lk_queuing: {
3570     __kmp_base_user_lock_size = sizeof(kmp_base_queuing_lock_t);
3571     __kmp_user_lock_size = sizeof(kmp_queuing_lock_t);
3572 
3573     __kmp_get_user_lock_owner_ =
3574         (kmp_int32(*)(kmp_user_lock_p))(&__kmp_get_queuing_lock_owner);
3575 
3576     if (__kmp_env_consistency_check) {
3577       KMP_BIND_USER_LOCK_WITH_CHECKS(queuing);
3578       KMP_BIND_NESTED_USER_LOCK_WITH_CHECKS(queuing);
3579     } else {
3580       KMP_BIND_USER_LOCK(queuing);
3581       KMP_BIND_NESTED_USER_LOCK(queuing);
3582     }
3583 
3584     __kmp_destroy_user_lock_ =
3585         (void (*)(kmp_user_lock_p))(&__kmp_destroy_queuing_lock);
3586 
3587     __kmp_is_user_lock_initialized_ =
3588         (int (*)(kmp_user_lock_p))(&__kmp_is_queuing_lock_initialized);
3589 
3590     __kmp_get_user_lock_location_ =
3591         (const ident_t *(*)(kmp_user_lock_p))(&__kmp_get_queuing_lock_location);
3592 
3593     __kmp_set_user_lock_location_ = (void (*)(
3594         kmp_user_lock_p, const ident_t *))(&__kmp_set_queuing_lock_location);
3595 
3596     __kmp_get_user_lock_flags_ =
3597         (kmp_lock_flags_t(*)(kmp_user_lock_p))(&__kmp_get_queuing_lock_flags);
3598 
3599     __kmp_set_user_lock_flags_ = (void (*)(kmp_user_lock_p, kmp_lock_flags_t))(
3600         &__kmp_set_queuing_lock_flags);
3601   } break;
3602 
3603 #if KMP_USE_ADAPTIVE_LOCKS
3604   case lk_adaptive: {
3605     __kmp_base_user_lock_size = sizeof(kmp_base_adaptive_lock_t);
3606     __kmp_user_lock_size = sizeof(kmp_adaptive_lock_t);
3607 
3608     __kmp_get_user_lock_owner_ =
3609         (kmp_int32(*)(kmp_user_lock_p))(&__kmp_get_queuing_lock_owner);
3610 
3611     if (__kmp_env_consistency_check) {
3612       KMP_BIND_USER_LOCK_WITH_CHECKS(adaptive);
3613     } else {
3614       KMP_BIND_USER_LOCK(adaptive);
3615     }
3616 
3617     __kmp_destroy_user_lock_ =
3618         (void (*)(kmp_user_lock_p))(&__kmp_destroy_adaptive_lock);
3619 
3620     __kmp_is_user_lock_initialized_ =
3621         (int (*)(kmp_user_lock_p))(&__kmp_is_queuing_lock_initialized);
3622 
3623     __kmp_get_user_lock_location_ =
3624         (const ident_t *(*)(kmp_user_lock_p))(&__kmp_get_queuing_lock_location);
3625 
3626     __kmp_set_user_lock_location_ = (void (*)(
3627         kmp_user_lock_p, const ident_t *))(&__kmp_set_queuing_lock_location);
3628 
3629     __kmp_get_user_lock_flags_ =
3630         (kmp_lock_flags_t(*)(kmp_user_lock_p))(&__kmp_get_queuing_lock_flags);
3631 
3632     __kmp_set_user_lock_flags_ = (void (*)(kmp_user_lock_p, kmp_lock_flags_t))(
3633         &__kmp_set_queuing_lock_flags);
3634 
3635   } break;
3636 #endif // KMP_USE_ADAPTIVE_LOCKS
3637 
3638   case lk_drdpa: {
3639     __kmp_base_user_lock_size = sizeof(kmp_base_drdpa_lock_t);
3640     __kmp_user_lock_size = sizeof(kmp_drdpa_lock_t);
3641 
3642     __kmp_get_user_lock_owner_ =
3643         (kmp_int32(*)(kmp_user_lock_p))(&__kmp_get_drdpa_lock_owner);
3644 
3645     if (__kmp_env_consistency_check) {
3646       KMP_BIND_USER_LOCK_WITH_CHECKS(drdpa);
3647       KMP_BIND_NESTED_USER_LOCK_WITH_CHECKS(drdpa);
3648     } else {
3649       KMP_BIND_USER_LOCK(drdpa);
3650       KMP_BIND_NESTED_USER_LOCK(drdpa);
3651     }
3652 
3653     __kmp_destroy_user_lock_ =
3654         (void (*)(kmp_user_lock_p))(&__kmp_destroy_drdpa_lock);
3655 
3656     __kmp_is_user_lock_initialized_ =
3657         (int (*)(kmp_user_lock_p))(&__kmp_is_drdpa_lock_initialized);
3658 
3659     __kmp_get_user_lock_location_ =
3660         (const ident_t *(*)(kmp_user_lock_p))(&__kmp_get_drdpa_lock_location);
3661 
3662     __kmp_set_user_lock_location_ = (void (*)(
3663         kmp_user_lock_p, const ident_t *))(&__kmp_set_drdpa_lock_location);
3664 
3665     __kmp_get_user_lock_flags_ =
3666         (kmp_lock_flags_t(*)(kmp_user_lock_p))(&__kmp_get_drdpa_lock_flags);
3667 
3668     __kmp_set_user_lock_flags_ = (void (*)(kmp_user_lock_p, kmp_lock_flags_t))(
3669         &__kmp_set_drdpa_lock_flags);
3670   } break;
3671   }
3672 }
3673 
3674 // ----------------------------------------------------------------------------
3675 // User lock table & lock allocation
3676 
3677 kmp_lock_table_t __kmp_user_lock_table = {1, 0, NULL};
3678 kmp_user_lock_p __kmp_lock_pool = NULL;
3679 
3680 // Lock block-allocation support.
3681 kmp_block_of_locks *__kmp_lock_blocks = NULL;
3682 int __kmp_num_locks_in_block = 1; // FIXME - tune this value
3683 
3684 static kmp_lock_index_t __kmp_lock_table_insert(kmp_user_lock_p lck) {
3685   // Assume that kmp_global_lock is held upon entry/exit.
3686   kmp_lock_index_t index;
3687   if (__kmp_user_lock_table.used >= __kmp_user_lock_table.allocated) {
3688     kmp_lock_index_t size;
3689     kmp_user_lock_p *table;
3690     // Reallocate lock table.
3691     if (__kmp_user_lock_table.allocated == 0) {
3692       size = 1024;
3693     } else {
3694       size = __kmp_user_lock_table.allocated * 2;
3695     }
3696     table = (kmp_user_lock_p *)__kmp_allocate(sizeof(kmp_user_lock_p) * size);
3697     KMP_MEMCPY(table + 1, __kmp_user_lock_table.table + 1,
3698                sizeof(kmp_user_lock_p) * (__kmp_user_lock_table.used - 1));
3699     table[0] = (kmp_user_lock_p)__kmp_user_lock_table.table;
3700     // We cannot free the previous table now, since it may be in use by other
3701     // threads. So save the pointer to the previous table in in the first
3702     // element of the new table. All the tables will be organized into a list,
3703     // and could be freed when library shutting down.
3704     __kmp_user_lock_table.table = table;
3705     __kmp_user_lock_table.allocated = size;
3706   }
3707   KMP_DEBUG_ASSERT(__kmp_user_lock_table.used <
3708                    __kmp_user_lock_table.allocated);
3709   index = __kmp_user_lock_table.used;
3710   __kmp_user_lock_table.table[index] = lck;
3711   ++__kmp_user_lock_table.used;
3712   return index;
3713 }
3714 
3715 static kmp_user_lock_p __kmp_lock_block_allocate() {
3716   // Assume that kmp_global_lock is held upon entry/exit.
3717   static int last_index = 0;
3718   if ((last_index >= __kmp_num_locks_in_block) || (__kmp_lock_blocks == NULL)) {
3719     // Restart the index.
3720     last_index = 0;
3721     // Need to allocate a new block.
3722     KMP_DEBUG_ASSERT(__kmp_user_lock_size > 0);
3723     size_t space_for_locks = __kmp_user_lock_size * __kmp_num_locks_in_block;
3724     char *buffer =
3725         (char *)__kmp_allocate(space_for_locks + sizeof(kmp_block_of_locks));
3726     // Set up the new block.
3727     kmp_block_of_locks *new_block =
3728         (kmp_block_of_locks *)(&buffer[space_for_locks]);
3729     new_block->next_block = __kmp_lock_blocks;
3730     new_block->locks = (void *)buffer;
3731     // Publish the new block.
3732     KMP_MB();
3733     __kmp_lock_blocks = new_block;
3734   }
3735   kmp_user_lock_p ret = (kmp_user_lock_p)(&(
3736       ((char *)(__kmp_lock_blocks->locks))[last_index * __kmp_user_lock_size]));
3737   last_index++;
3738   return ret;
3739 }
3740 
3741 // Get memory for a lock. It may be freshly allocated memory or reused memory
3742 // from lock pool.
3743 kmp_user_lock_p __kmp_user_lock_allocate(void **user_lock, kmp_int32 gtid,
3744                                          kmp_lock_flags_t flags) {
3745   kmp_user_lock_p lck;
3746   kmp_lock_index_t index;
3747   KMP_DEBUG_ASSERT(user_lock);
3748 
3749   __kmp_acquire_lock(&__kmp_global_lock, gtid);
3750 
3751   if (__kmp_lock_pool == NULL) {
3752     // Lock pool is empty. Allocate new memory.
3753 
3754     // ANNOTATION: Found no good way to express the syncronisation
3755     // between allocation and usage, so ignore the allocation
3756     ANNOTATE_IGNORE_WRITES_BEGIN();
3757     if (__kmp_num_locks_in_block <= 1) { // Tune this cutoff point.
3758       lck = (kmp_user_lock_p)__kmp_allocate(__kmp_user_lock_size);
3759     } else {
3760       lck = __kmp_lock_block_allocate();
3761     }
3762     ANNOTATE_IGNORE_WRITES_END();
3763 
3764     // Insert lock in the table so that it can be freed in __kmp_cleanup,
3765     // and debugger has info on all allocated locks.
3766     index = __kmp_lock_table_insert(lck);
3767   } else {
3768     // Pick up lock from pool.
3769     lck = __kmp_lock_pool;
3770     index = __kmp_lock_pool->pool.index;
3771     __kmp_lock_pool = __kmp_lock_pool->pool.next;
3772   }
3773 
3774   // We could potentially differentiate between nested and regular locks
3775   // here, and do the lock table lookup for regular locks only.
3776   if (OMP_LOCK_T_SIZE < sizeof(void *)) {
3777     *((kmp_lock_index_t *)user_lock) = index;
3778   } else {
3779     *((kmp_user_lock_p *)user_lock) = lck;
3780   }
3781 
3782   // mark the lock if it is critical section lock.
3783   __kmp_set_user_lock_flags(lck, flags);
3784 
3785   __kmp_release_lock(&__kmp_global_lock, gtid); // AC: TODO move this line upper
3786 
3787   return lck;
3788 }
3789 
3790 // Put lock's memory to pool for reusing.
3791 void __kmp_user_lock_free(void **user_lock, kmp_int32 gtid,
3792                           kmp_user_lock_p lck) {
3793   KMP_DEBUG_ASSERT(user_lock != NULL);
3794   KMP_DEBUG_ASSERT(lck != NULL);
3795 
3796   __kmp_acquire_lock(&__kmp_global_lock, gtid);
3797 
3798   lck->pool.next = __kmp_lock_pool;
3799   __kmp_lock_pool = lck;
3800   if (OMP_LOCK_T_SIZE < sizeof(void *)) {
3801     kmp_lock_index_t index = *((kmp_lock_index_t *)user_lock);
3802     KMP_DEBUG_ASSERT(0 < index && index <= __kmp_user_lock_table.used);
3803     lck->pool.index = index;
3804   }
3805 
3806   __kmp_release_lock(&__kmp_global_lock, gtid);
3807 }
3808 
3809 kmp_user_lock_p __kmp_lookup_user_lock(void **user_lock, char const *func) {
3810   kmp_user_lock_p lck = NULL;
3811 
3812   if (__kmp_env_consistency_check) {
3813     if (user_lock == NULL) {
3814       KMP_FATAL(LockIsUninitialized, func);
3815     }
3816   }
3817 
3818   if (OMP_LOCK_T_SIZE < sizeof(void *)) {
3819     kmp_lock_index_t index = *((kmp_lock_index_t *)user_lock);
3820     if (__kmp_env_consistency_check) {
3821       if (!(0 < index && index < __kmp_user_lock_table.used)) {
3822         KMP_FATAL(LockIsUninitialized, func);
3823       }
3824     }
3825     KMP_DEBUG_ASSERT(0 < index && index < __kmp_user_lock_table.used);
3826     KMP_DEBUG_ASSERT(__kmp_user_lock_size > 0);
3827     lck = __kmp_user_lock_table.table[index];
3828   } else {
3829     lck = *((kmp_user_lock_p *)user_lock);
3830   }
3831 
3832   if (__kmp_env_consistency_check) {
3833     if (lck == NULL) {
3834       KMP_FATAL(LockIsUninitialized, func);
3835     }
3836   }
3837 
3838   return lck;
3839 }
3840 
3841 void __kmp_cleanup_user_locks(void) {
3842   // Reset lock pool. Don't worry about lock in the pool--we will free them when
3843   // iterating through lock table (it includes all the locks, dead or alive).
3844   __kmp_lock_pool = NULL;
3845 
3846 #define IS_CRITICAL(lck)                                                       \
3847   ((__kmp_get_user_lock_flags_ != NULL) &&                                     \
3848    ((*__kmp_get_user_lock_flags_)(lck)&kmp_lf_critical_section))
3849 
3850   // Loop through lock table, free all locks.
3851   // Do not free item [0], it is reserved for lock tables list.
3852   //
3853   // FIXME - we are iterating through a list of (pointers to) objects of type
3854   // union kmp_user_lock, but we have no way of knowing whether the base type is
3855   // currently "pool" or whatever the global user lock type is.
3856   //
3857   // We are relying on the fact that for all of the user lock types
3858   // (except "tas"), the first field in the lock struct is the "initialized"
3859   // field, which is set to the address of the lock object itself when
3860   // the lock is initialized.  When the union is of type "pool", the
3861   // first field is a pointer to the next object in the free list, which
3862   // will not be the same address as the object itself.
3863   //
3864   // This means that the check (*__kmp_is_user_lock_initialized_)(lck) will fail
3865   // for "pool" objects on the free list.  This must happen as the "location"
3866   // field of real user locks overlaps the "index" field of "pool" objects.
3867   //
3868   // It would be better to run through the free list, and remove all "pool"
3869   // objects from the lock table before executing this loop.  However,
3870   // "pool" objects do not always have their index field set (only on
3871   // lin_32e), and I don't want to search the lock table for the address
3872   // of every "pool" object on the free list.
3873   while (__kmp_user_lock_table.used > 1) {
3874     const ident *loc;
3875 
3876     // reduce __kmp_user_lock_table.used before freeing the lock,
3877     // so that state of locks is consistent
3878     kmp_user_lock_p lck =
3879         __kmp_user_lock_table.table[--__kmp_user_lock_table.used];
3880 
3881     if ((__kmp_is_user_lock_initialized_ != NULL) &&
3882         (*__kmp_is_user_lock_initialized_)(lck)) {
3883       // Issue a warning if: KMP_CONSISTENCY_CHECK AND lock is initialized AND
3884       // it is NOT a critical section (user is not responsible for destroying
3885       // criticals) AND we know source location to report.
3886       if (__kmp_env_consistency_check && (!IS_CRITICAL(lck)) &&
3887           ((loc = __kmp_get_user_lock_location(lck)) != NULL) &&
3888           (loc->psource != NULL)) {
3889         kmp_str_loc_t str_loc = __kmp_str_loc_init(loc->psource, 0);
3890         KMP_WARNING(CnsLockNotDestroyed, str_loc.file, str_loc.line);
3891         __kmp_str_loc_free(&str_loc);
3892       }
3893 
3894 #ifdef KMP_DEBUG
3895       if (IS_CRITICAL(lck)) {
3896         KA_TRACE(
3897             20,
3898             ("__kmp_cleanup_user_locks: free critical section lock %p (%p)\n",
3899              lck, *(void **)lck));
3900       } else {
3901         KA_TRACE(20, ("__kmp_cleanup_user_locks: free lock %p (%p)\n", lck,
3902                       *(void **)lck));
3903       }
3904 #endif // KMP_DEBUG
3905 
3906       // Cleanup internal lock dynamic resources (for drdpa locks particularly).
3907       __kmp_destroy_user_lock(lck);
3908     }
3909 
3910     // Free the lock if block allocation of locks is not used.
3911     if (__kmp_lock_blocks == NULL) {
3912       __kmp_free(lck);
3913     }
3914   }
3915 
3916 #undef IS_CRITICAL
3917 
3918   // delete lock table(s).
3919   kmp_user_lock_p *table_ptr = __kmp_user_lock_table.table;
3920   __kmp_user_lock_table.table = NULL;
3921   __kmp_user_lock_table.allocated = 0;
3922 
3923   while (table_ptr != NULL) {
3924     // In the first element we saved the pointer to the previous
3925     // (smaller) lock table.
3926     kmp_user_lock_p *next = (kmp_user_lock_p *)(table_ptr[0]);
3927     __kmp_free(table_ptr);
3928     table_ptr = next;
3929   }
3930 
3931   // Free buffers allocated for blocks of locks.
3932   kmp_block_of_locks_t *block_ptr = __kmp_lock_blocks;
3933   __kmp_lock_blocks = NULL;
3934 
3935   while (block_ptr != NULL) {
3936     kmp_block_of_locks_t *next = block_ptr->next_block;
3937     __kmp_free(block_ptr->locks);
3938     // *block_ptr itself was allocated at the end of the locks vector.
3939     block_ptr = next;
3940   }
3941 
3942   TCW_4(__kmp_init_user_locks, FALSE);
3943 }
3944 
3945 #endif // KMP_USE_DYNAMIC_LOCK
3946