1 /*
2  * kmp_lock.cpp -- lock-related functions
3  */
4 
5 //===----------------------------------------------------------------------===//
6 //
7 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
8 // See https://llvm.org/LICENSE.txt for license information.
9 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #include <stddef.h>
14 #include <atomic>
15 
16 #include "kmp.h"
17 #include "kmp_i18n.h"
18 #include "kmp_io.h"
19 #include "kmp_itt.h"
20 #include "kmp_lock.h"
21 #include "kmp_wait_release.h"
22 #include "kmp_wrapper_getpid.h"
23 
24 #include "tsan_annotations.h"
25 
26 #if KMP_USE_FUTEX
27 #include <sys/syscall.h>
28 #include <unistd.h>
29 // We should really include <futex.h>, but that causes compatibility problems on
30 // different Linux* OS distributions that either require that you include (or
31 // break when you try to include) <pci/types.h>. Since all we need is the two
32 // macros below (which are part of the kernel ABI, so can't change) we just
33 // define the constants here and don't include <futex.h>
34 #ifndef FUTEX_WAIT
35 #define FUTEX_WAIT 0
36 #endif
37 #ifndef FUTEX_WAKE
38 #define FUTEX_WAKE 1
39 #endif
40 #endif
41 
42 /* Implement spin locks for internal library use.             */
43 /* The algorithm implemented is Lamport's bakery lock [1974]. */
44 
45 void __kmp_validate_locks(void) {
46   int i;
47   kmp_uint32 x, y;
48 
49   /* Check to make sure unsigned arithmetic does wraps properly */
50   x = ~((kmp_uint32)0) - 2;
51   y = x - 2;
52 
53   for (i = 0; i < 8; ++i, ++x, ++y) {
54     kmp_uint32 z = (x - y);
55     KMP_ASSERT(z == 2);
56   }
57 
58   KMP_ASSERT(offsetof(kmp_base_queuing_lock, tail_id) % 8 == 0);
59 }
60 
61 /* ------------------------------------------------------------------------ */
62 /* test and set locks */
63 
64 // For the non-nested locks, we can only assume that the first 4 bytes were
65 // allocated, since gcc only allocates 4 bytes for omp_lock_t, and the Intel
66 // compiler only allocates a 4 byte pointer on IA-32 architecture.  On
67 // Windows* OS on Intel(R) 64, we can assume that all 8 bytes were allocated.
68 //
69 // gcc reserves >= 8 bytes for nested locks, so we can assume that the
70 // entire 8 bytes were allocated for nested locks on all 64-bit platforms.
71 
72 static kmp_int32 __kmp_get_tas_lock_owner(kmp_tas_lock_t *lck) {
73   return KMP_LOCK_STRIP(KMP_ATOMIC_LD_RLX(&lck->lk.poll)) - 1;
74 }
75 
76 static inline bool __kmp_is_tas_lock_nestable(kmp_tas_lock_t *lck) {
77   return lck->lk.depth_locked != -1;
78 }
79 
80 __forceinline static int
81 __kmp_acquire_tas_lock_timed_template(kmp_tas_lock_t *lck, kmp_int32 gtid) {
82   KMP_MB();
83 
84 #ifdef USE_LOCK_PROFILE
85   kmp_uint32 curr = KMP_LOCK_STRIP(lck->lk.poll);
86   if ((curr != 0) && (curr != gtid + 1))
87     __kmp_printf("LOCK CONTENTION: %p\n", lck);
88 /* else __kmp_printf( "." );*/
89 #endif /* USE_LOCK_PROFILE */
90 
91   kmp_int32 tas_free = KMP_LOCK_FREE(tas);
92   kmp_int32 tas_busy = KMP_LOCK_BUSY(gtid + 1, tas);
93 
94   if (KMP_ATOMIC_LD_RLX(&lck->lk.poll) == tas_free &&
95       __kmp_atomic_compare_store_acq(&lck->lk.poll, tas_free, tas_busy)) {
96     KMP_FSYNC_ACQUIRED(lck);
97     return KMP_LOCK_ACQUIRED_FIRST;
98   }
99 
100   kmp_uint32 spins;
101   KMP_FSYNC_PREPARE(lck);
102   KMP_INIT_YIELD(spins);
103   kmp_backoff_t backoff = __kmp_spin_backoff_params;
104   do {
105     __kmp_spin_backoff(&backoff);
106     KMP_YIELD_OVERSUB_ELSE_SPIN(spins);
107   } while (KMP_ATOMIC_LD_RLX(&lck->lk.poll) != tas_free ||
108            !__kmp_atomic_compare_store_acq(&lck->lk.poll, tas_free, tas_busy));
109   KMP_FSYNC_ACQUIRED(lck);
110   return KMP_LOCK_ACQUIRED_FIRST;
111 }
112 
113 int __kmp_acquire_tas_lock(kmp_tas_lock_t *lck, kmp_int32 gtid) {
114   int retval = __kmp_acquire_tas_lock_timed_template(lck, gtid);
115   ANNOTATE_TAS_ACQUIRED(lck);
116   return retval;
117 }
118 
119 static int __kmp_acquire_tas_lock_with_checks(kmp_tas_lock_t *lck,
120                                               kmp_int32 gtid) {
121   char const *const func = "omp_set_lock";
122   if ((sizeof(kmp_tas_lock_t) <= OMP_LOCK_T_SIZE) &&
123       __kmp_is_tas_lock_nestable(lck)) {
124     KMP_FATAL(LockNestableUsedAsSimple, func);
125   }
126   if ((gtid >= 0) && (__kmp_get_tas_lock_owner(lck) == gtid)) {
127     KMP_FATAL(LockIsAlreadyOwned, func);
128   }
129   return __kmp_acquire_tas_lock(lck, gtid);
130 }
131 
132 int __kmp_test_tas_lock(kmp_tas_lock_t *lck, kmp_int32 gtid) {
133   kmp_int32 tas_free = KMP_LOCK_FREE(tas);
134   kmp_int32 tas_busy = KMP_LOCK_BUSY(gtid + 1, tas);
135   if (KMP_ATOMIC_LD_RLX(&lck->lk.poll) == tas_free &&
136       __kmp_atomic_compare_store_acq(&lck->lk.poll, tas_free, tas_busy)) {
137     KMP_FSYNC_ACQUIRED(lck);
138     return TRUE;
139   }
140   return FALSE;
141 }
142 
143 static int __kmp_test_tas_lock_with_checks(kmp_tas_lock_t *lck,
144                                            kmp_int32 gtid) {
145   char const *const func = "omp_test_lock";
146   if ((sizeof(kmp_tas_lock_t) <= OMP_LOCK_T_SIZE) &&
147       __kmp_is_tas_lock_nestable(lck)) {
148     KMP_FATAL(LockNestableUsedAsSimple, func);
149   }
150   return __kmp_test_tas_lock(lck, gtid);
151 }
152 
153 int __kmp_release_tas_lock(kmp_tas_lock_t *lck, kmp_int32 gtid) {
154   KMP_MB(); /* Flush all pending memory write invalidates.  */
155 
156   KMP_FSYNC_RELEASING(lck);
157   ANNOTATE_TAS_RELEASED(lck);
158   KMP_ATOMIC_ST_REL(&lck->lk.poll, KMP_LOCK_FREE(tas));
159   KMP_MB(); /* Flush all pending memory write invalidates.  */
160 
161   KMP_YIELD_OVERSUB();
162   return KMP_LOCK_RELEASED;
163 }
164 
165 static int __kmp_release_tas_lock_with_checks(kmp_tas_lock_t *lck,
166                                               kmp_int32 gtid) {
167   char const *const func = "omp_unset_lock";
168   KMP_MB(); /* in case another processor initialized lock */
169   if ((sizeof(kmp_tas_lock_t) <= OMP_LOCK_T_SIZE) &&
170       __kmp_is_tas_lock_nestable(lck)) {
171     KMP_FATAL(LockNestableUsedAsSimple, func);
172   }
173   if (__kmp_get_tas_lock_owner(lck) == -1) {
174     KMP_FATAL(LockUnsettingFree, func);
175   }
176   if ((gtid >= 0) && (__kmp_get_tas_lock_owner(lck) >= 0) &&
177       (__kmp_get_tas_lock_owner(lck) != gtid)) {
178     KMP_FATAL(LockUnsettingSetByAnother, func);
179   }
180   return __kmp_release_tas_lock(lck, gtid);
181 }
182 
183 void __kmp_init_tas_lock(kmp_tas_lock_t *lck) {
184   lck->lk.poll = KMP_LOCK_FREE(tas);
185 }
186 
187 void __kmp_destroy_tas_lock(kmp_tas_lock_t *lck) { lck->lk.poll = 0; }
188 
189 static void __kmp_destroy_tas_lock_with_checks(kmp_tas_lock_t *lck) {
190   char const *const func = "omp_destroy_lock";
191   if ((sizeof(kmp_tas_lock_t) <= OMP_LOCK_T_SIZE) &&
192       __kmp_is_tas_lock_nestable(lck)) {
193     KMP_FATAL(LockNestableUsedAsSimple, func);
194   }
195   if (__kmp_get_tas_lock_owner(lck) != -1) {
196     KMP_FATAL(LockStillOwned, func);
197   }
198   __kmp_destroy_tas_lock(lck);
199 }
200 
201 // nested test and set locks
202 
203 int __kmp_acquire_nested_tas_lock(kmp_tas_lock_t *lck, kmp_int32 gtid) {
204   KMP_DEBUG_ASSERT(gtid >= 0);
205 
206   if (__kmp_get_tas_lock_owner(lck) == gtid) {
207     lck->lk.depth_locked += 1;
208     return KMP_LOCK_ACQUIRED_NEXT;
209   } else {
210     __kmp_acquire_tas_lock_timed_template(lck, gtid);
211     ANNOTATE_TAS_ACQUIRED(lck);
212     lck->lk.depth_locked = 1;
213     return KMP_LOCK_ACQUIRED_FIRST;
214   }
215 }
216 
217 static int __kmp_acquire_nested_tas_lock_with_checks(kmp_tas_lock_t *lck,
218                                                      kmp_int32 gtid) {
219   char const *const func = "omp_set_nest_lock";
220   if (!__kmp_is_tas_lock_nestable(lck)) {
221     KMP_FATAL(LockSimpleUsedAsNestable, func);
222   }
223   return __kmp_acquire_nested_tas_lock(lck, gtid);
224 }
225 
226 int __kmp_test_nested_tas_lock(kmp_tas_lock_t *lck, kmp_int32 gtid) {
227   int retval;
228 
229   KMP_DEBUG_ASSERT(gtid >= 0);
230 
231   if (__kmp_get_tas_lock_owner(lck) == gtid) {
232     retval = ++lck->lk.depth_locked;
233   } else if (!__kmp_test_tas_lock(lck, gtid)) {
234     retval = 0;
235   } else {
236     KMP_MB();
237     retval = lck->lk.depth_locked = 1;
238   }
239   return retval;
240 }
241 
242 static int __kmp_test_nested_tas_lock_with_checks(kmp_tas_lock_t *lck,
243                                                   kmp_int32 gtid) {
244   char const *const func = "omp_test_nest_lock";
245   if (!__kmp_is_tas_lock_nestable(lck)) {
246     KMP_FATAL(LockSimpleUsedAsNestable, func);
247   }
248   return __kmp_test_nested_tas_lock(lck, gtid);
249 }
250 
251 int __kmp_release_nested_tas_lock(kmp_tas_lock_t *lck, kmp_int32 gtid) {
252   KMP_DEBUG_ASSERT(gtid >= 0);
253 
254   KMP_MB();
255   if (--(lck->lk.depth_locked) == 0) {
256     __kmp_release_tas_lock(lck, gtid);
257     return KMP_LOCK_RELEASED;
258   }
259   return KMP_LOCK_STILL_HELD;
260 }
261 
262 static int __kmp_release_nested_tas_lock_with_checks(kmp_tas_lock_t *lck,
263                                                      kmp_int32 gtid) {
264   char const *const func = "omp_unset_nest_lock";
265   KMP_MB(); /* in case another processor initialized lock */
266   if (!__kmp_is_tas_lock_nestable(lck)) {
267     KMP_FATAL(LockSimpleUsedAsNestable, func);
268   }
269   if (__kmp_get_tas_lock_owner(lck) == -1) {
270     KMP_FATAL(LockUnsettingFree, func);
271   }
272   if (__kmp_get_tas_lock_owner(lck) != gtid) {
273     KMP_FATAL(LockUnsettingSetByAnother, func);
274   }
275   return __kmp_release_nested_tas_lock(lck, gtid);
276 }
277 
278 void __kmp_init_nested_tas_lock(kmp_tas_lock_t *lck) {
279   __kmp_init_tas_lock(lck);
280   lck->lk.depth_locked = 0; // >= 0 for nestable locks, -1 for simple locks
281 }
282 
283 void __kmp_destroy_nested_tas_lock(kmp_tas_lock_t *lck) {
284   __kmp_destroy_tas_lock(lck);
285   lck->lk.depth_locked = 0;
286 }
287 
288 static void __kmp_destroy_nested_tas_lock_with_checks(kmp_tas_lock_t *lck) {
289   char const *const func = "omp_destroy_nest_lock";
290   if (!__kmp_is_tas_lock_nestable(lck)) {
291     KMP_FATAL(LockSimpleUsedAsNestable, func);
292   }
293   if (__kmp_get_tas_lock_owner(lck) != -1) {
294     KMP_FATAL(LockStillOwned, func);
295   }
296   __kmp_destroy_nested_tas_lock(lck);
297 }
298 
299 #if KMP_USE_FUTEX
300 
301 /* ------------------------------------------------------------------------ */
302 /* futex locks */
303 
304 // futex locks are really just test and set locks, with a different method
305 // of handling contention.  They take the same amount of space as test and
306 // set locks, and are allocated the same way (i.e. use the area allocated by
307 // the compiler for non-nested locks / allocate nested locks on the heap).
308 
309 static kmp_int32 __kmp_get_futex_lock_owner(kmp_futex_lock_t *lck) {
310   return KMP_LOCK_STRIP((TCR_4(lck->lk.poll) >> 1)) - 1;
311 }
312 
313 static inline bool __kmp_is_futex_lock_nestable(kmp_futex_lock_t *lck) {
314   return lck->lk.depth_locked != -1;
315 }
316 
317 __forceinline static int
318 __kmp_acquire_futex_lock_timed_template(kmp_futex_lock_t *lck, kmp_int32 gtid) {
319   kmp_int32 gtid_code = (gtid + 1) << 1;
320 
321   KMP_MB();
322 
323 #ifdef USE_LOCK_PROFILE
324   kmp_uint32 curr = KMP_LOCK_STRIP(TCR_4(lck->lk.poll));
325   if ((curr != 0) && (curr != gtid_code))
326     __kmp_printf("LOCK CONTENTION: %p\n", lck);
327 /* else __kmp_printf( "." );*/
328 #endif /* USE_LOCK_PROFILE */
329 
330   KMP_FSYNC_PREPARE(lck);
331   KA_TRACE(1000, ("__kmp_acquire_futex_lock: lck:%p(0x%x), T#%d entering\n",
332                   lck, lck->lk.poll, gtid));
333 
334   kmp_int32 poll_val;
335 
336   while ((poll_val = KMP_COMPARE_AND_STORE_RET32(
337               &(lck->lk.poll), KMP_LOCK_FREE(futex),
338               KMP_LOCK_BUSY(gtid_code, futex))) != KMP_LOCK_FREE(futex)) {
339 
340     kmp_int32 cond = KMP_LOCK_STRIP(poll_val) & 1;
341     KA_TRACE(
342         1000,
343         ("__kmp_acquire_futex_lock: lck:%p, T#%d poll_val = 0x%x cond = 0x%x\n",
344          lck, gtid, poll_val, cond));
345 
346     // NOTE: if you try to use the following condition for this branch
347     //
348     // if ( poll_val & 1 == 0 )
349     //
350     // Then the 12.0 compiler has a bug where the following block will
351     // always be skipped, regardless of the value of the LSB of poll_val.
352     if (!cond) {
353       // Try to set the lsb in the poll to indicate to the owner
354       // thread that they need to wake this thread up.
355       if (!KMP_COMPARE_AND_STORE_REL32(&(lck->lk.poll), poll_val,
356                                        poll_val | KMP_LOCK_BUSY(1, futex))) {
357         KA_TRACE(
358             1000,
359             ("__kmp_acquire_futex_lock: lck:%p(0x%x), T#%d can't set bit 0\n",
360              lck, lck->lk.poll, gtid));
361         continue;
362       }
363       poll_val |= KMP_LOCK_BUSY(1, futex);
364 
365       KA_TRACE(1000,
366                ("__kmp_acquire_futex_lock: lck:%p(0x%x), T#%d bit 0 set\n", lck,
367                 lck->lk.poll, gtid));
368     }
369 
370     KA_TRACE(
371         1000,
372         ("__kmp_acquire_futex_lock: lck:%p, T#%d before futex_wait(0x%x)\n",
373          lck, gtid, poll_val));
374 
375     long rc;
376     if ((rc = syscall(__NR_futex, &(lck->lk.poll), FUTEX_WAIT, poll_val, NULL,
377                       NULL, 0)) != 0) {
378       KA_TRACE(1000, ("__kmp_acquire_futex_lock: lck:%p, T#%d futex_wait(0x%x) "
379                       "failed (rc=%ld errno=%d)\n",
380                       lck, gtid, poll_val, rc, errno));
381       continue;
382     }
383 
384     KA_TRACE(1000,
385              ("__kmp_acquire_futex_lock: lck:%p, T#%d after futex_wait(0x%x)\n",
386               lck, gtid, poll_val));
387     // This thread has now done a successful futex wait call and was entered on
388     // the OS futex queue.  We must now perform a futex wake call when releasing
389     // the lock, as we have no idea how many other threads are in the queue.
390     gtid_code |= 1;
391   }
392 
393   KMP_FSYNC_ACQUIRED(lck);
394   KA_TRACE(1000, ("__kmp_acquire_futex_lock: lck:%p(0x%x), T#%d exiting\n", lck,
395                   lck->lk.poll, gtid));
396   return KMP_LOCK_ACQUIRED_FIRST;
397 }
398 
399 int __kmp_acquire_futex_lock(kmp_futex_lock_t *lck, kmp_int32 gtid) {
400   int retval = __kmp_acquire_futex_lock_timed_template(lck, gtid);
401   ANNOTATE_FUTEX_ACQUIRED(lck);
402   return retval;
403 }
404 
405 static int __kmp_acquire_futex_lock_with_checks(kmp_futex_lock_t *lck,
406                                                 kmp_int32 gtid) {
407   char const *const func = "omp_set_lock";
408   if ((sizeof(kmp_futex_lock_t) <= OMP_LOCK_T_SIZE) &&
409       __kmp_is_futex_lock_nestable(lck)) {
410     KMP_FATAL(LockNestableUsedAsSimple, func);
411   }
412   if ((gtid >= 0) && (__kmp_get_futex_lock_owner(lck) == gtid)) {
413     KMP_FATAL(LockIsAlreadyOwned, func);
414   }
415   return __kmp_acquire_futex_lock(lck, gtid);
416 }
417 
418 int __kmp_test_futex_lock(kmp_futex_lock_t *lck, kmp_int32 gtid) {
419   if (KMP_COMPARE_AND_STORE_ACQ32(&(lck->lk.poll), KMP_LOCK_FREE(futex),
420                                   KMP_LOCK_BUSY((gtid + 1) << 1, futex))) {
421     KMP_FSYNC_ACQUIRED(lck);
422     return TRUE;
423   }
424   return FALSE;
425 }
426 
427 static int __kmp_test_futex_lock_with_checks(kmp_futex_lock_t *lck,
428                                              kmp_int32 gtid) {
429   char const *const func = "omp_test_lock";
430   if ((sizeof(kmp_futex_lock_t) <= OMP_LOCK_T_SIZE) &&
431       __kmp_is_futex_lock_nestable(lck)) {
432     KMP_FATAL(LockNestableUsedAsSimple, func);
433   }
434   return __kmp_test_futex_lock(lck, gtid);
435 }
436 
437 int __kmp_release_futex_lock(kmp_futex_lock_t *lck, kmp_int32 gtid) {
438   KMP_MB(); /* Flush all pending memory write invalidates.  */
439 
440   KA_TRACE(1000, ("__kmp_release_futex_lock: lck:%p(0x%x), T#%d entering\n",
441                   lck, lck->lk.poll, gtid));
442 
443   KMP_FSYNC_RELEASING(lck);
444   ANNOTATE_FUTEX_RELEASED(lck);
445 
446   kmp_int32 poll_val = KMP_XCHG_FIXED32(&(lck->lk.poll), KMP_LOCK_FREE(futex));
447 
448   KA_TRACE(1000,
449            ("__kmp_release_futex_lock: lck:%p, T#%d released poll_val = 0x%x\n",
450             lck, gtid, poll_val));
451 
452   if (KMP_LOCK_STRIP(poll_val) & 1) {
453     KA_TRACE(1000,
454              ("__kmp_release_futex_lock: lck:%p, T#%d futex_wake 1 thread\n",
455               lck, gtid));
456     syscall(__NR_futex, &(lck->lk.poll), FUTEX_WAKE, KMP_LOCK_BUSY(1, futex),
457             NULL, NULL, 0);
458   }
459 
460   KMP_MB(); /* Flush all pending memory write invalidates.  */
461 
462   KA_TRACE(1000, ("__kmp_release_futex_lock: lck:%p(0x%x), T#%d exiting\n", lck,
463                   lck->lk.poll, gtid));
464 
465   KMP_YIELD_OVERSUB();
466   return KMP_LOCK_RELEASED;
467 }
468 
469 static int __kmp_release_futex_lock_with_checks(kmp_futex_lock_t *lck,
470                                                 kmp_int32 gtid) {
471   char const *const func = "omp_unset_lock";
472   KMP_MB(); /* in case another processor initialized lock */
473   if ((sizeof(kmp_futex_lock_t) <= OMP_LOCK_T_SIZE) &&
474       __kmp_is_futex_lock_nestable(lck)) {
475     KMP_FATAL(LockNestableUsedAsSimple, func);
476   }
477   if (__kmp_get_futex_lock_owner(lck) == -1) {
478     KMP_FATAL(LockUnsettingFree, func);
479   }
480   if ((gtid >= 0) && (__kmp_get_futex_lock_owner(lck) >= 0) &&
481       (__kmp_get_futex_lock_owner(lck) != gtid)) {
482     KMP_FATAL(LockUnsettingSetByAnother, func);
483   }
484   return __kmp_release_futex_lock(lck, gtid);
485 }
486 
487 void __kmp_init_futex_lock(kmp_futex_lock_t *lck) {
488   TCW_4(lck->lk.poll, KMP_LOCK_FREE(futex));
489 }
490 
491 void __kmp_destroy_futex_lock(kmp_futex_lock_t *lck) { lck->lk.poll = 0; }
492 
493 static void __kmp_destroy_futex_lock_with_checks(kmp_futex_lock_t *lck) {
494   char const *const func = "omp_destroy_lock";
495   if ((sizeof(kmp_futex_lock_t) <= OMP_LOCK_T_SIZE) &&
496       __kmp_is_futex_lock_nestable(lck)) {
497     KMP_FATAL(LockNestableUsedAsSimple, func);
498   }
499   if (__kmp_get_futex_lock_owner(lck) != -1) {
500     KMP_FATAL(LockStillOwned, func);
501   }
502   __kmp_destroy_futex_lock(lck);
503 }
504 
505 // nested futex locks
506 
507 int __kmp_acquire_nested_futex_lock(kmp_futex_lock_t *lck, kmp_int32 gtid) {
508   KMP_DEBUG_ASSERT(gtid >= 0);
509 
510   if (__kmp_get_futex_lock_owner(lck) == gtid) {
511     lck->lk.depth_locked += 1;
512     return KMP_LOCK_ACQUIRED_NEXT;
513   } else {
514     __kmp_acquire_futex_lock_timed_template(lck, gtid);
515     ANNOTATE_FUTEX_ACQUIRED(lck);
516     lck->lk.depth_locked = 1;
517     return KMP_LOCK_ACQUIRED_FIRST;
518   }
519 }
520 
521 static int __kmp_acquire_nested_futex_lock_with_checks(kmp_futex_lock_t *lck,
522                                                        kmp_int32 gtid) {
523   char const *const func = "omp_set_nest_lock";
524   if (!__kmp_is_futex_lock_nestable(lck)) {
525     KMP_FATAL(LockSimpleUsedAsNestable, func);
526   }
527   return __kmp_acquire_nested_futex_lock(lck, gtid);
528 }
529 
530 int __kmp_test_nested_futex_lock(kmp_futex_lock_t *lck, kmp_int32 gtid) {
531   int retval;
532 
533   KMP_DEBUG_ASSERT(gtid >= 0);
534 
535   if (__kmp_get_futex_lock_owner(lck) == gtid) {
536     retval = ++lck->lk.depth_locked;
537   } else if (!__kmp_test_futex_lock(lck, gtid)) {
538     retval = 0;
539   } else {
540     KMP_MB();
541     retval = lck->lk.depth_locked = 1;
542   }
543   return retval;
544 }
545 
546 static int __kmp_test_nested_futex_lock_with_checks(kmp_futex_lock_t *lck,
547                                                     kmp_int32 gtid) {
548   char const *const func = "omp_test_nest_lock";
549   if (!__kmp_is_futex_lock_nestable(lck)) {
550     KMP_FATAL(LockSimpleUsedAsNestable, func);
551   }
552   return __kmp_test_nested_futex_lock(lck, gtid);
553 }
554 
555 int __kmp_release_nested_futex_lock(kmp_futex_lock_t *lck, kmp_int32 gtid) {
556   KMP_DEBUG_ASSERT(gtid >= 0);
557 
558   KMP_MB();
559   if (--(lck->lk.depth_locked) == 0) {
560     __kmp_release_futex_lock(lck, gtid);
561     return KMP_LOCK_RELEASED;
562   }
563   return KMP_LOCK_STILL_HELD;
564 }
565 
566 static int __kmp_release_nested_futex_lock_with_checks(kmp_futex_lock_t *lck,
567                                                        kmp_int32 gtid) {
568   char const *const func = "omp_unset_nest_lock";
569   KMP_MB(); /* in case another processor initialized lock */
570   if (!__kmp_is_futex_lock_nestable(lck)) {
571     KMP_FATAL(LockSimpleUsedAsNestable, func);
572   }
573   if (__kmp_get_futex_lock_owner(lck) == -1) {
574     KMP_FATAL(LockUnsettingFree, func);
575   }
576   if (__kmp_get_futex_lock_owner(lck) != gtid) {
577     KMP_FATAL(LockUnsettingSetByAnother, func);
578   }
579   return __kmp_release_nested_futex_lock(lck, gtid);
580 }
581 
582 void __kmp_init_nested_futex_lock(kmp_futex_lock_t *lck) {
583   __kmp_init_futex_lock(lck);
584   lck->lk.depth_locked = 0; // >= 0 for nestable locks, -1 for simple locks
585 }
586 
587 void __kmp_destroy_nested_futex_lock(kmp_futex_lock_t *lck) {
588   __kmp_destroy_futex_lock(lck);
589   lck->lk.depth_locked = 0;
590 }
591 
592 static void __kmp_destroy_nested_futex_lock_with_checks(kmp_futex_lock_t *lck) {
593   char const *const func = "omp_destroy_nest_lock";
594   if (!__kmp_is_futex_lock_nestable(lck)) {
595     KMP_FATAL(LockSimpleUsedAsNestable, func);
596   }
597   if (__kmp_get_futex_lock_owner(lck) != -1) {
598     KMP_FATAL(LockStillOwned, func);
599   }
600   __kmp_destroy_nested_futex_lock(lck);
601 }
602 
603 #endif // KMP_USE_FUTEX
604 
605 /* ------------------------------------------------------------------------ */
606 /* ticket (bakery) locks */
607 
608 static kmp_int32 __kmp_get_ticket_lock_owner(kmp_ticket_lock_t *lck) {
609   return std::atomic_load_explicit(&lck->lk.owner_id,
610                                    std::memory_order_relaxed) -
611          1;
612 }
613 
614 static inline bool __kmp_is_ticket_lock_nestable(kmp_ticket_lock_t *lck) {
615   return std::atomic_load_explicit(&lck->lk.depth_locked,
616                                    std::memory_order_relaxed) != -1;
617 }
618 
619 static kmp_uint32 __kmp_bakery_check(void *now_serving, kmp_uint32 my_ticket) {
620   return std::atomic_load_explicit((std::atomic<unsigned> *)now_serving,
621                                    std::memory_order_acquire) == my_ticket;
622 }
623 
624 __forceinline static int
625 __kmp_acquire_ticket_lock_timed_template(kmp_ticket_lock_t *lck,
626                                          kmp_int32 gtid) {
627   kmp_uint32 my_ticket = std::atomic_fetch_add_explicit(
628       &lck->lk.next_ticket, 1U, std::memory_order_relaxed);
629 
630 #ifdef USE_LOCK_PROFILE
631   if (std::atomic_load_explicit(&lck->lk.now_serving,
632                                 std::memory_order_relaxed) != my_ticket)
633     __kmp_printf("LOCK CONTENTION: %p\n", lck);
634 /* else __kmp_printf( "." );*/
635 #endif /* USE_LOCK_PROFILE */
636 
637   if (std::atomic_load_explicit(&lck->lk.now_serving,
638                                 std::memory_order_acquire) == my_ticket) {
639     return KMP_LOCK_ACQUIRED_FIRST;
640   }
641   KMP_WAIT_PTR(&lck->lk.now_serving, my_ticket, __kmp_bakery_check, lck);
642   return KMP_LOCK_ACQUIRED_FIRST;
643 }
644 
645 int __kmp_acquire_ticket_lock(kmp_ticket_lock_t *lck, kmp_int32 gtid) {
646   int retval = __kmp_acquire_ticket_lock_timed_template(lck, gtid);
647   ANNOTATE_TICKET_ACQUIRED(lck);
648   return retval;
649 }
650 
651 static int __kmp_acquire_ticket_lock_with_checks(kmp_ticket_lock_t *lck,
652                                                  kmp_int32 gtid) {
653   char const *const func = "omp_set_lock";
654 
655   if (!std::atomic_load_explicit(&lck->lk.initialized,
656                                  std::memory_order_relaxed)) {
657     KMP_FATAL(LockIsUninitialized, func);
658   }
659   if (lck->lk.self != lck) {
660     KMP_FATAL(LockIsUninitialized, func);
661   }
662   if (__kmp_is_ticket_lock_nestable(lck)) {
663     KMP_FATAL(LockNestableUsedAsSimple, func);
664   }
665   if ((gtid >= 0) && (__kmp_get_ticket_lock_owner(lck) == gtid)) {
666     KMP_FATAL(LockIsAlreadyOwned, func);
667   }
668 
669   __kmp_acquire_ticket_lock(lck, gtid);
670 
671   std::atomic_store_explicit(&lck->lk.owner_id, gtid + 1,
672                              std::memory_order_relaxed);
673   return KMP_LOCK_ACQUIRED_FIRST;
674 }
675 
676 int __kmp_test_ticket_lock(kmp_ticket_lock_t *lck, kmp_int32 gtid) {
677   kmp_uint32 my_ticket = std::atomic_load_explicit(&lck->lk.next_ticket,
678                                                    std::memory_order_relaxed);
679 
680   if (std::atomic_load_explicit(&lck->lk.now_serving,
681                                 std::memory_order_relaxed) == my_ticket) {
682     kmp_uint32 next_ticket = my_ticket + 1;
683     if (std::atomic_compare_exchange_strong_explicit(
684             &lck->lk.next_ticket, &my_ticket, next_ticket,
685             std::memory_order_acquire, std::memory_order_acquire)) {
686       return TRUE;
687     }
688   }
689   return FALSE;
690 }
691 
692 static int __kmp_test_ticket_lock_with_checks(kmp_ticket_lock_t *lck,
693                                               kmp_int32 gtid) {
694   char const *const func = "omp_test_lock";
695 
696   if (!std::atomic_load_explicit(&lck->lk.initialized,
697                                  std::memory_order_relaxed)) {
698     KMP_FATAL(LockIsUninitialized, func);
699   }
700   if (lck->lk.self != lck) {
701     KMP_FATAL(LockIsUninitialized, func);
702   }
703   if (__kmp_is_ticket_lock_nestable(lck)) {
704     KMP_FATAL(LockNestableUsedAsSimple, func);
705   }
706 
707   int retval = __kmp_test_ticket_lock(lck, gtid);
708 
709   if (retval) {
710     std::atomic_store_explicit(&lck->lk.owner_id, gtid + 1,
711                                std::memory_order_relaxed);
712   }
713   return retval;
714 }
715 
716 int __kmp_release_ticket_lock(kmp_ticket_lock_t *lck, kmp_int32 gtid) {
717   kmp_uint32 distance = std::atomic_load_explicit(&lck->lk.next_ticket,
718                                                   std::memory_order_relaxed) -
719                         std::atomic_load_explicit(&lck->lk.now_serving,
720                                                   std::memory_order_relaxed);
721 
722   ANNOTATE_TICKET_RELEASED(lck);
723   std::atomic_fetch_add_explicit(&lck->lk.now_serving, 1U,
724                                  std::memory_order_release);
725 
726   KMP_YIELD(distance >
727             (kmp_uint32)(__kmp_avail_proc ? __kmp_avail_proc : __kmp_xproc));
728   return KMP_LOCK_RELEASED;
729 }
730 
731 static int __kmp_release_ticket_lock_with_checks(kmp_ticket_lock_t *lck,
732                                                  kmp_int32 gtid) {
733   char const *const func = "omp_unset_lock";
734 
735   if (!std::atomic_load_explicit(&lck->lk.initialized,
736                                  std::memory_order_relaxed)) {
737     KMP_FATAL(LockIsUninitialized, func);
738   }
739   if (lck->lk.self != lck) {
740     KMP_FATAL(LockIsUninitialized, func);
741   }
742   if (__kmp_is_ticket_lock_nestable(lck)) {
743     KMP_FATAL(LockNestableUsedAsSimple, func);
744   }
745   if (__kmp_get_ticket_lock_owner(lck) == -1) {
746     KMP_FATAL(LockUnsettingFree, func);
747   }
748   if ((gtid >= 0) && (__kmp_get_ticket_lock_owner(lck) >= 0) &&
749       (__kmp_get_ticket_lock_owner(lck) != gtid)) {
750     KMP_FATAL(LockUnsettingSetByAnother, func);
751   }
752   std::atomic_store_explicit(&lck->lk.owner_id, 0, std::memory_order_relaxed);
753   return __kmp_release_ticket_lock(lck, gtid);
754 }
755 
756 void __kmp_init_ticket_lock(kmp_ticket_lock_t *lck) {
757   lck->lk.location = NULL;
758   lck->lk.self = lck;
759   std::atomic_store_explicit(&lck->lk.next_ticket, 0U,
760                              std::memory_order_relaxed);
761   std::atomic_store_explicit(&lck->lk.now_serving, 0U,
762                              std::memory_order_relaxed);
763   std::atomic_store_explicit(
764       &lck->lk.owner_id, 0,
765       std::memory_order_relaxed); // no thread owns the lock.
766   std::atomic_store_explicit(
767       &lck->lk.depth_locked, -1,
768       std::memory_order_relaxed); // -1 => not a nested lock.
769   std::atomic_store_explicit(&lck->lk.initialized, true,
770                              std::memory_order_release);
771 }
772 
773 void __kmp_destroy_ticket_lock(kmp_ticket_lock_t *lck) {
774   std::atomic_store_explicit(&lck->lk.initialized, false,
775                              std::memory_order_release);
776   lck->lk.self = NULL;
777   lck->lk.location = NULL;
778   std::atomic_store_explicit(&lck->lk.next_ticket, 0U,
779                              std::memory_order_relaxed);
780   std::atomic_store_explicit(&lck->lk.now_serving, 0U,
781                              std::memory_order_relaxed);
782   std::atomic_store_explicit(&lck->lk.owner_id, 0, std::memory_order_relaxed);
783   std::atomic_store_explicit(&lck->lk.depth_locked, -1,
784                              std::memory_order_relaxed);
785 }
786 
787 static void __kmp_destroy_ticket_lock_with_checks(kmp_ticket_lock_t *lck) {
788   char const *const func = "omp_destroy_lock";
789 
790   if (!std::atomic_load_explicit(&lck->lk.initialized,
791                                  std::memory_order_relaxed)) {
792     KMP_FATAL(LockIsUninitialized, func);
793   }
794   if (lck->lk.self != lck) {
795     KMP_FATAL(LockIsUninitialized, func);
796   }
797   if (__kmp_is_ticket_lock_nestable(lck)) {
798     KMP_FATAL(LockNestableUsedAsSimple, func);
799   }
800   if (__kmp_get_ticket_lock_owner(lck) != -1) {
801     KMP_FATAL(LockStillOwned, func);
802   }
803   __kmp_destroy_ticket_lock(lck);
804 }
805 
806 // nested ticket locks
807 
808 int __kmp_acquire_nested_ticket_lock(kmp_ticket_lock_t *lck, kmp_int32 gtid) {
809   KMP_DEBUG_ASSERT(gtid >= 0);
810 
811   if (__kmp_get_ticket_lock_owner(lck) == gtid) {
812     std::atomic_fetch_add_explicit(&lck->lk.depth_locked, 1,
813                                    std::memory_order_relaxed);
814     return KMP_LOCK_ACQUIRED_NEXT;
815   } else {
816     __kmp_acquire_ticket_lock_timed_template(lck, gtid);
817     ANNOTATE_TICKET_ACQUIRED(lck);
818     std::atomic_store_explicit(&lck->lk.depth_locked, 1,
819                                std::memory_order_relaxed);
820     std::atomic_store_explicit(&lck->lk.owner_id, gtid + 1,
821                                std::memory_order_relaxed);
822     return KMP_LOCK_ACQUIRED_FIRST;
823   }
824 }
825 
826 static int __kmp_acquire_nested_ticket_lock_with_checks(kmp_ticket_lock_t *lck,
827                                                         kmp_int32 gtid) {
828   char const *const func = "omp_set_nest_lock";
829 
830   if (!std::atomic_load_explicit(&lck->lk.initialized,
831                                  std::memory_order_relaxed)) {
832     KMP_FATAL(LockIsUninitialized, func);
833   }
834   if (lck->lk.self != lck) {
835     KMP_FATAL(LockIsUninitialized, func);
836   }
837   if (!__kmp_is_ticket_lock_nestable(lck)) {
838     KMP_FATAL(LockSimpleUsedAsNestable, func);
839   }
840   return __kmp_acquire_nested_ticket_lock(lck, gtid);
841 }
842 
843 int __kmp_test_nested_ticket_lock(kmp_ticket_lock_t *lck, kmp_int32 gtid) {
844   int retval;
845 
846   KMP_DEBUG_ASSERT(gtid >= 0);
847 
848   if (__kmp_get_ticket_lock_owner(lck) == gtid) {
849     retval = std::atomic_fetch_add_explicit(&lck->lk.depth_locked, 1,
850                                             std::memory_order_relaxed) +
851              1;
852   } else if (!__kmp_test_ticket_lock(lck, gtid)) {
853     retval = 0;
854   } else {
855     std::atomic_store_explicit(&lck->lk.depth_locked, 1,
856                                std::memory_order_relaxed);
857     std::atomic_store_explicit(&lck->lk.owner_id, gtid + 1,
858                                std::memory_order_relaxed);
859     retval = 1;
860   }
861   return retval;
862 }
863 
864 static int __kmp_test_nested_ticket_lock_with_checks(kmp_ticket_lock_t *lck,
865                                                      kmp_int32 gtid) {
866   char const *const func = "omp_test_nest_lock";
867 
868   if (!std::atomic_load_explicit(&lck->lk.initialized,
869                                  std::memory_order_relaxed)) {
870     KMP_FATAL(LockIsUninitialized, func);
871   }
872   if (lck->lk.self != lck) {
873     KMP_FATAL(LockIsUninitialized, func);
874   }
875   if (!__kmp_is_ticket_lock_nestable(lck)) {
876     KMP_FATAL(LockSimpleUsedAsNestable, func);
877   }
878   return __kmp_test_nested_ticket_lock(lck, gtid);
879 }
880 
881 int __kmp_release_nested_ticket_lock(kmp_ticket_lock_t *lck, kmp_int32 gtid) {
882   KMP_DEBUG_ASSERT(gtid >= 0);
883 
884   if ((std::atomic_fetch_add_explicit(&lck->lk.depth_locked, -1,
885                                       std::memory_order_relaxed) -
886        1) == 0) {
887     std::atomic_store_explicit(&lck->lk.owner_id, 0, std::memory_order_relaxed);
888     __kmp_release_ticket_lock(lck, gtid);
889     return KMP_LOCK_RELEASED;
890   }
891   return KMP_LOCK_STILL_HELD;
892 }
893 
894 static int __kmp_release_nested_ticket_lock_with_checks(kmp_ticket_lock_t *lck,
895                                                         kmp_int32 gtid) {
896   char const *const func = "omp_unset_nest_lock";
897 
898   if (!std::atomic_load_explicit(&lck->lk.initialized,
899                                  std::memory_order_relaxed)) {
900     KMP_FATAL(LockIsUninitialized, func);
901   }
902   if (lck->lk.self != lck) {
903     KMP_FATAL(LockIsUninitialized, func);
904   }
905   if (!__kmp_is_ticket_lock_nestable(lck)) {
906     KMP_FATAL(LockSimpleUsedAsNestable, func);
907   }
908   if (__kmp_get_ticket_lock_owner(lck) == -1) {
909     KMP_FATAL(LockUnsettingFree, func);
910   }
911   if (__kmp_get_ticket_lock_owner(lck) != gtid) {
912     KMP_FATAL(LockUnsettingSetByAnother, func);
913   }
914   return __kmp_release_nested_ticket_lock(lck, gtid);
915 }
916 
917 void __kmp_init_nested_ticket_lock(kmp_ticket_lock_t *lck) {
918   __kmp_init_ticket_lock(lck);
919   std::atomic_store_explicit(&lck->lk.depth_locked, 0,
920                              std::memory_order_relaxed);
921   // >= 0 for nestable locks, -1 for simple locks
922 }
923 
924 void __kmp_destroy_nested_ticket_lock(kmp_ticket_lock_t *lck) {
925   __kmp_destroy_ticket_lock(lck);
926   std::atomic_store_explicit(&lck->lk.depth_locked, 0,
927                              std::memory_order_relaxed);
928 }
929 
930 static void
931 __kmp_destroy_nested_ticket_lock_with_checks(kmp_ticket_lock_t *lck) {
932   char const *const func = "omp_destroy_nest_lock";
933 
934   if (!std::atomic_load_explicit(&lck->lk.initialized,
935                                  std::memory_order_relaxed)) {
936     KMP_FATAL(LockIsUninitialized, func);
937   }
938   if (lck->lk.self != lck) {
939     KMP_FATAL(LockIsUninitialized, func);
940   }
941   if (!__kmp_is_ticket_lock_nestable(lck)) {
942     KMP_FATAL(LockSimpleUsedAsNestable, func);
943   }
944   if (__kmp_get_ticket_lock_owner(lck) != -1) {
945     KMP_FATAL(LockStillOwned, func);
946   }
947   __kmp_destroy_nested_ticket_lock(lck);
948 }
949 
950 // access functions to fields which don't exist for all lock kinds.
951 
952 static const ident_t *__kmp_get_ticket_lock_location(kmp_ticket_lock_t *lck) {
953   return lck->lk.location;
954 }
955 
956 static void __kmp_set_ticket_lock_location(kmp_ticket_lock_t *lck,
957                                            const ident_t *loc) {
958   lck->lk.location = loc;
959 }
960 
961 static kmp_lock_flags_t __kmp_get_ticket_lock_flags(kmp_ticket_lock_t *lck) {
962   return lck->lk.flags;
963 }
964 
965 static void __kmp_set_ticket_lock_flags(kmp_ticket_lock_t *lck,
966                                         kmp_lock_flags_t flags) {
967   lck->lk.flags = flags;
968 }
969 
970 /* ------------------------------------------------------------------------ */
971 /* queuing locks */
972 
973 /* First the states
974    (head,tail) =              0, 0  means lock is unheld, nobody on queue
975                  UINT_MAX or -1, 0  means lock is held, nobody on queue
976                               h, h  means lock held or about to transition,
977                                     1 element on queue
978                               h, t  h <> t, means lock is held or about to
979                                     transition, >1 elements on queue
980 
981    Now the transitions
982       Acquire(0,0)  = -1 ,0
983       Release(0,0)  = Error
984       Acquire(-1,0) =  h ,h    h > 0
985       Release(-1,0) =  0 ,0
986       Acquire(h,h)  =  h ,t    h > 0, t > 0, h <> t
987       Release(h,h)  = -1 ,0    h > 0
988       Acquire(h,t)  =  h ,t'   h > 0, t > 0, t' > 0, h <> t, h <> t', t <> t'
989       Release(h,t)  =  h',t    h > 0, t > 0, h <> t, h <> h', h' maybe = t
990 
991    And pictorially
992 
993            +-----+
994            | 0, 0|------- release -------> Error
995            +-----+
996              |  ^
997       acquire|  |release
998              |  |
999              |  |
1000              v  |
1001            +-----+
1002            |-1, 0|
1003            +-----+
1004              |  ^
1005       acquire|  |release
1006              |  |
1007              |  |
1008              v  |
1009            +-----+
1010            | h, h|
1011            +-----+
1012              |  ^
1013       acquire|  |release
1014              |  |
1015              |  |
1016              v  |
1017            +-----+
1018            | h, t|----- acquire, release loopback ---+
1019            +-----+                                   |
1020                 ^                                    |
1021                 |                                    |
1022                 +------------------------------------+
1023  */
1024 
1025 #ifdef DEBUG_QUEUING_LOCKS
1026 
1027 /* Stuff for circular trace buffer */
1028 #define TRACE_BUF_ELE 1024
1029 static char traces[TRACE_BUF_ELE][128] = {0};
1030 static int tc = 0;
1031 #define TRACE_LOCK(X, Y)                                                       \
1032   KMP_SNPRINTF(traces[tc++ % TRACE_BUF_ELE], 128, "t%d at %s\n", X, Y);
1033 #define TRACE_LOCK_T(X, Y, Z)                                                  \
1034   KMP_SNPRINTF(traces[tc++ % TRACE_BUF_ELE], 128, "t%d at %s%d\n", X, Y, Z);
1035 #define TRACE_LOCK_HT(X, Y, Z, Q)                                              \
1036   KMP_SNPRINTF(traces[tc++ % TRACE_BUF_ELE], 128, "t%d at %s %d,%d\n", X, Y,   \
1037                Z, Q);
1038 
1039 static void __kmp_dump_queuing_lock(kmp_info_t *this_thr, kmp_int32 gtid,
1040                                     kmp_queuing_lock_t *lck, kmp_int32 head_id,
1041                                     kmp_int32 tail_id) {
1042   kmp_int32 t, i;
1043 
1044   __kmp_printf_no_lock("\n__kmp_dump_queuing_lock: TRACE BEGINS HERE! \n");
1045 
1046   i = tc % TRACE_BUF_ELE;
1047   __kmp_printf_no_lock("%s\n", traces[i]);
1048   i = (i + 1) % TRACE_BUF_ELE;
1049   while (i != (tc % TRACE_BUF_ELE)) {
1050     __kmp_printf_no_lock("%s", traces[i]);
1051     i = (i + 1) % TRACE_BUF_ELE;
1052   }
1053   __kmp_printf_no_lock("\n");
1054 
1055   __kmp_printf_no_lock("\n__kmp_dump_queuing_lock: gtid+1:%d, spin_here:%d, "
1056                        "next_wait:%d, head_id:%d, tail_id:%d\n",
1057                        gtid + 1, this_thr->th.th_spin_here,
1058                        this_thr->th.th_next_waiting, head_id, tail_id);
1059 
1060   __kmp_printf_no_lock("\t\thead: %d ", lck->lk.head_id);
1061 
1062   if (lck->lk.head_id >= 1) {
1063     t = __kmp_threads[lck->lk.head_id - 1]->th.th_next_waiting;
1064     while (t > 0) {
1065       __kmp_printf_no_lock("-> %d ", t);
1066       t = __kmp_threads[t - 1]->th.th_next_waiting;
1067     }
1068   }
1069   __kmp_printf_no_lock(";  tail: %d ", lck->lk.tail_id);
1070   __kmp_printf_no_lock("\n\n");
1071 }
1072 
1073 #endif /* DEBUG_QUEUING_LOCKS */
1074 
1075 static kmp_int32 __kmp_get_queuing_lock_owner(kmp_queuing_lock_t *lck) {
1076   return TCR_4(lck->lk.owner_id) - 1;
1077 }
1078 
1079 static inline bool __kmp_is_queuing_lock_nestable(kmp_queuing_lock_t *lck) {
1080   return lck->lk.depth_locked != -1;
1081 }
1082 
1083 /* Acquire a lock using a the queuing lock implementation */
1084 template <bool takeTime>
1085 /* [TLW] The unused template above is left behind because of what BEB believes
1086    is a potential compiler problem with __forceinline. */
1087 __forceinline static int
1088 __kmp_acquire_queuing_lock_timed_template(kmp_queuing_lock_t *lck,
1089                                           kmp_int32 gtid) {
1090   kmp_info_t *this_thr = __kmp_thread_from_gtid(gtid);
1091   volatile kmp_int32 *head_id_p = &lck->lk.head_id;
1092   volatile kmp_int32 *tail_id_p = &lck->lk.tail_id;
1093   volatile kmp_uint32 *spin_here_p;
1094 
1095 #if OMPT_SUPPORT
1096   ompt_state_t prev_state = ompt_state_undefined;
1097 #endif
1098 
1099   KA_TRACE(1000,
1100            ("__kmp_acquire_queuing_lock: lck:%p, T#%d entering\n", lck, gtid));
1101 
1102   KMP_FSYNC_PREPARE(lck);
1103   KMP_DEBUG_ASSERT(this_thr != NULL);
1104   spin_here_p = &this_thr->th.th_spin_here;
1105 
1106 #ifdef DEBUG_QUEUING_LOCKS
1107   TRACE_LOCK(gtid + 1, "acq ent");
1108   if (*spin_here_p)
1109     __kmp_dump_queuing_lock(this_thr, gtid, lck, *head_id_p, *tail_id_p);
1110   if (this_thr->th.th_next_waiting != 0)
1111     __kmp_dump_queuing_lock(this_thr, gtid, lck, *head_id_p, *tail_id_p);
1112 #endif
1113   KMP_DEBUG_ASSERT(!*spin_here_p);
1114   KMP_DEBUG_ASSERT(this_thr->th.th_next_waiting == 0);
1115 
1116   /* The following st.rel to spin_here_p needs to precede the cmpxchg.acq to
1117      head_id_p that may follow, not just in execution order, but also in
1118      visibility order. This way, when a releasing thread observes the changes to
1119      the queue by this thread, it can rightly assume that spin_here_p has
1120      already been set to TRUE, so that when it sets spin_here_p to FALSE, it is
1121      not premature.  If the releasing thread sets spin_here_p to FALSE before
1122      this thread sets it to TRUE, this thread will hang. */
1123   *spin_here_p = TRUE; /* before enqueuing to prevent race */
1124 
1125   while (1) {
1126     kmp_int32 enqueued;
1127     kmp_int32 head;
1128     kmp_int32 tail;
1129 
1130     head = *head_id_p;
1131 
1132     switch (head) {
1133 
1134     case -1: {
1135 #ifdef DEBUG_QUEUING_LOCKS
1136       tail = *tail_id_p;
1137       TRACE_LOCK_HT(gtid + 1, "acq read: ", head, tail);
1138 #endif
1139       tail = 0; /* to make sure next link asynchronously read is not set
1140                 accidentally; this assignment prevents us from entering the
1141                 if ( t > 0 ) condition in the enqueued case below, which is not
1142                 necessary for this state transition */
1143 
1144       /* try (-1,0)->(tid,tid) */
1145       enqueued = KMP_COMPARE_AND_STORE_ACQ64((volatile kmp_int64 *)tail_id_p,
1146                                              KMP_PACK_64(-1, 0),
1147                                              KMP_PACK_64(gtid + 1, gtid + 1));
1148 #ifdef DEBUG_QUEUING_LOCKS
1149       if (enqueued)
1150         TRACE_LOCK(gtid + 1, "acq enq: (-1,0)->(tid,tid)");
1151 #endif
1152     } break;
1153 
1154     default: {
1155       tail = *tail_id_p;
1156       KMP_DEBUG_ASSERT(tail != gtid + 1);
1157 
1158 #ifdef DEBUG_QUEUING_LOCKS
1159       TRACE_LOCK_HT(gtid + 1, "acq read: ", head, tail);
1160 #endif
1161 
1162       if (tail == 0) {
1163         enqueued = FALSE;
1164       } else {
1165         /* try (h,t) or (h,h)->(h,tid) */
1166         enqueued = KMP_COMPARE_AND_STORE_ACQ32(tail_id_p, tail, gtid + 1);
1167 
1168 #ifdef DEBUG_QUEUING_LOCKS
1169         if (enqueued)
1170           TRACE_LOCK(gtid + 1, "acq enq: (h,t)->(h,tid)");
1171 #endif
1172       }
1173     } break;
1174 
1175     case 0: /* empty queue */
1176     {
1177       kmp_int32 grabbed_lock;
1178 
1179 #ifdef DEBUG_QUEUING_LOCKS
1180       tail = *tail_id_p;
1181       TRACE_LOCK_HT(gtid + 1, "acq read: ", head, tail);
1182 #endif
1183       /* try (0,0)->(-1,0) */
1184 
1185       /* only legal transition out of head = 0 is head = -1 with no change to
1186        * tail */
1187       grabbed_lock = KMP_COMPARE_AND_STORE_ACQ32(head_id_p, 0, -1);
1188 
1189       if (grabbed_lock) {
1190 
1191         *spin_here_p = FALSE;
1192 
1193         KA_TRACE(
1194             1000,
1195             ("__kmp_acquire_queuing_lock: lck:%p, T#%d exiting: no queuing\n",
1196              lck, gtid));
1197 #ifdef DEBUG_QUEUING_LOCKS
1198         TRACE_LOCK_HT(gtid + 1, "acq exit: ", head, 0);
1199 #endif
1200 
1201 #if OMPT_SUPPORT
1202         if (ompt_enabled.enabled && prev_state != ompt_state_undefined) {
1203           /* change the state before clearing wait_id */
1204           this_thr->th.ompt_thread_info.state = prev_state;
1205           this_thr->th.ompt_thread_info.wait_id = 0;
1206         }
1207 #endif
1208 
1209         KMP_FSYNC_ACQUIRED(lck);
1210         return KMP_LOCK_ACQUIRED_FIRST; /* lock holder cannot be on queue */
1211       }
1212       enqueued = FALSE;
1213     } break;
1214     }
1215 
1216 #if OMPT_SUPPORT
1217     if (ompt_enabled.enabled && prev_state == ompt_state_undefined) {
1218       /* this thread will spin; set wait_id before entering wait state */
1219       prev_state = this_thr->th.ompt_thread_info.state;
1220       this_thr->th.ompt_thread_info.wait_id = (uint64_t)lck;
1221       this_thr->th.ompt_thread_info.state = ompt_state_wait_lock;
1222     }
1223 #endif
1224 
1225     if (enqueued) {
1226       if (tail > 0) {
1227         kmp_info_t *tail_thr = __kmp_thread_from_gtid(tail - 1);
1228         KMP_ASSERT(tail_thr != NULL);
1229         tail_thr->th.th_next_waiting = gtid + 1;
1230         /* corresponding wait for this write in release code */
1231       }
1232       KA_TRACE(1000,
1233                ("__kmp_acquire_queuing_lock: lck:%p, T#%d waiting for lock\n",
1234                 lck, gtid));
1235 
1236       KMP_MB();
1237       // ToDo: Use __kmp_wait_sleep or similar when blocktime != inf
1238       KMP_WAIT(spin_here_p, FALSE, KMP_EQ, lck);
1239       // Synchronize writes to both runtime thread structures
1240       // and writes in user code.
1241       KMP_MB();
1242 
1243 #ifdef DEBUG_QUEUING_LOCKS
1244       TRACE_LOCK(gtid + 1, "acq spin");
1245 
1246       if (this_thr->th.th_next_waiting != 0)
1247         __kmp_dump_queuing_lock(this_thr, gtid, lck, *head_id_p, *tail_id_p);
1248 #endif
1249       KMP_DEBUG_ASSERT(this_thr->th.th_next_waiting == 0);
1250       KA_TRACE(1000, ("__kmp_acquire_queuing_lock: lck:%p, T#%d exiting: after "
1251                       "waiting on queue\n",
1252                       lck, gtid));
1253 
1254 #ifdef DEBUG_QUEUING_LOCKS
1255       TRACE_LOCK(gtid + 1, "acq exit 2");
1256 #endif
1257 
1258 #if OMPT_SUPPORT
1259       /* change the state before clearing wait_id */
1260       this_thr->th.ompt_thread_info.state = prev_state;
1261       this_thr->th.ompt_thread_info.wait_id = 0;
1262 #endif
1263 
1264       /* got lock, we were dequeued by the thread that released lock */
1265       return KMP_LOCK_ACQUIRED_FIRST;
1266     }
1267 
1268     /* Yield if number of threads > number of logical processors */
1269     /* ToDo: Not sure why this should only be in oversubscription case,
1270        maybe should be traditional YIELD_INIT/YIELD_WHEN loop */
1271     KMP_YIELD_OVERSUB();
1272 
1273 #ifdef DEBUG_QUEUING_LOCKS
1274     TRACE_LOCK(gtid + 1, "acq retry");
1275 #endif
1276   }
1277   KMP_ASSERT2(0, "should not get here");
1278   return KMP_LOCK_ACQUIRED_FIRST;
1279 }
1280 
1281 int __kmp_acquire_queuing_lock(kmp_queuing_lock_t *lck, kmp_int32 gtid) {
1282   KMP_DEBUG_ASSERT(gtid >= 0);
1283 
1284   int retval = __kmp_acquire_queuing_lock_timed_template<false>(lck, gtid);
1285   ANNOTATE_QUEUING_ACQUIRED(lck);
1286   return retval;
1287 }
1288 
1289 static int __kmp_acquire_queuing_lock_with_checks(kmp_queuing_lock_t *lck,
1290                                                   kmp_int32 gtid) {
1291   char const *const func = "omp_set_lock";
1292   if (lck->lk.initialized != lck) {
1293     KMP_FATAL(LockIsUninitialized, func);
1294   }
1295   if (__kmp_is_queuing_lock_nestable(lck)) {
1296     KMP_FATAL(LockNestableUsedAsSimple, func);
1297   }
1298   if (__kmp_get_queuing_lock_owner(lck) == gtid) {
1299     KMP_FATAL(LockIsAlreadyOwned, func);
1300   }
1301 
1302   __kmp_acquire_queuing_lock(lck, gtid);
1303 
1304   lck->lk.owner_id = gtid + 1;
1305   return KMP_LOCK_ACQUIRED_FIRST;
1306 }
1307 
1308 int __kmp_test_queuing_lock(kmp_queuing_lock_t *lck, kmp_int32 gtid) {
1309   volatile kmp_int32 *head_id_p = &lck->lk.head_id;
1310   kmp_int32 head;
1311 #ifdef KMP_DEBUG
1312   kmp_info_t *this_thr;
1313 #endif
1314 
1315   KA_TRACE(1000, ("__kmp_test_queuing_lock: T#%d entering\n", gtid));
1316   KMP_DEBUG_ASSERT(gtid >= 0);
1317 #ifdef KMP_DEBUG
1318   this_thr = __kmp_thread_from_gtid(gtid);
1319   KMP_DEBUG_ASSERT(this_thr != NULL);
1320   KMP_DEBUG_ASSERT(!this_thr->th.th_spin_here);
1321 #endif
1322 
1323   head = *head_id_p;
1324 
1325   if (head == 0) { /* nobody on queue, nobody holding */
1326     /* try (0,0)->(-1,0) */
1327     if (KMP_COMPARE_AND_STORE_ACQ32(head_id_p, 0, -1)) {
1328       KA_TRACE(1000,
1329                ("__kmp_test_queuing_lock: T#%d exiting: holding lock\n", gtid));
1330       KMP_FSYNC_ACQUIRED(lck);
1331       ANNOTATE_QUEUING_ACQUIRED(lck);
1332       return TRUE;
1333     }
1334   }
1335 
1336   KA_TRACE(1000,
1337            ("__kmp_test_queuing_lock: T#%d exiting: without lock\n", gtid));
1338   return FALSE;
1339 }
1340 
1341 static int __kmp_test_queuing_lock_with_checks(kmp_queuing_lock_t *lck,
1342                                                kmp_int32 gtid) {
1343   char const *const func = "omp_test_lock";
1344   if (lck->lk.initialized != lck) {
1345     KMP_FATAL(LockIsUninitialized, func);
1346   }
1347   if (__kmp_is_queuing_lock_nestable(lck)) {
1348     KMP_FATAL(LockNestableUsedAsSimple, func);
1349   }
1350 
1351   int retval = __kmp_test_queuing_lock(lck, gtid);
1352 
1353   if (retval) {
1354     lck->lk.owner_id = gtid + 1;
1355   }
1356   return retval;
1357 }
1358 
1359 int __kmp_release_queuing_lock(kmp_queuing_lock_t *lck, kmp_int32 gtid) {
1360   kmp_info_t *this_thr;
1361   volatile kmp_int32 *head_id_p = &lck->lk.head_id;
1362   volatile kmp_int32 *tail_id_p = &lck->lk.tail_id;
1363 
1364   KA_TRACE(1000,
1365            ("__kmp_release_queuing_lock: lck:%p, T#%d entering\n", lck, gtid));
1366   KMP_DEBUG_ASSERT(gtid >= 0);
1367   this_thr = __kmp_thread_from_gtid(gtid);
1368   KMP_DEBUG_ASSERT(this_thr != NULL);
1369 #ifdef DEBUG_QUEUING_LOCKS
1370   TRACE_LOCK(gtid + 1, "rel ent");
1371 
1372   if (this_thr->th.th_spin_here)
1373     __kmp_dump_queuing_lock(this_thr, gtid, lck, *head_id_p, *tail_id_p);
1374   if (this_thr->th.th_next_waiting != 0)
1375     __kmp_dump_queuing_lock(this_thr, gtid, lck, *head_id_p, *tail_id_p);
1376 #endif
1377   KMP_DEBUG_ASSERT(!this_thr->th.th_spin_here);
1378   KMP_DEBUG_ASSERT(this_thr->th.th_next_waiting == 0);
1379 
1380   KMP_FSYNC_RELEASING(lck);
1381   ANNOTATE_QUEUING_RELEASED(lck);
1382 
1383   while (1) {
1384     kmp_int32 dequeued;
1385     kmp_int32 head;
1386     kmp_int32 tail;
1387 
1388     head = *head_id_p;
1389 
1390 #ifdef DEBUG_QUEUING_LOCKS
1391     tail = *tail_id_p;
1392     TRACE_LOCK_HT(gtid + 1, "rel read: ", head, tail);
1393     if (head == 0)
1394       __kmp_dump_queuing_lock(this_thr, gtid, lck, head, tail);
1395 #endif
1396     KMP_DEBUG_ASSERT(head !=
1397                      0); /* holding the lock, head must be -1 or queue head */
1398 
1399     if (head == -1) { /* nobody on queue */
1400       /* try (-1,0)->(0,0) */
1401       if (KMP_COMPARE_AND_STORE_REL32(head_id_p, -1, 0)) {
1402         KA_TRACE(
1403             1000,
1404             ("__kmp_release_queuing_lock: lck:%p, T#%d exiting: queue empty\n",
1405              lck, gtid));
1406 #ifdef DEBUG_QUEUING_LOCKS
1407         TRACE_LOCK_HT(gtid + 1, "rel exit: ", 0, 0);
1408 #endif
1409 
1410 #if OMPT_SUPPORT
1411 /* nothing to do - no other thread is trying to shift blame */
1412 #endif
1413         return KMP_LOCK_RELEASED;
1414       }
1415       dequeued = FALSE;
1416     } else {
1417       KMP_MB();
1418       tail = *tail_id_p;
1419       if (head == tail) { /* only one thread on the queue */
1420 #ifdef DEBUG_QUEUING_LOCKS
1421         if (head <= 0)
1422           __kmp_dump_queuing_lock(this_thr, gtid, lck, head, tail);
1423 #endif
1424         KMP_DEBUG_ASSERT(head > 0);
1425 
1426         /* try (h,h)->(-1,0) */
1427         dequeued = KMP_COMPARE_AND_STORE_REL64(
1428             RCAST(volatile kmp_int64 *, tail_id_p), KMP_PACK_64(head, head),
1429             KMP_PACK_64(-1, 0));
1430 #ifdef DEBUG_QUEUING_LOCKS
1431         TRACE_LOCK(gtid + 1, "rel deq: (h,h)->(-1,0)");
1432 #endif
1433 
1434       } else {
1435         volatile kmp_int32 *waiting_id_p;
1436         kmp_info_t *head_thr = __kmp_thread_from_gtid(head - 1);
1437         KMP_DEBUG_ASSERT(head_thr != NULL);
1438         waiting_id_p = &head_thr->th.th_next_waiting;
1439 
1440 /* Does this require synchronous reads? */
1441 #ifdef DEBUG_QUEUING_LOCKS
1442         if (head <= 0 || tail <= 0)
1443           __kmp_dump_queuing_lock(this_thr, gtid, lck, head, tail);
1444 #endif
1445         KMP_DEBUG_ASSERT(head > 0 && tail > 0);
1446 
1447         /* try (h,t)->(h',t) or (t,t) */
1448         KMP_MB();
1449         /* make sure enqueuing thread has time to update next waiting thread
1450          * field */
1451         *head_id_p =
1452             KMP_WAIT((volatile kmp_uint32 *)waiting_id_p, 0, KMP_NEQ, NULL);
1453 #ifdef DEBUG_QUEUING_LOCKS
1454         TRACE_LOCK(gtid + 1, "rel deq: (h,t)->(h',t)");
1455 #endif
1456         dequeued = TRUE;
1457       }
1458     }
1459 
1460     if (dequeued) {
1461       kmp_info_t *head_thr = __kmp_thread_from_gtid(head - 1);
1462       KMP_DEBUG_ASSERT(head_thr != NULL);
1463 
1464 /* Does this require synchronous reads? */
1465 #ifdef DEBUG_QUEUING_LOCKS
1466       if (head <= 0 || tail <= 0)
1467         __kmp_dump_queuing_lock(this_thr, gtid, lck, head, tail);
1468 #endif
1469       KMP_DEBUG_ASSERT(head > 0 && tail > 0);
1470 
1471       /* For clean code only. Thread not released until next statement prevents
1472          race with acquire code. */
1473       head_thr->th.th_next_waiting = 0;
1474 #ifdef DEBUG_QUEUING_LOCKS
1475       TRACE_LOCK_T(gtid + 1, "rel nw=0 for t=", head);
1476 #endif
1477 
1478       KMP_MB();
1479       /* reset spin value */
1480       head_thr->th.th_spin_here = FALSE;
1481 
1482       KA_TRACE(1000, ("__kmp_release_queuing_lock: lck:%p, T#%d exiting: after "
1483                       "dequeuing\n",
1484                       lck, gtid));
1485 #ifdef DEBUG_QUEUING_LOCKS
1486       TRACE_LOCK(gtid + 1, "rel exit 2");
1487 #endif
1488       return KMP_LOCK_RELEASED;
1489     }
1490     /* KMP_CPU_PAUSE(); don't want to make releasing thread hold up acquiring
1491        threads */
1492 
1493 #ifdef DEBUG_QUEUING_LOCKS
1494     TRACE_LOCK(gtid + 1, "rel retry");
1495 #endif
1496 
1497   } /* while */
1498   KMP_ASSERT2(0, "should not get here");
1499   return KMP_LOCK_RELEASED;
1500 }
1501 
1502 static int __kmp_release_queuing_lock_with_checks(kmp_queuing_lock_t *lck,
1503                                                   kmp_int32 gtid) {
1504   char const *const func = "omp_unset_lock";
1505   KMP_MB(); /* in case another processor initialized lock */
1506   if (lck->lk.initialized != lck) {
1507     KMP_FATAL(LockIsUninitialized, func);
1508   }
1509   if (__kmp_is_queuing_lock_nestable(lck)) {
1510     KMP_FATAL(LockNestableUsedAsSimple, func);
1511   }
1512   if (__kmp_get_queuing_lock_owner(lck) == -1) {
1513     KMP_FATAL(LockUnsettingFree, func);
1514   }
1515   if (__kmp_get_queuing_lock_owner(lck) != gtid) {
1516     KMP_FATAL(LockUnsettingSetByAnother, func);
1517   }
1518   lck->lk.owner_id = 0;
1519   return __kmp_release_queuing_lock(lck, gtid);
1520 }
1521 
1522 void __kmp_init_queuing_lock(kmp_queuing_lock_t *lck) {
1523   lck->lk.location = NULL;
1524   lck->lk.head_id = 0;
1525   lck->lk.tail_id = 0;
1526   lck->lk.next_ticket = 0;
1527   lck->lk.now_serving = 0;
1528   lck->lk.owner_id = 0; // no thread owns the lock.
1529   lck->lk.depth_locked = -1; // >= 0 for nestable locks, -1 for simple locks.
1530   lck->lk.initialized = lck;
1531 
1532   KA_TRACE(1000, ("__kmp_init_queuing_lock: lock %p initialized\n", lck));
1533 }
1534 
1535 void __kmp_destroy_queuing_lock(kmp_queuing_lock_t *lck) {
1536   lck->lk.initialized = NULL;
1537   lck->lk.location = NULL;
1538   lck->lk.head_id = 0;
1539   lck->lk.tail_id = 0;
1540   lck->lk.next_ticket = 0;
1541   lck->lk.now_serving = 0;
1542   lck->lk.owner_id = 0;
1543   lck->lk.depth_locked = -1;
1544 }
1545 
1546 static void __kmp_destroy_queuing_lock_with_checks(kmp_queuing_lock_t *lck) {
1547   char const *const func = "omp_destroy_lock";
1548   if (lck->lk.initialized != lck) {
1549     KMP_FATAL(LockIsUninitialized, func);
1550   }
1551   if (__kmp_is_queuing_lock_nestable(lck)) {
1552     KMP_FATAL(LockNestableUsedAsSimple, func);
1553   }
1554   if (__kmp_get_queuing_lock_owner(lck) != -1) {
1555     KMP_FATAL(LockStillOwned, func);
1556   }
1557   __kmp_destroy_queuing_lock(lck);
1558 }
1559 
1560 // nested queuing locks
1561 
1562 int __kmp_acquire_nested_queuing_lock(kmp_queuing_lock_t *lck, kmp_int32 gtid) {
1563   KMP_DEBUG_ASSERT(gtid >= 0);
1564 
1565   if (__kmp_get_queuing_lock_owner(lck) == gtid) {
1566     lck->lk.depth_locked += 1;
1567     return KMP_LOCK_ACQUIRED_NEXT;
1568   } else {
1569     __kmp_acquire_queuing_lock_timed_template<false>(lck, gtid);
1570     ANNOTATE_QUEUING_ACQUIRED(lck);
1571     KMP_MB();
1572     lck->lk.depth_locked = 1;
1573     KMP_MB();
1574     lck->lk.owner_id = gtid + 1;
1575     return KMP_LOCK_ACQUIRED_FIRST;
1576   }
1577 }
1578 
1579 static int
1580 __kmp_acquire_nested_queuing_lock_with_checks(kmp_queuing_lock_t *lck,
1581                                               kmp_int32 gtid) {
1582   char const *const func = "omp_set_nest_lock";
1583   if (lck->lk.initialized != lck) {
1584     KMP_FATAL(LockIsUninitialized, func);
1585   }
1586   if (!__kmp_is_queuing_lock_nestable(lck)) {
1587     KMP_FATAL(LockSimpleUsedAsNestable, func);
1588   }
1589   return __kmp_acquire_nested_queuing_lock(lck, gtid);
1590 }
1591 
1592 int __kmp_test_nested_queuing_lock(kmp_queuing_lock_t *lck, kmp_int32 gtid) {
1593   int retval;
1594 
1595   KMP_DEBUG_ASSERT(gtid >= 0);
1596 
1597   if (__kmp_get_queuing_lock_owner(lck) == gtid) {
1598     retval = ++lck->lk.depth_locked;
1599   } else if (!__kmp_test_queuing_lock(lck, gtid)) {
1600     retval = 0;
1601   } else {
1602     KMP_MB();
1603     retval = lck->lk.depth_locked = 1;
1604     KMP_MB();
1605     lck->lk.owner_id = gtid + 1;
1606   }
1607   return retval;
1608 }
1609 
1610 static int __kmp_test_nested_queuing_lock_with_checks(kmp_queuing_lock_t *lck,
1611                                                       kmp_int32 gtid) {
1612   char const *const func = "omp_test_nest_lock";
1613   if (lck->lk.initialized != lck) {
1614     KMP_FATAL(LockIsUninitialized, func);
1615   }
1616   if (!__kmp_is_queuing_lock_nestable(lck)) {
1617     KMP_FATAL(LockSimpleUsedAsNestable, func);
1618   }
1619   return __kmp_test_nested_queuing_lock(lck, gtid);
1620 }
1621 
1622 int __kmp_release_nested_queuing_lock(kmp_queuing_lock_t *lck, kmp_int32 gtid) {
1623   KMP_DEBUG_ASSERT(gtid >= 0);
1624 
1625   KMP_MB();
1626   if (--(lck->lk.depth_locked) == 0) {
1627     KMP_MB();
1628     lck->lk.owner_id = 0;
1629     __kmp_release_queuing_lock(lck, gtid);
1630     return KMP_LOCK_RELEASED;
1631   }
1632   return KMP_LOCK_STILL_HELD;
1633 }
1634 
1635 static int
1636 __kmp_release_nested_queuing_lock_with_checks(kmp_queuing_lock_t *lck,
1637                                               kmp_int32 gtid) {
1638   char const *const func = "omp_unset_nest_lock";
1639   KMP_MB(); /* in case another processor initialized lock */
1640   if (lck->lk.initialized != lck) {
1641     KMP_FATAL(LockIsUninitialized, func);
1642   }
1643   if (!__kmp_is_queuing_lock_nestable(lck)) {
1644     KMP_FATAL(LockSimpleUsedAsNestable, func);
1645   }
1646   if (__kmp_get_queuing_lock_owner(lck) == -1) {
1647     KMP_FATAL(LockUnsettingFree, func);
1648   }
1649   if (__kmp_get_queuing_lock_owner(lck) != gtid) {
1650     KMP_FATAL(LockUnsettingSetByAnother, func);
1651   }
1652   return __kmp_release_nested_queuing_lock(lck, gtid);
1653 }
1654 
1655 void __kmp_init_nested_queuing_lock(kmp_queuing_lock_t *lck) {
1656   __kmp_init_queuing_lock(lck);
1657   lck->lk.depth_locked = 0; // >= 0 for nestable locks, -1 for simple locks
1658 }
1659 
1660 void __kmp_destroy_nested_queuing_lock(kmp_queuing_lock_t *lck) {
1661   __kmp_destroy_queuing_lock(lck);
1662   lck->lk.depth_locked = 0;
1663 }
1664 
1665 static void
1666 __kmp_destroy_nested_queuing_lock_with_checks(kmp_queuing_lock_t *lck) {
1667   char const *const func = "omp_destroy_nest_lock";
1668   if (lck->lk.initialized != lck) {
1669     KMP_FATAL(LockIsUninitialized, func);
1670   }
1671   if (!__kmp_is_queuing_lock_nestable(lck)) {
1672     KMP_FATAL(LockSimpleUsedAsNestable, func);
1673   }
1674   if (__kmp_get_queuing_lock_owner(lck) != -1) {
1675     KMP_FATAL(LockStillOwned, func);
1676   }
1677   __kmp_destroy_nested_queuing_lock(lck);
1678 }
1679 
1680 // access functions to fields which don't exist for all lock kinds.
1681 
1682 static const ident_t *__kmp_get_queuing_lock_location(kmp_queuing_lock_t *lck) {
1683   return lck->lk.location;
1684 }
1685 
1686 static void __kmp_set_queuing_lock_location(kmp_queuing_lock_t *lck,
1687                                             const ident_t *loc) {
1688   lck->lk.location = loc;
1689 }
1690 
1691 static kmp_lock_flags_t __kmp_get_queuing_lock_flags(kmp_queuing_lock_t *lck) {
1692   return lck->lk.flags;
1693 }
1694 
1695 static void __kmp_set_queuing_lock_flags(kmp_queuing_lock_t *lck,
1696                                          kmp_lock_flags_t flags) {
1697   lck->lk.flags = flags;
1698 }
1699 
1700 #if KMP_USE_ADAPTIVE_LOCKS
1701 
1702 /* RTM Adaptive locks */
1703 
1704 #if KMP_HAVE_RTM_INTRINSICS
1705 #include <immintrin.h>
1706 #define SOFT_ABORT_MASK (_XABORT_RETRY | _XABORT_CONFLICT | _XABORT_EXPLICIT)
1707 
1708 #else
1709 
1710 // Values from the status register after failed speculation.
1711 #define _XBEGIN_STARTED (~0u)
1712 #define _XABORT_EXPLICIT (1 << 0)
1713 #define _XABORT_RETRY (1 << 1)
1714 #define _XABORT_CONFLICT (1 << 2)
1715 #define _XABORT_CAPACITY (1 << 3)
1716 #define _XABORT_DEBUG (1 << 4)
1717 #define _XABORT_NESTED (1 << 5)
1718 #define _XABORT_CODE(x) ((unsigned char)(((x) >> 24) & 0xFF))
1719 
1720 // Aborts for which it's worth trying again immediately
1721 #define SOFT_ABORT_MASK (_XABORT_RETRY | _XABORT_CONFLICT | _XABORT_EXPLICIT)
1722 
1723 #define STRINGIZE_INTERNAL(arg) #arg
1724 #define STRINGIZE(arg) STRINGIZE_INTERNAL(arg)
1725 
1726 // Access to RTM instructions
1727 /*A version of XBegin which returns -1 on speculation, and the value of EAX on
1728   an abort. This is the same definition as the compiler intrinsic that will be
1729   supported at some point. */
1730 static __inline int _xbegin() {
1731   int res = -1;
1732 
1733 #if KMP_OS_WINDOWS
1734 #if KMP_ARCH_X86_64
1735   _asm {
1736         _emit 0xC7
1737         _emit 0xF8
1738         _emit 2
1739         _emit 0
1740         _emit 0
1741         _emit 0
1742         jmp   L2
1743         mov   res, eax
1744     L2:
1745   }
1746 #else /* IA32 */
1747   _asm {
1748         _emit 0xC7
1749         _emit 0xF8
1750         _emit 2
1751         _emit 0
1752         _emit 0
1753         _emit 0
1754         jmp   L2
1755         mov   res, eax
1756     L2:
1757   }
1758 #endif // KMP_ARCH_X86_64
1759 #else
1760   /* Note that %eax must be noted as killed (clobbered), because the XSR is
1761      returned in %eax(%rax) on abort.  Other register values are restored, so
1762      don't need to be killed.
1763 
1764      We must also mark 'res' as an input and an output, since otherwise
1765      'res=-1' may be dropped as being dead, whereas we do need the assignment on
1766      the successful (i.e., non-abort) path. */
1767   __asm__ volatile("1: .byte  0xC7; .byte 0xF8;\n"
1768                    "   .long  1f-1b-6\n"
1769                    "    jmp   2f\n"
1770                    "1:  movl  %%eax,%0\n"
1771                    "2:"
1772                    : "+r"(res)::"memory", "%eax");
1773 #endif // KMP_OS_WINDOWS
1774   return res;
1775 }
1776 
1777 /* Transaction end */
1778 static __inline void _xend() {
1779 #if KMP_OS_WINDOWS
1780   __asm {
1781         _emit 0x0f
1782         _emit 0x01
1783         _emit 0xd5
1784   }
1785 #else
1786   __asm__ volatile(".byte 0x0f; .byte 0x01; .byte 0xd5" ::: "memory");
1787 #endif
1788 }
1789 
1790 /* This is a macro, the argument must be a single byte constant which can be
1791    evaluated by the inline assembler, since it is emitted as a byte into the
1792    assembly code. */
1793 // clang-format off
1794 #if KMP_OS_WINDOWS
1795 #define _xabort(ARG) _asm _emit 0xc6 _asm _emit 0xf8 _asm _emit ARG
1796 #else
1797 #define _xabort(ARG)                                                           \
1798   __asm__ volatile(".byte 0xC6; .byte 0xF8; .byte " STRINGIZE(ARG):::"memory");
1799 #endif
1800 // clang-format on
1801 #endif // KMP_COMPILER_ICC && __INTEL_COMPILER >= 1300
1802 
1803 // Statistics is collected for testing purpose
1804 #if KMP_DEBUG_ADAPTIVE_LOCKS
1805 
1806 // We accumulate speculative lock statistics when the lock is destroyed. We
1807 // keep locks that haven't been destroyed in the liveLocks list so that we can
1808 // grab their statistics too.
1809 static kmp_adaptive_lock_statistics_t destroyedStats;
1810 
1811 // To hold the list of live locks.
1812 static kmp_adaptive_lock_info_t liveLocks;
1813 
1814 // A lock so we can safely update the list of locks.
1815 static kmp_bootstrap_lock_t chain_lock =
1816     KMP_BOOTSTRAP_LOCK_INITIALIZER(chain_lock);
1817 
1818 // Initialize the list of stats.
1819 void __kmp_init_speculative_stats() {
1820   kmp_adaptive_lock_info_t *lck = &liveLocks;
1821 
1822   memset(CCAST(kmp_adaptive_lock_statistics_t *, &(lck->stats)), 0,
1823          sizeof(lck->stats));
1824   lck->stats.next = lck;
1825   lck->stats.prev = lck;
1826 
1827   KMP_ASSERT(lck->stats.next->stats.prev == lck);
1828   KMP_ASSERT(lck->stats.prev->stats.next == lck);
1829 
1830   __kmp_init_bootstrap_lock(&chain_lock);
1831 }
1832 
1833 // Insert the lock into the circular list
1834 static void __kmp_remember_lock(kmp_adaptive_lock_info_t *lck) {
1835   __kmp_acquire_bootstrap_lock(&chain_lock);
1836 
1837   lck->stats.next = liveLocks.stats.next;
1838   lck->stats.prev = &liveLocks;
1839 
1840   liveLocks.stats.next = lck;
1841   lck->stats.next->stats.prev = lck;
1842 
1843   KMP_ASSERT(lck->stats.next->stats.prev == lck);
1844   KMP_ASSERT(lck->stats.prev->stats.next == lck);
1845 
1846   __kmp_release_bootstrap_lock(&chain_lock);
1847 }
1848 
1849 static void __kmp_forget_lock(kmp_adaptive_lock_info_t *lck) {
1850   KMP_ASSERT(lck->stats.next->stats.prev == lck);
1851   KMP_ASSERT(lck->stats.prev->stats.next == lck);
1852 
1853   kmp_adaptive_lock_info_t *n = lck->stats.next;
1854   kmp_adaptive_lock_info_t *p = lck->stats.prev;
1855 
1856   n->stats.prev = p;
1857   p->stats.next = n;
1858 }
1859 
1860 static void __kmp_zero_speculative_stats(kmp_adaptive_lock_info_t *lck) {
1861   memset(CCAST(kmp_adaptive_lock_statistics_t *, &lck->stats), 0,
1862          sizeof(lck->stats));
1863   __kmp_remember_lock(lck);
1864 }
1865 
1866 static void __kmp_add_stats(kmp_adaptive_lock_statistics_t *t,
1867                             kmp_adaptive_lock_info_t *lck) {
1868   kmp_adaptive_lock_statistics_t volatile *s = &lck->stats;
1869 
1870   t->nonSpeculativeAcquireAttempts += lck->acquire_attempts;
1871   t->successfulSpeculations += s->successfulSpeculations;
1872   t->hardFailedSpeculations += s->hardFailedSpeculations;
1873   t->softFailedSpeculations += s->softFailedSpeculations;
1874   t->nonSpeculativeAcquires += s->nonSpeculativeAcquires;
1875   t->lemmingYields += s->lemmingYields;
1876 }
1877 
1878 static void __kmp_accumulate_speculative_stats(kmp_adaptive_lock_info_t *lck) {
1879   __kmp_acquire_bootstrap_lock(&chain_lock);
1880 
1881   __kmp_add_stats(&destroyedStats, lck);
1882   __kmp_forget_lock(lck);
1883 
1884   __kmp_release_bootstrap_lock(&chain_lock);
1885 }
1886 
1887 static float percent(kmp_uint32 count, kmp_uint32 total) {
1888   return (total == 0) ? 0.0 : (100.0 * count) / total;
1889 }
1890 
1891 void __kmp_print_speculative_stats() {
1892   kmp_adaptive_lock_statistics_t total = destroyedStats;
1893   kmp_adaptive_lock_info_t *lck;
1894 
1895   for (lck = liveLocks.stats.next; lck != &liveLocks; lck = lck->stats.next) {
1896     __kmp_add_stats(&total, lck);
1897   }
1898   kmp_adaptive_lock_statistics_t *t = &total;
1899   kmp_uint32 totalSections =
1900       t->nonSpeculativeAcquires + t->successfulSpeculations;
1901   kmp_uint32 totalSpeculations = t->successfulSpeculations +
1902                                  t->hardFailedSpeculations +
1903                                  t->softFailedSpeculations;
1904   if (totalSections <= 0)
1905     return;
1906 
1907   kmp_safe_raii_file_t statsFile;
1908   if (strcmp(__kmp_speculative_statsfile, "-") == 0) {
1909     statsFile.set_stdout();
1910   } else {
1911     size_t buffLen = KMP_STRLEN(__kmp_speculative_statsfile) + 20;
1912     char buffer[buffLen];
1913     KMP_SNPRINTF(&buffer[0], buffLen, __kmp_speculative_statsfile,
1914                  (kmp_int32)getpid());
1915     statsFile.open(buffer, "w");
1916   }
1917 
1918   fprintf(statsFile, "Speculative lock statistics (all approximate!)\n");
1919   fprintf(statsFile,
1920           " Lock parameters: \n"
1921           "   max_soft_retries               : %10d\n"
1922           "   max_badness                    : %10d\n",
1923           __kmp_adaptive_backoff_params.max_soft_retries,
1924           __kmp_adaptive_backoff_params.max_badness);
1925   fprintf(statsFile, " Non-speculative acquire attempts : %10d\n",
1926           t->nonSpeculativeAcquireAttempts);
1927   fprintf(statsFile, " Total critical sections          : %10d\n",
1928           totalSections);
1929   fprintf(statsFile, " Successful speculations          : %10d (%5.1f%%)\n",
1930           t->successfulSpeculations,
1931           percent(t->successfulSpeculations, totalSections));
1932   fprintf(statsFile, " Non-speculative acquires         : %10d (%5.1f%%)\n",
1933           t->nonSpeculativeAcquires,
1934           percent(t->nonSpeculativeAcquires, totalSections));
1935   fprintf(statsFile, " Lemming yields                   : %10d\n\n",
1936           t->lemmingYields);
1937 
1938   fprintf(statsFile, " Speculative acquire attempts     : %10d\n",
1939           totalSpeculations);
1940   fprintf(statsFile, " Successes                        : %10d (%5.1f%%)\n",
1941           t->successfulSpeculations,
1942           percent(t->successfulSpeculations, totalSpeculations));
1943   fprintf(statsFile, " Soft failures                    : %10d (%5.1f%%)\n",
1944           t->softFailedSpeculations,
1945           percent(t->softFailedSpeculations, totalSpeculations));
1946   fprintf(statsFile, " Hard failures                    : %10d (%5.1f%%)\n",
1947           t->hardFailedSpeculations,
1948           percent(t->hardFailedSpeculations, totalSpeculations));
1949 }
1950 
1951 #define KMP_INC_STAT(lck, stat) (lck->lk.adaptive.stats.stat++)
1952 #else
1953 #define KMP_INC_STAT(lck, stat)
1954 
1955 #endif // KMP_DEBUG_ADAPTIVE_LOCKS
1956 
1957 static inline bool __kmp_is_unlocked_queuing_lock(kmp_queuing_lock_t *lck) {
1958   // It is enough to check that the head_id is zero.
1959   // We don't also need to check the tail.
1960   bool res = lck->lk.head_id == 0;
1961 
1962 // We need a fence here, since we must ensure that no memory operations
1963 // from later in this thread float above that read.
1964 #if KMP_COMPILER_ICC
1965   _mm_mfence();
1966 #else
1967   __sync_synchronize();
1968 #endif
1969 
1970   return res;
1971 }
1972 
1973 // Functions for manipulating the badness
1974 static __inline void
1975 __kmp_update_badness_after_success(kmp_adaptive_lock_t *lck) {
1976   // Reset the badness to zero so we eagerly try to speculate again
1977   lck->lk.adaptive.badness = 0;
1978   KMP_INC_STAT(lck, successfulSpeculations);
1979 }
1980 
1981 // Create a bit mask with one more set bit.
1982 static __inline void __kmp_step_badness(kmp_adaptive_lock_t *lck) {
1983   kmp_uint32 newBadness = (lck->lk.adaptive.badness << 1) | 1;
1984   if (newBadness > lck->lk.adaptive.max_badness) {
1985     return;
1986   } else {
1987     lck->lk.adaptive.badness = newBadness;
1988   }
1989 }
1990 
1991 // Check whether speculation should be attempted.
1992 KMP_ATTRIBUTE_TARGET_RTM
1993 static __inline int __kmp_should_speculate(kmp_adaptive_lock_t *lck,
1994                                            kmp_int32 gtid) {
1995   kmp_uint32 badness = lck->lk.adaptive.badness;
1996   kmp_uint32 attempts = lck->lk.adaptive.acquire_attempts;
1997   int res = (attempts & badness) == 0;
1998   return res;
1999 }
2000 
2001 // Attempt to acquire only the speculative lock.
2002 // Does not back off to the non-speculative lock.
2003 KMP_ATTRIBUTE_TARGET_RTM
2004 static int __kmp_test_adaptive_lock_only(kmp_adaptive_lock_t *lck,
2005                                          kmp_int32 gtid) {
2006   int retries = lck->lk.adaptive.max_soft_retries;
2007 
2008   // We don't explicitly count the start of speculation, rather we record the
2009   // results (success, hard fail, soft fail). The sum of all of those is the
2010   // total number of times we started speculation since all speculations must
2011   // end one of those ways.
2012   do {
2013     kmp_uint32 status = _xbegin();
2014     // Switch this in to disable actual speculation but exercise at least some
2015     // of the rest of the code. Useful for debugging...
2016     // kmp_uint32 status = _XABORT_NESTED;
2017 
2018     if (status == _XBEGIN_STARTED) {
2019       /* We have successfully started speculation. Check that no-one acquired
2020          the lock for real between when we last looked and now. This also gets
2021          the lock cache line into our read-set, which we need so that we'll
2022          abort if anyone later claims it for real. */
2023       if (!__kmp_is_unlocked_queuing_lock(GET_QLK_PTR(lck))) {
2024         // Lock is now visibly acquired, so someone beat us to it. Abort the
2025         // transaction so we'll restart from _xbegin with the failure status.
2026         _xabort(0x01);
2027         KMP_ASSERT2(0, "should not get here");
2028       }
2029       return 1; // Lock has been acquired (speculatively)
2030     } else {
2031       // We have aborted, update the statistics
2032       if (status & SOFT_ABORT_MASK) {
2033         KMP_INC_STAT(lck, softFailedSpeculations);
2034         // and loop round to retry.
2035       } else {
2036         KMP_INC_STAT(lck, hardFailedSpeculations);
2037         // Give up if we had a hard failure.
2038         break;
2039       }
2040     }
2041   } while (retries--); // Loop while we have retries, and didn't fail hard.
2042 
2043   // Either we had a hard failure or we didn't succeed softly after
2044   // the full set of attempts, so back off the badness.
2045   __kmp_step_badness(lck);
2046   return 0;
2047 }
2048 
2049 // Attempt to acquire the speculative lock, or back off to the non-speculative
2050 // one if the speculative lock cannot be acquired.
2051 // We can succeed speculatively, non-speculatively, or fail.
2052 static int __kmp_test_adaptive_lock(kmp_adaptive_lock_t *lck, kmp_int32 gtid) {
2053   // First try to acquire the lock speculatively
2054   if (__kmp_should_speculate(lck, gtid) &&
2055       __kmp_test_adaptive_lock_only(lck, gtid))
2056     return 1;
2057 
2058   // Speculative acquisition failed, so try to acquire it non-speculatively.
2059   // Count the non-speculative acquire attempt
2060   lck->lk.adaptive.acquire_attempts++;
2061 
2062   // Use base, non-speculative lock.
2063   if (__kmp_test_queuing_lock(GET_QLK_PTR(lck), gtid)) {
2064     KMP_INC_STAT(lck, nonSpeculativeAcquires);
2065     return 1; // Lock is acquired (non-speculatively)
2066   } else {
2067     return 0; // Failed to acquire the lock, it's already visibly locked.
2068   }
2069 }
2070 
2071 static int __kmp_test_adaptive_lock_with_checks(kmp_adaptive_lock_t *lck,
2072                                                 kmp_int32 gtid) {
2073   char const *const func = "omp_test_lock";
2074   if (lck->lk.qlk.initialized != GET_QLK_PTR(lck)) {
2075     KMP_FATAL(LockIsUninitialized, func);
2076   }
2077 
2078   int retval = __kmp_test_adaptive_lock(lck, gtid);
2079 
2080   if (retval) {
2081     lck->lk.qlk.owner_id = gtid + 1;
2082   }
2083   return retval;
2084 }
2085 
2086 // Block until we can acquire a speculative, adaptive lock. We check whether we
2087 // should be trying to speculate. If we should be, we check the real lock to see
2088 // if it is free, and, if not, pause without attempting to acquire it until it
2089 // is. Then we try the speculative acquire. This means that although we suffer
2090 // from lemmings a little (because all we can't acquire the lock speculatively
2091 // until the queue of threads waiting has cleared), we don't get into a state
2092 // where we can never acquire the lock speculatively (because we force the queue
2093 // to clear by preventing new arrivals from entering the queue). This does mean
2094 // that when we're trying to break lemmings, the lock is no longer fair. However
2095 // OpenMP makes no guarantee that its locks are fair, so this isn't a real
2096 // problem.
2097 static void __kmp_acquire_adaptive_lock(kmp_adaptive_lock_t *lck,
2098                                         kmp_int32 gtid) {
2099   if (__kmp_should_speculate(lck, gtid)) {
2100     if (__kmp_is_unlocked_queuing_lock(GET_QLK_PTR(lck))) {
2101       if (__kmp_test_adaptive_lock_only(lck, gtid))
2102         return;
2103       // We tried speculation and failed, so give up.
2104     } else {
2105       // We can't try speculation until the lock is free, so we pause here
2106       // (without suspending on the queueing lock, to allow it to drain, then
2107       // try again. All other threads will also see the same result for
2108       // shouldSpeculate, so will be doing the same if they try to claim the
2109       // lock from now on.
2110       while (!__kmp_is_unlocked_queuing_lock(GET_QLK_PTR(lck))) {
2111         KMP_INC_STAT(lck, lemmingYields);
2112         KMP_YIELD(TRUE);
2113       }
2114 
2115       if (__kmp_test_adaptive_lock_only(lck, gtid))
2116         return;
2117     }
2118   }
2119 
2120   // Speculative acquisition failed, so acquire it non-speculatively.
2121   // Count the non-speculative acquire attempt
2122   lck->lk.adaptive.acquire_attempts++;
2123 
2124   __kmp_acquire_queuing_lock_timed_template<FALSE>(GET_QLK_PTR(lck), gtid);
2125   // We have acquired the base lock, so count that.
2126   KMP_INC_STAT(lck, nonSpeculativeAcquires);
2127   ANNOTATE_QUEUING_ACQUIRED(lck);
2128 }
2129 
2130 static void __kmp_acquire_adaptive_lock_with_checks(kmp_adaptive_lock_t *lck,
2131                                                     kmp_int32 gtid) {
2132   char const *const func = "omp_set_lock";
2133   if (lck->lk.qlk.initialized != GET_QLK_PTR(lck)) {
2134     KMP_FATAL(LockIsUninitialized, func);
2135   }
2136   if (__kmp_get_queuing_lock_owner(GET_QLK_PTR(lck)) == gtid) {
2137     KMP_FATAL(LockIsAlreadyOwned, func);
2138   }
2139 
2140   __kmp_acquire_adaptive_lock(lck, gtid);
2141 
2142   lck->lk.qlk.owner_id = gtid + 1;
2143 }
2144 
2145 KMP_ATTRIBUTE_TARGET_RTM
2146 static int __kmp_release_adaptive_lock(kmp_adaptive_lock_t *lck,
2147                                        kmp_int32 gtid) {
2148   if (__kmp_is_unlocked_queuing_lock(GET_QLK_PTR(
2149           lck))) { // If the lock doesn't look claimed we must be speculating.
2150     // (Or the user's code is buggy and they're releasing without locking;
2151     // if we had XTEST we'd be able to check that case...)
2152     _xend(); // Exit speculation
2153     __kmp_update_badness_after_success(lck);
2154   } else { // Since the lock *is* visibly locked we're not speculating,
2155     // so should use the underlying lock's release scheme.
2156     __kmp_release_queuing_lock(GET_QLK_PTR(lck), gtid);
2157   }
2158   return KMP_LOCK_RELEASED;
2159 }
2160 
2161 static int __kmp_release_adaptive_lock_with_checks(kmp_adaptive_lock_t *lck,
2162                                                    kmp_int32 gtid) {
2163   char const *const func = "omp_unset_lock";
2164   KMP_MB(); /* in case another processor initialized lock */
2165   if (lck->lk.qlk.initialized != GET_QLK_PTR(lck)) {
2166     KMP_FATAL(LockIsUninitialized, func);
2167   }
2168   if (__kmp_get_queuing_lock_owner(GET_QLK_PTR(lck)) == -1) {
2169     KMP_FATAL(LockUnsettingFree, func);
2170   }
2171   if (__kmp_get_queuing_lock_owner(GET_QLK_PTR(lck)) != gtid) {
2172     KMP_FATAL(LockUnsettingSetByAnother, func);
2173   }
2174   lck->lk.qlk.owner_id = 0;
2175   __kmp_release_adaptive_lock(lck, gtid);
2176   return KMP_LOCK_RELEASED;
2177 }
2178 
2179 static void __kmp_init_adaptive_lock(kmp_adaptive_lock_t *lck) {
2180   __kmp_init_queuing_lock(GET_QLK_PTR(lck));
2181   lck->lk.adaptive.badness = 0;
2182   lck->lk.adaptive.acquire_attempts = 0; // nonSpeculativeAcquireAttempts = 0;
2183   lck->lk.adaptive.max_soft_retries =
2184       __kmp_adaptive_backoff_params.max_soft_retries;
2185   lck->lk.adaptive.max_badness = __kmp_adaptive_backoff_params.max_badness;
2186 #if KMP_DEBUG_ADAPTIVE_LOCKS
2187   __kmp_zero_speculative_stats(&lck->lk.adaptive);
2188 #endif
2189   KA_TRACE(1000, ("__kmp_init_adaptive_lock: lock %p initialized\n", lck));
2190 }
2191 
2192 static void __kmp_destroy_adaptive_lock(kmp_adaptive_lock_t *lck) {
2193 #if KMP_DEBUG_ADAPTIVE_LOCKS
2194   __kmp_accumulate_speculative_stats(&lck->lk.adaptive);
2195 #endif
2196   __kmp_destroy_queuing_lock(GET_QLK_PTR(lck));
2197   // Nothing needed for the speculative part.
2198 }
2199 
2200 static void __kmp_destroy_adaptive_lock_with_checks(kmp_adaptive_lock_t *lck) {
2201   char const *const func = "omp_destroy_lock";
2202   if (lck->lk.qlk.initialized != GET_QLK_PTR(lck)) {
2203     KMP_FATAL(LockIsUninitialized, func);
2204   }
2205   if (__kmp_get_queuing_lock_owner(GET_QLK_PTR(lck)) != -1) {
2206     KMP_FATAL(LockStillOwned, func);
2207   }
2208   __kmp_destroy_adaptive_lock(lck);
2209 }
2210 
2211 #endif // KMP_USE_ADAPTIVE_LOCKS
2212 
2213 /* ------------------------------------------------------------------------ */
2214 /* DRDPA ticket locks                                                */
2215 /* "DRDPA" means Dynamically Reconfigurable Distributed Polling Area */
2216 
2217 static kmp_int32 __kmp_get_drdpa_lock_owner(kmp_drdpa_lock_t *lck) {
2218   return lck->lk.owner_id - 1;
2219 }
2220 
2221 static inline bool __kmp_is_drdpa_lock_nestable(kmp_drdpa_lock_t *lck) {
2222   return lck->lk.depth_locked != -1;
2223 }
2224 
2225 __forceinline static int
2226 __kmp_acquire_drdpa_lock_timed_template(kmp_drdpa_lock_t *lck, kmp_int32 gtid) {
2227   kmp_uint64 ticket = KMP_ATOMIC_INC(&lck->lk.next_ticket);
2228   kmp_uint64 mask = lck->lk.mask; // atomic load
2229   std::atomic<kmp_uint64> *polls = lck->lk.polls;
2230 
2231 #ifdef USE_LOCK_PROFILE
2232   if (polls[ticket & mask] != ticket)
2233     __kmp_printf("LOCK CONTENTION: %p\n", lck);
2234 /* else __kmp_printf( "." );*/
2235 #endif /* USE_LOCK_PROFILE */
2236 
2237   // Now spin-wait, but reload the polls pointer and mask, in case the
2238   // polling area has been reconfigured.  Unless it is reconfigured, the
2239   // reloads stay in L1 cache and are cheap.
2240   //
2241   // Keep this code in sync with KMP_WAIT, in kmp_dispatch.cpp !!!
2242   // The current implementation of KMP_WAIT doesn't allow for mask
2243   // and poll to be re-read every spin iteration.
2244   kmp_uint32 spins;
2245   KMP_FSYNC_PREPARE(lck);
2246   KMP_INIT_YIELD(spins);
2247   while (polls[ticket & mask] < ticket) { // atomic load
2248     KMP_YIELD_OVERSUB_ELSE_SPIN(spins);
2249     // Re-read the mask and the poll pointer from the lock structure.
2250     //
2251     // Make certain that "mask" is read before "polls" !!!
2252     //
2253     // If another thread picks reconfigures the polling area and updates their
2254     // values, and we get the new value of mask and the old polls pointer, we
2255     // could access memory beyond the end of the old polling area.
2256     mask = lck->lk.mask; // atomic load
2257     polls = lck->lk.polls; // atomic load
2258   }
2259 
2260   // Critical section starts here
2261   KMP_FSYNC_ACQUIRED(lck);
2262   KA_TRACE(1000, ("__kmp_acquire_drdpa_lock: ticket #%lld acquired lock %p\n",
2263                   ticket, lck));
2264   lck->lk.now_serving = ticket; // non-volatile store
2265 
2266   // Deallocate a garbage polling area if we know that we are the last
2267   // thread that could possibly access it.
2268   //
2269   // The >= check is in case __kmp_test_drdpa_lock() allocated the cleanup
2270   // ticket.
2271   if ((lck->lk.old_polls != NULL) && (ticket >= lck->lk.cleanup_ticket)) {
2272     __kmp_free(lck->lk.old_polls);
2273     lck->lk.old_polls = NULL;
2274     lck->lk.cleanup_ticket = 0;
2275   }
2276 
2277   // Check to see if we should reconfigure the polling area.
2278   // If there is still a garbage polling area to be deallocated from a
2279   // previous reconfiguration, let a later thread reconfigure it.
2280   if (lck->lk.old_polls == NULL) {
2281     bool reconfigure = false;
2282     std::atomic<kmp_uint64> *old_polls = polls;
2283     kmp_uint32 num_polls = TCR_4(lck->lk.num_polls);
2284 
2285     if (TCR_4(__kmp_nth) >
2286         (__kmp_avail_proc ? __kmp_avail_proc : __kmp_xproc)) {
2287       // We are in oversubscription mode.  Contract the polling area
2288       // down to a single location, if that hasn't been done already.
2289       if (num_polls > 1) {
2290         reconfigure = true;
2291         num_polls = TCR_4(lck->lk.num_polls);
2292         mask = 0;
2293         num_polls = 1;
2294         polls = (std::atomic<kmp_uint64> *)__kmp_allocate(num_polls *
2295                                                           sizeof(*polls));
2296         polls[0] = ticket;
2297       }
2298     } else {
2299       // We are in under/fully subscribed mode.  Check the number of
2300       // threads waiting on the lock.  The size of the polling area
2301       // should be at least the number of threads waiting.
2302       kmp_uint64 num_waiting = TCR_8(lck->lk.next_ticket) - ticket - 1;
2303       if (num_waiting > num_polls) {
2304         kmp_uint32 old_num_polls = num_polls;
2305         reconfigure = true;
2306         do {
2307           mask = (mask << 1) | 1;
2308           num_polls *= 2;
2309         } while (num_polls <= num_waiting);
2310 
2311         // Allocate the new polling area, and copy the relevant portion
2312         // of the old polling area to the new area.  __kmp_allocate()
2313         // zeroes the memory it allocates, and most of the old area is
2314         // just zero padding, so we only copy the release counters.
2315         polls = (std::atomic<kmp_uint64> *)__kmp_allocate(num_polls *
2316                                                           sizeof(*polls));
2317         kmp_uint32 i;
2318         for (i = 0; i < old_num_polls; i++) {
2319           polls[i].store(old_polls[i]);
2320         }
2321       }
2322     }
2323 
2324     if (reconfigure) {
2325       // Now write the updated fields back to the lock structure.
2326       //
2327       // Make certain that "polls" is written before "mask" !!!
2328       //
2329       // If another thread picks up the new value of mask and the old polls
2330       // pointer , it could access memory beyond the end of the old polling
2331       // area.
2332       //
2333       // On x86, we need memory fences.
2334       KA_TRACE(1000, ("__kmp_acquire_drdpa_lock: ticket #%lld reconfiguring "
2335                       "lock %p to %d polls\n",
2336                       ticket, lck, num_polls));
2337 
2338       lck->lk.old_polls = old_polls;
2339       lck->lk.polls = polls; // atomic store
2340 
2341       KMP_MB();
2342 
2343       lck->lk.num_polls = num_polls;
2344       lck->lk.mask = mask; // atomic store
2345 
2346       KMP_MB();
2347 
2348       // Only after the new polling area and mask have been flushed
2349       // to main memory can we update the cleanup ticket field.
2350       //
2351       // volatile load / non-volatile store
2352       lck->lk.cleanup_ticket = lck->lk.next_ticket;
2353     }
2354   }
2355   return KMP_LOCK_ACQUIRED_FIRST;
2356 }
2357 
2358 int __kmp_acquire_drdpa_lock(kmp_drdpa_lock_t *lck, kmp_int32 gtid) {
2359   int retval = __kmp_acquire_drdpa_lock_timed_template(lck, gtid);
2360   ANNOTATE_DRDPA_ACQUIRED(lck);
2361   return retval;
2362 }
2363 
2364 static int __kmp_acquire_drdpa_lock_with_checks(kmp_drdpa_lock_t *lck,
2365                                                 kmp_int32 gtid) {
2366   char const *const func = "omp_set_lock";
2367   if (lck->lk.initialized != lck) {
2368     KMP_FATAL(LockIsUninitialized, func);
2369   }
2370   if (__kmp_is_drdpa_lock_nestable(lck)) {
2371     KMP_FATAL(LockNestableUsedAsSimple, func);
2372   }
2373   if ((gtid >= 0) && (__kmp_get_drdpa_lock_owner(lck) == gtid)) {
2374     KMP_FATAL(LockIsAlreadyOwned, func);
2375   }
2376 
2377   __kmp_acquire_drdpa_lock(lck, gtid);
2378 
2379   lck->lk.owner_id = gtid + 1;
2380   return KMP_LOCK_ACQUIRED_FIRST;
2381 }
2382 
2383 int __kmp_test_drdpa_lock(kmp_drdpa_lock_t *lck, kmp_int32 gtid) {
2384   // First get a ticket, then read the polls pointer and the mask.
2385   // The polls pointer must be read before the mask!!! (See above)
2386   kmp_uint64 ticket = lck->lk.next_ticket; // atomic load
2387   std::atomic<kmp_uint64> *polls = lck->lk.polls;
2388   kmp_uint64 mask = lck->lk.mask; // atomic load
2389   if (polls[ticket & mask] == ticket) {
2390     kmp_uint64 next_ticket = ticket + 1;
2391     if (__kmp_atomic_compare_store_acq(&lck->lk.next_ticket, ticket,
2392                                        next_ticket)) {
2393       KMP_FSYNC_ACQUIRED(lck);
2394       KA_TRACE(1000, ("__kmp_test_drdpa_lock: ticket #%lld acquired lock %p\n",
2395                       ticket, lck));
2396       lck->lk.now_serving = ticket; // non-volatile store
2397 
2398       // Since no threads are waiting, there is no possibility that we would
2399       // want to reconfigure the polling area.  We might have the cleanup ticket
2400       // value (which says that it is now safe to deallocate old_polls), but
2401       // we'll let a later thread which calls __kmp_acquire_lock do that - this
2402       // routine isn't supposed to block, and we would risk blocks if we called
2403       // __kmp_free() to do the deallocation.
2404       return TRUE;
2405     }
2406   }
2407   return FALSE;
2408 }
2409 
2410 static int __kmp_test_drdpa_lock_with_checks(kmp_drdpa_lock_t *lck,
2411                                              kmp_int32 gtid) {
2412   char const *const func = "omp_test_lock";
2413   if (lck->lk.initialized != lck) {
2414     KMP_FATAL(LockIsUninitialized, func);
2415   }
2416   if (__kmp_is_drdpa_lock_nestable(lck)) {
2417     KMP_FATAL(LockNestableUsedAsSimple, func);
2418   }
2419 
2420   int retval = __kmp_test_drdpa_lock(lck, gtid);
2421 
2422   if (retval) {
2423     lck->lk.owner_id = gtid + 1;
2424   }
2425   return retval;
2426 }
2427 
2428 int __kmp_release_drdpa_lock(kmp_drdpa_lock_t *lck, kmp_int32 gtid) {
2429   // Read the ticket value from the lock data struct, then the polls pointer and
2430   // the mask.  The polls pointer must be read before the mask!!! (See above)
2431   kmp_uint64 ticket = lck->lk.now_serving + 1; // non-atomic load
2432   std::atomic<kmp_uint64> *polls = lck->lk.polls; // atomic load
2433   kmp_uint64 mask = lck->lk.mask; // atomic load
2434   KA_TRACE(1000, ("__kmp_release_drdpa_lock: ticket #%lld released lock %p\n",
2435                   ticket - 1, lck));
2436   KMP_FSYNC_RELEASING(lck);
2437   ANNOTATE_DRDPA_RELEASED(lck);
2438   polls[ticket & mask] = ticket; // atomic store
2439   return KMP_LOCK_RELEASED;
2440 }
2441 
2442 static int __kmp_release_drdpa_lock_with_checks(kmp_drdpa_lock_t *lck,
2443                                                 kmp_int32 gtid) {
2444   char const *const func = "omp_unset_lock";
2445   KMP_MB(); /* in case another processor initialized lock */
2446   if (lck->lk.initialized != lck) {
2447     KMP_FATAL(LockIsUninitialized, func);
2448   }
2449   if (__kmp_is_drdpa_lock_nestable(lck)) {
2450     KMP_FATAL(LockNestableUsedAsSimple, func);
2451   }
2452   if (__kmp_get_drdpa_lock_owner(lck) == -1) {
2453     KMP_FATAL(LockUnsettingFree, func);
2454   }
2455   if ((gtid >= 0) && (__kmp_get_drdpa_lock_owner(lck) >= 0) &&
2456       (__kmp_get_drdpa_lock_owner(lck) != gtid)) {
2457     KMP_FATAL(LockUnsettingSetByAnother, func);
2458   }
2459   lck->lk.owner_id = 0;
2460   return __kmp_release_drdpa_lock(lck, gtid);
2461 }
2462 
2463 void __kmp_init_drdpa_lock(kmp_drdpa_lock_t *lck) {
2464   lck->lk.location = NULL;
2465   lck->lk.mask = 0;
2466   lck->lk.num_polls = 1;
2467   lck->lk.polls = (std::atomic<kmp_uint64> *)__kmp_allocate(
2468       lck->lk.num_polls * sizeof(*(lck->lk.polls)));
2469   lck->lk.cleanup_ticket = 0;
2470   lck->lk.old_polls = NULL;
2471   lck->lk.next_ticket = 0;
2472   lck->lk.now_serving = 0;
2473   lck->lk.owner_id = 0; // no thread owns the lock.
2474   lck->lk.depth_locked = -1; // >= 0 for nestable locks, -1 for simple locks.
2475   lck->lk.initialized = lck;
2476 
2477   KA_TRACE(1000, ("__kmp_init_drdpa_lock: lock %p initialized\n", lck));
2478 }
2479 
2480 void __kmp_destroy_drdpa_lock(kmp_drdpa_lock_t *lck) {
2481   lck->lk.initialized = NULL;
2482   lck->lk.location = NULL;
2483   if (lck->lk.polls.load() != NULL) {
2484     __kmp_free(lck->lk.polls.load());
2485     lck->lk.polls = NULL;
2486   }
2487   if (lck->lk.old_polls != NULL) {
2488     __kmp_free(lck->lk.old_polls);
2489     lck->lk.old_polls = NULL;
2490   }
2491   lck->lk.mask = 0;
2492   lck->lk.num_polls = 0;
2493   lck->lk.cleanup_ticket = 0;
2494   lck->lk.next_ticket = 0;
2495   lck->lk.now_serving = 0;
2496   lck->lk.owner_id = 0;
2497   lck->lk.depth_locked = -1;
2498 }
2499 
2500 static void __kmp_destroy_drdpa_lock_with_checks(kmp_drdpa_lock_t *lck) {
2501   char const *const func = "omp_destroy_lock";
2502   if (lck->lk.initialized != lck) {
2503     KMP_FATAL(LockIsUninitialized, func);
2504   }
2505   if (__kmp_is_drdpa_lock_nestable(lck)) {
2506     KMP_FATAL(LockNestableUsedAsSimple, func);
2507   }
2508   if (__kmp_get_drdpa_lock_owner(lck) != -1) {
2509     KMP_FATAL(LockStillOwned, func);
2510   }
2511   __kmp_destroy_drdpa_lock(lck);
2512 }
2513 
2514 // nested drdpa ticket locks
2515 
2516 int __kmp_acquire_nested_drdpa_lock(kmp_drdpa_lock_t *lck, kmp_int32 gtid) {
2517   KMP_DEBUG_ASSERT(gtid >= 0);
2518 
2519   if (__kmp_get_drdpa_lock_owner(lck) == gtid) {
2520     lck->lk.depth_locked += 1;
2521     return KMP_LOCK_ACQUIRED_NEXT;
2522   } else {
2523     __kmp_acquire_drdpa_lock_timed_template(lck, gtid);
2524     ANNOTATE_DRDPA_ACQUIRED(lck);
2525     KMP_MB();
2526     lck->lk.depth_locked = 1;
2527     KMP_MB();
2528     lck->lk.owner_id = gtid + 1;
2529     return KMP_LOCK_ACQUIRED_FIRST;
2530   }
2531 }
2532 
2533 static void __kmp_acquire_nested_drdpa_lock_with_checks(kmp_drdpa_lock_t *lck,
2534                                                         kmp_int32 gtid) {
2535   char const *const func = "omp_set_nest_lock";
2536   if (lck->lk.initialized != lck) {
2537     KMP_FATAL(LockIsUninitialized, func);
2538   }
2539   if (!__kmp_is_drdpa_lock_nestable(lck)) {
2540     KMP_FATAL(LockSimpleUsedAsNestable, func);
2541   }
2542   __kmp_acquire_nested_drdpa_lock(lck, gtid);
2543 }
2544 
2545 int __kmp_test_nested_drdpa_lock(kmp_drdpa_lock_t *lck, kmp_int32 gtid) {
2546   int retval;
2547 
2548   KMP_DEBUG_ASSERT(gtid >= 0);
2549 
2550   if (__kmp_get_drdpa_lock_owner(lck) == gtid) {
2551     retval = ++lck->lk.depth_locked;
2552   } else if (!__kmp_test_drdpa_lock(lck, gtid)) {
2553     retval = 0;
2554   } else {
2555     KMP_MB();
2556     retval = lck->lk.depth_locked = 1;
2557     KMP_MB();
2558     lck->lk.owner_id = gtid + 1;
2559   }
2560   return retval;
2561 }
2562 
2563 static int __kmp_test_nested_drdpa_lock_with_checks(kmp_drdpa_lock_t *lck,
2564                                                     kmp_int32 gtid) {
2565   char const *const func = "omp_test_nest_lock";
2566   if (lck->lk.initialized != lck) {
2567     KMP_FATAL(LockIsUninitialized, func);
2568   }
2569   if (!__kmp_is_drdpa_lock_nestable(lck)) {
2570     KMP_FATAL(LockSimpleUsedAsNestable, func);
2571   }
2572   return __kmp_test_nested_drdpa_lock(lck, gtid);
2573 }
2574 
2575 int __kmp_release_nested_drdpa_lock(kmp_drdpa_lock_t *lck, kmp_int32 gtid) {
2576   KMP_DEBUG_ASSERT(gtid >= 0);
2577 
2578   KMP_MB();
2579   if (--(lck->lk.depth_locked) == 0) {
2580     KMP_MB();
2581     lck->lk.owner_id = 0;
2582     __kmp_release_drdpa_lock(lck, gtid);
2583     return KMP_LOCK_RELEASED;
2584   }
2585   return KMP_LOCK_STILL_HELD;
2586 }
2587 
2588 static int __kmp_release_nested_drdpa_lock_with_checks(kmp_drdpa_lock_t *lck,
2589                                                        kmp_int32 gtid) {
2590   char const *const func = "omp_unset_nest_lock";
2591   KMP_MB(); /* in case another processor initialized lock */
2592   if (lck->lk.initialized != lck) {
2593     KMP_FATAL(LockIsUninitialized, func);
2594   }
2595   if (!__kmp_is_drdpa_lock_nestable(lck)) {
2596     KMP_FATAL(LockSimpleUsedAsNestable, func);
2597   }
2598   if (__kmp_get_drdpa_lock_owner(lck) == -1) {
2599     KMP_FATAL(LockUnsettingFree, func);
2600   }
2601   if (__kmp_get_drdpa_lock_owner(lck) != gtid) {
2602     KMP_FATAL(LockUnsettingSetByAnother, func);
2603   }
2604   return __kmp_release_nested_drdpa_lock(lck, gtid);
2605 }
2606 
2607 void __kmp_init_nested_drdpa_lock(kmp_drdpa_lock_t *lck) {
2608   __kmp_init_drdpa_lock(lck);
2609   lck->lk.depth_locked = 0; // >= 0 for nestable locks, -1 for simple locks
2610 }
2611 
2612 void __kmp_destroy_nested_drdpa_lock(kmp_drdpa_lock_t *lck) {
2613   __kmp_destroy_drdpa_lock(lck);
2614   lck->lk.depth_locked = 0;
2615 }
2616 
2617 static void __kmp_destroy_nested_drdpa_lock_with_checks(kmp_drdpa_lock_t *lck) {
2618   char const *const func = "omp_destroy_nest_lock";
2619   if (lck->lk.initialized != lck) {
2620     KMP_FATAL(LockIsUninitialized, func);
2621   }
2622   if (!__kmp_is_drdpa_lock_nestable(lck)) {
2623     KMP_FATAL(LockSimpleUsedAsNestable, func);
2624   }
2625   if (__kmp_get_drdpa_lock_owner(lck) != -1) {
2626     KMP_FATAL(LockStillOwned, func);
2627   }
2628   __kmp_destroy_nested_drdpa_lock(lck);
2629 }
2630 
2631 // access functions to fields which don't exist for all lock kinds.
2632 
2633 static const ident_t *__kmp_get_drdpa_lock_location(kmp_drdpa_lock_t *lck) {
2634   return lck->lk.location;
2635 }
2636 
2637 static void __kmp_set_drdpa_lock_location(kmp_drdpa_lock_t *lck,
2638                                           const ident_t *loc) {
2639   lck->lk.location = loc;
2640 }
2641 
2642 static kmp_lock_flags_t __kmp_get_drdpa_lock_flags(kmp_drdpa_lock_t *lck) {
2643   return lck->lk.flags;
2644 }
2645 
2646 static void __kmp_set_drdpa_lock_flags(kmp_drdpa_lock_t *lck,
2647                                        kmp_lock_flags_t flags) {
2648   lck->lk.flags = flags;
2649 }
2650 
2651 // Time stamp counter
2652 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
2653 #define __kmp_tsc() __kmp_hardware_timestamp()
2654 // Runtime's default backoff parameters
2655 kmp_backoff_t __kmp_spin_backoff_params = {1, 4096, 100};
2656 #else
2657 // Use nanoseconds for other platforms
2658 extern kmp_uint64 __kmp_now_nsec();
2659 kmp_backoff_t __kmp_spin_backoff_params = {1, 256, 100};
2660 #define __kmp_tsc() __kmp_now_nsec()
2661 #endif
2662 
2663 // A useful predicate for dealing with timestamps that may wrap.
2664 // Is a before b? Since the timestamps may wrap, this is asking whether it's
2665 // shorter to go clockwise from a to b around the clock-face, or anti-clockwise.
2666 // Times where going clockwise is less distance than going anti-clockwise
2667 // are in the future, others are in the past. e.g. a = MAX-1, b = MAX+1 (=0),
2668 // then a > b (true) does not mean a reached b; whereas signed(a) = -2,
2669 // signed(b) = 0 captures the actual difference
2670 static inline bool before(kmp_uint64 a, kmp_uint64 b) {
2671   return ((kmp_int64)b - (kmp_int64)a) > 0;
2672 }
2673 
2674 // Truncated binary exponential backoff function
2675 void __kmp_spin_backoff(kmp_backoff_t *boff) {
2676   // We could flatten this loop, but making it a nested loop gives better result
2677   kmp_uint32 i;
2678   for (i = boff->step; i > 0; i--) {
2679     kmp_uint64 goal = __kmp_tsc() + boff->min_tick;
2680     do {
2681       KMP_CPU_PAUSE();
2682     } while (before(__kmp_tsc(), goal));
2683   }
2684   boff->step = (boff->step << 1 | 1) & (boff->max_backoff - 1);
2685 }
2686 
2687 #if KMP_USE_DYNAMIC_LOCK
2688 
2689 // Direct lock initializers. It simply writes a tag to the low 8 bits of the
2690 // lock word.
2691 static void __kmp_init_direct_lock(kmp_dyna_lock_t *lck,
2692                                    kmp_dyna_lockseq_t seq) {
2693   TCW_4(*lck, KMP_GET_D_TAG(seq));
2694   KA_TRACE(
2695       20,
2696       ("__kmp_init_direct_lock: initialized direct lock with type#%d\n", seq));
2697 }
2698 
2699 #if KMP_USE_TSX
2700 
2701 // HLE lock functions - imported from the testbed runtime.
2702 #define HLE_ACQUIRE ".byte 0xf2;"
2703 #define HLE_RELEASE ".byte 0xf3;"
2704 
2705 static inline kmp_uint32 swap4(kmp_uint32 volatile *p, kmp_uint32 v) {
2706   __asm__ volatile(HLE_ACQUIRE "xchg %1,%0" : "+r"(v), "+m"(*p) : : "memory");
2707   return v;
2708 }
2709 
2710 static void __kmp_destroy_hle_lock(kmp_dyna_lock_t *lck) { TCW_4(*lck, 0); }
2711 
2712 static void __kmp_destroy_hle_lock_with_checks(kmp_dyna_lock_t *lck) {
2713   TCW_4(*lck, 0);
2714 }
2715 
2716 static void __kmp_acquire_hle_lock(kmp_dyna_lock_t *lck, kmp_int32 gtid) {
2717   // Use gtid for KMP_LOCK_BUSY if necessary
2718   if (swap4(lck, KMP_LOCK_BUSY(1, hle)) != KMP_LOCK_FREE(hle)) {
2719     int delay = 1;
2720     do {
2721       while (*(kmp_uint32 volatile *)lck != KMP_LOCK_FREE(hle)) {
2722         for (int i = delay; i != 0; --i)
2723           KMP_CPU_PAUSE();
2724         delay = ((delay << 1) | 1) & 7;
2725       }
2726     } while (swap4(lck, KMP_LOCK_BUSY(1, hle)) != KMP_LOCK_FREE(hle));
2727   }
2728 }
2729 
2730 static void __kmp_acquire_hle_lock_with_checks(kmp_dyna_lock_t *lck,
2731                                                kmp_int32 gtid) {
2732   __kmp_acquire_hle_lock(lck, gtid); // TODO: add checks
2733 }
2734 
2735 static int __kmp_release_hle_lock(kmp_dyna_lock_t *lck, kmp_int32 gtid) {
2736   __asm__ volatile(HLE_RELEASE "movl %1,%0"
2737                    : "=m"(*lck)
2738                    : "r"(KMP_LOCK_FREE(hle))
2739                    : "memory");
2740   return KMP_LOCK_RELEASED;
2741 }
2742 
2743 static int __kmp_release_hle_lock_with_checks(kmp_dyna_lock_t *lck,
2744                                               kmp_int32 gtid) {
2745   return __kmp_release_hle_lock(lck, gtid); // TODO: add checks
2746 }
2747 
2748 static int __kmp_test_hle_lock(kmp_dyna_lock_t *lck, kmp_int32 gtid) {
2749   return swap4(lck, KMP_LOCK_BUSY(1, hle)) == KMP_LOCK_FREE(hle);
2750 }
2751 
2752 static int __kmp_test_hle_lock_with_checks(kmp_dyna_lock_t *lck,
2753                                            kmp_int32 gtid) {
2754   return __kmp_test_hle_lock(lck, gtid); // TODO: add checks
2755 }
2756 
2757 static void __kmp_init_rtm_queuing_lock(kmp_queuing_lock_t *lck) {
2758   __kmp_init_queuing_lock(lck);
2759 }
2760 
2761 static void __kmp_destroy_rtm_queuing_lock(kmp_queuing_lock_t *lck) {
2762   __kmp_destroy_queuing_lock(lck);
2763 }
2764 
2765 static void
2766 __kmp_destroy_rtm_queuing_lock_with_checks(kmp_queuing_lock_t *lck) {
2767   __kmp_destroy_queuing_lock_with_checks(lck);
2768 }
2769 
2770 KMP_ATTRIBUTE_TARGET_RTM
2771 static void __kmp_acquire_rtm_queuing_lock(kmp_queuing_lock_t *lck,
2772                                            kmp_int32 gtid) {
2773   unsigned retries = 3, status;
2774   do {
2775     status = _xbegin();
2776     if (status == _XBEGIN_STARTED) {
2777       if (__kmp_is_unlocked_queuing_lock(lck))
2778         return;
2779       _xabort(0xff);
2780     }
2781     if ((status & _XABORT_EXPLICIT) && _XABORT_CODE(status) == 0xff) {
2782       // Wait until lock becomes free
2783       while (!__kmp_is_unlocked_queuing_lock(lck)) {
2784         KMP_YIELD(TRUE);
2785       }
2786     } else if (!(status & _XABORT_RETRY))
2787       break;
2788   } while (retries--);
2789 
2790   // Fall-back non-speculative lock (xchg)
2791   __kmp_acquire_queuing_lock(lck, gtid);
2792 }
2793 
2794 static void __kmp_acquire_rtm_queuing_lock_with_checks(kmp_queuing_lock_t *lck,
2795                                                        kmp_int32 gtid) {
2796   __kmp_acquire_rtm_queuing_lock(lck, gtid);
2797 }
2798 
2799 KMP_ATTRIBUTE_TARGET_RTM
2800 static int __kmp_release_rtm_queuing_lock(kmp_queuing_lock_t *lck,
2801                                           kmp_int32 gtid) {
2802   if (__kmp_is_unlocked_queuing_lock(lck)) {
2803     // Releasing from speculation
2804     _xend();
2805   } else {
2806     // Releasing from a real lock
2807     __kmp_release_queuing_lock(lck, gtid);
2808   }
2809   return KMP_LOCK_RELEASED;
2810 }
2811 
2812 static int __kmp_release_rtm_queuing_lock_with_checks(kmp_queuing_lock_t *lck,
2813                                                       kmp_int32 gtid) {
2814   return __kmp_release_rtm_queuing_lock(lck, gtid);
2815 }
2816 
2817 KMP_ATTRIBUTE_TARGET_RTM
2818 static int __kmp_test_rtm_queuing_lock(kmp_queuing_lock_t *lck,
2819                                        kmp_int32 gtid) {
2820   unsigned retries = 3, status;
2821   do {
2822     status = _xbegin();
2823     if (status == _XBEGIN_STARTED && __kmp_is_unlocked_queuing_lock(lck)) {
2824       return 1;
2825     }
2826     if (!(status & _XABORT_RETRY))
2827       break;
2828   } while (retries--);
2829 
2830   return __kmp_test_queuing_lock(lck, gtid);
2831 }
2832 
2833 static int __kmp_test_rtm_queuing_lock_with_checks(kmp_queuing_lock_t *lck,
2834                                                    kmp_int32 gtid) {
2835   return __kmp_test_rtm_queuing_lock(lck, gtid);
2836 }
2837 
2838 // Reuse kmp_tas_lock_t for TSX lock which use RTM with fall-back spin lock.
2839 typedef kmp_tas_lock_t kmp_rtm_spin_lock_t;
2840 
2841 static void __kmp_destroy_rtm_spin_lock(kmp_rtm_spin_lock_t *lck) {
2842   KMP_ATOMIC_ST_REL(&lck->lk.poll, 0);
2843 }
2844 
2845 static void __kmp_destroy_rtm_spin_lock_with_checks(kmp_rtm_spin_lock_t *lck) {
2846   __kmp_destroy_rtm_spin_lock(lck);
2847 }
2848 
2849 KMP_ATTRIBUTE_TARGET_RTM
2850 static int __kmp_acquire_rtm_spin_lock(kmp_rtm_spin_lock_t *lck,
2851                                        kmp_int32 gtid) {
2852   unsigned retries = 3, status;
2853   kmp_int32 lock_free = KMP_LOCK_FREE(rtm_spin);
2854   kmp_int32 lock_busy = KMP_LOCK_BUSY(1, rtm_spin);
2855   do {
2856     status = _xbegin();
2857     if (status == _XBEGIN_STARTED) {
2858       if (KMP_ATOMIC_LD_RLX(&lck->lk.poll) == lock_free)
2859         return KMP_LOCK_ACQUIRED_FIRST;
2860       _xabort(0xff);
2861     }
2862     if ((status & _XABORT_EXPLICIT) && _XABORT_CODE(status) == 0xff) {
2863       // Wait until lock becomes free
2864       while (KMP_ATOMIC_LD_RLX(&lck->lk.poll) != lock_free) {
2865         KMP_YIELD(TRUE);
2866       }
2867     } else if (!(status & _XABORT_RETRY))
2868       break;
2869   } while (retries--);
2870 
2871   // Fall-back spin lock
2872   KMP_FSYNC_PREPARE(lck);
2873   kmp_backoff_t backoff = __kmp_spin_backoff_params;
2874   while (KMP_ATOMIC_LD_RLX(&lck->lk.poll) != lock_free ||
2875          !__kmp_atomic_compare_store_acq(&lck->lk.poll, lock_free, lock_busy)) {
2876     __kmp_spin_backoff(&backoff);
2877   }
2878   KMP_FSYNC_ACQUIRED(lck);
2879   return KMP_LOCK_ACQUIRED_FIRST;
2880 }
2881 
2882 static int __kmp_acquire_rtm_spin_lock_with_checks(kmp_rtm_spin_lock_t *lck,
2883                                                    kmp_int32 gtid) {
2884   return __kmp_acquire_rtm_spin_lock(lck, gtid);
2885 }
2886 
2887 KMP_ATTRIBUTE_TARGET_RTM
2888 static int __kmp_release_rtm_spin_lock(kmp_rtm_spin_lock_t *lck,
2889                                        kmp_int32 gtid) {
2890   if (KMP_ATOMIC_LD_RLX(&lck->lk.poll) == KMP_LOCK_FREE(rtm_spin)) {
2891     // Releasing from speculation
2892     _xend();
2893   } else {
2894     // Releasing from a real lock
2895     KMP_FSYNC_RELEASING(lck);
2896     KMP_ATOMIC_ST_REL(&lck->lk.poll, KMP_LOCK_FREE(rtm_spin));
2897   }
2898   return KMP_LOCK_RELEASED;
2899 }
2900 
2901 static int __kmp_release_rtm_spin_lock_with_checks(kmp_rtm_spin_lock_t *lck,
2902                                                    kmp_int32 gtid) {
2903   return __kmp_release_rtm_spin_lock(lck, gtid);
2904 }
2905 
2906 KMP_ATTRIBUTE_TARGET_RTM
2907 static int __kmp_test_rtm_spin_lock(kmp_rtm_spin_lock_t *lck, kmp_int32 gtid) {
2908   unsigned retries = 3, status;
2909   kmp_int32 lock_free = KMP_LOCK_FREE(rtm_spin);
2910   kmp_int32 lock_busy = KMP_LOCK_BUSY(1, rtm_spin);
2911   do {
2912     status = _xbegin();
2913     if (status == _XBEGIN_STARTED &&
2914         KMP_ATOMIC_LD_RLX(&lck->lk.poll) == lock_free) {
2915       return TRUE;
2916     }
2917     if (!(status & _XABORT_RETRY))
2918       break;
2919   } while (retries--);
2920 
2921   if (KMP_ATOMIC_LD_RLX(&lck->lk.poll) == lock_free &&
2922       __kmp_atomic_compare_store_acq(&lck->lk.poll, lock_free, lock_busy)) {
2923     KMP_FSYNC_ACQUIRED(lck);
2924     return TRUE;
2925   }
2926   return FALSE;
2927 }
2928 
2929 static int __kmp_test_rtm_spin_lock_with_checks(kmp_rtm_spin_lock_t *lck,
2930                                                 kmp_int32 gtid) {
2931   return __kmp_test_rtm_spin_lock(lck, gtid);
2932 }
2933 
2934 #endif // KMP_USE_TSX
2935 
2936 // Entry functions for indirect locks (first element of direct lock jump tables)
2937 static void __kmp_init_indirect_lock(kmp_dyna_lock_t *l,
2938                                      kmp_dyna_lockseq_t tag);
2939 static void __kmp_destroy_indirect_lock(kmp_dyna_lock_t *lock);
2940 static int __kmp_set_indirect_lock(kmp_dyna_lock_t *lock, kmp_int32);
2941 static int __kmp_unset_indirect_lock(kmp_dyna_lock_t *lock, kmp_int32);
2942 static int __kmp_test_indirect_lock(kmp_dyna_lock_t *lock, kmp_int32);
2943 static int __kmp_set_indirect_lock_with_checks(kmp_dyna_lock_t *lock,
2944                                                kmp_int32);
2945 static int __kmp_unset_indirect_lock_with_checks(kmp_dyna_lock_t *lock,
2946                                                  kmp_int32);
2947 static int __kmp_test_indirect_lock_with_checks(kmp_dyna_lock_t *lock,
2948                                                 kmp_int32);
2949 
2950 // Lock function definitions for the union parameter type
2951 #define KMP_FOREACH_LOCK_KIND(m, a) m(ticket, a) m(queuing, a) m(drdpa, a)
2952 
2953 #define expand1(lk, op)                                                        \
2954   static void __kmp_##op##_##lk##_##lock(kmp_user_lock_p lock) {               \
2955     __kmp_##op##_##lk##_##lock(&lock->lk);                                     \
2956   }
2957 #define expand2(lk, op)                                                        \
2958   static int __kmp_##op##_##lk##_##lock(kmp_user_lock_p lock,                  \
2959                                         kmp_int32 gtid) {                      \
2960     return __kmp_##op##_##lk##_##lock(&lock->lk, gtid);                        \
2961   }
2962 #define expand3(lk, op)                                                        \
2963   static void __kmp_set_##lk##_##lock_flags(kmp_user_lock_p lock,              \
2964                                             kmp_lock_flags_t flags) {          \
2965     __kmp_set_##lk##_lock_flags(&lock->lk, flags);                             \
2966   }
2967 #define expand4(lk, op)                                                        \
2968   static void __kmp_set_##lk##_##lock_location(kmp_user_lock_p lock,           \
2969                                                const ident_t *loc) {           \
2970     __kmp_set_##lk##_lock_location(&lock->lk, loc);                            \
2971   }
2972 
2973 KMP_FOREACH_LOCK_KIND(expand1, init)
2974 KMP_FOREACH_LOCK_KIND(expand1, init_nested)
2975 KMP_FOREACH_LOCK_KIND(expand1, destroy)
2976 KMP_FOREACH_LOCK_KIND(expand1, destroy_nested)
2977 KMP_FOREACH_LOCK_KIND(expand2, acquire)
2978 KMP_FOREACH_LOCK_KIND(expand2, acquire_nested)
2979 KMP_FOREACH_LOCK_KIND(expand2, release)
2980 KMP_FOREACH_LOCK_KIND(expand2, release_nested)
2981 KMP_FOREACH_LOCK_KIND(expand2, test)
2982 KMP_FOREACH_LOCK_KIND(expand2, test_nested)
2983 KMP_FOREACH_LOCK_KIND(expand3, )
2984 KMP_FOREACH_LOCK_KIND(expand4, )
2985 
2986 #undef expand1
2987 #undef expand2
2988 #undef expand3
2989 #undef expand4
2990 
2991 // Jump tables for the indirect lock functions
2992 // Only fill in the odd entries, that avoids the need to shift out the low bit
2993 
2994 // init functions
2995 #define expand(l, op) 0, __kmp_init_direct_lock,
2996 void (*__kmp_direct_init[])(kmp_dyna_lock_t *, kmp_dyna_lockseq_t) = {
2997     __kmp_init_indirect_lock, 0, KMP_FOREACH_D_LOCK(expand, init)};
2998 #undef expand
2999 
3000 // destroy functions
3001 #define expand(l, op) 0, (void (*)(kmp_dyna_lock_t *))__kmp_##op##_##l##_lock,
3002 static void (*direct_destroy[])(kmp_dyna_lock_t *) = {
3003     __kmp_destroy_indirect_lock, 0, KMP_FOREACH_D_LOCK(expand, destroy)};
3004 #undef expand
3005 #define expand(l, op)                                                          \
3006   0, (void (*)(kmp_dyna_lock_t *))__kmp_destroy_##l##_lock_with_checks,
3007 static void (*direct_destroy_check[])(kmp_dyna_lock_t *) = {
3008     __kmp_destroy_indirect_lock, 0, KMP_FOREACH_D_LOCK(expand, destroy)};
3009 #undef expand
3010 
3011 // set/acquire functions
3012 #define expand(l, op)                                                          \
3013   0, (int (*)(kmp_dyna_lock_t *, kmp_int32))__kmp_##op##_##l##_lock,
3014 static int (*direct_set[])(kmp_dyna_lock_t *, kmp_int32) = {
3015     __kmp_set_indirect_lock, 0, KMP_FOREACH_D_LOCK(expand, acquire)};
3016 #undef expand
3017 #define expand(l, op)                                                          \
3018   0, (int (*)(kmp_dyna_lock_t *, kmp_int32))__kmp_##op##_##l##_lock_with_checks,
3019 static int (*direct_set_check[])(kmp_dyna_lock_t *, kmp_int32) = {
3020     __kmp_set_indirect_lock_with_checks, 0,
3021     KMP_FOREACH_D_LOCK(expand, acquire)};
3022 #undef expand
3023 
3024 // unset/release and test functions
3025 #define expand(l, op)                                                          \
3026   0, (int (*)(kmp_dyna_lock_t *, kmp_int32))__kmp_##op##_##l##_lock,
3027 static int (*direct_unset[])(kmp_dyna_lock_t *, kmp_int32) = {
3028     __kmp_unset_indirect_lock, 0, KMP_FOREACH_D_LOCK(expand, release)};
3029 static int (*direct_test[])(kmp_dyna_lock_t *, kmp_int32) = {
3030     __kmp_test_indirect_lock, 0, KMP_FOREACH_D_LOCK(expand, test)};
3031 #undef expand
3032 #define expand(l, op)                                                          \
3033   0, (int (*)(kmp_dyna_lock_t *, kmp_int32))__kmp_##op##_##l##_lock_with_checks,
3034 static int (*direct_unset_check[])(kmp_dyna_lock_t *, kmp_int32) = {
3035     __kmp_unset_indirect_lock_with_checks, 0,
3036     KMP_FOREACH_D_LOCK(expand, release)};
3037 static int (*direct_test_check[])(kmp_dyna_lock_t *, kmp_int32) = {
3038     __kmp_test_indirect_lock_with_checks, 0, KMP_FOREACH_D_LOCK(expand, test)};
3039 #undef expand
3040 
3041 // Exposes only one set of jump tables (*lock or *lock_with_checks).
3042 void (**__kmp_direct_destroy)(kmp_dyna_lock_t *) = 0;
3043 int (**__kmp_direct_set)(kmp_dyna_lock_t *, kmp_int32) = 0;
3044 int (**__kmp_direct_unset)(kmp_dyna_lock_t *, kmp_int32) = 0;
3045 int (**__kmp_direct_test)(kmp_dyna_lock_t *, kmp_int32) = 0;
3046 
3047 // Jump tables for the indirect lock functions
3048 #define expand(l, op) (void (*)(kmp_user_lock_p)) __kmp_##op##_##l##_##lock,
3049 void (*__kmp_indirect_init[])(kmp_user_lock_p) = {
3050     KMP_FOREACH_I_LOCK(expand, init)};
3051 #undef expand
3052 
3053 #define expand(l, op) (void (*)(kmp_user_lock_p)) __kmp_##op##_##l##_##lock,
3054 static void (*indirect_destroy[])(kmp_user_lock_p) = {
3055     KMP_FOREACH_I_LOCK(expand, destroy)};
3056 #undef expand
3057 #define expand(l, op)                                                          \
3058   (void (*)(kmp_user_lock_p)) __kmp_##op##_##l##_##lock_with_checks,
3059 static void (*indirect_destroy_check[])(kmp_user_lock_p) = {
3060     KMP_FOREACH_I_LOCK(expand, destroy)};
3061 #undef expand
3062 
3063 // set/acquire functions
3064 #define expand(l, op)                                                          \
3065   (int (*)(kmp_user_lock_p, kmp_int32)) __kmp_##op##_##l##_##lock,
3066 static int (*indirect_set[])(kmp_user_lock_p,
3067                              kmp_int32) = {KMP_FOREACH_I_LOCK(expand, acquire)};
3068 #undef expand
3069 #define expand(l, op)                                                          \
3070   (int (*)(kmp_user_lock_p, kmp_int32)) __kmp_##op##_##l##_##lock_with_checks,
3071 static int (*indirect_set_check[])(kmp_user_lock_p, kmp_int32) = {
3072     KMP_FOREACH_I_LOCK(expand, acquire)};
3073 #undef expand
3074 
3075 // unset/release and test functions
3076 #define expand(l, op)                                                          \
3077   (int (*)(kmp_user_lock_p, kmp_int32)) __kmp_##op##_##l##_##lock,
3078 static int (*indirect_unset[])(kmp_user_lock_p, kmp_int32) = {
3079     KMP_FOREACH_I_LOCK(expand, release)};
3080 static int (*indirect_test[])(kmp_user_lock_p,
3081                               kmp_int32) = {KMP_FOREACH_I_LOCK(expand, test)};
3082 #undef expand
3083 #define expand(l, op)                                                          \
3084   (int (*)(kmp_user_lock_p, kmp_int32)) __kmp_##op##_##l##_##lock_with_checks,
3085 static int (*indirect_unset_check[])(kmp_user_lock_p, kmp_int32) = {
3086     KMP_FOREACH_I_LOCK(expand, release)};
3087 static int (*indirect_test_check[])(kmp_user_lock_p, kmp_int32) = {
3088     KMP_FOREACH_I_LOCK(expand, test)};
3089 #undef expand
3090 
3091 // Exposes only one jump tables (*lock or *lock_with_checks).
3092 void (**__kmp_indirect_destroy)(kmp_user_lock_p) = 0;
3093 int (**__kmp_indirect_set)(kmp_user_lock_p, kmp_int32) = 0;
3094 int (**__kmp_indirect_unset)(kmp_user_lock_p, kmp_int32) = 0;
3095 int (**__kmp_indirect_test)(kmp_user_lock_p, kmp_int32) = 0;
3096 
3097 // Lock index table.
3098 kmp_indirect_lock_table_t __kmp_i_lock_table;
3099 
3100 // Size of indirect locks.
3101 static kmp_uint32 __kmp_indirect_lock_size[KMP_NUM_I_LOCKS] = {0};
3102 
3103 // Jump tables for lock accessor/modifier.
3104 void (*__kmp_indirect_set_location[KMP_NUM_I_LOCKS])(kmp_user_lock_p,
3105                                                      const ident_t *) = {0};
3106 void (*__kmp_indirect_set_flags[KMP_NUM_I_LOCKS])(kmp_user_lock_p,
3107                                                   kmp_lock_flags_t) = {0};
3108 const ident_t *(*__kmp_indirect_get_location[KMP_NUM_I_LOCKS])(
3109     kmp_user_lock_p) = {0};
3110 kmp_lock_flags_t (*__kmp_indirect_get_flags[KMP_NUM_I_LOCKS])(
3111     kmp_user_lock_p) = {0};
3112 
3113 // Use different lock pools for different lock types.
3114 static kmp_indirect_lock_t *__kmp_indirect_lock_pool[KMP_NUM_I_LOCKS] = {0};
3115 
3116 // User lock allocator for dynamically dispatched indirect locks. Every entry of
3117 // the indirect lock table holds the address and type of the allocated indirect
3118 // lock (kmp_indirect_lock_t), and the size of the table doubles when it is
3119 // full. A destroyed indirect lock object is returned to the reusable pool of
3120 // locks, unique to each lock type.
3121 kmp_indirect_lock_t *__kmp_allocate_indirect_lock(void **user_lock,
3122                                                   kmp_int32 gtid,
3123                                                   kmp_indirect_locktag_t tag) {
3124   kmp_indirect_lock_t *lck;
3125   kmp_lock_index_t idx;
3126 
3127   __kmp_acquire_lock(&__kmp_global_lock, gtid);
3128 
3129   if (__kmp_indirect_lock_pool[tag] != NULL) {
3130     // Reuse the allocated and destroyed lock object
3131     lck = __kmp_indirect_lock_pool[tag];
3132     if (OMP_LOCK_T_SIZE < sizeof(void *))
3133       idx = lck->lock->pool.index;
3134     __kmp_indirect_lock_pool[tag] = (kmp_indirect_lock_t *)lck->lock->pool.next;
3135     KA_TRACE(20, ("__kmp_allocate_indirect_lock: reusing an existing lock %p\n",
3136                   lck));
3137   } else {
3138     idx = __kmp_i_lock_table.next;
3139     // Check capacity and double the size if it is full
3140     if (idx == __kmp_i_lock_table.size) {
3141       // Double up the space for block pointers
3142       int row = __kmp_i_lock_table.size / KMP_I_LOCK_CHUNK;
3143       kmp_indirect_lock_t **new_table = (kmp_indirect_lock_t **)__kmp_allocate(
3144           2 * row * sizeof(kmp_indirect_lock_t *));
3145       KMP_MEMCPY(new_table, __kmp_i_lock_table.table,
3146                  row * sizeof(kmp_indirect_lock_t *));
3147       kmp_indirect_lock_t **old_table = __kmp_i_lock_table.table;
3148       __kmp_i_lock_table.table = new_table;
3149       __kmp_free(old_table);
3150       // Allocate new objects in the new blocks
3151       for (int i = row; i < 2 * row; ++i)
3152         *(__kmp_i_lock_table.table + i) = (kmp_indirect_lock_t *)__kmp_allocate(
3153             KMP_I_LOCK_CHUNK * sizeof(kmp_indirect_lock_t));
3154       __kmp_i_lock_table.size = 2 * idx;
3155     }
3156     __kmp_i_lock_table.next++;
3157     lck = KMP_GET_I_LOCK(idx);
3158     // Allocate a new base lock object
3159     lck->lock = (kmp_user_lock_p)__kmp_allocate(__kmp_indirect_lock_size[tag]);
3160     KA_TRACE(20,
3161              ("__kmp_allocate_indirect_lock: allocated a new lock %p\n", lck));
3162   }
3163 
3164   __kmp_release_lock(&__kmp_global_lock, gtid);
3165 
3166   lck->type = tag;
3167 
3168   if (OMP_LOCK_T_SIZE < sizeof(void *)) {
3169     *((kmp_lock_index_t *)user_lock) = idx
3170                                        << 1; // indirect lock word must be even
3171   } else {
3172     *((kmp_indirect_lock_t **)user_lock) = lck;
3173   }
3174 
3175   return lck;
3176 }
3177 
3178 // User lock lookup for dynamically dispatched locks.
3179 static __forceinline kmp_indirect_lock_t *
3180 __kmp_lookup_indirect_lock(void **user_lock, const char *func) {
3181   if (__kmp_env_consistency_check) {
3182     kmp_indirect_lock_t *lck = NULL;
3183     if (user_lock == NULL) {
3184       KMP_FATAL(LockIsUninitialized, func);
3185     }
3186     if (OMP_LOCK_T_SIZE < sizeof(void *)) {
3187       kmp_lock_index_t idx = KMP_EXTRACT_I_INDEX(user_lock);
3188       if (idx >= __kmp_i_lock_table.size) {
3189         KMP_FATAL(LockIsUninitialized, func);
3190       }
3191       lck = KMP_GET_I_LOCK(idx);
3192     } else {
3193       lck = *((kmp_indirect_lock_t **)user_lock);
3194     }
3195     if (lck == NULL) {
3196       KMP_FATAL(LockIsUninitialized, func);
3197     }
3198     return lck;
3199   } else {
3200     if (OMP_LOCK_T_SIZE < sizeof(void *)) {
3201       return KMP_GET_I_LOCK(KMP_EXTRACT_I_INDEX(user_lock));
3202     } else {
3203       return *((kmp_indirect_lock_t **)user_lock);
3204     }
3205   }
3206 }
3207 
3208 static void __kmp_init_indirect_lock(kmp_dyna_lock_t *lock,
3209                                      kmp_dyna_lockseq_t seq) {
3210 #if KMP_USE_ADAPTIVE_LOCKS
3211   if (seq == lockseq_adaptive && !__kmp_cpuinfo.rtm) {
3212     KMP_WARNING(AdaptiveNotSupported, "kmp_lockseq_t", "adaptive");
3213     seq = lockseq_queuing;
3214   }
3215 #endif
3216 #if KMP_USE_TSX
3217   if (seq == lockseq_rtm_queuing && !__kmp_cpuinfo.rtm) {
3218     seq = lockseq_queuing;
3219   }
3220 #endif
3221   kmp_indirect_locktag_t tag = KMP_GET_I_TAG(seq);
3222   kmp_indirect_lock_t *l =
3223       __kmp_allocate_indirect_lock((void **)lock, __kmp_entry_gtid(), tag);
3224   KMP_I_LOCK_FUNC(l, init)(l->lock);
3225   KA_TRACE(
3226       20, ("__kmp_init_indirect_lock: initialized indirect lock with type#%d\n",
3227            seq));
3228 }
3229 
3230 static void __kmp_destroy_indirect_lock(kmp_dyna_lock_t *lock) {
3231   kmp_uint32 gtid = __kmp_entry_gtid();
3232   kmp_indirect_lock_t *l =
3233       __kmp_lookup_indirect_lock((void **)lock, "omp_destroy_lock");
3234   KMP_I_LOCK_FUNC(l, destroy)(l->lock);
3235   kmp_indirect_locktag_t tag = l->type;
3236 
3237   __kmp_acquire_lock(&__kmp_global_lock, gtid);
3238 
3239   // Use the base lock's space to keep the pool chain.
3240   l->lock->pool.next = (kmp_user_lock_p)__kmp_indirect_lock_pool[tag];
3241   if (OMP_LOCK_T_SIZE < sizeof(void *)) {
3242     l->lock->pool.index = KMP_EXTRACT_I_INDEX(lock);
3243   }
3244   __kmp_indirect_lock_pool[tag] = l;
3245 
3246   __kmp_release_lock(&__kmp_global_lock, gtid);
3247 }
3248 
3249 static int __kmp_set_indirect_lock(kmp_dyna_lock_t *lock, kmp_int32 gtid) {
3250   kmp_indirect_lock_t *l = KMP_LOOKUP_I_LOCK(lock);
3251   return KMP_I_LOCK_FUNC(l, set)(l->lock, gtid);
3252 }
3253 
3254 static int __kmp_unset_indirect_lock(kmp_dyna_lock_t *lock, kmp_int32 gtid) {
3255   kmp_indirect_lock_t *l = KMP_LOOKUP_I_LOCK(lock);
3256   return KMP_I_LOCK_FUNC(l, unset)(l->lock, gtid);
3257 }
3258 
3259 static int __kmp_test_indirect_lock(kmp_dyna_lock_t *lock, kmp_int32 gtid) {
3260   kmp_indirect_lock_t *l = KMP_LOOKUP_I_LOCK(lock);
3261   return KMP_I_LOCK_FUNC(l, test)(l->lock, gtid);
3262 }
3263 
3264 static int __kmp_set_indirect_lock_with_checks(kmp_dyna_lock_t *lock,
3265                                                kmp_int32 gtid) {
3266   kmp_indirect_lock_t *l =
3267       __kmp_lookup_indirect_lock((void **)lock, "omp_set_lock");
3268   return KMP_I_LOCK_FUNC(l, set)(l->lock, gtid);
3269 }
3270 
3271 static int __kmp_unset_indirect_lock_with_checks(kmp_dyna_lock_t *lock,
3272                                                  kmp_int32 gtid) {
3273   kmp_indirect_lock_t *l =
3274       __kmp_lookup_indirect_lock((void **)lock, "omp_unset_lock");
3275   return KMP_I_LOCK_FUNC(l, unset)(l->lock, gtid);
3276 }
3277 
3278 static int __kmp_test_indirect_lock_with_checks(kmp_dyna_lock_t *lock,
3279                                                 kmp_int32 gtid) {
3280   kmp_indirect_lock_t *l =
3281       __kmp_lookup_indirect_lock((void **)lock, "omp_test_lock");
3282   return KMP_I_LOCK_FUNC(l, test)(l->lock, gtid);
3283 }
3284 
3285 kmp_dyna_lockseq_t __kmp_user_lock_seq = lockseq_queuing;
3286 
3287 // This is used only in kmp_error.cpp when consistency checking is on.
3288 kmp_int32 __kmp_get_user_lock_owner(kmp_user_lock_p lck, kmp_uint32 seq) {
3289   switch (seq) {
3290   case lockseq_tas:
3291   case lockseq_nested_tas:
3292     return __kmp_get_tas_lock_owner((kmp_tas_lock_t *)lck);
3293 #if KMP_USE_FUTEX
3294   case lockseq_futex:
3295   case lockseq_nested_futex:
3296     return __kmp_get_futex_lock_owner((kmp_futex_lock_t *)lck);
3297 #endif
3298   case lockseq_ticket:
3299   case lockseq_nested_ticket:
3300     return __kmp_get_ticket_lock_owner((kmp_ticket_lock_t *)lck);
3301   case lockseq_queuing:
3302   case lockseq_nested_queuing:
3303 #if KMP_USE_ADAPTIVE_LOCKS
3304   case lockseq_adaptive:
3305 #endif
3306     return __kmp_get_queuing_lock_owner((kmp_queuing_lock_t *)lck);
3307   case lockseq_drdpa:
3308   case lockseq_nested_drdpa:
3309     return __kmp_get_drdpa_lock_owner((kmp_drdpa_lock_t *)lck);
3310   default:
3311     return 0;
3312   }
3313 }
3314 
3315 // Initializes data for dynamic user locks.
3316 void __kmp_init_dynamic_user_locks() {
3317   // Initialize jump table for the lock functions
3318   if (__kmp_env_consistency_check) {
3319     __kmp_direct_set = direct_set_check;
3320     __kmp_direct_unset = direct_unset_check;
3321     __kmp_direct_test = direct_test_check;
3322     __kmp_direct_destroy = direct_destroy_check;
3323     __kmp_indirect_set = indirect_set_check;
3324     __kmp_indirect_unset = indirect_unset_check;
3325     __kmp_indirect_test = indirect_test_check;
3326     __kmp_indirect_destroy = indirect_destroy_check;
3327   } else {
3328     __kmp_direct_set = direct_set;
3329     __kmp_direct_unset = direct_unset;
3330     __kmp_direct_test = direct_test;
3331     __kmp_direct_destroy = direct_destroy;
3332     __kmp_indirect_set = indirect_set;
3333     __kmp_indirect_unset = indirect_unset;
3334     __kmp_indirect_test = indirect_test;
3335     __kmp_indirect_destroy = indirect_destroy;
3336   }
3337   // If the user locks have already been initialized, then return. Allow the
3338   // switch between different KMP_CONSISTENCY_CHECK values, but do not allocate
3339   // new lock tables if they have already been allocated.
3340   if (__kmp_init_user_locks)
3341     return;
3342 
3343   // Initialize lock index table
3344   __kmp_i_lock_table.size = KMP_I_LOCK_CHUNK;
3345   __kmp_i_lock_table.table =
3346       (kmp_indirect_lock_t **)__kmp_allocate(sizeof(kmp_indirect_lock_t *));
3347   *(__kmp_i_lock_table.table) = (kmp_indirect_lock_t *)__kmp_allocate(
3348       KMP_I_LOCK_CHUNK * sizeof(kmp_indirect_lock_t));
3349   __kmp_i_lock_table.next = 0;
3350 
3351   // Indirect lock size
3352   __kmp_indirect_lock_size[locktag_ticket] = sizeof(kmp_ticket_lock_t);
3353   __kmp_indirect_lock_size[locktag_queuing] = sizeof(kmp_queuing_lock_t);
3354 #if KMP_USE_ADAPTIVE_LOCKS
3355   __kmp_indirect_lock_size[locktag_adaptive] = sizeof(kmp_adaptive_lock_t);
3356 #endif
3357   __kmp_indirect_lock_size[locktag_drdpa] = sizeof(kmp_drdpa_lock_t);
3358 #if KMP_USE_TSX
3359   __kmp_indirect_lock_size[locktag_rtm_queuing] = sizeof(kmp_queuing_lock_t);
3360 #endif
3361   __kmp_indirect_lock_size[locktag_nested_tas] = sizeof(kmp_tas_lock_t);
3362 #if KMP_USE_FUTEX
3363   __kmp_indirect_lock_size[locktag_nested_futex] = sizeof(kmp_futex_lock_t);
3364 #endif
3365   __kmp_indirect_lock_size[locktag_nested_ticket] = sizeof(kmp_ticket_lock_t);
3366   __kmp_indirect_lock_size[locktag_nested_queuing] = sizeof(kmp_queuing_lock_t);
3367   __kmp_indirect_lock_size[locktag_nested_drdpa] = sizeof(kmp_drdpa_lock_t);
3368 
3369 // Initialize lock accessor/modifier
3370 #define fill_jumps(table, expand, sep)                                         \
3371   {                                                                            \
3372     table[locktag##sep##ticket] = expand(ticket);                              \
3373     table[locktag##sep##queuing] = expand(queuing);                            \
3374     table[locktag##sep##drdpa] = expand(drdpa);                                \
3375   }
3376 
3377 #if KMP_USE_ADAPTIVE_LOCKS
3378 #define fill_table(table, expand)                                              \
3379   {                                                                            \
3380     fill_jumps(table, expand, _);                                              \
3381     table[locktag_adaptive] = expand(queuing);                                 \
3382     fill_jumps(table, expand, _nested_);                                       \
3383   }
3384 #else
3385 #define fill_table(table, expand)                                              \
3386   {                                                                            \
3387     fill_jumps(table, expand, _);                                              \
3388     fill_jumps(table, expand, _nested_);                                       \
3389   }
3390 #endif // KMP_USE_ADAPTIVE_LOCKS
3391 
3392 #define expand(l)                                                              \
3393   (void (*)(kmp_user_lock_p, const ident_t *)) __kmp_set_##l##_lock_location
3394   fill_table(__kmp_indirect_set_location, expand);
3395 #undef expand
3396 #define expand(l)                                                              \
3397   (void (*)(kmp_user_lock_p, kmp_lock_flags_t)) __kmp_set_##l##_lock_flags
3398   fill_table(__kmp_indirect_set_flags, expand);
3399 #undef expand
3400 #define expand(l)                                                              \
3401   (const ident_t *(*)(kmp_user_lock_p)) __kmp_get_##l##_lock_location
3402   fill_table(__kmp_indirect_get_location, expand);
3403 #undef expand
3404 #define expand(l)                                                              \
3405   (kmp_lock_flags_t(*)(kmp_user_lock_p)) __kmp_get_##l##_lock_flags
3406   fill_table(__kmp_indirect_get_flags, expand);
3407 #undef expand
3408 
3409   __kmp_init_user_locks = TRUE;
3410 }
3411 
3412 // Clean up the lock table.
3413 void __kmp_cleanup_indirect_user_locks() {
3414   kmp_lock_index_t i;
3415   int k;
3416 
3417   // Clean up locks in the pools first (they were already destroyed before going
3418   // into the pools).
3419   for (k = 0; k < KMP_NUM_I_LOCKS; ++k) {
3420     kmp_indirect_lock_t *l = __kmp_indirect_lock_pool[k];
3421     while (l != NULL) {
3422       kmp_indirect_lock_t *ll = l;
3423       l = (kmp_indirect_lock_t *)l->lock->pool.next;
3424       KA_TRACE(20, ("__kmp_cleanup_indirect_user_locks: freeing %p from pool\n",
3425                     ll));
3426       __kmp_free(ll->lock);
3427       ll->lock = NULL;
3428     }
3429     __kmp_indirect_lock_pool[k] = NULL;
3430   }
3431   // Clean up the remaining undestroyed locks.
3432   for (i = 0; i < __kmp_i_lock_table.next; i++) {
3433     kmp_indirect_lock_t *l = KMP_GET_I_LOCK(i);
3434     if (l->lock != NULL) {
3435       // Locks not destroyed explicitly need to be destroyed here.
3436       KMP_I_LOCK_FUNC(l, destroy)(l->lock);
3437       KA_TRACE(
3438           20,
3439           ("__kmp_cleanup_indirect_user_locks: destroy/freeing %p from table\n",
3440            l));
3441       __kmp_free(l->lock);
3442     }
3443   }
3444   // Free the table
3445   for (i = 0; i < __kmp_i_lock_table.size / KMP_I_LOCK_CHUNK; i++)
3446     __kmp_free(__kmp_i_lock_table.table[i]);
3447   __kmp_free(__kmp_i_lock_table.table);
3448 
3449   __kmp_init_user_locks = FALSE;
3450 }
3451 
3452 enum kmp_lock_kind __kmp_user_lock_kind = lk_default;
3453 int __kmp_num_locks_in_block = 1; // FIXME - tune this value
3454 
3455 #else // KMP_USE_DYNAMIC_LOCK
3456 
3457 static void __kmp_init_tas_lock_with_checks(kmp_tas_lock_t *lck) {
3458   __kmp_init_tas_lock(lck);
3459 }
3460 
3461 static void __kmp_init_nested_tas_lock_with_checks(kmp_tas_lock_t *lck) {
3462   __kmp_init_nested_tas_lock(lck);
3463 }
3464 
3465 #if KMP_USE_FUTEX
3466 static void __kmp_init_futex_lock_with_checks(kmp_futex_lock_t *lck) {
3467   __kmp_init_futex_lock(lck);
3468 }
3469 
3470 static void __kmp_init_nested_futex_lock_with_checks(kmp_futex_lock_t *lck) {
3471   __kmp_init_nested_futex_lock(lck);
3472 }
3473 #endif
3474 
3475 static int __kmp_is_ticket_lock_initialized(kmp_ticket_lock_t *lck) {
3476   return lck == lck->lk.self;
3477 }
3478 
3479 static void __kmp_init_ticket_lock_with_checks(kmp_ticket_lock_t *lck) {
3480   __kmp_init_ticket_lock(lck);
3481 }
3482 
3483 static void __kmp_init_nested_ticket_lock_with_checks(kmp_ticket_lock_t *lck) {
3484   __kmp_init_nested_ticket_lock(lck);
3485 }
3486 
3487 static int __kmp_is_queuing_lock_initialized(kmp_queuing_lock_t *lck) {
3488   return lck == lck->lk.initialized;
3489 }
3490 
3491 static void __kmp_init_queuing_lock_with_checks(kmp_queuing_lock_t *lck) {
3492   __kmp_init_queuing_lock(lck);
3493 }
3494 
3495 static void
3496 __kmp_init_nested_queuing_lock_with_checks(kmp_queuing_lock_t *lck) {
3497   __kmp_init_nested_queuing_lock(lck);
3498 }
3499 
3500 #if KMP_USE_ADAPTIVE_LOCKS
3501 static void __kmp_init_adaptive_lock_with_checks(kmp_adaptive_lock_t *lck) {
3502   __kmp_init_adaptive_lock(lck);
3503 }
3504 #endif
3505 
3506 static int __kmp_is_drdpa_lock_initialized(kmp_drdpa_lock_t *lck) {
3507   return lck == lck->lk.initialized;
3508 }
3509 
3510 static void __kmp_init_drdpa_lock_with_checks(kmp_drdpa_lock_t *lck) {
3511   __kmp_init_drdpa_lock(lck);
3512 }
3513 
3514 static void __kmp_init_nested_drdpa_lock_with_checks(kmp_drdpa_lock_t *lck) {
3515   __kmp_init_nested_drdpa_lock(lck);
3516 }
3517 
3518 /* user locks
3519  * They are implemented as a table of function pointers which are set to the
3520  * lock functions of the appropriate kind, once that has been determined. */
3521 
3522 enum kmp_lock_kind __kmp_user_lock_kind = lk_default;
3523 
3524 size_t __kmp_base_user_lock_size = 0;
3525 size_t __kmp_user_lock_size = 0;
3526 
3527 kmp_int32 (*__kmp_get_user_lock_owner_)(kmp_user_lock_p lck) = NULL;
3528 int (*__kmp_acquire_user_lock_with_checks_)(kmp_user_lock_p lck,
3529                                             kmp_int32 gtid) = NULL;
3530 
3531 int (*__kmp_test_user_lock_with_checks_)(kmp_user_lock_p lck,
3532                                          kmp_int32 gtid) = NULL;
3533 int (*__kmp_release_user_lock_with_checks_)(kmp_user_lock_p lck,
3534                                             kmp_int32 gtid) = NULL;
3535 void (*__kmp_init_user_lock_with_checks_)(kmp_user_lock_p lck) = NULL;
3536 void (*__kmp_destroy_user_lock_)(kmp_user_lock_p lck) = NULL;
3537 void (*__kmp_destroy_user_lock_with_checks_)(kmp_user_lock_p lck) = NULL;
3538 int (*__kmp_acquire_nested_user_lock_with_checks_)(kmp_user_lock_p lck,
3539                                                    kmp_int32 gtid) = NULL;
3540 
3541 int (*__kmp_test_nested_user_lock_with_checks_)(kmp_user_lock_p lck,
3542                                                 kmp_int32 gtid) = NULL;
3543 int (*__kmp_release_nested_user_lock_with_checks_)(kmp_user_lock_p lck,
3544                                                    kmp_int32 gtid) = NULL;
3545 void (*__kmp_init_nested_user_lock_with_checks_)(kmp_user_lock_p lck) = NULL;
3546 void (*__kmp_destroy_nested_user_lock_with_checks_)(kmp_user_lock_p lck) = NULL;
3547 
3548 int (*__kmp_is_user_lock_initialized_)(kmp_user_lock_p lck) = NULL;
3549 const ident_t *(*__kmp_get_user_lock_location_)(kmp_user_lock_p lck) = NULL;
3550 void (*__kmp_set_user_lock_location_)(kmp_user_lock_p lck,
3551                                       const ident_t *loc) = NULL;
3552 kmp_lock_flags_t (*__kmp_get_user_lock_flags_)(kmp_user_lock_p lck) = NULL;
3553 void (*__kmp_set_user_lock_flags_)(kmp_user_lock_p lck,
3554                                    kmp_lock_flags_t flags) = NULL;
3555 
3556 void __kmp_set_user_lock_vptrs(kmp_lock_kind_t user_lock_kind) {
3557   switch (user_lock_kind) {
3558   case lk_default:
3559   default:
3560     KMP_ASSERT(0);
3561 
3562   case lk_tas: {
3563     __kmp_base_user_lock_size = sizeof(kmp_base_tas_lock_t);
3564     __kmp_user_lock_size = sizeof(kmp_tas_lock_t);
3565 
3566     __kmp_get_user_lock_owner_ =
3567         (kmp_int32(*)(kmp_user_lock_p))(&__kmp_get_tas_lock_owner);
3568 
3569     if (__kmp_env_consistency_check) {
3570       KMP_BIND_USER_LOCK_WITH_CHECKS(tas);
3571       KMP_BIND_NESTED_USER_LOCK_WITH_CHECKS(tas);
3572     } else {
3573       KMP_BIND_USER_LOCK(tas);
3574       KMP_BIND_NESTED_USER_LOCK(tas);
3575     }
3576 
3577     __kmp_destroy_user_lock_ =
3578         (void (*)(kmp_user_lock_p))(&__kmp_destroy_tas_lock);
3579 
3580     __kmp_is_user_lock_initialized_ = (int (*)(kmp_user_lock_p))NULL;
3581 
3582     __kmp_get_user_lock_location_ = (const ident_t *(*)(kmp_user_lock_p))NULL;
3583 
3584     __kmp_set_user_lock_location_ =
3585         (void (*)(kmp_user_lock_p, const ident_t *))NULL;
3586 
3587     __kmp_get_user_lock_flags_ = (kmp_lock_flags_t(*)(kmp_user_lock_p))NULL;
3588 
3589     __kmp_set_user_lock_flags_ =
3590         (void (*)(kmp_user_lock_p, kmp_lock_flags_t))NULL;
3591   } break;
3592 
3593 #if KMP_USE_FUTEX
3594 
3595   case lk_futex: {
3596     __kmp_base_user_lock_size = sizeof(kmp_base_futex_lock_t);
3597     __kmp_user_lock_size = sizeof(kmp_futex_lock_t);
3598 
3599     __kmp_get_user_lock_owner_ =
3600         (kmp_int32(*)(kmp_user_lock_p))(&__kmp_get_futex_lock_owner);
3601 
3602     if (__kmp_env_consistency_check) {
3603       KMP_BIND_USER_LOCK_WITH_CHECKS(futex);
3604       KMP_BIND_NESTED_USER_LOCK_WITH_CHECKS(futex);
3605     } else {
3606       KMP_BIND_USER_LOCK(futex);
3607       KMP_BIND_NESTED_USER_LOCK(futex);
3608     }
3609 
3610     __kmp_destroy_user_lock_ =
3611         (void (*)(kmp_user_lock_p))(&__kmp_destroy_futex_lock);
3612 
3613     __kmp_is_user_lock_initialized_ = (int (*)(kmp_user_lock_p))NULL;
3614 
3615     __kmp_get_user_lock_location_ = (const ident_t *(*)(kmp_user_lock_p))NULL;
3616 
3617     __kmp_set_user_lock_location_ =
3618         (void (*)(kmp_user_lock_p, const ident_t *))NULL;
3619 
3620     __kmp_get_user_lock_flags_ = (kmp_lock_flags_t(*)(kmp_user_lock_p))NULL;
3621 
3622     __kmp_set_user_lock_flags_ =
3623         (void (*)(kmp_user_lock_p, kmp_lock_flags_t))NULL;
3624   } break;
3625 
3626 #endif // KMP_USE_FUTEX
3627 
3628   case lk_ticket: {
3629     __kmp_base_user_lock_size = sizeof(kmp_base_ticket_lock_t);
3630     __kmp_user_lock_size = sizeof(kmp_ticket_lock_t);
3631 
3632     __kmp_get_user_lock_owner_ =
3633         (kmp_int32(*)(kmp_user_lock_p))(&__kmp_get_ticket_lock_owner);
3634 
3635     if (__kmp_env_consistency_check) {
3636       KMP_BIND_USER_LOCK_WITH_CHECKS(ticket);
3637       KMP_BIND_NESTED_USER_LOCK_WITH_CHECKS(ticket);
3638     } else {
3639       KMP_BIND_USER_LOCK(ticket);
3640       KMP_BIND_NESTED_USER_LOCK(ticket);
3641     }
3642 
3643     __kmp_destroy_user_lock_ =
3644         (void (*)(kmp_user_lock_p))(&__kmp_destroy_ticket_lock);
3645 
3646     __kmp_is_user_lock_initialized_ =
3647         (int (*)(kmp_user_lock_p))(&__kmp_is_ticket_lock_initialized);
3648 
3649     __kmp_get_user_lock_location_ =
3650         (const ident_t *(*)(kmp_user_lock_p))(&__kmp_get_ticket_lock_location);
3651 
3652     __kmp_set_user_lock_location_ = (void (*)(
3653         kmp_user_lock_p, const ident_t *))(&__kmp_set_ticket_lock_location);
3654 
3655     __kmp_get_user_lock_flags_ =
3656         (kmp_lock_flags_t(*)(kmp_user_lock_p))(&__kmp_get_ticket_lock_flags);
3657 
3658     __kmp_set_user_lock_flags_ = (void (*)(kmp_user_lock_p, kmp_lock_flags_t))(
3659         &__kmp_set_ticket_lock_flags);
3660   } break;
3661 
3662   case lk_queuing: {
3663     __kmp_base_user_lock_size = sizeof(kmp_base_queuing_lock_t);
3664     __kmp_user_lock_size = sizeof(kmp_queuing_lock_t);
3665 
3666     __kmp_get_user_lock_owner_ =
3667         (kmp_int32(*)(kmp_user_lock_p))(&__kmp_get_queuing_lock_owner);
3668 
3669     if (__kmp_env_consistency_check) {
3670       KMP_BIND_USER_LOCK_WITH_CHECKS(queuing);
3671       KMP_BIND_NESTED_USER_LOCK_WITH_CHECKS(queuing);
3672     } else {
3673       KMP_BIND_USER_LOCK(queuing);
3674       KMP_BIND_NESTED_USER_LOCK(queuing);
3675     }
3676 
3677     __kmp_destroy_user_lock_ =
3678         (void (*)(kmp_user_lock_p))(&__kmp_destroy_queuing_lock);
3679 
3680     __kmp_is_user_lock_initialized_ =
3681         (int (*)(kmp_user_lock_p))(&__kmp_is_queuing_lock_initialized);
3682 
3683     __kmp_get_user_lock_location_ =
3684         (const ident_t *(*)(kmp_user_lock_p))(&__kmp_get_queuing_lock_location);
3685 
3686     __kmp_set_user_lock_location_ = (void (*)(
3687         kmp_user_lock_p, const ident_t *))(&__kmp_set_queuing_lock_location);
3688 
3689     __kmp_get_user_lock_flags_ =
3690         (kmp_lock_flags_t(*)(kmp_user_lock_p))(&__kmp_get_queuing_lock_flags);
3691 
3692     __kmp_set_user_lock_flags_ = (void (*)(kmp_user_lock_p, kmp_lock_flags_t))(
3693         &__kmp_set_queuing_lock_flags);
3694   } break;
3695 
3696 #if KMP_USE_ADAPTIVE_LOCKS
3697   case lk_adaptive: {
3698     __kmp_base_user_lock_size = sizeof(kmp_base_adaptive_lock_t);
3699     __kmp_user_lock_size = sizeof(kmp_adaptive_lock_t);
3700 
3701     __kmp_get_user_lock_owner_ =
3702         (kmp_int32(*)(kmp_user_lock_p))(&__kmp_get_queuing_lock_owner);
3703 
3704     if (__kmp_env_consistency_check) {
3705       KMP_BIND_USER_LOCK_WITH_CHECKS(adaptive);
3706     } else {
3707       KMP_BIND_USER_LOCK(adaptive);
3708     }
3709 
3710     __kmp_destroy_user_lock_ =
3711         (void (*)(kmp_user_lock_p))(&__kmp_destroy_adaptive_lock);
3712 
3713     __kmp_is_user_lock_initialized_ =
3714         (int (*)(kmp_user_lock_p))(&__kmp_is_queuing_lock_initialized);
3715 
3716     __kmp_get_user_lock_location_ =
3717         (const ident_t *(*)(kmp_user_lock_p))(&__kmp_get_queuing_lock_location);
3718 
3719     __kmp_set_user_lock_location_ = (void (*)(
3720         kmp_user_lock_p, const ident_t *))(&__kmp_set_queuing_lock_location);
3721 
3722     __kmp_get_user_lock_flags_ =
3723         (kmp_lock_flags_t(*)(kmp_user_lock_p))(&__kmp_get_queuing_lock_flags);
3724 
3725     __kmp_set_user_lock_flags_ = (void (*)(kmp_user_lock_p, kmp_lock_flags_t))(
3726         &__kmp_set_queuing_lock_flags);
3727 
3728   } break;
3729 #endif // KMP_USE_ADAPTIVE_LOCKS
3730 
3731   case lk_drdpa: {
3732     __kmp_base_user_lock_size = sizeof(kmp_base_drdpa_lock_t);
3733     __kmp_user_lock_size = sizeof(kmp_drdpa_lock_t);
3734 
3735     __kmp_get_user_lock_owner_ =
3736         (kmp_int32(*)(kmp_user_lock_p))(&__kmp_get_drdpa_lock_owner);
3737 
3738     if (__kmp_env_consistency_check) {
3739       KMP_BIND_USER_LOCK_WITH_CHECKS(drdpa);
3740       KMP_BIND_NESTED_USER_LOCK_WITH_CHECKS(drdpa);
3741     } else {
3742       KMP_BIND_USER_LOCK(drdpa);
3743       KMP_BIND_NESTED_USER_LOCK(drdpa);
3744     }
3745 
3746     __kmp_destroy_user_lock_ =
3747         (void (*)(kmp_user_lock_p))(&__kmp_destroy_drdpa_lock);
3748 
3749     __kmp_is_user_lock_initialized_ =
3750         (int (*)(kmp_user_lock_p))(&__kmp_is_drdpa_lock_initialized);
3751 
3752     __kmp_get_user_lock_location_ =
3753         (const ident_t *(*)(kmp_user_lock_p))(&__kmp_get_drdpa_lock_location);
3754 
3755     __kmp_set_user_lock_location_ = (void (*)(
3756         kmp_user_lock_p, const ident_t *))(&__kmp_set_drdpa_lock_location);
3757 
3758     __kmp_get_user_lock_flags_ =
3759         (kmp_lock_flags_t(*)(kmp_user_lock_p))(&__kmp_get_drdpa_lock_flags);
3760 
3761     __kmp_set_user_lock_flags_ = (void (*)(kmp_user_lock_p, kmp_lock_flags_t))(
3762         &__kmp_set_drdpa_lock_flags);
3763   } break;
3764   }
3765 }
3766 
3767 // ----------------------------------------------------------------------------
3768 // User lock table & lock allocation
3769 
3770 kmp_lock_table_t __kmp_user_lock_table = {1, 0, NULL};
3771 kmp_user_lock_p __kmp_lock_pool = NULL;
3772 
3773 // Lock block-allocation support.
3774 kmp_block_of_locks *__kmp_lock_blocks = NULL;
3775 int __kmp_num_locks_in_block = 1; // FIXME - tune this value
3776 
3777 static kmp_lock_index_t __kmp_lock_table_insert(kmp_user_lock_p lck) {
3778   // Assume that kmp_global_lock is held upon entry/exit.
3779   kmp_lock_index_t index;
3780   if (__kmp_user_lock_table.used >= __kmp_user_lock_table.allocated) {
3781     kmp_lock_index_t size;
3782     kmp_user_lock_p *table;
3783     // Reallocate lock table.
3784     if (__kmp_user_lock_table.allocated == 0) {
3785       size = 1024;
3786     } else {
3787       size = __kmp_user_lock_table.allocated * 2;
3788     }
3789     table = (kmp_user_lock_p *)__kmp_allocate(sizeof(kmp_user_lock_p) * size);
3790     KMP_MEMCPY(table + 1, __kmp_user_lock_table.table + 1,
3791                sizeof(kmp_user_lock_p) * (__kmp_user_lock_table.used - 1));
3792     table[0] = (kmp_user_lock_p)__kmp_user_lock_table.table;
3793     // We cannot free the previous table now, since it may be in use by other
3794     // threads. So save the pointer to the previous table in in the first
3795     // element of the new table. All the tables will be organized into a list,
3796     // and could be freed when library shutting down.
3797     __kmp_user_lock_table.table = table;
3798     __kmp_user_lock_table.allocated = size;
3799   }
3800   KMP_DEBUG_ASSERT(__kmp_user_lock_table.used <
3801                    __kmp_user_lock_table.allocated);
3802   index = __kmp_user_lock_table.used;
3803   __kmp_user_lock_table.table[index] = lck;
3804   ++__kmp_user_lock_table.used;
3805   return index;
3806 }
3807 
3808 static kmp_user_lock_p __kmp_lock_block_allocate() {
3809   // Assume that kmp_global_lock is held upon entry/exit.
3810   static int last_index = 0;
3811   if ((last_index >= __kmp_num_locks_in_block) || (__kmp_lock_blocks == NULL)) {
3812     // Restart the index.
3813     last_index = 0;
3814     // Need to allocate a new block.
3815     KMP_DEBUG_ASSERT(__kmp_user_lock_size > 0);
3816     size_t space_for_locks = __kmp_user_lock_size * __kmp_num_locks_in_block;
3817     char *buffer =
3818         (char *)__kmp_allocate(space_for_locks + sizeof(kmp_block_of_locks));
3819     // Set up the new block.
3820     kmp_block_of_locks *new_block =
3821         (kmp_block_of_locks *)(&buffer[space_for_locks]);
3822     new_block->next_block = __kmp_lock_blocks;
3823     new_block->locks = (void *)buffer;
3824     // Publish the new block.
3825     KMP_MB();
3826     __kmp_lock_blocks = new_block;
3827   }
3828   kmp_user_lock_p ret = (kmp_user_lock_p)(&(
3829       ((char *)(__kmp_lock_blocks->locks))[last_index * __kmp_user_lock_size]));
3830   last_index++;
3831   return ret;
3832 }
3833 
3834 // Get memory for a lock. It may be freshly allocated memory or reused memory
3835 // from lock pool.
3836 kmp_user_lock_p __kmp_user_lock_allocate(void **user_lock, kmp_int32 gtid,
3837                                          kmp_lock_flags_t flags) {
3838   kmp_user_lock_p lck;
3839   kmp_lock_index_t index;
3840   KMP_DEBUG_ASSERT(user_lock);
3841 
3842   __kmp_acquire_lock(&__kmp_global_lock, gtid);
3843 
3844   if (__kmp_lock_pool == NULL) {
3845     // Lock pool is empty. Allocate new memory.
3846 
3847     // ANNOTATION: Found no good way to express the syncronisation
3848     // between allocation and usage, so ignore the allocation
3849     ANNOTATE_IGNORE_WRITES_BEGIN();
3850     if (__kmp_num_locks_in_block <= 1) { // Tune this cutoff point.
3851       lck = (kmp_user_lock_p)__kmp_allocate(__kmp_user_lock_size);
3852     } else {
3853       lck = __kmp_lock_block_allocate();
3854     }
3855     ANNOTATE_IGNORE_WRITES_END();
3856 
3857     // Insert lock in the table so that it can be freed in __kmp_cleanup,
3858     // and debugger has info on all allocated locks.
3859     index = __kmp_lock_table_insert(lck);
3860   } else {
3861     // Pick up lock from pool.
3862     lck = __kmp_lock_pool;
3863     index = __kmp_lock_pool->pool.index;
3864     __kmp_lock_pool = __kmp_lock_pool->pool.next;
3865   }
3866 
3867   // We could potentially differentiate between nested and regular locks
3868   // here, and do the lock table lookup for regular locks only.
3869   if (OMP_LOCK_T_SIZE < sizeof(void *)) {
3870     *((kmp_lock_index_t *)user_lock) = index;
3871   } else {
3872     *((kmp_user_lock_p *)user_lock) = lck;
3873   }
3874 
3875   // mark the lock if it is critical section lock.
3876   __kmp_set_user_lock_flags(lck, flags);
3877 
3878   __kmp_release_lock(&__kmp_global_lock, gtid); // AC: TODO move this line upper
3879 
3880   return lck;
3881 }
3882 
3883 // Put lock's memory to pool for reusing.
3884 void __kmp_user_lock_free(void **user_lock, kmp_int32 gtid,
3885                           kmp_user_lock_p lck) {
3886   KMP_DEBUG_ASSERT(user_lock != NULL);
3887   KMP_DEBUG_ASSERT(lck != NULL);
3888 
3889   __kmp_acquire_lock(&__kmp_global_lock, gtid);
3890 
3891   lck->pool.next = __kmp_lock_pool;
3892   __kmp_lock_pool = lck;
3893   if (OMP_LOCK_T_SIZE < sizeof(void *)) {
3894     kmp_lock_index_t index = *((kmp_lock_index_t *)user_lock);
3895     KMP_DEBUG_ASSERT(0 < index && index <= __kmp_user_lock_table.used);
3896     lck->pool.index = index;
3897   }
3898 
3899   __kmp_release_lock(&__kmp_global_lock, gtid);
3900 }
3901 
3902 kmp_user_lock_p __kmp_lookup_user_lock(void **user_lock, char const *func) {
3903   kmp_user_lock_p lck = NULL;
3904 
3905   if (__kmp_env_consistency_check) {
3906     if (user_lock == NULL) {
3907       KMP_FATAL(LockIsUninitialized, func);
3908     }
3909   }
3910 
3911   if (OMP_LOCK_T_SIZE < sizeof(void *)) {
3912     kmp_lock_index_t index = *((kmp_lock_index_t *)user_lock);
3913     if (__kmp_env_consistency_check) {
3914       if (!(0 < index && index < __kmp_user_lock_table.used)) {
3915         KMP_FATAL(LockIsUninitialized, func);
3916       }
3917     }
3918     KMP_DEBUG_ASSERT(0 < index && index < __kmp_user_lock_table.used);
3919     KMP_DEBUG_ASSERT(__kmp_user_lock_size > 0);
3920     lck = __kmp_user_lock_table.table[index];
3921   } else {
3922     lck = *((kmp_user_lock_p *)user_lock);
3923   }
3924 
3925   if (__kmp_env_consistency_check) {
3926     if (lck == NULL) {
3927       KMP_FATAL(LockIsUninitialized, func);
3928     }
3929   }
3930 
3931   return lck;
3932 }
3933 
3934 void __kmp_cleanup_user_locks(void) {
3935   // Reset lock pool. Don't worry about lock in the pool--we will free them when
3936   // iterating through lock table (it includes all the locks, dead or alive).
3937   __kmp_lock_pool = NULL;
3938 
3939 #define IS_CRITICAL(lck)                                                       \
3940   ((__kmp_get_user_lock_flags_ != NULL) &&                                     \
3941    ((*__kmp_get_user_lock_flags_)(lck)&kmp_lf_critical_section))
3942 
3943   // Loop through lock table, free all locks.
3944   // Do not free item [0], it is reserved for lock tables list.
3945   //
3946   // FIXME - we are iterating through a list of (pointers to) objects of type
3947   // union kmp_user_lock, but we have no way of knowing whether the base type is
3948   // currently "pool" or whatever the global user lock type is.
3949   //
3950   // We are relying on the fact that for all of the user lock types
3951   // (except "tas"), the first field in the lock struct is the "initialized"
3952   // field, which is set to the address of the lock object itself when
3953   // the lock is initialized.  When the union is of type "pool", the
3954   // first field is a pointer to the next object in the free list, which
3955   // will not be the same address as the object itself.
3956   //
3957   // This means that the check (*__kmp_is_user_lock_initialized_)(lck) will fail
3958   // for "pool" objects on the free list.  This must happen as the "location"
3959   // field of real user locks overlaps the "index" field of "pool" objects.
3960   //
3961   // It would be better to run through the free list, and remove all "pool"
3962   // objects from the lock table before executing this loop.  However,
3963   // "pool" objects do not always have their index field set (only on
3964   // lin_32e), and I don't want to search the lock table for the address
3965   // of every "pool" object on the free list.
3966   while (__kmp_user_lock_table.used > 1) {
3967     const ident *loc;
3968 
3969     // reduce __kmp_user_lock_table.used before freeing the lock,
3970     // so that state of locks is consistent
3971     kmp_user_lock_p lck =
3972         __kmp_user_lock_table.table[--__kmp_user_lock_table.used];
3973 
3974     if ((__kmp_is_user_lock_initialized_ != NULL) &&
3975         (*__kmp_is_user_lock_initialized_)(lck)) {
3976       // Issue a warning if: KMP_CONSISTENCY_CHECK AND lock is initialized AND
3977       // it is NOT a critical section (user is not responsible for destroying
3978       // criticals) AND we know source location to report.
3979       if (__kmp_env_consistency_check && (!IS_CRITICAL(lck)) &&
3980           ((loc = __kmp_get_user_lock_location(lck)) != NULL) &&
3981           (loc->psource != NULL)) {
3982         kmp_str_loc_t str_loc = __kmp_str_loc_init(loc->psource, false);
3983         KMP_WARNING(CnsLockNotDestroyed, str_loc.file, str_loc.line);
3984         __kmp_str_loc_free(&str_loc);
3985       }
3986 
3987 #ifdef KMP_DEBUG
3988       if (IS_CRITICAL(lck)) {
3989         KA_TRACE(
3990             20,
3991             ("__kmp_cleanup_user_locks: free critical section lock %p (%p)\n",
3992              lck, *(void **)lck));
3993       } else {
3994         KA_TRACE(20, ("__kmp_cleanup_user_locks: free lock %p (%p)\n", lck,
3995                       *(void **)lck));
3996       }
3997 #endif // KMP_DEBUG
3998 
3999       // Cleanup internal lock dynamic resources (for drdpa locks particularly).
4000       __kmp_destroy_user_lock(lck);
4001     }
4002 
4003     // Free the lock if block allocation of locks is not used.
4004     if (__kmp_lock_blocks == NULL) {
4005       __kmp_free(lck);
4006     }
4007   }
4008 
4009 #undef IS_CRITICAL
4010 
4011   // delete lock table(s).
4012   kmp_user_lock_p *table_ptr = __kmp_user_lock_table.table;
4013   __kmp_user_lock_table.table = NULL;
4014   __kmp_user_lock_table.allocated = 0;
4015 
4016   while (table_ptr != NULL) {
4017     // In the first element we saved the pointer to the previous
4018     // (smaller) lock table.
4019     kmp_user_lock_p *next = (kmp_user_lock_p *)(table_ptr[0]);
4020     __kmp_free(table_ptr);
4021     table_ptr = next;
4022   }
4023 
4024   // Free buffers allocated for blocks of locks.
4025   kmp_block_of_locks_t *block_ptr = __kmp_lock_blocks;
4026   __kmp_lock_blocks = NULL;
4027 
4028   while (block_ptr != NULL) {
4029     kmp_block_of_locks_t *next = block_ptr->next_block;
4030     __kmp_free(block_ptr->locks);
4031     // *block_ptr itself was allocated at the end of the locks vector.
4032     block_ptr = next;
4033   }
4034 
4035   TCW_4(__kmp_init_user_locks, FALSE);
4036 }
4037 
4038 #endif // KMP_USE_DYNAMIC_LOCK
4039