LLVM OpenMP* Runtime Library
kmp_lock.cpp
1 /*
2  * kmp_lock.cpp -- lock-related functions
3  */
4 
5 //===----------------------------------------------------------------------===//
6 //
7 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
8 // See https://llvm.org/LICENSE.txt for license information.
9 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #include <stddef.h>
14 #include <atomic>
15 
16 #include "kmp.h"
17 #include "kmp_i18n.h"
18 #include "kmp_io.h"
19 #include "kmp_itt.h"
20 #include "kmp_lock.h"
21 #include "kmp_wait_release.h"
22 #include "kmp_wrapper_getpid.h"
23 
24 #include "tsan_annotations.h"
25 
26 #if KMP_USE_FUTEX
27 #include <sys/syscall.h>
28 #include <unistd.h>
29 // We should really include <futex.h>, but that causes compatibility problems on
30 // different Linux* OS distributions that either require that you include (or
31 // break when you try to include) <pci/types.h>. Since all we need is the two
32 // macros below (which are part of the kernel ABI, so can't change) we just
33 // define the constants here and don't include <futex.h>
34 #ifndef FUTEX_WAIT
35 #define FUTEX_WAIT 0
36 #endif
37 #ifndef FUTEX_WAKE
38 #define FUTEX_WAKE 1
39 #endif
40 #endif
41 
42 /* Implement spin locks for internal library use. */
43 /* The algorithm implemented is Lamport's bakery lock [1974]. */
44 
45 void __kmp_validate_locks(void) {
46  int i;
47  kmp_uint32 x, y;
48 
49  /* Check to make sure unsigned arithmetic does wraps properly */
50  x = ~((kmp_uint32)0) - 2;
51  y = x - 2;
52 
53  for (i = 0; i < 8; ++i, ++x, ++y) {
54  kmp_uint32 z = (x - y);
55  KMP_ASSERT(z == 2);
56  }
57 
58  KMP_ASSERT(offsetof(kmp_base_queuing_lock, tail_id) % 8 == 0);
59 }
60 
61 /* ------------------------------------------------------------------------ */
62 /* test and set locks */
63 
64 // For the non-nested locks, we can only assume that the first 4 bytes were
65 // allocated, since gcc only allocates 4 bytes for omp_lock_t, and the Intel
66 // compiler only allocates a 4 byte pointer on IA-32 architecture. On
67 // Windows* OS on Intel(R) 64, we can assume that all 8 bytes were allocated.
68 //
69 // gcc reserves >= 8 bytes for nested locks, so we can assume that the
70 // entire 8 bytes were allocated for nested locks on all 64-bit platforms.
71 
72 static kmp_int32 __kmp_get_tas_lock_owner(kmp_tas_lock_t *lck) {
73  return KMP_LOCK_STRIP(KMP_ATOMIC_LD_RLX(&lck->lk.poll)) - 1;
74 }
75 
76 static inline bool __kmp_is_tas_lock_nestable(kmp_tas_lock_t *lck) {
77  return lck->lk.depth_locked != -1;
78 }
79 
80 __forceinline static int
81 __kmp_acquire_tas_lock_timed_template(kmp_tas_lock_t *lck, kmp_int32 gtid) {
82  KMP_MB();
83 
84 #ifdef USE_LOCK_PROFILE
85  kmp_uint32 curr = KMP_LOCK_STRIP(lck->lk.poll);
86  if ((curr != 0) && (curr != gtid + 1))
87  __kmp_printf("LOCK CONTENTION: %p\n", lck);
88 /* else __kmp_printf( "." );*/
89 #endif /* USE_LOCK_PROFILE */
90 
91  kmp_int32 tas_free = KMP_LOCK_FREE(tas);
92  kmp_int32 tas_busy = KMP_LOCK_BUSY(gtid + 1, tas);
93 
94  if (KMP_ATOMIC_LD_RLX(&lck->lk.poll) == tas_free &&
95  __kmp_atomic_compare_store_acq(&lck->lk.poll, tas_free, tas_busy)) {
96  KMP_FSYNC_ACQUIRED(lck);
97  return KMP_LOCK_ACQUIRED_FIRST;
98  }
99 
100  kmp_uint32 spins;
101  KMP_FSYNC_PREPARE(lck);
102  KMP_INIT_YIELD(spins);
103  kmp_backoff_t backoff = __kmp_spin_backoff_params;
104  do {
105  __kmp_spin_backoff(&backoff);
106  KMP_YIELD_OVERSUB_ELSE_SPIN(spins);
107  } while (KMP_ATOMIC_LD_RLX(&lck->lk.poll) != tas_free ||
108  !__kmp_atomic_compare_store_acq(&lck->lk.poll, tas_free, tas_busy));
109  KMP_FSYNC_ACQUIRED(lck);
110  return KMP_LOCK_ACQUIRED_FIRST;
111 }
112 
113 int __kmp_acquire_tas_lock(kmp_tas_lock_t *lck, kmp_int32 gtid) {
114  int retval = __kmp_acquire_tas_lock_timed_template(lck, gtid);
115  ANNOTATE_TAS_ACQUIRED(lck);
116  return retval;
117 }
118 
119 static int __kmp_acquire_tas_lock_with_checks(kmp_tas_lock_t *lck,
120  kmp_int32 gtid) {
121  char const *const func = "omp_set_lock";
122  if ((sizeof(kmp_tas_lock_t) <= OMP_LOCK_T_SIZE) &&
123  __kmp_is_tas_lock_nestable(lck)) {
124  KMP_FATAL(LockNestableUsedAsSimple, func);
125  }
126  if ((gtid >= 0) && (__kmp_get_tas_lock_owner(lck) == gtid)) {
127  KMP_FATAL(LockIsAlreadyOwned, func);
128  }
129  return __kmp_acquire_tas_lock(lck, gtid);
130 }
131 
132 int __kmp_test_tas_lock(kmp_tas_lock_t *lck, kmp_int32 gtid) {
133  kmp_int32 tas_free = KMP_LOCK_FREE(tas);
134  kmp_int32 tas_busy = KMP_LOCK_BUSY(gtid + 1, tas);
135  if (KMP_ATOMIC_LD_RLX(&lck->lk.poll) == tas_free &&
136  __kmp_atomic_compare_store_acq(&lck->lk.poll, tas_free, tas_busy)) {
137  KMP_FSYNC_ACQUIRED(lck);
138  return TRUE;
139  }
140  return FALSE;
141 }
142 
143 static int __kmp_test_tas_lock_with_checks(kmp_tas_lock_t *lck,
144  kmp_int32 gtid) {
145  char const *const func = "omp_test_lock";
146  if ((sizeof(kmp_tas_lock_t) <= OMP_LOCK_T_SIZE) &&
147  __kmp_is_tas_lock_nestable(lck)) {
148  KMP_FATAL(LockNestableUsedAsSimple, func);
149  }
150  return __kmp_test_tas_lock(lck, gtid);
151 }
152 
153 int __kmp_release_tas_lock(kmp_tas_lock_t *lck, kmp_int32 gtid) {
154  KMP_MB(); /* Flush all pending memory write invalidates. */
155 
156  KMP_FSYNC_RELEASING(lck);
157  ANNOTATE_TAS_RELEASED(lck);
158  KMP_ATOMIC_ST_REL(&lck->lk.poll, KMP_LOCK_FREE(tas));
159  KMP_MB(); /* Flush all pending memory write invalidates. */
160 
161  KMP_YIELD_OVERSUB();
162  return KMP_LOCK_RELEASED;
163 }
164 
165 static int __kmp_release_tas_lock_with_checks(kmp_tas_lock_t *lck,
166  kmp_int32 gtid) {
167  char const *const func = "omp_unset_lock";
168  KMP_MB(); /* in case another processor initialized lock */
169  if ((sizeof(kmp_tas_lock_t) <= OMP_LOCK_T_SIZE) &&
170  __kmp_is_tas_lock_nestable(lck)) {
171  KMP_FATAL(LockNestableUsedAsSimple, func);
172  }
173  if (__kmp_get_tas_lock_owner(lck) == -1) {
174  KMP_FATAL(LockUnsettingFree, func);
175  }
176  if ((gtid >= 0) && (__kmp_get_tas_lock_owner(lck) >= 0) &&
177  (__kmp_get_tas_lock_owner(lck) != gtid)) {
178  KMP_FATAL(LockUnsettingSetByAnother, func);
179  }
180  return __kmp_release_tas_lock(lck, gtid);
181 }
182 
183 void __kmp_init_tas_lock(kmp_tas_lock_t *lck) {
184  lck->lk.poll = KMP_LOCK_FREE(tas);
185 }
186 
187 void __kmp_destroy_tas_lock(kmp_tas_lock_t *lck) { lck->lk.poll = 0; }
188 
189 static void __kmp_destroy_tas_lock_with_checks(kmp_tas_lock_t *lck) {
190  char const *const func = "omp_destroy_lock";
191  if ((sizeof(kmp_tas_lock_t) <= OMP_LOCK_T_SIZE) &&
192  __kmp_is_tas_lock_nestable(lck)) {
193  KMP_FATAL(LockNestableUsedAsSimple, func);
194  }
195  if (__kmp_get_tas_lock_owner(lck) != -1) {
196  KMP_FATAL(LockStillOwned, func);
197  }
198  __kmp_destroy_tas_lock(lck);
199 }
200 
201 // nested test and set locks
202 
203 int __kmp_acquire_nested_tas_lock(kmp_tas_lock_t *lck, kmp_int32 gtid) {
204  KMP_DEBUG_ASSERT(gtid >= 0);
205 
206  if (__kmp_get_tas_lock_owner(lck) == gtid) {
207  lck->lk.depth_locked += 1;
208  return KMP_LOCK_ACQUIRED_NEXT;
209  } else {
210  __kmp_acquire_tas_lock_timed_template(lck, gtid);
211  ANNOTATE_TAS_ACQUIRED(lck);
212  lck->lk.depth_locked = 1;
213  return KMP_LOCK_ACQUIRED_FIRST;
214  }
215 }
216 
217 static int __kmp_acquire_nested_tas_lock_with_checks(kmp_tas_lock_t *lck,
218  kmp_int32 gtid) {
219  char const *const func = "omp_set_nest_lock";
220  if (!__kmp_is_tas_lock_nestable(lck)) {
221  KMP_FATAL(LockSimpleUsedAsNestable, func);
222  }
223  return __kmp_acquire_nested_tas_lock(lck, gtid);
224 }
225 
226 int __kmp_test_nested_tas_lock(kmp_tas_lock_t *lck, kmp_int32 gtid) {
227  int retval;
228 
229  KMP_DEBUG_ASSERT(gtid >= 0);
230 
231  if (__kmp_get_tas_lock_owner(lck) == gtid) {
232  retval = ++lck->lk.depth_locked;
233  } else if (!__kmp_test_tas_lock(lck, gtid)) {
234  retval = 0;
235  } else {
236  KMP_MB();
237  retval = lck->lk.depth_locked = 1;
238  }
239  return retval;
240 }
241 
242 static int __kmp_test_nested_tas_lock_with_checks(kmp_tas_lock_t *lck,
243  kmp_int32 gtid) {
244  char const *const func = "omp_test_nest_lock";
245  if (!__kmp_is_tas_lock_nestable(lck)) {
246  KMP_FATAL(LockSimpleUsedAsNestable, func);
247  }
248  return __kmp_test_nested_tas_lock(lck, gtid);
249 }
250 
251 int __kmp_release_nested_tas_lock(kmp_tas_lock_t *lck, kmp_int32 gtid) {
252  KMP_DEBUG_ASSERT(gtid >= 0);
253 
254  KMP_MB();
255  if (--(lck->lk.depth_locked) == 0) {
256  __kmp_release_tas_lock(lck, gtid);
257  return KMP_LOCK_RELEASED;
258  }
259  return KMP_LOCK_STILL_HELD;
260 }
261 
262 static int __kmp_release_nested_tas_lock_with_checks(kmp_tas_lock_t *lck,
263  kmp_int32 gtid) {
264  char const *const func = "omp_unset_nest_lock";
265  KMP_MB(); /* in case another processor initialized lock */
266  if (!__kmp_is_tas_lock_nestable(lck)) {
267  KMP_FATAL(LockSimpleUsedAsNestable, func);
268  }
269  if (__kmp_get_tas_lock_owner(lck) == -1) {
270  KMP_FATAL(LockUnsettingFree, func);
271  }
272  if (__kmp_get_tas_lock_owner(lck) != gtid) {
273  KMP_FATAL(LockUnsettingSetByAnother, func);
274  }
275  return __kmp_release_nested_tas_lock(lck, gtid);
276 }
277 
278 void __kmp_init_nested_tas_lock(kmp_tas_lock_t *lck) {
279  __kmp_init_tas_lock(lck);
280  lck->lk.depth_locked = 0; // >= 0 for nestable locks, -1 for simple locks
281 }
282 
283 void __kmp_destroy_nested_tas_lock(kmp_tas_lock_t *lck) {
284  __kmp_destroy_tas_lock(lck);
285  lck->lk.depth_locked = 0;
286 }
287 
288 static void __kmp_destroy_nested_tas_lock_with_checks(kmp_tas_lock_t *lck) {
289  char const *const func = "omp_destroy_nest_lock";
290  if (!__kmp_is_tas_lock_nestable(lck)) {
291  KMP_FATAL(LockSimpleUsedAsNestable, func);
292  }
293  if (__kmp_get_tas_lock_owner(lck) != -1) {
294  KMP_FATAL(LockStillOwned, func);
295  }
296  __kmp_destroy_nested_tas_lock(lck);
297 }
298 
299 #if KMP_USE_FUTEX
300 
301 /* ------------------------------------------------------------------------ */
302 /* futex locks */
303 
304 // futex locks are really just test and set locks, with a different method
305 // of handling contention. They take the same amount of space as test and
306 // set locks, and are allocated the same way (i.e. use the area allocated by
307 // the compiler for non-nested locks / allocate nested locks on the heap).
308 
309 static kmp_int32 __kmp_get_futex_lock_owner(kmp_futex_lock_t *lck) {
310  return KMP_LOCK_STRIP((TCR_4(lck->lk.poll) >> 1)) - 1;
311 }
312 
313 static inline bool __kmp_is_futex_lock_nestable(kmp_futex_lock_t *lck) {
314  return lck->lk.depth_locked != -1;
315 }
316 
317 __forceinline static int
318 __kmp_acquire_futex_lock_timed_template(kmp_futex_lock_t *lck, kmp_int32 gtid) {
319  kmp_int32 gtid_code = (gtid + 1) << 1;
320 
321  KMP_MB();
322 
323 #ifdef USE_LOCK_PROFILE
324  kmp_uint32 curr = KMP_LOCK_STRIP(TCR_4(lck->lk.poll));
325  if ((curr != 0) && (curr != gtid_code))
326  __kmp_printf("LOCK CONTENTION: %p\n", lck);
327 /* else __kmp_printf( "." );*/
328 #endif /* USE_LOCK_PROFILE */
329 
330  KMP_FSYNC_PREPARE(lck);
331  KA_TRACE(1000, ("__kmp_acquire_futex_lock: lck:%p(0x%x), T#%d entering\n",
332  lck, lck->lk.poll, gtid));
333 
334  kmp_int32 poll_val;
335 
336  while ((poll_val = KMP_COMPARE_AND_STORE_RET32(
337  &(lck->lk.poll), KMP_LOCK_FREE(futex),
338  KMP_LOCK_BUSY(gtid_code, futex))) != KMP_LOCK_FREE(futex)) {
339 
340  kmp_int32 cond = KMP_LOCK_STRIP(poll_val) & 1;
341  KA_TRACE(
342  1000,
343  ("__kmp_acquire_futex_lock: lck:%p, T#%d poll_val = 0x%x cond = 0x%x\n",
344  lck, gtid, poll_val, cond));
345 
346  // NOTE: if you try to use the following condition for this branch
347  //
348  // if ( poll_val & 1 == 0 )
349  //
350  // Then the 12.0 compiler has a bug where the following block will
351  // always be skipped, regardless of the value of the LSB of poll_val.
352  if (!cond) {
353  // Try to set the lsb in the poll to indicate to the owner
354  // thread that they need to wake this thread up.
355  if (!KMP_COMPARE_AND_STORE_REL32(&(lck->lk.poll), poll_val,
356  poll_val | KMP_LOCK_BUSY(1, futex))) {
357  KA_TRACE(
358  1000,
359  ("__kmp_acquire_futex_lock: lck:%p(0x%x), T#%d can't set bit 0\n",
360  lck, lck->lk.poll, gtid));
361  continue;
362  }
363  poll_val |= KMP_LOCK_BUSY(1, futex);
364 
365  KA_TRACE(1000,
366  ("__kmp_acquire_futex_lock: lck:%p(0x%x), T#%d bit 0 set\n", lck,
367  lck->lk.poll, gtid));
368  }
369 
370  KA_TRACE(
371  1000,
372  ("__kmp_acquire_futex_lock: lck:%p, T#%d before futex_wait(0x%x)\n",
373  lck, gtid, poll_val));
374 
375  long rc;
376  if ((rc = syscall(__NR_futex, &(lck->lk.poll), FUTEX_WAIT, poll_val, NULL,
377  NULL, 0)) != 0) {
378  KA_TRACE(1000, ("__kmp_acquire_futex_lock: lck:%p, T#%d futex_wait(0x%x) "
379  "failed (rc=%ld errno=%d)\n",
380  lck, gtid, poll_val, rc, errno));
381  continue;
382  }
383 
384  KA_TRACE(1000,
385  ("__kmp_acquire_futex_lock: lck:%p, T#%d after futex_wait(0x%x)\n",
386  lck, gtid, poll_val));
387  // This thread has now done a successful futex wait call and was entered on
388  // the OS futex queue. We must now perform a futex wake call when releasing
389  // the lock, as we have no idea how many other threads are in the queue.
390  gtid_code |= 1;
391  }
392 
393  KMP_FSYNC_ACQUIRED(lck);
394  KA_TRACE(1000, ("__kmp_acquire_futex_lock: lck:%p(0x%x), T#%d exiting\n", lck,
395  lck->lk.poll, gtid));
396  return KMP_LOCK_ACQUIRED_FIRST;
397 }
398 
399 int __kmp_acquire_futex_lock(kmp_futex_lock_t *lck, kmp_int32 gtid) {
400  int retval = __kmp_acquire_futex_lock_timed_template(lck, gtid);
401  ANNOTATE_FUTEX_ACQUIRED(lck);
402  return retval;
403 }
404 
405 static int __kmp_acquire_futex_lock_with_checks(kmp_futex_lock_t *lck,
406  kmp_int32 gtid) {
407  char const *const func = "omp_set_lock";
408  if ((sizeof(kmp_futex_lock_t) <= OMP_LOCK_T_SIZE) &&
409  __kmp_is_futex_lock_nestable(lck)) {
410  KMP_FATAL(LockNestableUsedAsSimple, func);
411  }
412  if ((gtid >= 0) && (__kmp_get_futex_lock_owner(lck) == gtid)) {
413  KMP_FATAL(LockIsAlreadyOwned, func);
414  }
415  return __kmp_acquire_futex_lock(lck, gtid);
416 }
417 
418 int __kmp_test_futex_lock(kmp_futex_lock_t *lck, kmp_int32 gtid) {
419  if (KMP_COMPARE_AND_STORE_ACQ32(&(lck->lk.poll), KMP_LOCK_FREE(futex),
420  KMP_LOCK_BUSY((gtid + 1) << 1, futex))) {
421  KMP_FSYNC_ACQUIRED(lck);
422  return TRUE;
423  }
424  return FALSE;
425 }
426 
427 static int __kmp_test_futex_lock_with_checks(kmp_futex_lock_t *lck,
428  kmp_int32 gtid) {
429  char const *const func = "omp_test_lock";
430  if ((sizeof(kmp_futex_lock_t) <= OMP_LOCK_T_SIZE) &&
431  __kmp_is_futex_lock_nestable(lck)) {
432  KMP_FATAL(LockNestableUsedAsSimple, func);
433  }
434  return __kmp_test_futex_lock(lck, gtid);
435 }
436 
437 int __kmp_release_futex_lock(kmp_futex_lock_t *lck, kmp_int32 gtid) {
438  KMP_MB(); /* Flush all pending memory write invalidates. */
439 
440  KA_TRACE(1000, ("__kmp_release_futex_lock: lck:%p(0x%x), T#%d entering\n",
441  lck, lck->lk.poll, gtid));
442 
443  KMP_FSYNC_RELEASING(lck);
444  ANNOTATE_FUTEX_RELEASED(lck);
445 
446  kmp_int32 poll_val = KMP_XCHG_FIXED32(&(lck->lk.poll), KMP_LOCK_FREE(futex));
447 
448  KA_TRACE(1000,
449  ("__kmp_release_futex_lock: lck:%p, T#%d released poll_val = 0x%x\n",
450  lck, gtid, poll_val));
451 
452  if (KMP_LOCK_STRIP(poll_val) & 1) {
453  KA_TRACE(1000,
454  ("__kmp_release_futex_lock: lck:%p, T#%d futex_wake 1 thread\n",
455  lck, gtid));
456  syscall(__NR_futex, &(lck->lk.poll), FUTEX_WAKE, KMP_LOCK_BUSY(1, futex),
457  NULL, NULL, 0);
458  }
459 
460  KMP_MB(); /* Flush all pending memory write invalidates. */
461 
462  KA_TRACE(1000, ("__kmp_release_futex_lock: lck:%p(0x%x), T#%d exiting\n", lck,
463  lck->lk.poll, gtid));
464 
465  KMP_YIELD_OVERSUB();
466  return KMP_LOCK_RELEASED;
467 }
468 
469 static int __kmp_release_futex_lock_with_checks(kmp_futex_lock_t *lck,
470  kmp_int32 gtid) {
471  char const *const func = "omp_unset_lock";
472  KMP_MB(); /* in case another processor initialized lock */
473  if ((sizeof(kmp_futex_lock_t) <= OMP_LOCK_T_SIZE) &&
474  __kmp_is_futex_lock_nestable(lck)) {
475  KMP_FATAL(LockNestableUsedAsSimple, func);
476  }
477  if (__kmp_get_futex_lock_owner(lck) == -1) {
478  KMP_FATAL(LockUnsettingFree, func);
479  }
480  if ((gtid >= 0) && (__kmp_get_futex_lock_owner(lck) >= 0) &&
481  (__kmp_get_futex_lock_owner(lck) != gtid)) {
482  KMP_FATAL(LockUnsettingSetByAnother, func);
483  }
484  return __kmp_release_futex_lock(lck, gtid);
485 }
486 
487 void __kmp_init_futex_lock(kmp_futex_lock_t *lck) {
488  TCW_4(lck->lk.poll, KMP_LOCK_FREE(futex));
489 }
490 
491 void __kmp_destroy_futex_lock(kmp_futex_lock_t *lck) { lck->lk.poll = 0; }
492 
493 static void __kmp_destroy_futex_lock_with_checks(kmp_futex_lock_t *lck) {
494  char const *const func = "omp_destroy_lock";
495  if ((sizeof(kmp_futex_lock_t) <= OMP_LOCK_T_SIZE) &&
496  __kmp_is_futex_lock_nestable(lck)) {
497  KMP_FATAL(LockNestableUsedAsSimple, func);
498  }
499  if (__kmp_get_futex_lock_owner(lck) != -1) {
500  KMP_FATAL(LockStillOwned, func);
501  }
502  __kmp_destroy_futex_lock(lck);
503 }
504 
505 // nested futex locks
506 
507 int __kmp_acquire_nested_futex_lock(kmp_futex_lock_t *lck, kmp_int32 gtid) {
508  KMP_DEBUG_ASSERT(gtid >= 0);
509 
510  if (__kmp_get_futex_lock_owner(lck) == gtid) {
511  lck->lk.depth_locked += 1;
512  return KMP_LOCK_ACQUIRED_NEXT;
513  } else {
514  __kmp_acquire_futex_lock_timed_template(lck, gtid);
515  ANNOTATE_FUTEX_ACQUIRED(lck);
516  lck->lk.depth_locked = 1;
517  return KMP_LOCK_ACQUIRED_FIRST;
518  }
519 }
520 
521 static int __kmp_acquire_nested_futex_lock_with_checks(kmp_futex_lock_t *lck,
522  kmp_int32 gtid) {
523  char const *const func = "omp_set_nest_lock";
524  if (!__kmp_is_futex_lock_nestable(lck)) {
525  KMP_FATAL(LockSimpleUsedAsNestable, func);
526  }
527  return __kmp_acquire_nested_futex_lock(lck, gtid);
528 }
529 
530 int __kmp_test_nested_futex_lock(kmp_futex_lock_t *lck, kmp_int32 gtid) {
531  int retval;
532 
533  KMP_DEBUG_ASSERT(gtid >= 0);
534 
535  if (__kmp_get_futex_lock_owner(lck) == gtid) {
536  retval = ++lck->lk.depth_locked;
537  } else if (!__kmp_test_futex_lock(lck, gtid)) {
538  retval = 0;
539  } else {
540  KMP_MB();
541  retval = lck->lk.depth_locked = 1;
542  }
543  return retval;
544 }
545 
546 static int __kmp_test_nested_futex_lock_with_checks(kmp_futex_lock_t *lck,
547  kmp_int32 gtid) {
548  char const *const func = "omp_test_nest_lock";
549  if (!__kmp_is_futex_lock_nestable(lck)) {
550  KMP_FATAL(LockSimpleUsedAsNestable, func);
551  }
552  return __kmp_test_nested_futex_lock(lck, gtid);
553 }
554 
555 int __kmp_release_nested_futex_lock(kmp_futex_lock_t *lck, kmp_int32 gtid) {
556  KMP_DEBUG_ASSERT(gtid >= 0);
557 
558  KMP_MB();
559  if (--(lck->lk.depth_locked) == 0) {
560  __kmp_release_futex_lock(lck, gtid);
561  return KMP_LOCK_RELEASED;
562  }
563  return KMP_LOCK_STILL_HELD;
564 }
565 
566 static int __kmp_release_nested_futex_lock_with_checks(kmp_futex_lock_t *lck,
567  kmp_int32 gtid) {
568  char const *const func = "omp_unset_nest_lock";
569  KMP_MB(); /* in case another processor initialized lock */
570  if (!__kmp_is_futex_lock_nestable(lck)) {
571  KMP_FATAL(LockSimpleUsedAsNestable, func);
572  }
573  if (__kmp_get_futex_lock_owner(lck) == -1) {
574  KMP_FATAL(LockUnsettingFree, func);
575  }
576  if (__kmp_get_futex_lock_owner(lck) != gtid) {
577  KMP_FATAL(LockUnsettingSetByAnother, func);
578  }
579  return __kmp_release_nested_futex_lock(lck, gtid);
580 }
581 
582 void __kmp_init_nested_futex_lock(kmp_futex_lock_t *lck) {
583  __kmp_init_futex_lock(lck);
584  lck->lk.depth_locked = 0; // >= 0 for nestable locks, -1 for simple locks
585 }
586 
587 void __kmp_destroy_nested_futex_lock(kmp_futex_lock_t *lck) {
588  __kmp_destroy_futex_lock(lck);
589  lck->lk.depth_locked = 0;
590 }
591 
592 static void __kmp_destroy_nested_futex_lock_with_checks(kmp_futex_lock_t *lck) {
593  char const *const func = "omp_destroy_nest_lock";
594  if (!__kmp_is_futex_lock_nestable(lck)) {
595  KMP_FATAL(LockSimpleUsedAsNestable, func);
596  }
597  if (__kmp_get_futex_lock_owner(lck) != -1) {
598  KMP_FATAL(LockStillOwned, func);
599  }
600  __kmp_destroy_nested_futex_lock(lck);
601 }
602 
603 #endif // KMP_USE_FUTEX
604 
605 /* ------------------------------------------------------------------------ */
606 /* ticket (bakery) locks */
607 
608 static kmp_int32 __kmp_get_ticket_lock_owner(kmp_ticket_lock_t *lck) {
609  return std::atomic_load_explicit(&lck->lk.owner_id,
610  std::memory_order_relaxed) -
611  1;
612 }
613 
614 static inline bool __kmp_is_ticket_lock_nestable(kmp_ticket_lock_t *lck) {
615  return std::atomic_load_explicit(&lck->lk.depth_locked,
616  std::memory_order_relaxed) != -1;
617 }
618 
619 static kmp_uint32 __kmp_bakery_check(void *now_serving, kmp_uint32 my_ticket) {
620  return std::atomic_load_explicit((std::atomic<unsigned> *)now_serving,
621  std::memory_order_acquire) == my_ticket;
622 }
623 
624 __forceinline static int
625 __kmp_acquire_ticket_lock_timed_template(kmp_ticket_lock_t *lck,
626  kmp_int32 gtid) {
627  kmp_uint32 my_ticket = std::atomic_fetch_add_explicit(
628  &lck->lk.next_ticket, 1U, std::memory_order_relaxed);
629 
630 #ifdef USE_LOCK_PROFILE
631  if (std::atomic_load_explicit(&lck->lk.now_serving,
632  std::memory_order_relaxed) != my_ticket)
633  __kmp_printf("LOCK CONTENTION: %p\n", lck);
634 /* else __kmp_printf( "." );*/
635 #endif /* USE_LOCK_PROFILE */
636 
637  if (std::atomic_load_explicit(&lck->lk.now_serving,
638  std::memory_order_acquire) == my_ticket) {
639  return KMP_LOCK_ACQUIRED_FIRST;
640  }
641  KMP_WAIT_PTR(&lck->lk.now_serving, my_ticket, __kmp_bakery_check, lck);
642  return KMP_LOCK_ACQUIRED_FIRST;
643 }
644 
645 int __kmp_acquire_ticket_lock(kmp_ticket_lock_t *lck, kmp_int32 gtid) {
646  int retval = __kmp_acquire_ticket_lock_timed_template(lck, gtid);
647  ANNOTATE_TICKET_ACQUIRED(lck);
648  return retval;
649 }
650 
651 static int __kmp_acquire_ticket_lock_with_checks(kmp_ticket_lock_t *lck,
652  kmp_int32 gtid) {
653  char const *const func = "omp_set_lock";
654 
655  if (!std::atomic_load_explicit(&lck->lk.initialized,
656  std::memory_order_relaxed)) {
657  KMP_FATAL(LockIsUninitialized, func);
658  }
659  if (lck->lk.self != lck) {
660  KMP_FATAL(LockIsUninitialized, func);
661  }
662  if (__kmp_is_ticket_lock_nestable(lck)) {
663  KMP_FATAL(LockNestableUsedAsSimple, func);
664  }
665  if ((gtid >= 0) && (__kmp_get_ticket_lock_owner(lck) == gtid)) {
666  KMP_FATAL(LockIsAlreadyOwned, func);
667  }
668 
669  __kmp_acquire_ticket_lock(lck, gtid);
670 
671  std::atomic_store_explicit(&lck->lk.owner_id, gtid + 1,
672  std::memory_order_relaxed);
673  return KMP_LOCK_ACQUIRED_FIRST;
674 }
675 
676 int __kmp_test_ticket_lock(kmp_ticket_lock_t *lck, kmp_int32 gtid) {
677  kmp_uint32 my_ticket = std::atomic_load_explicit(&lck->lk.next_ticket,
678  std::memory_order_relaxed);
679 
680  if (std::atomic_load_explicit(&lck->lk.now_serving,
681  std::memory_order_relaxed) == my_ticket) {
682  kmp_uint32 next_ticket = my_ticket + 1;
683  if (std::atomic_compare_exchange_strong_explicit(
684  &lck->lk.next_ticket, &my_ticket, next_ticket,
685  std::memory_order_acquire, std::memory_order_acquire)) {
686  return TRUE;
687  }
688  }
689  return FALSE;
690 }
691 
692 static int __kmp_test_ticket_lock_with_checks(kmp_ticket_lock_t *lck,
693  kmp_int32 gtid) {
694  char const *const func = "omp_test_lock";
695 
696  if (!std::atomic_load_explicit(&lck->lk.initialized,
697  std::memory_order_relaxed)) {
698  KMP_FATAL(LockIsUninitialized, func);
699  }
700  if (lck->lk.self != lck) {
701  KMP_FATAL(LockIsUninitialized, func);
702  }
703  if (__kmp_is_ticket_lock_nestable(lck)) {
704  KMP_FATAL(LockNestableUsedAsSimple, func);
705  }
706 
707  int retval = __kmp_test_ticket_lock(lck, gtid);
708 
709  if (retval) {
710  std::atomic_store_explicit(&lck->lk.owner_id, gtid + 1,
711  std::memory_order_relaxed);
712  }
713  return retval;
714 }
715 
716 int __kmp_release_ticket_lock(kmp_ticket_lock_t *lck, kmp_int32 gtid) {
717  kmp_uint32 distance = std::atomic_load_explicit(&lck->lk.next_ticket,
718  std::memory_order_relaxed) -
719  std::atomic_load_explicit(&lck->lk.now_serving,
720  std::memory_order_relaxed);
721 
722  ANNOTATE_TICKET_RELEASED(lck);
723  std::atomic_fetch_add_explicit(&lck->lk.now_serving, 1U,
724  std::memory_order_release);
725 
726  KMP_YIELD(distance >
727  (kmp_uint32)(__kmp_avail_proc ? __kmp_avail_proc : __kmp_xproc));
728  return KMP_LOCK_RELEASED;
729 }
730 
731 static int __kmp_release_ticket_lock_with_checks(kmp_ticket_lock_t *lck,
732  kmp_int32 gtid) {
733  char const *const func = "omp_unset_lock";
734 
735  if (!std::atomic_load_explicit(&lck->lk.initialized,
736  std::memory_order_relaxed)) {
737  KMP_FATAL(LockIsUninitialized, func);
738  }
739  if (lck->lk.self != lck) {
740  KMP_FATAL(LockIsUninitialized, func);
741  }
742  if (__kmp_is_ticket_lock_nestable(lck)) {
743  KMP_FATAL(LockNestableUsedAsSimple, func);
744  }
745  if (__kmp_get_ticket_lock_owner(lck) == -1) {
746  KMP_FATAL(LockUnsettingFree, func);
747  }
748  if ((gtid >= 0) && (__kmp_get_ticket_lock_owner(lck) >= 0) &&
749  (__kmp_get_ticket_lock_owner(lck) != gtid)) {
750  KMP_FATAL(LockUnsettingSetByAnother, func);
751  }
752  std::atomic_store_explicit(&lck->lk.owner_id, 0, std::memory_order_relaxed);
753  return __kmp_release_ticket_lock(lck, gtid);
754 }
755 
756 void __kmp_init_ticket_lock(kmp_ticket_lock_t *lck) {
757  lck->lk.location = NULL;
758  lck->lk.self = lck;
759  std::atomic_store_explicit(&lck->lk.next_ticket, 0U,
760  std::memory_order_relaxed);
761  std::atomic_store_explicit(&lck->lk.now_serving, 0U,
762  std::memory_order_relaxed);
763  std::atomic_store_explicit(
764  &lck->lk.owner_id, 0,
765  std::memory_order_relaxed); // no thread owns the lock.
766  std::atomic_store_explicit(
767  &lck->lk.depth_locked, -1,
768  std::memory_order_relaxed); // -1 => not a nested lock.
769  std::atomic_store_explicit(&lck->lk.initialized, true,
770  std::memory_order_release);
771 }
772 
773 void __kmp_destroy_ticket_lock(kmp_ticket_lock_t *lck) {
774  std::atomic_store_explicit(&lck->lk.initialized, false,
775  std::memory_order_release);
776  lck->lk.self = NULL;
777  lck->lk.location = NULL;
778  std::atomic_store_explicit(&lck->lk.next_ticket, 0U,
779  std::memory_order_relaxed);
780  std::atomic_store_explicit(&lck->lk.now_serving, 0U,
781  std::memory_order_relaxed);
782  std::atomic_store_explicit(&lck->lk.owner_id, 0, std::memory_order_relaxed);
783  std::atomic_store_explicit(&lck->lk.depth_locked, -1,
784  std::memory_order_relaxed);
785 }
786 
787 static void __kmp_destroy_ticket_lock_with_checks(kmp_ticket_lock_t *lck) {
788  char const *const func = "omp_destroy_lock";
789 
790  if (!std::atomic_load_explicit(&lck->lk.initialized,
791  std::memory_order_relaxed)) {
792  KMP_FATAL(LockIsUninitialized, func);
793  }
794  if (lck->lk.self != lck) {
795  KMP_FATAL(LockIsUninitialized, func);
796  }
797  if (__kmp_is_ticket_lock_nestable(lck)) {
798  KMP_FATAL(LockNestableUsedAsSimple, func);
799  }
800  if (__kmp_get_ticket_lock_owner(lck) != -1) {
801  KMP_FATAL(LockStillOwned, func);
802  }
803  __kmp_destroy_ticket_lock(lck);
804 }
805 
806 // nested ticket locks
807 
808 int __kmp_acquire_nested_ticket_lock(kmp_ticket_lock_t *lck, kmp_int32 gtid) {
809  KMP_DEBUG_ASSERT(gtid >= 0);
810 
811  if (__kmp_get_ticket_lock_owner(lck) == gtid) {
812  std::atomic_fetch_add_explicit(&lck->lk.depth_locked, 1,
813  std::memory_order_relaxed);
814  return KMP_LOCK_ACQUIRED_NEXT;
815  } else {
816  __kmp_acquire_ticket_lock_timed_template(lck, gtid);
817  ANNOTATE_TICKET_ACQUIRED(lck);
818  std::atomic_store_explicit(&lck->lk.depth_locked, 1,
819  std::memory_order_relaxed);
820  std::atomic_store_explicit(&lck->lk.owner_id, gtid + 1,
821  std::memory_order_relaxed);
822  return KMP_LOCK_ACQUIRED_FIRST;
823  }
824 }
825 
826 static int __kmp_acquire_nested_ticket_lock_with_checks(kmp_ticket_lock_t *lck,
827  kmp_int32 gtid) {
828  char const *const func = "omp_set_nest_lock";
829 
830  if (!std::atomic_load_explicit(&lck->lk.initialized,
831  std::memory_order_relaxed)) {
832  KMP_FATAL(LockIsUninitialized, func);
833  }
834  if (lck->lk.self != lck) {
835  KMP_FATAL(LockIsUninitialized, func);
836  }
837  if (!__kmp_is_ticket_lock_nestable(lck)) {
838  KMP_FATAL(LockSimpleUsedAsNestable, func);
839  }
840  return __kmp_acquire_nested_ticket_lock(lck, gtid);
841 }
842 
843 int __kmp_test_nested_ticket_lock(kmp_ticket_lock_t *lck, kmp_int32 gtid) {
844  int retval;
845 
846  KMP_DEBUG_ASSERT(gtid >= 0);
847 
848  if (__kmp_get_ticket_lock_owner(lck) == gtid) {
849  retval = std::atomic_fetch_add_explicit(&lck->lk.depth_locked, 1,
850  std::memory_order_relaxed) +
851  1;
852  } else if (!__kmp_test_ticket_lock(lck, gtid)) {
853  retval = 0;
854  } else {
855  std::atomic_store_explicit(&lck->lk.depth_locked, 1,
856  std::memory_order_relaxed);
857  std::atomic_store_explicit(&lck->lk.owner_id, gtid + 1,
858  std::memory_order_relaxed);
859  retval = 1;
860  }
861  return retval;
862 }
863 
864 static int __kmp_test_nested_ticket_lock_with_checks(kmp_ticket_lock_t *lck,
865  kmp_int32 gtid) {
866  char const *const func = "omp_test_nest_lock";
867 
868  if (!std::atomic_load_explicit(&lck->lk.initialized,
869  std::memory_order_relaxed)) {
870  KMP_FATAL(LockIsUninitialized, func);
871  }
872  if (lck->lk.self != lck) {
873  KMP_FATAL(LockIsUninitialized, func);
874  }
875  if (!__kmp_is_ticket_lock_nestable(lck)) {
876  KMP_FATAL(LockSimpleUsedAsNestable, func);
877  }
878  return __kmp_test_nested_ticket_lock(lck, gtid);
879 }
880 
881 int __kmp_release_nested_ticket_lock(kmp_ticket_lock_t *lck, kmp_int32 gtid) {
882  KMP_DEBUG_ASSERT(gtid >= 0);
883 
884  if ((std::atomic_fetch_add_explicit(&lck->lk.depth_locked, -1,
885  std::memory_order_relaxed) -
886  1) == 0) {
887  std::atomic_store_explicit(&lck->lk.owner_id, 0, std::memory_order_relaxed);
888  __kmp_release_ticket_lock(lck, gtid);
889  return KMP_LOCK_RELEASED;
890  }
891  return KMP_LOCK_STILL_HELD;
892 }
893 
894 static int __kmp_release_nested_ticket_lock_with_checks(kmp_ticket_lock_t *lck,
895  kmp_int32 gtid) {
896  char const *const func = "omp_unset_nest_lock";
897 
898  if (!std::atomic_load_explicit(&lck->lk.initialized,
899  std::memory_order_relaxed)) {
900  KMP_FATAL(LockIsUninitialized, func);
901  }
902  if (lck->lk.self != lck) {
903  KMP_FATAL(LockIsUninitialized, func);
904  }
905  if (!__kmp_is_ticket_lock_nestable(lck)) {
906  KMP_FATAL(LockSimpleUsedAsNestable, func);
907  }
908  if (__kmp_get_ticket_lock_owner(lck) == -1) {
909  KMP_FATAL(LockUnsettingFree, func);
910  }
911  if (__kmp_get_ticket_lock_owner(lck) != gtid) {
912  KMP_FATAL(LockUnsettingSetByAnother, func);
913  }
914  return __kmp_release_nested_ticket_lock(lck, gtid);
915 }
916 
917 void __kmp_init_nested_ticket_lock(kmp_ticket_lock_t *lck) {
918  __kmp_init_ticket_lock(lck);
919  std::atomic_store_explicit(&lck->lk.depth_locked, 0,
920  std::memory_order_relaxed);
921  // >= 0 for nestable locks, -1 for simple locks
922 }
923 
924 void __kmp_destroy_nested_ticket_lock(kmp_ticket_lock_t *lck) {
925  __kmp_destroy_ticket_lock(lck);
926  std::atomic_store_explicit(&lck->lk.depth_locked, 0,
927  std::memory_order_relaxed);
928 }
929 
930 static void
931 __kmp_destroy_nested_ticket_lock_with_checks(kmp_ticket_lock_t *lck) {
932  char const *const func = "omp_destroy_nest_lock";
933 
934  if (!std::atomic_load_explicit(&lck->lk.initialized,
935  std::memory_order_relaxed)) {
936  KMP_FATAL(LockIsUninitialized, func);
937  }
938  if (lck->lk.self != lck) {
939  KMP_FATAL(LockIsUninitialized, func);
940  }
941  if (!__kmp_is_ticket_lock_nestable(lck)) {
942  KMP_FATAL(LockSimpleUsedAsNestable, func);
943  }
944  if (__kmp_get_ticket_lock_owner(lck) != -1) {
945  KMP_FATAL(LockStillOwned, func);
946  }
947  __kmp_destroy_nested_ticket_lock(lck);
948 }
949 
950 // access functions to fields which don't exist for all lock kinds.
951 
952 static const ident_t *__kmp_get_ticket_lock_location(kmp_ticket_lock_t *lck) {
953  return lck->lk.location;
954 }
955 
956 static void __kmp_set_ticket_lock_location(kmp_ticket_lock_t *lck,
957  const ident_t *loc) {
958  lck->lk.location = loc;
959 }
960 
961 static kmp_lock_flags_t __kmp_get_ticket_lock_flags(kmp_ticket_lock_t *lck) {
962  return lck->lk.flags;
963 }
964 
965 static void __kmp_set_ticket_lock_flags(kmp_ticket_lock_t *lck,
966  kmp_lock_flags_t flags) {
967  lck->lk.flags = flags;
968 }
969 
970 /* ------------------------------------------------------------------------ */
971 /* queuing locks */
972 
973 /* First the states
974  (head,tail) = 0, 0 means lock is unheld, nobody on queue
975  UINT_MAX or -1, 0 means lock is held, nobody on queue
976  h, h means lock held or about to transition,
977  1 element on queue
978  h, t h <> t, means lock is held or about to
979  transition, >1 elements on queue
980 
981  Now the transitions
982  Acquire(0,0) = -1 ,0
983  Release(0,0) = Error
984  Acquire(-1,0) = h ,h h > 0
985  Release(-1,0) = 0 ,0
986  Acquire(h,h) = h ,t h > 0, t > 0, h <> t
987  Release(h,h) = -1 ,0 h > 0
988  Acquire(h,t) = h ,t' h > 0, t > 0, t' > 0, h <> t, h <> t', t <> t'
989  Release(h,t) = h',t h > 0, t > 0, h <> t, h <> h', h' maybe = t
990 
991  And pictorially
992 
993  +-----+
994  | 0, 0|------- release -------> Error
995  +-----+
996  | ^
997  acquire| |release
998  | |
999  | |
1000  v |
1001  +-----+
1002  |-1, 0|
1003  +-----+
1004  | ^
1005  acquire| |release
1006  | |
1007  | |
1008  v |
1009  +-----+
1010  | h, h|
1011  +-----+
1012  | ^
1013  acquire| |release
1014  | |
1015  | |
1016  v |
1017  +-----+
1018  | h, t|----- acquire, release loopback ---+
1019  +-----+ |
1020  ^ |
1021  | |
1022  +------------------------------------+
1023  */
1024 
1025 #ifdef DEBUG_QUEUING_LOCKS
1026 
1027 /* Stuff for circular trace buffer */
1028 #define TRACE_BUF_ELE 1024
1029 static char traces[TRACE_BUF_ELE][128] = {0};
1030 static int tc = 0;
1031 #define TRACE_LOCK(X, Y) \
1032  KMP_SNPRINTF(traces[tc++ % TRACE_BUF_ELE], 128, "t%d at %s\n", X, Y);
1033 #define TRACE_LOCK_T(X, Y, Z) \
1034  KMP_SNPRINTF(traces[tc++ % TRACE_BUF_ELE], 128, "t%d at %s%d\n", X, Y, Z);
1035 #define TRACE_LOCK_HT(X, Y, Z, Q) \
1036  KMP_SNPRINTF(traces[tc++ % TRACE_BUF_ELE], 128, "t%d at %s %d,%d\n", X, Y, \
1037  Z, Q);
1038 
1039 static void __kmp_dump_queuing_lock(kmp_info_t *this_thr, kmp_int32 gtid,
1040  kmp_queuing_lock_t *lck, kmp_int32 head_id,
1041  kmp_int32 tail_id) {
1042  kmp_int32 t, i;
1043 
1044  __kmp_printf_no_lock("\n__kmp_dump_queuing_lock: TRACE BEGINS HERE! \n");
1045 
1046  i = tc % TRACE_BUF_ELE;
1047  __kmp_printf_no_lock("%s\n", traces[i]);
1048  i = (i + 1) % TRACE_BUF_ELE;
1049  while (i != (tc % TRACE_BUF_ELE)) {
1050  __kmp_printf_no_lock("%s", traces[i]);
1051  i = (i + 1) % TRACE_BUF_ELE;
1052  }
1053  __kmp_printf_no_lock("\n");
1054 
1055  __kmp_printf_no_lock("\n__kmp_dump_queuing_lock: gtid+1:%d, spin_here:%d, "
1056  "next_wait:%d, head_id:%d, tail_id:%d\n",
1057  gtid + 1, this_thr->th.th_spin_here,
1058  this_thr->th.th_next_waiting, head_id, tail_id);
1059 
1060  __kmp_printf_no_lock("\t\thead: %d ", lck->lk.head_id);
1061 
1062  if (lck->lk.head_id >= 1) {
1063  t = __kmp_threads[lck->lk.head_id - 1]->th.th_next_waiting;
1064  while (t > 0) {
1065  __kmp_printf_no_lock("-> %d ", t);
1066  t = __kmp_threads[t - 1]->th.th_next_waiting;
1067  }
1068  }
1069  __kmp_printf_no_lock("; tail: %d ", lck->lk.tail_id);
1070  __kmp_printf_no_lock("\n\n");
1071 }
1072 
1073 #endif /* DEBUG_QUEUING_LOCKS */
1074 
1075 static kmp_int32 __kmp_get_queuing_lock_owner(kmp_queuing_lock_t *lck) {
1076  return TCR_4(lck->lk.owner_id) - 1;
1077 }
1078 
1079 static inline bool __kmp_is_queuing_lock_nestable(kmp_queuing_lock_t *lck) {
1080  return lck->lk.depth_locked != -1;
1081 }
1082 
1083 /* Acquire a lock using a the queuing lock implementation */
1084 template <bool takeTime>
1085 /* [TLW] The unused template above is left behind because of what BEB believes
1086  is a potential compiler problem with __forceinline. */
1087 __forceinline static int
1088 __kmp_acquire_queuing_lock_timed_template(kmp_queuing_lock_t *lck,
1089  kmp_int32 gtid) {
1090  kmp_info_t *this_thr = __kmp_thread_from_gtid(gtid);
1091  volatile kmp_int32 *head_id_p = &lck->lk.head_id;
1092  volatile kmp_int32 *tail_id_p = &lck->lk.tail_id;
1093  volatile kmp_uint32 *spin_here_p;
1094  kmp_int32 need_mf = 1;
1095 
1096 #if OMPT_SUPPORT
1097  ompt_state_t prev_state = ompt_state_undefined;
1098 #endif
1099 
1100  KA_TRACE(1000,
1101  ("__kmp_acquire_queuing_lock: lck:%p, T#%d entering\n", lck, gtid));
1102 
1103  KMP_FSYNC_PREPARE(lck);
1104  KMP_DEBUG_ASSERT(this_thr != NULL);
1105  spin_here_p = &this_thr->th.th_spin_here;
1106 
1107 #ifdef DEBUG_QUEUING_LOCKS
1108  TRACE_LOCK(gtid + 1, "acq ent");
1109  if (*spin_here_p)
1110  __kmp_dump_queuing_lock(this_thr, gtid, lck, *head_id_p, *tail_id_p);
1111  if (this_thr->th.th_next_waiting != 0)
1112  __kmp_dump_queuing_lock(this_thr, gtid, lck, *head_id_p, *tail_id_p);
1113 #endif
1114  KMP_DEBUG_ASSERT(!*spin_here_p);
1115  KMP_DEBUG_ASSERT(this_thr->th.th_next_waiting == 0);
1116 
1117  /* The following st.rel to spin_here_p needs to precede the cmpxchg.acq to
1118  head_id_p that may follow, not just in execution order, but also in
1119  visibility order. This way, when a releasing thread observes the changes to
1120  the queue by this thread, it can rightly assume that spin_here_p has
1121  already been set to TRUE, so that when it sets spin_here_p to FALSE, it is
1122  not premature. If the releasing thread sets spin_here_p to FALSE before
1123  this thread sets it to TRUE, this thread will hang. */
1124  *spin_here_p = TRUE; /* before enqueuing to prevent race */
1125 
1126  while (1) {
1127  kmp_int32 enqueued;
1128  kmp_int32 head;
1129  kmp_int32 tail;
1130 
1131  head = *head_id_p;
1132 
1133  switch (head) {
1134 
1135  case -1: {
1136 #ifdef DEBUG_QUEUING_LOCKS
1137  tail = *tail_id_p;
1138  TRACE_LOCK_HT(gtid + 1, "acq read: ", head, tail);
1139 #endif
1140  tail = 0; /* to make sure next link asynchronously read is not set
1141  accidentally; this assignment prevents us from entering the
1142  if ( t > 0 ) condition in the enqueued case below, which is not
1143  necessary for this state transition */
1144 
1145  need_mf = 0;
1146  /* try (-1,0)->(tid,tid) */
1147  enqueued = KMP_COMPARE_AND_STORE_ACQ64((volatile kmp_int64 *)tail_id_p,
1148  KMP_PACK_64(-1, 0),
1149  KMP_PACK_64(gtid + 1, gtid + 1));
1150 #ifdef DEBUG_QUEUING_LOCKS
1151  if (enqueued)
1152  TRACE_LOCK(gtid + 1, "acq enq: (-1,0)->(tid,tid)");
1153 #endif
1154  } break;
1155 
1156  default: {
1157  tail = *tail_id_p;
1158  KMP_DEBUG_ASSERT(tail != gtid + 1);
1159 
1160 #ifdef DEBUG_QUEUING_LOCKS
1161  TRACE_LOCK_HT(gtid + 1, "acq read: ", head, tail);
1162 #endif
1163 
1164  if (tail == 0) {
1165  enqueued = FALSE;
1166  } else {
1167  need_mf = 0;
1168  /* try (h,t) or (h,h)->(h,tid) */
1169  enqueued = KMP_COMPARE_AND_STORE_ACQ32(tail_id_p, tail, gtid + 1);
1170 
1171 #ifdef DEBUG_QUEUING_LOCKS
1172  if (enqueued)
1173  TRACE_LOCK(gtid + 1, "acq enq: (h,t)->(h,tid)");
1174 #endif
1175  }
1176  } break;
1177 
1178  case 0: /* empty queue */
1179  {
1180  kmp_int32 grabbed_lock;
1181 
1182 #ifdef DEBUG_QUEUING_LOCKS
1183  tail = *tail_id_p;
1184  TRACE_LOCK_HT(gtid + 1, "acq read: ", head, tail);
1185 #endif
1186  /* try (0,0)->(-1,0) */
1187 
1188  /* only legal transition out of head = 0 is head = -1 with no change to
1189  * tail */
1190  grabbed_lock = KMP_COMPARE_AND_STORE_ACQ32(head_id_p, 0, -1);
1191 
1192  if (grabbed_lock) {
1193 
1194  *spin_here_p = FALSE;
1195 
1196  KA_TRACE(
1197  1000,
1198  ("__kmp_acquire_queuing_lock: lck:%p, T#%d exiting: no queuing\n",
1199  lck, gtid));
1200 #ifdef DEBUG_QUEUING_LOCKS
1201  TRACE_LOCK_HT(gtid + 1, "acq exit: ", head, 0);
1202 #endif
1203 
1204 #if OMPT_SUPPORT
1205  if (ompt_enabled.enabled && prev_state != ompt_state_undefined) {
1206  /* change the state before clearing wait_id */
1207  this_thr->th.ompt_thread_info.state = prev_state;
1208  this_thr->th.ompt_thread_info.wait_id = 0;
1209  }
1210 #endif
1211 
1212  KMP_FSYNC_ACQUIRED(lck);
1213  return KMP_LOCK_ACQUIRED_FIRST; /* lock holder cannot be on queue */
1214  }
1215  enqueued = FALSE;
1216  } break;
1217  }
1218 
1219 #if OMPT_SUPPORT
1220  if (ompt_enabled.enabled && prev_state == ompt_state_undefined) {
1221  /* this thread will spin; set wait_id before entering wait state */
1222  prev_state = this_thr->th.ompt_thread_info.state;
1223  this_thr->th.ompt_thread_info.wait_id = (uint64_t)lck;
1224  this_thr->th.ompt_thread_info.state = ompt_state_wait_lock;
1225  }
1226 #endif
1227 
1228  if (enqueued) {
1229  if (tail > 0) {
1230  kmp_info_t *tail_thr = __kmp_thread_from_gtid(tail - 1);
1231  KMP_ASSERT(tail_thr != NULL);
1232  tail_thr->th.th_next_waiting = gtid + 1;
1233  /* corresponding wait for this write in release code */
1234  }
1235  KA_TRACE(1000,
1236  ("__kmp_acquire_queuing_lock: lck:%p, T#%d waiting for lock\n",
1237  lck, gtid));
1238 
1239  KMP_MB();
1240  // ToDo: Use __kmp_wait_sleep or similar when blocktime != inf
1241  KMP_WAIT(spin_here_p, FALSE, KMP_EQ, lck);
1242  // Synchronize writes to both runtime thread structures
1243  // and writes in user code.
1244  KMP_MB();
1245 
1246 #ifdef DEBUG_QUEUING_LOCKS
1247  TRACE_LOCK(gtid + 1, "acq spin");
1248 
1249  if (this_thr->th.th_next_waiting != 0)
1250  __kmp_dump_queuing_lock(this_thr, gtid, lck, *head_id_p, *tail_id_p);
1251 #endif
1252  KMP_DEBUG_ASSERT(this_thr->th.th_next_waiting == 0);
1253  KA_TRACE(1000, ("__kmp_acquire_queuing_lock: lck:%p, T#%d exiting: after "
1254  "waiting on queue\n",
1255  lck, gtid));
1256 
1257 #ifdef DEBUG_QUEUING_LOCKS
1258  TRACE_LOCK(gtid + 1, "acq exit 2");
1259 #endif
1260 
1261 #if OMPT_SUPPORT
1262  /* change the state before clearing wait_id */
1263  this_thr->th.ompt_thread_info.state = prev_state;
1264  this_thr->th.ompt_thread_info.wait_id = 0;
1265 #endif
1266 
1267  /* got lock, we were dequeued by the thread that released lock */
1268  return KMP_LOCK_ACQUIRED_FIRST;
1269  }
1270 
1271  /* Yield if number of threads > number of logical processors */
1272  /* ToDo: Not sure why this should only be in oversubscription case,
1273  maybe should be traditional YIELD_INIT/YIELD_WHEN loop */
1274  KMP_YIELD_OVERSUB();
1275 
1276 #ifdef DEBUG_QUEUING_LOCKS
1277  TRACE_LOCK(gtid + 1, "acq retry");
1278 #endif
1279  }
1280  KMP_ASSERT2(0, "should not get here");
1281  return KMP_LOCK_ACQUIRED_FIRST;
1282 }
1283 
1284 int __kmp_acquire_queuing_lock(kmp_queuing_lock_t *lck, kmp_int32 gtid) {
1285  KMP_DEBUG_ASSERT(gtid >= 0);
1286 
1287  int retval = __kmp_acquire_queuing_lock_timed_template<false>(lck, gtid);
1288  ANNOTATE_QUEUING_ACQUIRED(lck);
1289  return retval;
1290 }
1291 
1292 static int __kmp_acquire_queuing_lock_with_checks(kmp_queuing_lock_t *lck,
1293  kmp_int32 gtid) {
1294  char const *const func = "omp_set_lock";
1295  if (lck->lk.initialized != lck) {
1296  KMP_FATAL(LockIsUninitialized, func);
1297  }
1298  if (__kmp_is_queuing_lock_nestable(lck)) {
1299  KMP_FATAL(LockNestableUsedAsSimple, func);
1300  }
1301  if (__kmp_get_queuing_lock_owner(lck) == gtid) {
1302  KMP_FATAL(LockIsAlreadyOwned, func);
1303  }
1304 
1305  __kmp_acquire_queuing_lock(lck, gtid);
1306 
1307  lck->lk.owner_id = gtid + 1;
1308  return KMP_LOCK_ACQUIRED_FIRST;
1309 }
1310 
1311 int __kmp_test_queuing_lock(kmp_queuing_lock_t *lck, kmp_int32 gtid) {
1312  volatile kmp_int32 *head_id_p = &lck->lk.head_id;
1313  kmp_int32 head;
1314 #ifdef KMP_DEBUG
1315  kmp_info_t *this_thr;
1316 #endif
1317 
1318  KA_TRACE(1000, ("__kmp_test_queuing_lock: T#%d entering\n", gtid));
1319  KMP_DEBUG_ASSERT(gtid >= 0);
1320 #ifdef KMP_DEBUG
1321  this_thr = __kmp_thread_from_gtid(gtid);
1322  KMP_DEBUG_ASSERT(this_thr != NULL);
1323  KMP_DEBUG_ASSERT(!this_thr->th.th_spin_here);
1324 #endif
1325 
1326  head = *head_id_p;
1327 
1328  if (head == 0) { /* nobody on queue, nobody holding */
1329  /* try (0,0)->(-1,0) */
1330  if (KMP_COMPARE_AND_STORE_ACQ32(head_id_p, 0, -1)) {
1331  KA_TRACE(1000,
1332  ("__kmp_test_queuing_lock: T#%d exiting: holding lock\n", gtid));
1333  KMP_FSYNC_ACQUIRED(lck);
1334  ANNOTATE_QUEUING_ACQUIRED(lck);
1335  return TRUE;
1336  }
1337  }
1338 
1339  KA_TRACE(1000,
1340  ("__kmp_test_queuing_lock: T#%d exiting: without lock\n", gtid));
1341  return FALSE;
1342 }
1343 
1344 static int __kmp_test_queuing_lock_with_checks(kmp_queuing_lock_t *lck,
1345  kmp_int32 gtid) {
1346  char const *const func = "omp_test_lock";
1347  if (lck->lk.initialized != lck) {
1348  KMP_FATAL(LockIsUninitialized, func);
1349  }
1350  if (__kmp_is_queuing_lock_nestable(lck)) {
1351  KMP_FATAL(LockNestableUsedAsSimple, func);
1352  }
1353 
1354  int retval = __kmp_test_queuing_lock(lck, gtid);
1355 
1356  if (retval) {
1357  lck->lk.owner_id = gtid + 1;
1358  }
1359  return retval;
1360 }
1361 
1362 int __kmp_release_queuing_lock(kmp_queuing_lock_t *lck, kmp_int32 gtid) {
1363  kmp_info_t *this_thr;
1364  volatile kmp_int32 *head_id_p = &lck->lk.head_id;
1365  volatile kmp_int32 *tail_id_p = &lck->lk.tail_id;
1366 
1367  KA_TRACE(1000,
1368  ("__kmp_release_queuing_lock: lck:%p, T#%d entering\n", lck, gtid));
1369  KMP_DEBUG_ASSERT(gtid >= 0);
1370  this_thr = __kmp_thread_from_gtid(gtid);
1371  KMP_DEBUG_ASSERT(this_thr != NULL);
1372 #ifdef DEBUG_QUEUING_LOCKS
1373  TRACE_LOCK(gtid + 1, "rel ent");
1374 
1375  if (this_thr->th.th_spin_here)
1376  __kmp_dump_queuing_lock(this_thr, gtid, lck, *head_id_p, *tail_id_p);
1377  if (this_thr->th.th_next_waiting != 0)
1378  __kmp_dump_queuing_lock(this_thr, gtid, lck, *head_id_p, *tail_id_p);
1379 #endif
1380  KMP_DEBUG_ASSERT(!this_thr->th.th_spin_here);
1381  KMP_DEBUG_ASSERT(this_thr->th.th_next_waiting == 0);
1382 
1383  KMP_FSYNC_RELEASING(lck);
1384  ANNOTATE_QUEUING_RELEASED(lck);
1385 
1386  while (1) {
1387  kmp_int32 dequeued;
1388  kmp_int32 head;
1389  kmp_int32 tail;
1390 
1391  head = *head_id_p;
1392 
1393 #ifdef DEBUG_QUEUING_LOCKS
1394  tail = *tail_id_p;
1395  TRACE_LOCK_HT(gtid + 1, "rel read: ", head, tail);
1396  if (head == 0)
1397  __kmp_dump_queuing_lock(this_thr, gtid, lck, head, tail);
1398 #endif
1399  KMP_DEBUG_ASSERT(head !=
1400  0); /* holding the lock, head must be -1 or queue head */
1401 
1402  if (head == -1) { /* nobody on queue */
1403  /* try (-1,0)->(0,0) */
1404  if (KMP_COMPARE_AND_STORE_REL32(head_id_p, -1, 0)) {
1405  KA_TRACE(
1406  1000,
1407  ("__kmp_release_queuing_lock: lck:%p, T#%d exiting: queue empty\n",
1408  lck, gtid));
1409 #ifdef DEBUG_QUEUING_LOCKS
1410  TRACE_LOCK_HT(gtid + 1, "rel exit: ", 0, 0);
1411 #endif
1412 
1413 #if OMPT_SUPPORT
1414 /* nothing to do - no other thread is trying to shift blame */
1415 #endif
1416  return KMP_LOCK_RELEASED;
1417  }
1418  dequeued = FALSE;
1419  } else {
1420  KMP_MB();
1421  tail = *tail_id_p;
1422  if (head == tail) { /* only one thread on the queue */
1423 #ifdef DEBUG_QUEUING_LOCKS
1424  if (head <= 0)
1425  __kmp_dump_queuing_lock(this_thr, gtid, lck, head, tail);
1426 #endif
1427  KMP_DEBUG_ASSERT(head > 0);
1428 
1429  /* try (h,h)->(-1,0) */
1430  dequeued = KMP_COMPARE_AND_STORE_REL64(
1431  RCAST(volatile kmp_int64 *, tail_id_p), KMP_PACK_64(head, head),
1432  KMP_PACK_64(-1, 0));
1433 #ifdef DEBUG_QUEUING_LOCKS
1434  TRACE_LOCK(gtid + 1, "rel deq: (h,h)->(-1,0)");
1435 #endif
1436 
1437  } else {
1438  volatile kmp_int32 *waiting_id_p;
1439  kmp_info_t *head_thr = __kmp_thread_from_gtid(head - 1);
1440  KMP_DEBUG_ASSERT(head_thr != NULL);
1441  waiting_id_p = &head_thr->th.th_next_waiting;
1442 
1443 /* Does this require synchronous reads? */
1444 #ifdef DEBUG_QUEUING_LOCKS
1445  if (head <= 0 || tail <= 0)
1446  __kmp_dump_queuing_lock(this_thr, gtid, lck, head, tail);
1447 #endif
1448  KMP_DEBUG_ASSERT(head > 0 && tail > 0);
1449 
1450  /* try (h,t)->(h',t) or (t,t) */
1451  KMP_MB();
1452  /* make sure enqueuing thread has time to update next waiting thread
1453  * field */
1454  *head_id_p =
1455  KMP_WAIT((volatile kmp_uint32 *)waiting_id_p, 0, KMP_NEQ, NULL);
1456 #ifdef DEBUG_QUEUING_LOCKS
1457  TRACE_LOCK(gtid + 1, "rel deq: (h,t)->(h',t)");
1458 #endif
1459  dequeued = TRUE;
1460  }
1461  }
1462 
1463  if (dequeued) {
1464  kmp_info_t *head_thr = __kmp_thread_from_gtid(head - 1);
1465  KMP_DEBUG_ASSERT(head_thr != NULL);
1466 
1467 /* Does this require synchronous reads? */
1468 #ifdef DEBUG_QUEUING_LOCKS
1469  if (head <= 0 || tail <= 0)
1470  __kmp_dump_queuing_lock(this_thr, gtid, lck, head, tail);
1471 #endif
1472  KMP_DEBUG_ASSERT(head > 0 && tail > 0);
1473 
1474  /* For clean code only. Thread not released until next statement prevents
1475  race with acquire code. */
1476  head_thr->th.th_next_waiting = 0;
1477 #ifdef DEBUG_QUEUING_LOCKS
1478  TRACE_LOCK_T(gtid + 1, "rel nw=0 for t=", head);
1479 #endif
1480 
1481  KMP_MB();
1482  /* reset spin value */
1483  head_thr->th.th_spin_here = FALSE;
1484 
1485  KA_TRACE(1000, ("__kmp_release_queuing_lock: lck:%p, T#%d exiting: after "
1486  "dequeuing\n",
1487  lck, gtid));
1488 #ifdef DEBUG_QUEUING_LOCKS
1489  TRACE_LOCK(gtid + 1, "rel exit 2");
1490 #endif
1491  return KMP_LOCK_RELEASED;
1492  }
1493  /* KMP_CPU_PAUSE(); don't want to make releasing thread hold up acquiring
1494  threads */
1495 
1496 #ifdef DEBUG_QUEUING_LOCKS
1497  TRACE_LOCK(gtid + 1, "rel retry");
1498 #endif
1499 
1500  } /* while */
1501  KMP_ASSERT2(0, "should not get here");
1502  return KMP_LOCK_RELEASED;
1503 }
1504 
1505 static int __kmp_release_queuing_lock_with_checks(kmp_queuing_lock_t *lck,
1506  kmp_int32 gtid) {
1507  char const *const func = "omp_unset_lock";
1508  KMP_MB(); /* in case another processor initialized lock */
1509  if (lck->lk.initialized != lck) {
1510  KMP_FATAL(LockIsUninitialized, func);
1511  }
1512  if (__kmp_is_queuing_lock_nestable(lck)) {
1513  KMP_FATAL(LockNestableUsedAsSimple, func);
1514  }
1515  if (__kmp_get_queuing_lock_owner(lck) == -1) {
1516  KMP_FATAL(LockUnsettingFree, func);
1517  }
1518  if (__kmp_get_queuing_lock_owner(lck) != gtid) {
1519  KMP_FATAL(LockUnsettingSetByAnother, func);
1520  }
1521  lck->lk.owner_id = 0;
1522  return __kmp_release_queuing_lock(lck, gtid);
1523 }
1524 
1525 void __kmp_init_queuing_lock(kmp_queuing_lock_t *lck) {
1526  lck->lk.location = NULL;
1527  lck->lk.head_id = 0;
1528  lck->lk.tail_id = 0;
1529  lck->lk.next_ticket = 0;
1530  lck->lk.now_serving = 0;
1531  lck->lk.owner_id = 0; // no thread owns the lock.
1532  lck->lk.depth_locked = -1; // >= 0 for nestable locks, -1 for simple locks.
1533  lck->lk.initialized = lck;
1534 
1535  KA_TRACE(1000, ("__kmp_init_queuing_lock: lock %p initialized\n", lck));
1536 }
1537 
1538 void __kmp_destroy_queuing_lock(kmp_queuing_lock_t *lck) {
1539  lck->lk.initialized = NULL;
1540  lck->lk.location = NULL;
1541  lck->lk.head_id = 0;
1542  lck->lk.tail_id = 0;
1543  lck->lk.next_ticket = 0;
1544  lck->lk.now_serving = 0;
1545  lck->lk.owner_id = 0;
1546  lck->lk.depth_locked = -1;
1547 }
1548 
1549 static void __kmp_destroy_queuing_lock_with_checks(kmp_queuing_lock_t *lck) {
1550  char const *const func = "omp_destroy_lock";
1551  if (lck->lk.initialized != lck) {
1552  KMP_FATAL(LockIsUninitialized, func);
1553  }
1554  if (__kmp_is_queuing_lock_nestable(lck)) {
1555  KMP_FATAL(LockNestableUsedAsSimple, func);
1556  }
1557  if (__kmp_get_queuing_lock_owner(lck) != -1) {
1558  KMP_FATAL(LockStillOwned, func);
1559  }
1560  __kmp_destroy_queuing_lock(lck);
1561 }
1562 
1563 // nested queuing locks
1564 
1565 int __kmp_acquire_nested_queuing_lock(kmp_queuing_lock_t *lck, kmp_int32 gtid) {
1566  KMP_DEBUG_ASSERT(gtid >= 0);
1567 
1568  if (__kmp_get_queuing_lock_owner(lck) == gtid) {
1569  lck->lk.depth_locked += 1;
1570  return KMP_LOCK_ACQUIRED_NEXT;
1571  } else {
1572  __kmp_acquire_queuing_lock_timed_template<false>(lck, gtid);
1573  ANNOTATE_QUEUING_ACQUIRED(lck);
1574  KMP_MB();
1575  lck->lk.depth_locked = 1;
1576  KMP_MB();
1577  lck->lk.owner_id = gtid + 1;
1578  return KMP_LOCK_ACQUIRED_FIRST;
1579  }
1580 }
1581 
1582 static int
1583 __kmp_acquire_nested_queuing_lock_with_checks(kmp_queuing_lock_t *lck,
1584  kmp_int32 gtid) {
1585  char const *const func = "omp_set_nest_lock";
1586  if (lck->lk.initialized != lck) {
1587  KMP_FATAL(LockIsUninitialized, func);
1588  }
1589  if (!__kmp_is_queuing_lock_nestable(lck)) {
1590  KMP_FATAL(LockSimpleUsedAsNestable, func);
1591  }
1592  return __kmp_acquire_nested_queuing_lock(lck, gtid);
1593 }
1594 
1595 int __kmp_test_nested_queuing_lock(kmp_queuing_lock_t *lck, kmp_int32 gtid) {
1596  int retval;
1597 
1598  KMP_DEBUG_ASSERT(gtid >= 0);
1599 
1600  if (__kmp_get_queuing_lock_owner(lck) == gtid) {
1601  retval = ++lck->lk.depth_locked;
1602  } else if (!__kmp_test_queuing_lock(lck, gtid)) {
1603  retval = 0;
1604  } else {
1605  KMP_MB();
1606  retval = lck->lk.depth_locked = 1;
1607  KMP_MB();
1608  lck->lk.owner_id = gtid + 1;
1609  }
1610  return retval;
1611 }
1612 
1613 static int __kmp_test_nested_queuing_lock_with_checks(kmp_queuing_lock_t *lck,
1614  kmp_int32 gtid) {
1615  char const *const func = "omp_test_nest_lock";
1616  if (lck->lk.initialized != lck) {
1617  KMP_FATAL(LockIsUninitialized, func);
1618  }
1619  if (!__kmp_is_queuing_lock_nestable(lck)) {
1620  KMP_FATAL(LockSimpleUsedAsNestable, func);
1621  }
1622  return __kmp_test_nested_queuing_lock(lck, gtid);
1623 }
1624 
1625 int __kmp_release_nested_queuing_lock(kmp_queuing_lock_t *lck, kmp_int32 gtid) {
1626  KMP_DEBUG_ASSERT(gtid >= 0);
1627 
1628  KMP_MB();
1629  if (--(lck->lk.depth_locked) == 0) {
1630  KMP_MB();
1631  lck->lk.owner_id = 0;
1632  __kmp_release_queuing_lock(lck, gtid);
1633  return KMP_LOCK_RELEASED;
1634  }
1635  return KMP_LOCK_STILL_HELD;
1636 }
1637 
1638 static int
1639 __kmp_release_nested_queuing_lock_with_checks(kmp_queuing_lock_t *lck,
1640  kmp_int32 gtid) {
1641  char const *const func = "omp_unset_nest_lock";
1642  KMP_MB(); /* in case another processor initialized lock */
1643  if (lck->lk.initialized != lck) {
1644  KMP_FATAL(LockIsUninitialized, func);
1645  }
1646  if (!__kmp_is_queuing_lock_nestable(lck)) {
1647  KMP_FATAL(LockSimpleUsedAsNestable, func);
1648  }
1649  if (__kmp_get_queuing_lock_owner(lck) == -1) {
1650  KMP_FATAL(LockUnsettingFree, func);
1651  }
1652  if (__kmp_get_queuing_lock_owner(lck) != gtid) {
1653  KMP_FATAL(LockUnsettingSetByAnother, func);
1654  }
1655  return __kmp_release_nested_queuing_lock(lck, gtid);
1656 }
1657 
1658 void __kmp_init_nested_queuing_lock(kmp_queuing_lock_t *lck) {
1659  __kmp_init_queuing_lock(lck);
1660  lck->lk.depth_locked = 0; // >= 0 for nestable locks, -1 for simple locks
1661 }
1662 
1663 void __kmp_destroy_nested_queuing_lock(kmp_queuing_lock_t *lck) {
1664  __kmp_destroy_queuing_lock(lck);
1665  lck->lk.depth_locked = 0;
1666 }
1667 
1668 static void
1669 __kmp_destroy_nested_queuing_lock_with_checks(kmp_queuing_lock_t *lck) {
1670  char const *const func = "omp_destroy_nest_lock";
1671  if (lck->lk.initialized != lck) {
1672  KMP_FATAL(LockIsUninitialized, func);
1673  }
1674  if (!__kmp_is_queuing_lock_nestable(lck)) {
1675  KMP_FATAL(LockSimpleUsedAsNestable, func);
1676  }
1677  if (__kmp_get_queuing_lock_owner(lck) != -1) {
1678  KMP_FATAL(LockStillOwned, func);
1679  }
1680  __kmp_destroy_nested_queuing_lock(lck);
1681 }
1682 
1683 // access functions to fields which don't exist for all lock kinds.
1684 
1685 static const ident_t *__kmp_get_queuing_lock_location(kmp_queuing_lock_t *lck) {
1686  return lck->lk.location;
1687 }
1688 
1689 static void __kmp_set_queuing_lock_location(kmp_queuing_lock_t *lck,
1690  const ident_t *loc) {
1691  lck->lk.location = loc;
1692 }
1693 
1694 static kmp_lock_flags_t __kmp_get_queuing_lock_flags(kmp_queuing_lock_t *lck) {
1695  return lck->lk.flags;
1696 }
1697 
1698 static void __kmp_set_queuing_lock_flags(kmp_queuing_lock_t *lck,
1699  kmp_lock_flags_t flags) {
1700  lck->lk.flags = flags;
1701 }
1702 
1703 #if KMP_USE_ADAPTIVE_LOCKS
1704 
1705 /* RTM Adaptive locks */
1706 
1707 #if KMP_HAVE_RTM_INTRINSICS
1708 #include <immintrin.h>
1709 #define SOFT_ABORT_MASK (_XABORT_RETRY | _XABORT_CONFLICT | _XABORT_EXPLICIT)
1710 
1711 #else
1712 
1713 // Values from the status register after failed speculation.
1714 #define _XBEGIN_STARTED (~0u)
1715 #define _XABORT_EXPLICIT (1 << 0)
1716 #define _XABORT_RETRY (1 << 1)
1717 #define _XABORT_CONFLICT (1 << 2)
1718 #define _XABORT_CAPACITY (1 << 3)
1719 #define _XABORT_DEBUG (1 << 4)
1720 #define _XABORT_NESTED (1 << 5)
1721 #define _XABORT_CODE(x) ((unsigned char)(((x) >> 24) & 0xFF))
1722 
1723 // Aborts for which it's worth trying again immediately
1724 #define SOFT_ABORT_MASK (_XABORT_RETRY | _XABORT_CONFLICT | _XABORT_EXPLICIT)
1725 
1726 #define STRINGIZE_INTERNAL(arg) #arg
1727 #define STRINGIZE(arg) STRINGIZE_INTERNAL(arg)
1728 
1729 // Access to RTM instructions
1730 /*A version of XBegin which returns -1 on speculation, and the value of EAX on
1731  an abort. This is the same definition as the compiler intrinsic that will be
1732  supported at some point. */
1733 static __inline int _xbegin() {
1734  int res = -1;
1735 
1736 #if KMP_OS_WINDOWS
1737 #if KMP_ARCH_X86_64
1738  _asm {
1739  _emit 0xC7
1740  _emit 0xF8
1741  _emit 2
1742  _emit 0
1743  _emit 0
1744  _emit 0
1745  jmp L2
1746  mov res, eax
1747  L2:
1748  }
1749 #else /* IA32 */
1750  _asm {
1751  _emit 0xC7
1752  _emit 0xF8
1753  _emit 2
1754  _emit 0
1755  _emit 0
1756  _emit 0
1757  jmp L2
1758  mov res, eax
1759  L2:
1760  }
1761 #endif // KMP_ARCH_X86_64
1762 #else
1763  /* Note that %eax must be noted as killed (clobbered), because the XSR is
1764  returned in %eax(%rax) on abort. Other register values are restored, so
1765  don't need to be killed.
1766 
1767  We must also mark 'res' as an input and an output, since otherwise
1768  'res=-1' may be dropped as being dead, whereas we do need the assignment on
1769  the successful (i.e., non-abort) path. */
1770  __asm__ volatile("1: .byte 0xC7; .byte 0xF8;\n"
1771  " .long 1f-1b-6\n"
1772  " jmp 2f\n"
1773  "1: movl %%eax,%0\n"
1774  "2:"
1775  : "+r"(res)::"memory", "%eax");
1776 #endif // KMP_OS_WINDOWS
1777  return res;
1778 }
1779 
1780 /* Transaction end */
1781 static __inline void _xend() {
1782 #if KMP_OS_WINDOWS
1783  __asm {
1784  _emit 0x0f
1785  _emit 0x01
1786  _emit 0xd5
1787  }
1788 #else
1789  __asm__ volatile(".byte 0x0f; .byte 0x01; .byte 0xd5" ::: "memory");
1790 #endif
1791 }
1792 
1793 /* This is a macro, the argument must be a single byte constant which can be
1794  evaluated by the inline assembler, since it is emitted as a byte into the
1795  assembly code. */
1796 // clang-format off
1797 #if KMP_OS_WINDOWS
1798 #define _xabort(ARG) _asm _emit 0xc6 _asm _emit 0xf8 _asm _emit ARG
1799 #else
1800 #define _xabort(ARG) \
1801  __asm__ volatile(".byte 0xC6; .byte 0xF8; .byte " STRINGIZE(ARG):::"memory");
1802 #endif
1803 // clang-format on
1804 #endif // KMP_COMPILER_ICC && __INTEL_COMPILER >= 1300
1805 
1806 // Statistics is collected for testing purpose
1807 #if KMP_DEBUG_ADAPTIVE_LOCKS
1808 
1809 // We accumulate speculative lock statistics when the lock is destroyed. We
1810 // keep locks that haven't been destroyed in the liveLocks list so that we can
1811 // grab their statistics too.
1812 static kmp_adaptive_lock_statistics_t destroyedStats;
1813 
1814 // To hold the list of live locks.
1815 static kmp_adaptive_lock_info_t liveLocks;
1816 
1817 // A lock so we can safely update the list of locks.
1818 static kmp_bootstrap_lock_t chain_lock =
1819  KMP_BOOTSTRAP_LOCK_INITIALIZER(chain_lock);
1820 
1821 // Initialize the list of stats.
1822 void __kmp_init_speculative_stats() {
1823  kmp_adaptive_lock_info_t *lck = &liveLocks;
1824 
1825  memset(CCAST(kmp_adaptive_lock_statistics_t *, &(lck->stats)), 0,
1826  sizeof(lck->stats));
1827  lck->stats.next = lck;
1828  lck->stats.prev = lck;
1829 
1830  KMP_ASSERT(lck->stats.next->stats.prev == lck);
1831  KMP_ASSERT(lck->stats.prev->stats.next == lck);
1832 
1833  __kmp_init_bootstrap_lock(&chain_lock);
1834 }
1835 
1836 // Insert the lock into the circular list
1837 static void __kmp_remember_lock(kmp_adaptive_lock_info_t *lck) {
1838  __kmp_acquire_bootstrap_lock(&chain_lock);
1839 
1840  lck->stats.next = liveLocks.stats.next;
1841  lck->stats.prev = &liveLocks;
1842 
1843  liveLocks.stats.next = lck;
1844  lck->stats.next->stats.prev = lck;
1845 
1846  KMP_ASSERT(lck->stats.next->stats.prev == lck);
1847  KMP_ASSERT(lck->stats.prev->stats.next == lck);
1848 
1849  __kmp_release_bootstrap_lock(&chain_lock);
1850 }
1851 
1852 static void __kmp_forget_lock(kmp_adaptive_lock_info_t *lck) {
1853  KMP_ASSERT(lck->stats.next->stats.prev == lck);
1854  KMP_ASSERT(lck->stats.prev->stats.next == lck);
1855 
1856  kmp_adaptive_lock_info_t *n = lck->stats.next;
1857  kmp_adaptive_lock_info_t *p = lck->stats.prev;
1858 
1859  n->stats.prev = p;
1860  p->stats.next = n;
1861 }
1862 
1863 static void __kmp_zero_speculative_stats(kmp_adaptive_lock_info_t *lck) {
1864  memset(CCAST(kmp_adaptive_lock_statistics_t *, &lck->stats), 0,
1865  sizeof(lck->stats));
1866  __kmp_remember_lock(lck);
1867 }
1868 
1869 static void __kmp_add_stats(kmp_adaptive_lock_statistics_t *t,
1870  kmp_adaptive_lock_info_t *lck) {
1871  kmp_adaptive_lock_statistics_t volatile *s = &lck->stats;
1872 
1873  t->nonSpeculativeAcquireAttempts += lck->acquire_attempts;
1874  t->successfulSpeculations += s->successfulSpeculations;
1875  t->hardFailedSpeculations += s->hardFailedSpeculations;
1876  t->softFailedSpeculations += s->softFailedSpeculations;
1877  t->nonSpeculativeAcquires += s->nonSpeculativeAcquires;
1878  t->lemmingYields += s->lemmingYields;
1879 }
1880 
1881 static void __kmp_accumulate_speculative_stats(kmp_adaptive_lock_info_t *lck) {
1882  __kmp_acquire_bootstrap_lock(&chain_lock);
1883 
1884  __kmp_add_stats(&destroyedStats, lck);
1885  __kmp_forget_lock(lck);
1886 
1887  __kmp_release_bootstrap_lock(&chain_lock);
1888 }
1889 
1890 static float percent(kmp_uint32 count, kmp_uint32 total) {
1891  return (total == 0) ? 0.0 : (100.0 * count) / total;
1892 }
1893 
1894 void __kmp_print_speculative_stats() {
1895  kmp_adaptive_lock_statistics_t total = destroyedStats;
1896  kmp_adaptive_lock_info_t *lck;
1897 
1898  for (lck = liveLocks.stats.next; lck != &liveLocks; lck = lck->stats.next) {
1899  __kmp_add_stats(&total, lck);
1900  }
1901  kmp_adaptive_lock_statistics_t *t = &total;
1902  kmp_uint32 totalSections =
1903  t->nonSpeculativeAcquires + t->successfulSpeculations;
1904  kmp_uint32 totalSpeculations = t->successfulSpeculations +
1905  t->hardFailedSpeculations +
1906  t->softFailedSpeculations;
1907  if (totalSections <= 0)
1908  return;
1909 
1910  kmp_safe_raii_file_t statsFile;
1911  if (strcmp(__kmp_speculative_statsfile, "-") == 0) {
1912  statsFile.set_stdout();
1913  } else {
1914  size_t buffLen = KMP_STRLEN(__kmp_speculative_statsfile) + 20;
1915  char buffer[buffLen];
1916  KMP_SNPRINTF(&buffer[0], buffLen, __kmp_speculative_statsfile,
1917  (kmp_int32)getpid());
1918  statsFile.open(buffer, "w");
1919  }
1920 
1921  fprintf(statsFile, "Speculative lock statistics (all approximate!)\n");
1922  fprintf(statsFile,
1923  " Lock parameters: \n"
1924  " max_soft_retries : %10d\n"
1925  " max_badness : %10d\n",
1926  __kmp_adaptive_backoff_params.max_soft_retries,
1927  __kmp_adaptive_backoff_params.max_badness);
1928  fprintf(statsFile, " Non-speculative acquire attempts : %10d\n",
1929  t->nonSpeculativeAcquireAttempts);
1930  fprintf(statsFile, " Total critical sections : %10d\n",
1931  totalSections);
1932  fprintf(statsFile, " Successful speculations : %10d (%5.1f%%)\n",
1933  t->successfulSpeculations,
1934  percent(t->successfulSpeculations, totalSections));
1935  fprintf(statsFile, " Non-speculative acquires : %10d (%5.1f%%)\n",
1936  t->nonSpeculativeAcquires,
1937  percent(t->nonSpeculativeAcquires, totalSections));
1938  fprintf(statsFile, " Lemming yields : %10d\n\n",
1939  t->lemmingYields);
1940 
1941  fprintf(statsFile, " Speculative acquire attempts : %10d\n",
1942  totalSpeculations);
1943  fprintf(statsFile, " Successes : %10d (%5.1f%%)\n",
1944  t->successfulSpeculations,
1945  percent(t->successfulSpeculations, totalSpeculations));
1946  fprintf(statsFile, " Soft failures : %10d (%5.1f%%)\n",
1947  t->softFailedSpeculations,
1948  percent(t->softFailedSpeculations, totalSpeculations));
1949  fprintf(statsFile, " Hard failures : %10d (%5.1f%%)\n",
1950  t->hardFailedSpeculations,
1951  percent(t->hardFailedSpeculations, totalSpeculations));
1952 }
1953 
1954 #define KMP_INC_STAT(lck, stat) (lck->lk.adaptive.stats.stat++)
1955 #else
1956 #define KMP_INC_STAT(lck, stat)
1957 
1958 #endif // KMP_DEBUG_ADAPTIVE_LOCKS
1959 
1960 static inline bool __kmp_is_unlocked_queuing_lock(kmp_queuing_lock_t *lck) {
1961  // It is enough to check that the head_id is zero.
1962  // We don't also need to check the tail.
1963  bool res = lck->lk.head_id == 0;
1964 
1965 // We need a fence here, since we must ensure that no memory operations
1966 // from later in this thread float above that read.
1967 #if KMP_COMPILER_ICC
1968  _mm_mfence();
1969 #else
1970  __sync_synchronize();
1971 #endif
1972 
1973  return res;
1974 }
1975 
1976 // Functions for manipulating the badness
1977 static __inline void
1978 __kmp_update_badness_after_success(kmp_adaptive_lock_t *lck) {
1979  // Reset the badness to zero so we eagerly try to speculate again
1980  lck->lk.adaptive.badness = 0;
1981  KMP_INC_STAT(lck, successfulSpeculations);
1982 }
1983 
1984 // Create a bit mask with one more set bit.
1985 static __inline void __kmp_step_badness(kmp_adaptive_lock_t *lck) {
1986  kmp_uint32 newBadness = (lck->lk.adaptive.badness << 1) | 1;
1987  if (newBadness > lck->lk.adaptive.max_badness) {
1988  return;
1989  } else {
1990  lck->lk.adaptive.badness = newBadness;
1991  }
1992 }
1993 
1994 // Check whether speculation should be attempted.
1995 KMP_ATTRIBUTE_TARGET_RTM
1996 static __inline int __kmp_should_speculate(kmp_adaptive_lock_t *lck,
1997  kmp_int32 gtid) {
1998  kmp_uint32 badness = lck->lk.adaptive.badness;
1999  kmp_uint32 attempts = lck->lk.adaptive.acquire_attempts;
2000  int res = (attempts & badness) == 0;
2001  return res;
2002 }
2003 
2004 // Attempt to acquire only the speculative lock.
2005 // Does not back off to the non-speculative lock.
2006 KMP_ATTRIBUTE_TARGET_RTM
2007 static int __kmp_test_adaptive_lock_only(kmp_adaptive_lock_t *lck,
2008  kmp_int32 gtid) {
2009  int retries = lck->lk.adaptive.max_soft_retries;
2010 
2011  // We don't explicitly count the start of speculation, rather we record the
2012  // results (success, hard fail, soft fail). The sum of all of those is the
2013  // total number of times we started speculation since all speculations must
2014  // end one of those ways.
2015  do {
2016  kmp_uint32 status = _xbegin();
2017  // Switch this in to disable actual speculation but exercise at least some
2018  // of the rest of the code. Useful for debugging...
2019  // kmp_uint32 status = _XABORT_NESTED;
2020 
2021  if (status == _XBEGIN_STARTED) {
2022  /* We have successfully started speculation. Check that no-one acquired
2023  the lock for real between when we last looked and now. This also gets
2024  the lock cache line into our read-set, which we need so that we'll
2025  abort if anyone later claims it for real. */
2026  if (!__kmp_is_unlocked_queuing_lock(GET_QLK_PTR(lck))) {
2027  // Lock is now visibly acquired, so someone beat us to it. Abort the
2028  // transaction so we'll restart from _xbegin with the failure status.
2029  _xabort(0x01);
2030  KMP_ASSERT2(0, "should not get here");
2031  }
2032  return 1; // Lock has been acquired (speculatively)
2033  } else {
2034  // We have aborted, update the statistics
2035  if (status & SOFT_ABORT_MASK) {
2036  KMP_INC_STAT(lck, softFailedSpeculations);
2037  // and loop round to retry.
2038  } else {
2039  KMP_INC_STAT(lck, hardFailedSpeculations);
2040  // Give up if we had a hard failure.
2041  break;
2042  }
2043  }
2044  } while (retries--); // Loop while we have retries, and didn't fail hard.
2045 
2046  // Either we had a hard failure or we didn't succeed softly after
2047  // the full set of attempts, so back off the badness.
2048  __kmp_step_badness(lck);
2049  return 0;
2050 }
2051 
2052 // Attempt to acquire the speculative lock, or back off to the non-speculative
2053 // one if the speculative lock cannot be acquired.
2054 // We can succeed speculatively, non-speculatively, or fail.
2055 static int __kmp_test_adaptive_lock(kmp_adaptive_lock_t *lck, kmp_int32 gtid) {
2056  // First try to acquire the lock speculatively
2057  if (__kmp_should_speculate(lck, gtid) &&
2058  __kmp_test_adaptive_lock_only(lck, gtid))
2059  return 1;
2060 
2061  // Speculative acquisition failed, so try to acquire it non-speculatively.
2062  // Count the non-speculative acquire attempt
2063  lck->lk.adaptive.acquire_attempts++;
2064 
2065  // Use base, non-speculative lock.
2066  if (__kmp_test_queuing_lock(GET_QLK_PTR(lck), gtid)) {
2067  KMP_INC_STAT(lck, nonSpeculativeAcquires);
2068  return 1; // Lock is acquired (non-speculatively)
2069  } else {
2070  return 0; // Failed to acquire the lock, it's already visibly locked.
2071  }
2072 }
2073 
2074 static int __kmp_test_adaptive_lock_with_checks(kmp_adaptive_lock_t *lck,
2075  kmp_int32 gtid) {
2076  char const *const func = "omp_test_lock";
2077  if (lck->lk.qlk.initialized != GET_QLK_PTR(lck)) {
2078  KMP_FATAL(LockIsUninitialized, func);
2079  }
2080 
2081  int retval = __kmp_test_adaptive_lock(lck, gtid);
2082 
2083  if (retval) {
2084  lck->lk.qlk.owner_id = gtid + 1;
2085  }
2086  return retval;
2087 }
2088 
2089 // Block until we can acquire a speculative, adaptive lock. We check whether we
2090 // should be trying to speculate. If we should be, we check the real lock to see
2091 // if it is free, and, if not, pause without attempting to acquire it until it
2092 // is. Then we try the speculative acquire. This means that although we suffer
2093 // from lemmings a little (because all we can't acquire the lock speculatively
2094 // until the queue of threads waiting has cleared), we don't get into a state
2095 // where we can never acquire the lock speculatively (because we force the queue
2096 // to clear by preventing new arrivals from entering the queue). This does mean
2097 // that when we're trying to break lemmings, the lock is no longer fair. However
2098 // OpenMP makes no guarantee that its locks are fair, so this isn't a real
2099 // problem.
2100 static void __kmp_acquire_adaptive_lock(kmp_adaptive_lock_t *lck,
2101  kmp_int32 gtid) {
2102  if (__kmp_should_speculate(lck, gtid)) {
2103  if (__kmp_is_unlocked_queuing_lock(GET_QLK_PTR(lck))) {
2104  if (__kmp_test_adaptive_lock_only(lck, gtid))
2105  return;
2106  // We tried speculation and failed, so give up.
2107  } else {
2108  // We can't try speculation until the lock is free, so we pause here
2109  // (without suspending on the queueing lock, to allow it to drain, then
2110  // try again. All other threads will also see the same result for
2111  // shouldSpeculate, so will be doing the same if they try to claim the
2112  // lock from now on.
2113  while (!__kmp_is_unlocked_queuing_lock(GET_QLK_PTR(lck))) {
2114  KMP_INC_STAT(lck, lemmingYields);
2115  KMP_YIELD(TRUE);
2116  }
2117 
2118  if (__kmp_test_adaptive_lock_only(lck, gtid))
2119  return;
2120  }
2121  }
2122 
2123  // Speculative acquisition failed, so acquire it non-speculatively.
2124  // Count the non-speculative acquire attempt
2125  lck->lk.adaptive.acquire_attempts++;
2126 
2127  __kmp_acquire_queuing_lock_timed_template<FALSE>(GET_QLK_PTR(lck), gtid);
2128  // We have acquired the base lock, so count that.
2129  KMP_INC_STAT(lck, nonSpeculativeAcquires);
2130  ANNOTATE_QUEUING_ACQUIRED(lck);
2131 }
2132 
2133 static void __kmp_acquire_adaptive_lock_with_checks(kmp_adaptive_lock_t *lck,
2134  kmp_int32 gtid) {
2135  char const *const func = "omp_set_lock";
2136  if (lck->lk.qlk.initialized != GET_QLK_PTR(lck)) {
2137  KMP_FATAL(LockIsUninitialized, func);
2138  }
2139  if (__kmp_get_queuing_lock_owner(GET_QLK_PTR(lck)) == gtid) {
2140  KMP_FATAL(LockIsAlreadyOwned, func);
2141  }
2142 
2143  __kmp_acquire_adaptive_lock(lck, gtid);
2144 
2145  lck->lk.qlk.owner_id = gtid + 1;
2146 }
2147 
2148 KMP_ATTRIBUTE_TARGET_RTM
2149 static int __kmp_release_adaptive_lock(kmp_adaptive_lock_t *lck,
2150  kmp_int32 gtid) {
2151  if (__kmp_is_unlocked_queuing_lock(GET_QLK_PTR(
2152  lck))) { // If the lock doesn't look claimed we must be speculating.
2153  // (Or the user's code is buggy and they're releasing without locking;
2154  // if we had XTEST we'd be able to check that case...)
2155  _xend(); // Exit speculation
2156  __kmp_update_badness_after_success(lck);
2157  } else { // Since the lock *is* visibly locked we're not speculating,
2158  // so should use the underlying lock's release scheme.
2159  __kmp_release_queuing_lock(GET_QLK_PTR(lck), gtid);
2160  }
2161  return KMP_LOCK_RELEASED;
2162 }
2163 
2164 static int __kmp_release_adaptive_lock_with_checks(kmp_adaptive_lock_t *lck,
2165  kmp_int32 gtid) {
2166  char const *const func = "omp_unset_lock";
2167  KMP_MB(); /* in case another processor initialized lock */
2168  if (lck->lk.qlk.initialized != GET_QLK_PTR(lck)) {
2169  KMP_FATAL(LockIsUninitialized, func);
2170  }
2171  if (__kmp_get_queuing_lock_owner(GET_QLK_PTR(lck)) == -1) {
2172  KMP_FATAL(LockUnsettingFree, func);
2173  }
2174  if (__kmp_get_queuing_lock_owner(GET_QLK_PTR(lck)) != gtid) {
2175  KMP_FATAL(LockUnsettingSetByAnother, func);
2176  }
2177  lck->lk.qlk.owner_id = 0;
2178  __kmp_release_adaptive_lock(lck, gtid);
2179  return KMP_LOCK_RELEASED;
2180 }
2181 
2182 static void __kmp_init_adaptive_lock(kmp_adaptive_lock_t *lck) {
2183  __kmp_init_queuing_lock(GET_QLK_PTR(lck));
2184  lck->lk.adaptive.badness = 0;
2185  lck->lk.adaptive.acquire_attempts = 0; // nonSpeculativeAcquireAttempts = 0;
2186  lck->lk.adaptive.max_soft_retries =
2187  __kmp_adaptive_backoff_params.max_soft_retries;
2188  lck->lk.adaptive.max_badness = __kmp_adaptive_backoff_params.max_badness;
2189 #if KMP_DEBUG_ADAPTIVE_LOCKS
2190  __kmp_zero_speculative_stats(&lck->lk.adaptive);
2191 #endif
2192  KA_TRACE(1000, ("__kmp_init_adaptive_lock: lock %p initialized\n", lck));
2193 }
2194 
2195 static void __kmp_destroy_adaptive_lock(kmp_adaptive_lock_t *lck) {
2196 #if KMP_DEBUG_ADAPTIVE_LOCKS
2197  __kmp_accumulate_speculative_stats(&lck->lk.adaptive);
2198 #endif
2199  __kmp_destroy_queuing_lock(GET_QLK_PTR(lck));
2200  // Nothing needed for the speculative part.
2201 }
2202 
2203 static void __kmp_destroy_adaptive_lock_with_checks(kmp_adaptive_lock_t *lck) {
2204  char const *const func = "omp_destroy_lock";
2205  if (lck->lk.qlk.initialized != GET_QLK_PTR(lck)) {
2206  KMP_FATAL(LockIsUninitialized, func);
2207  }
2208  if (__kmp_get_queuing_lock_owner(GET_QLK_PTR(lck)) != -1) {
2209  KMP_FATAL(LockStillOwned, func);
2210  }
2211  __kmp_destroy_adaptive_lock(lck);
2212 }
2213 
2214 #endif // KMP_USE_ADAPTIVE_LOCKS
2215 
2216 /* ------------------------------------------------------------------------ */
2217 /* DRDPA ticket locks */
2218 /* "DRDPA" means Dynamically Reconfigurable Distributed Polling Area */
2219 
2220 static kmp_int32 __kmp_get_drdpa_lock_owner(kmp_drdpa_lock_t *lck) {
2221  return lck->lk.owner_id - 1;
2222 }
2223 
2224 static inline bool __kmp_is_drdpa_lock_nestable(kmp_drdpa_lock_t *lck) {
2225  return lck->lk.depth_locked != -1;
2226 }
2227 
2228 __forceinline static int
2229 __kmp_acquire_drdpa_lock_timed_template(kmp_drdpa_lock_t *lck, kmp_int32 gtid) {
2230  kmp_uint64 ticket = KMP_ATOMIC_INC(&lck->lk.next_ticket);
2231  kmp_uint64 mask = lck->lk.mask; // atomic load
2232  std::atomic<kmp_uint64> *polls = lck->lk.polls;
2233 
2234 #ifdef USE_LOCK_PROFILE
2235  if (polls[ticket & mask] != ticket)
2236  __kmp_printf("LOCK CONTENTION: %p\n", lck);
2237 /* else __kmp_printf( "." );*/
2238 #endif /* USE_LOCK_PROFILE */
2239 
2240  // Now spin-wait, but reload the polls pointer and mask, in case the
2241  // polling area has been reconfigured. Unless it is reconfigured, the
2242  // reloads stay in L1 cache and are cheap.
2243  //
2244  // Keep this code in sync with KMP_WAIT, in kmp_dispatch.cpp !!!
2245  // The current implementation of KMP_WAIT doesn't allow for mask
2246  // and poll to be re-read every spin iteration.
2247  kmp_uint32 spins;
2248  KMP_FSYNC_PREPARE(lck);
2249  KMP_INIT_YIELD(spins);
2250  while (polls[ticket & mask] < ticket) { // atomic load
2251  KMP_YIELD_OVERSUB_ELSE_SPIN(spins);
2252  // Re-read the mask and the poll pointer from the lock structure.
2253  //
2254  // Make certain that "mask" is read before "polls" !!!
2255  //
2256  // If another thread picks reconfigures the polling area and updates their
2257  // values, and we get the new value of mask and the old polls pointer, we
2258  // could access memory beyond the end of the old polling area.
2259  mask = lck->lk.mask; // atomic load
2260  polls = lck->lk.polls; // atomic load
2261  }
2262 
2263  // Critical section starts here
2264  KMP_FSYNC_ACQUIRED(lck);
2265  KA_TRACE(1000, ("__kmp_acquire_drdpa_lock: ticket #%lld acquired lock %p\n",
2266  ticket, lck));
2267  lck->lk.now_serving = ticket; // non-volatile store
2268 
2269  // Deallocate a garbage polling area if we know that we are the last
2270  // thread that could possibly access it.
2271  //
2272  // The >= check is in case __kmp_test_drdpa_lock() allocated the cleanup
2273  // ticket.
2274  if ((lck->lk.old_polls != NULL) && (ticket >= lck->lk.cleanup_ticket)) {
2275  __kmp_free(lck->lk.old_polls);
2276  lck->lk.old_polls = NULL;
2277  lck->lk.cleanup_ticket = 0;
2278  }
2279 
2280  // Check to see if we should reconfigure the polling area.
2281  // If there is still a garbage polling area to be deallocated from a
2282  // previous reconfiguration, let a later thread reconfigure it.
2283  if (lck->lk.old_polls == NULL) {
2284  bool reconfigure = false;
2285  std::atomic<kmp_uint64> *old_polls = polls;
2286  kmp_uint32 num_polls = TCR_4(lck->lk.num_polls);
2287 
2288  if (TCR_4(__kmp_nth) >
2289  (__kmp_avail_proc ? __kmp_avail_proc : __kmp_xproc)) {
2290  // We are in oversubscription mode. Contract the polling area
2291  // down to a single location, if that hasn't been done already.
2292  if (num_polls > 1) {
2293  reconfigure = true;
2294  num_polls = TCR_4(lck->lk.num_polls);
2295  mask = 0;
2296  num_polls = 1;
2297  polls = (std::atomic<kmp_uint64> *)__kmp_allocate(num_polls *
2298  sizeof(*polls));
2299  polls[0] = ticket;
2300  }
2301  } else {
2302  // We are in under/fully subscribed mode. Check the number of
2303  // threads waiting on the lock. The size of the polling area
2304  // should be at least the number of threads waiting.
2305  kmp_uint64 num_waiting = TCR_8(lck->lk.next_ticket) - ticket - 1;
2306  if (num_waiting > num_polls) {
2307  kmp_uint32 old_num_polls = num_polls;
2308  reconfigure = true;
2309  do {
2310  mask = (mask << 1) | 1;
2311  num_polls *= 2;
2312  } while (num_polls <= num_waiting);
2313 
2314  // Allocate the new polling area, and copy the relevant portion
2315  // of the old polling area to the new area. __kmp_allocate()
2316  // zeroes the memory it allocates, and most of the old area is
2317  // just zero padding, so we only copy the release counters.
2318  polls = (std::atomic<kmp_uint64> *)__kmp_allocate(num_polls *
2319  sizeof(*polls));
2320  kmp_uint32 i;
2321  for (i = 0; i < old_num_polls; i++) {
2322  polls[i].store(old_polls[i]);
2323  }
2324  }
2325  }
2326 
2327  if (reconfigure) {
2328  // Now write the updated fields back to the lock structure.
2329  //
2330  // Make certain that "polls" is written before "mask" !!!
2331  //
2332  // If another thread picks up the new value of mask and the old polls
2333  // pointer , it could access memory beyond the end of the old polling
2334  // area.
2335  //
2336  // On x86, we need memory fences.
2337  KA_TRACE(1000, ("__kmp_acquire_drdpa_lock: ticket #%lld reconfiguring "
2338  "lock %p to %d polls\n",
2339  ticket, lck, num_polls));
2340 
2341  lck->lk.old_polls = old_polls;
2342  lck->lk.polls = polls; // atomic store
2343 
2344  KMP_MB();
2345 
2346  lck->lk.num_polls = num_polls;
2347  lck->lk.mask = mask; // atomic store
2348 
2349  KMP_MB();
2350 
2351  // Only after the new polling area and mask have been flushed
2352  // to main memory can we update the cleanup ticket field.
2353  //
2354  // volatile load / non-volatile store
2355  lck->lk.cleanup_ticket = lck->lk.next_ticket;
2356  }
2357  }
2358  return KMP_LOCK_ACQUIRED_FIRST;
2359 }
2360 
2361 int __kmp_acquire_drdpa_lock(kmp_drdpa_lock_t *lck, kmp_int32 gtid) {
2362  int retval = __kmp_acquire_drdpa_lock_timed_template(lck, gtid);
2363  ANNOTATE_DRDPA_ACQUIRED(lck);
2364  return retval;
2365 }
2366 
2367 static int __kmp_acquire_drdpa_lock_with_checks(kmp_drdpa_lock_t *lck,
2368  kmp_int32 gtid) {
2369  char const *const func = "omp_set_lock";
2370  if (lck->lk.initialized != lck) {
2371  KMP_FATAL(LockIsUninitialized, func);
2372  }
2373  if (__kmp_is_drdpa_lock_nestable(lck)) {
2374  KMP_FATAL(LockNestableUsedAsSimple, func);
2375  }
2376  if ((gtid >= 0) && (__kmp_get_drdpa_lock_owner(lck) == gtid)) {
2377  KMP_FATAL(LockIsAlreadyOwned, func);
2378  }
2379 
2380  __kmp_acquire_drdpa_lock(lck, gtid);
2381 
2382  lck->lk.owner_id = gtid + 1;
2383  return KMP_LOCK_ACQUIRED_FIRST;
2384 }
2385 
2386 int __kmp_test_drdpa_lock(kmp_drdpa_lock_t *lck, kmp_int32 gtid) {
2387  // First get a ticket, then read the polls pointer and the mask.
2388  // The polls pointer must be read before the mask!!! (See above)
2389  kmp_uint64 ticket = lck->lk.next_ticket; // atomic load
2390  std::atomic<kmp_uint64> *polls = lck->lk.polls;
2391  kmp_uint64 mask = lck->lk.mask; // atomic load
2392  if (polls[ticket & mask] == ticket) {
2393  kmp_uint64 next_ticket = ticket + 1;
2394  if (__kmp_atomic_compare_store_acq(&lck->lk.next_ticket, ticket,
2395  next_ticket)) {
2396  KMP_FSYNC_ACQUIRED(lck);
2397  KA_TRACE(1000, ("__kmp_test_drdpa_lock: ticket #%lld acquired lock %p\n",
2398  ticket, lck));
2399  lck->lk.now_serving = ticket; // non-volatile store
2400 
2401  // Since no threads are waiting, there is no possibility that we would
2402  // want to reconfigure the polling area. We might have the cleanup ticket
2403  // value (which says that it is now safe to deallocate old_polls), but
2404  // we'll let a later thread which calls __kmp_acquire_lock do that - this
2405  // routine isn't supposed to block, and we would risk blocks if we called
2406  // __kmp_free() to do the deallocation.
2407  return TRUE;
2408  }
2409  }
2410  return FALSE;
2411 }
2412 
2413 static int __kmp_test_drdpa_lock_with_checks(kmp_drdpa_lock_t *lck,
2414  kmp_int32 gtid) {
2415  char const *const func = "omp_test_lock";
2416  if (lck->lk.initialized != lck) {
2417  KMP_FATAL(LockIsUninitialized, func);
2418  }
2419  if (__kmp_is_drdpa_lock_nestable(lck)) {
2420  KMP_FATAL(LockNestableUsedAsSimple, func);
2421  }
2422 
2423  int retval = __kmp_test_drdpa_lock(lck, gtid);
2424 
2425  if (retval) {
2426  lck->lk.owner_id = gtid + 1;
2427  }
2428  return retval;
2429 }
2430 
2431 int __kmp_release_drdpa_lock(kmp_drdpa_lock_t *lck, kmp_int32 gtid) {
2432  // Read the ticket value from the lock data struct, then the polls pointer and
2433  // the mask. The polls pointer must be read before the mask!!! (See above)
2434  kmp_uint64 ticket = lck->lk.now_serving + 1; // non-atomic load
2435  std::atomic<kmp_uint64> *polls = lck->lk.polls; // atomic load
2436  kmp_uint64 mask = lck->lk.mask; // atomic load
2437  KA_TRACE(1000, ("__kmp_release_drdpa_lock: ticket #%lld released lock %p\n",
2438  ticket - 1, lck));
2439  KMP_FSYNC_RELEASING(lck);
2440  ANNOTATE_DRDPA_RELEASED(lck);
2441  polls[ticket & mask] = ticket; // atomic store
2442  return KMP_LOCK_RELEASED;
2443 }
2444 
2445 static int __kmp_release_drdpa_lock_with_checks(kmp_drdpa_lock_t *lck,
2446  kmp_int32 gtid) {
2447  char const *const func = "omp_unset_lock";
2448  KMP_MB(); /* in case another processor initialized lock */
2449  if (lck->lk.initialized != lck) {
2450  KMP_FATAL(LockIsUninitialized, func);
2451  }
2452  if (__kmp_is_drdpa_lock_nestable(lck)) {
2453  KMP_FATAL(LockNestableUsedAsSimple, func);
2454  }
2455  if (__kmp_get_drdpa_lock_owner(lck) == -1) {
2456  KMP_FATAL(LockUnsettingFree, func);
2457  }
2458  if ((gtid >= 0) && (__kmp_get_drdpa_lock_owner(lck) >= 0) &&
2459  (__kmp_get_drdpa_lock_owner(lck) != gtid)) {
2460  KMP_FATAL(LockUnsettingSetByAnother, func);
2461  }
2462  lck->lk.owner_id = 0;
2463  return __kmp_release_drdpa_lock(lck, gtid);
2464 }
2465 
2466 void __kmp_init_drdpa_lock(kmp_drdpa_lock_t *lck) {
2467  lck->lk.location = NULL;
2468  lck->lk.mask = 0;
2469  lck->lk.num_polls = 1;
2470  lck->lk.polls = (std::atomic<kmp_uint64> *)__kmp_allocate(
2471  lck->lk.num_polls * sizeof(*(lck->lk.polls)));
2472  lck->lk.cleanup_ticket = 0;
2473  lck->lk.old_polls = NULL;
2474  lck->lk.next_ticket = 0;
2475  lck->lk.now_serving = 0;
2476  lck->lk.owner_id = 0; // no thread owns the lock.
2477  lck->lk.depth_locked = -1; // >= 0 for nestable locks, -1 for simple locks.
2478  lck->lk.initialized = lck;
2479 
2480  KA_TRACE(1000, ("__kmp_init_drdpa_lock: lock %p initialized\n", lck));
2481 }
2482 
2483 void __kmp_destroy_drdpa_lock(kmp_drdpa_lock_t *lck) {
2484  lck->lk.initialized = NULL;
2485  lck->lk.location = NULL;
2486  if (lck->lk.polls.load() != NULL) {
2487  __kmp_free(lck->lk.polls.load());
2488  lck->lk.polls = NULL;
2489  }
2490  if (lck->lk.old_polls != NULL) {
2491  __kmp_free(lck->lk.old_polls);
2492  lck->lk.old_polls = NULL;
2493  }
2494  lck->lk.mask = 0;
2495  lck->lk.num_polls = 0;
2496  lck->lk.cleanup_ticket = 0;
2497  lck->lk.next_ticket = 0;
2498  lck->lk.now_serving = 0;
2499  lck->lk.owner_id = 0;
2500  lck->lk.depth_locked = -1;
2501 }
2502 
2503 static void __kmp_destroy_drdpa_lock_with_checks(kmp_drdpa_lock_t *lck) {
2504  char const *const func = "omp_destroy_lock";
2505  if (lck->lk.initialized != lck) {
2506  KMP_FATAL(LockIsUninitialized, func);
2507  }
2508  if (__kmp_is_drdpa_lock_nestable(lck)) {
2509  KMP_FATAL(LockNestableUsedAsSimple, func);
2510  }
2511  if (__kmp_get_drdpa_lock_owner(lck) != -1) {
2512  KMP_FATAL(LockStillOwned, func);
2513  }
2514  __kmp_destroy_drdpa_lock(lck);
2515 }
2516 
2517 // nested drdpa ticket locks
2518 
2519 int __kmp_acquire_nested_drdpa_lock(kmp_drdpa_lock_t *lck, kmp_int32 gtid) {
2520  KMP_DEBUG_ASSERT(gtid >= 0);
2521 
2522  if (__kmp_get_drdpa_lock_owner(lck) == gtid) {
2523  lck->lk.depth_locked += 1;
2524  return KMP_LOCK_ACQUIRED_NEXT;
2525  } else {
2526  __kmp_acquire_drdpa_lock_timed_template(lck, gtid);
2527  ANNOTATE_DRDPA_ACQUIRED(lck);
2528  KMP_MB();
2529  lck->lk.depth_locked = 1;
2530  KMP_MB();
2531  lck->lk.owner_id = gtid + 1;
2532  return KMP_LOCK_ACQUIRED_FIRST;
2533  }
2534 }
2535 
2536 static void __kmp_acquire_nested_drdpa_lock_with_checks(kmp_drdpa_lock_t *lck,
2537  kmp_int32 gtid) {
2538  char const *const func = "omp_set_nest_lock";
2539  if (lck->lk.initialized != lck) {
2540  KMP_FATAL(LockIsUninitialized, func);
2541  }
2542  if (!__kmp_is_drdpa_lock_nestable(lck)) {
2543  KMP_FATAL(LockSimpleUsedAsNestable, func);
2544  }
2545  __kmp_acquire_nested_drdpa_lock(lck, gtid);
2546 }
2547 
2548 int __kmp_test_nested_drdpa_lock(kmp_drdpa_lock_t *lck, kmp_int32 gtid) {
2549  int retval;
2550 
2551  KMP_DEBUG_ASSERT(gtid >= 0);
2552 
2553  if (__kmp_get_drdpa_lock_owner(lck) == gtid) {
2554  retval = ++lck->lk.depth_locked;
2555  } else if (!__kmp_test_drdpa_lock(lck, gtid)) {
2556  retval = 0;
2557  } else {
2558  KMP_MB();
2559  retval = lck->lk.depth_locked = 1;
2560  KMP_MB();
2561  lck->lk.owner_id = gtid + 1;
2562  }
2563  return retval;
2564 }
2565 
2566 static int __kmp_test_nested_drdpa_lock_with_checks(kmp_drdpa_lock_t *lck,
2567  kmp_int32 gtid) {
2568  char const *const func = "omp_test_nest_lock";
2569  if (lck->lk.initialized != lck) {
2570  KMP_FATAL(LockIsUninitialized, func);
2571  }
2572  if (!__kmp_is_drdpa_lock_nestable(lck)) {
2573  KMP_FATAL(LockSimpleUsedAsNestable, func);
2574  }
2575  return __kmp_test_nested_drdpa_lock(lck, gtid);
2576 }
2577 
2578 int __kmp_release_nested_drdpa_lock(kmp_drdpa_lock_t *lck, kmp_int32 gtid) {
2579  KMP_DEBUG_ASSERT(gtid >= 0);
2580 
2581  KMP_MB();
2582  if (--(lck->lk.depth_locked) == 0) {
2583  KMP_MB();
2584  lck->lk.owner_id = 0;
2585  __kmp_release_drdpa_lock(lck, gtid);
2586  return KMP_LOCK_RELEASED;
2587  }
2588  return KMP_LOCK_STILL_HELD;
2589 }
2590 
2591 static int __kmp_release_nested_drdpa_lock_with_checks(kmp_drdpa_lock_t *lck,
2592  kmp_int32 gtid) {
2593  char const *const func = "omp_unset_nest_lock";
2594  KMP_MB(); /* in case another processor initialized lock */
2595  if (lck->lk.initialized != lck) {
2596  KMP_FATAL(LockIsUninitialized, func);
2597  }
2598  if (!__kmp_is_drdpa_lock_nestable(lck)) {
2599  KMP_FATAL(LockSimpleUsedAsNestable, func);
2600  }
2601  if (__kmp_get_drdpa_lock_owner(lck) == -1) {
2602  KMP_FATAL(LockUnsettingFree, func);
2603  }
2604  if (__kmp_get_drdpa_lock_owner(lck) != gtid) {
2605  KMP_FATAL(LockUnsettingSetByAnother, func);
2606  }
2607  return __kmp_release_nested_drdpa_lock(lck, gtid);
2608 }
2609 
2610 void __kmp_init_nested_drdpa_lock(kmp_drdpa_lock_t *lck) {
2611  __kmp_init_drdpa_lock(lck);
2612  lck->lk.depth_locked = 0; // >= 0 for nestable locks, -1 for simple locks
2613 }
2614 
2615 void __kmp_destroy_nested_drdpa_lock(kmp_drdpa_lock_t *lck) {
2616  __kmp_destroy_drdpa_lock(lck);
2617  lck->lk.depth_locked = 0;
2618 }
2619 
2620 static void __kmp_destroy_nested_drdpa_lock_with_checks(kmp_drdpa_lock_t *lck) {
2621  char const *const func = "omp_destroy_nest_lock";
2622  if (lck->lk.initialized != lck) {
2623  KMP_FATAL(LockIsUninitialized, func);
2624  }
2625  if (!__kmp_is_drdpa_lock_nestable(lck)) {
2626  KMP_FATAL(LockSimpleUsedAsNestable, func);
2627  }
2628  if (__kmp_get_drdpa_lock_owner(lck) != -1) {
2629  KMP_FATAL(LockStillOwned, func);
2630  }
2631  __kmp_destroy_nested_drdpa_lock(lck);
2632 }
2633 
2634 // access functions to fields which don't exist for all lock kinds.
2635 
2636 static const ident_t *__kmp_get_drdpa_lock_location(kmp_drdpa_lock_t *lck) {
2637  return lck->lk.location;
2638 }
2639 
2640 static void __kmp_set_drdpa_lock_location(kmp_drdpa_lock_t *lck,
2641  const ident_t *loc) {
2642  lck->lk.location = loc;
2643 }
2644 
2645 static kmp_lock_flags_t __kmp_get_drdpa_lock_flags(kmp_drdpa_lock_t *lck) {
2646  return lck->lk.flags;
2647 }
2648 
2649 static void __kmp_set_drdpa_lock_flags(kmp_drdpa_lock_t *lck,
2650  kmp_lock_flags_t flags) {
2651  lck->lk.flags = flags;
2652 }
2653 
2654 // Time stamp counter
2655 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
2656 #define __kmp_tsc() __kmp_hardware_timestamp()
2657 // Runtime's default backoff parameters
2658 kmp_backoff_t __kmp_spin_backoff_params = {1, 4096, 100};
2659 #else
2660 // Use nanoseconds for other platforms
2661 extern kmp_uint64 __kmp_now_nsec();
2662 kmp_backoff_t __kmp_spin_backoff_params = {1, 256, 100};
2663 #define __kmp_tsc() __kmp_now_nsec()
2664 #endif
2665 
2666 // A useful predicate for dealing with timestamps that may wrap.
2667 // Is a before b? Since the timestamps may wrap, this is asking whether it's
2668 // shorter to go clockwise from a to b around the clock-face, or anti-clockwise.
2669 // Times where going clockwise is less distance than going anti-clockwise
2670 // are in the future, others are in the past. e.g. a = MAX-1, b = MAX+1 (=0),
2671 // then a > b (true) does not mean a reached b; whereas signed(a) = -2,
2672 // signed(b) = 0 captures the actual difference
2673 static inline bool before(kmp_uint64 a, kmp_uint64 b) {
2674  return ((kmp_int64)b - (kmp_int64)a) > 0;
2675 }
2676 
2677 // Truncated binary exponential backoff function
2678 void __kmp_spin_backoff(kmp_backoff_t *boff) {
2679  // We could flatten this loop, but making it a nested loop gives better result
2680  kmp_uint32 i;
2681  for (i = boff->step; i > 0; i--) {
2682  kmp_uint64 goal = __kmp_tsc() + boff->min_tick;
2683  do {
2684  KMP_CPU_PAUSE();
2685  } while (before(__kmp_tsc(), goal));
2686  }
2687  boff->step = (boff->step << 1 | 1) & (boff->max_backoff - 1);
2688 }
2689 
2690 #if KMP_USE_DYNAMIC_LOCK
2691 
2692 // Direct lock initializers. It simply writes a tag to the low 8 bits of the
2693 // lock word.
2694 static void __kmp_init_direct_lock(kmp_dyna_lock_t *lck,
2695  kmp_dyna_lockseq_t seq) {
2696  TCW_4(*lck, KMP_GET_D_TAG(seq));
2697  KA_TRACE(
2698  20,
2699  ("__kmp_init_direct_lock: initialized direct lock with type#%d\n", seq));
2700 }
2701 
2702 #if KMP_USE_TSX
2703 
2704 // HLE lock functions - imported from the testbed runtime.
2705 #define HLE_ACQUIRE ".byte 0xf2;"
2706 #define HLE_RELEASE ".byte 0xf3;"
2707 
2708 static inline kmp_uint32 swap4(kmp_uint32 volatile *p, kmp_uint32 v) {
2709  __asm__ volatile(HLE_ACQUIRE "xchg %1,%0" : "+r"(v), "+m"(*p) : : "memory");
2710  return v;
2711 }
2712 
2713 static void __kmp_destroy_hle_lock(kmp_dyna_lock_t *lck) { TCW_4(*lck, 0); }
2714 
2715 static void __kmp_destroy_hle_lock_with_checks(kmp_dyna_lock_t *lck) {
2716  TCW_4(*lck, 0);
2717 }
2718 
2719 static void __kmp_acquire_hle_lock(kmp_dyna_lock_t *lck, kmp_int32 gtid) {
2720  // Use gtid for KMP_LOCK_BUSY if necessary
2721  if (swap4(lck, KMP_LOCK_BUSY(1, hle)) != KMP_LOCK_FREE(hle)) {
2722  int delay = 1;
2723  do {
2724  while (*(kmp_uint32 volatile *)lck != KMP_LOCK_FREE(hle)) {
2725  for (int i = delay; i != 0; --i)
2726  KMP_CPU_PAUSE();
2727  delay = ((delay << 1) | 1) & 7;
2728  }
2729  } while (swap4(lck, KMP_LOCK_BUSY(1, hle)) != KMP_LOCK_FREE(hle));
2730  }
2731 }
2732 
2733 static void __kmp_acquire_hle_lock_with_checks(kmp_dyna_lock_t *lck,
2734  kmp_int32 gtid) {
2735  __kmp_acquire_hle_lock(lck, gtid); // TODO: add checks
2736 }
2737 
2738 static int __kmp_release_hle_lock(kmp_dyna_lock_t *lck, kmp_int32 gtid) {
2739  __asm__ volatile(HLE_RELEASE "movl %1,%0"
2740  : "=m"(*lck)
2741  : "r"(KMP_LOCK_FREE(hle))
2742  : "memory");
2743  return KMP_LOCK_RELEASED;
2744 }
2745 
2746 static int __kmp_release_hle_lock_with_checks(kmp_dyna_lock_t *lck,
2747  kmp_int32 gtid) {
2748  return __kmp_release_hle_lock(lck, gtid); // TODO: add checks
2749 }
2750 
2751 static int __kmp_test_hle_lock(kmp_dyna_lock_t *lck, kmp_int32 gtid) {
2752  return swap4(lck, KMP_LOCK_BUSY(1, hle)) == KMP_LOCK_FREE(hle);
2753 }
2754 
2755 static int __kmp_test_hle_lock_with_checks(kmp_dyna_lock_t *lck,
2756  kmp_int32 gtid) {
2757  return __kmp_test_hle_lock(lck, gtid); // TODO: add checks
2758 }
2759 
2760 static void __kmp_init_rtm_queuing_lock(kmp_queuing_lock_t *lck) {
2761  __kmp_init_queuing_lock(lck);
2762 }
2763 
2764 static void __kmp_destroy_rtm_queuing_lock(kmp_queuing_lock_t *lck) {
2765  __kmp_destroy_queuing_lock(lck);
2766 }
2767 
2768 static void
2769 __kmp_destroy_rtm_queuing_lock_with_checks(kmp_queuing_lock_t *lck) {
2770  __kmp_destroy_queuing_lock_with_checks(lck);
2771 }
2772 
2773 KMP_ATTRIBUTE_TARGET_RTM
2774 static void __kmp_acquire_rtm_queuing_lock(kmp_queuing_lock_t *lck,
2775  kmp_int32 gtid) {
2776  unsigned retries = 3, status;
2777  do {
2778  status = _xbegin();
2779  if (status == _XBEGIN_STARTED) {
2780  if (__kmp_is_unlocked_queuing_lock(lck))
2781  return;
2782  _xabort(0xff);
2783  }
2784  if ((status & _XABORT_EXPLICIT) && _XABORT_CODE(status) == 0xff) {
2785  // Wait until lock becomes free
2786  while (!__kmp_is_unlocked_queuing_lock(lck)) {
2787  KMP_YIELD(TRUE);
2788  }
2789  } else if (!(status & _XABORT_RETRY))
2790  break;
2791  } while (retries--);
2792 
2793  // Fall-back non-speculative lock (xchg)
2794  __kmp_acquire_queuing_lock(lck, gtid);
2795 }
2796 
2797 static void __kmp_acquire_rtm_queuing_lock_with_checks(kmp_queuing_lock_t *lck,
2798  kmp_int32 gtid) {
2799  __kmp_acquire_rtm_queuing_lock(lck, gtid);
2800 }
2801 
2802 KMP_ATTRIBUTE_TARGET_RTM
2803 static int __kmp_release_rtm_queuing_lock(kmp_queuing_lock_t *lck,
2804  kmp_int32 gtid) {
2805  if (__kmp_is_unlocked_queuing_lock(lck)) {
2806  // Releasing from speculation
2807  _xend();
2808  } else {
2809  // Releasing from a real lock
2810  __kmp_release_queuing_lock(lck, gtid);
2811  }
2812  return KMP_LOCK_RELEASED;
2813 }
2814 
2815 static int __kmp_release_rtm_queuing_lock_with_checks(kmp_queuing_lock_t *lck,
2816  kmp_int32 gtid) {
2817  return __kmp_release_rtm_queuing_lock(lck, gtid);
2818 }
2819 
2820 KMP_ATTRIBUTE_TARGET_RTM
2821 static int __kmp_test_rtm_queuing_lock(kmp_queuing_lock_t *lck,
2822  kmp_int32 gtid) {
2823  unsigned retries = 3, status;
2824  do {
2825  status = _xbegin();
2826  if (status == _XBEGIN_STARTED && __kmp_is_unlocked_queuing_lock(lck)) {
2827  return 1;
2828  }
2829  if (!(status & _XABORT_RETRY))
2830  break;
2831  } while (retries--);
2832 
2833  return __kmp_test_queuing_lock(lck, gtid);
2834 }
2835 
2836 static int __kmp_test_rtm_queuing_lock_with_checks(kmp_queuing_lock_t *lck,
2837  kmp_int32 gtid) {
2838  return __kmp_test_rtm_queuing_lock(lck, gtid);
2839 }
2840 
2841 // Reuse kmp_tas_lock_t for TSX lock which use RTM with fall-back spin lock.
2842 typedef kmp_tas_lock_t kmp_rtm_spin_lock_t;
2843 
2844 static void __kmp_destroy_rtm_spin_lock(kmp_rtm_spin_lock_t *lck) {
2845  KMP_ATOMIC_ST_REL(&lck->lk.poll, 0);
2846 }
2847 
2848 static void __kmp_destroy_rtm_spin_lock_with_checks(kmp_rtm_spin_lock_t *lck) {
2849  __kmp_destroy_rtm_spin_lock(lck);
2850 }
2851 
2852 KMP_ATTRIBUTE_TARGET_RTM
2853 static int __kmp_acquire_rtm_spin_lock(kmp_rtm_spin_lock_t *lck,
2854  kmp_int32 gtid) {
2855  unsigned retries = 3, status;
2856  kmp_int32 lock_free = KMP_LOCK_FREE(rtm_spin);
2857  kmp_int32 lock_busy = KMP_LOCK_BUSY(1, rtm_spin);
2858  do {
2859  status = _xbegin();
2860  if (status == _XBEGIN_STARTED) {
2861  if (KMP_ATOMIC_LD_RLX(&lck->lk.poll) == lock_free)
2862  return KMP_LOCK_ACQUIRED_FIRST;
2863  _xabort(0xff);
2864  }
2865  if ((status & _XABORT_EXPLICIT) && _XABORT_CODE(status) == 0xff) {
2866  // Wait until lock becomes free
2867  while (KMP_ATOMIC_LD_RLX(&lck->lk.poll) != lock_free) {
2868  KMP_YIELD(TRUE);
2869  }
2870  } else if (!(status & _XABORT_RETRY))
2871  break;
2872  } while (retries--);
2873 
2874  // Fall-back spin lock
2875  KMP_FSYNC_PREPARE(lck);
2876  kmp_backoff_t backoff = __kmp_spin_backoff_params;
2877  while (KMP_ATOMIC_LD_RLX(&lck->lk.poll) != lock_free ||
2878  !__kmp_atomic_compare_store_acq(&lck->lk.poll, lock_free, lock_busy)) {
2879  __kmp_spin_backoff(&backoff);
2880  }
2881  KMP_FSYNC_ACQUIRED(lck);
2882  return KMP_LOCK_ACQUIRED_FIRST;
2883 }
2884 
2885 static int __kmp_acquire_rtm_spin_lock_with_checks(kmp_rtm_spin_lock_t *lck,
2886  kmp_int32 gtid) {
2887  return __kmp_acquire_rtm_spin_lock(lck, gtid);
2888 }
2889 
2890 KMP_ATTRIBUTE_TARGET_RTM
2891 static int __kmp_release_rtm_spin_lock(kmp_rtm_spin_lock_t *lck,
2892  kmp_int32 gtid) {
2893  if (KMP_ATOMIC_LD_RLX(&lck->lk.poll) == KMP_LOCK_FREE(rtm_spin)) {
2894  // Releasing from speculation
2895  _xend();
2896  } else {
2897  // Releasing from a real lock
2898  KMP_FSYNC_RELEASING(lck);
2899  KMP_ATOMIC_ST_REL(&lck->lk.poll, KMP_LOCK_FREE(rtm_spin));
2900  }
2901  return KMP_LOCK_RELEASED;
2902 }
2903 
2904 static int __kmp_release_rtm_spin_lock_with_checks(kmp_rtm_spin_lock_t *lck,
2905  kmp_int32 gtid) {
2906  return __kmp_release_rtm_spin_lock(lck, gtid);
2907 }
2908 
2909 KMP_ATTRIBUTE_TARGET_RTM
2910 static int __kmp_test_rtm_spin_lock(kmp_rtm_spin_lock_t *lck, kmp_int32 gtid) {
2911  unsigned retries = 3, status;
2912  kmp_int32 lock_free = KMP_LOCK_FREE(rtm_spin);
2913  kmp_int32 lock_busy = KMP_LOCK_BUSY(1, rtm_spin);
2914  do {
2915  status = _xbegin();
2916  if (status == _XBEGIN_STARTED &&
2917  KMP_ATOMIC_LD_RLX(&lck->lk.poll) == lock_free) {
2918  return TRUE;
2919  }
2920  if (!(status & _XABORT_RETRY))
2921  break;
2922  } while (retries--);
2923 
2924  if (KMP_ATOMIC_LD_RLX(&lck->lk.poll) == lock_free &&
2925  __kmp_atomic_compare_store_acq(&lck->lk.poll, lock_free, lock_busy)) {
2926  KMP_FSYNC_ACQUIRED(lck);
2927  return TRUE;
2928  }
2929  return FALSE;
2930 }
2931 
2932 static int __kmp_test_rtm_spin_lock_with_checks(kmp_rtm_spin_lock_t *lck,
2933  kmp_int32 gtid) {
2934  return __kmp_test_rtm_spin_lock(lck, gtid);
2935 }
2936 
2937 #endif // KMP_USE_TSX
2938 
2939 // Entry functions for indirect locks (first element of direct lock jump tables)
2940 static void __kmp_init_indirect_lock(kmp_dyna_lock_t *l,
2941  kmp_dyna_lockseq_t tag);
2942 static void __kmp_destroy_indirect_lock(kmp_dyna_lock_t *lock);
2943 static int __kmp_set_indirect_lock(kmp_dyna_lock_t *lock, kmp_int32);
2944 static int __kmp_unset_indirect_lock(kmp_dyna_lock_t *lock, kmp_int32);
2945 static int __kmp_test_indirect_lock(kmp_dyna_lock_t *lock, kmp_int32);
2946 static int __kmp_set_indirect_lock_with_checks(kmp_dyna_lock_t *lock,
2947  kmp_int32);
2948 static int __kmp_unset_indirect_lock_with_checks(kmp_dyna_lock_t *lock,
2949  kmp_int32);
2950 static int __kmp_test_indirect_lock_with_checks(kmp_dyna_lock_t *lock,
2951  kmp_int32);
2952 
2953 // Lock function definitions for the union parameter type
2954 #define KMP_FOREACH_LOCK_KIND(m, a) m(ticket, a) m(queuing, a) m(drdpa, a)
2955 
2956 #define expand1(lk, op) \
2957  static void __kmp_##op##_##lk##_##lock(kmp_user_lock_p lock) { \
2958  __kmp_##op##_##lk##_##lock(&lock->lk); \
2959  }
2960 #define expand2(lk, op) \
2961  static int __kmp_##op##_##lk##_##lock(kmp_user_lock_p lock, \
2962  kmp_int32 gtid) { \
2963  return __kmp_##op##_##lk##_##lock(&lock->lk, gtid); \
2964  }
2965 #define expand3(lk, op) \
2966  static void __kmp_set_##lk##_##lock_flags(kmp_user_lock_p lock, \
2967  kmp_lock_flags_t flags) { \
2968  __kmp_set_##lk##_lock_flags(&lock->lk, flags); \
2969  }
2970 #define expand4(lk, op) \
2971  static void __kmp_set_##lk##_##lock_location(kmp_user_lock_p lock, \
2972  const ident_t *loc) { \
2973  __kmp_set_##lk##_lock_location(&lock->lk, loc); \
2974  }
2975 
2976 KMP_FOREACH_LOCK_KIND(expand1, init)
2977 KMP_FOREACH_LOCK_KIND(expand1, init_nested)
2978 KMP_FOREACH_LOCK_KIND(expand1, destroy)
2979 KMP_FOREACH_LOCK_KIND(expand1, destroy_nested)
2980 KMP_FOREACH_LOCK_KIND(expand2, acquire)
2981 KMP_FOREACH_LOCK_KIND(expand2, acquire_nested)
2982 KMP_FOREACH_LOCK_KIND(expand2, release)
2983 KMP_FOREACH_LOCK_KIND(expand2, release_nested)
2984 KMP_FOREACH_LOCK_KIND(expand2, test)
2985 KMP_FOREACH_LOCK_KIND(expand2, test_nested)
2986 KMP_FOREACH_LOCK_KIND(expand3, )
2987 KMP_FOREACH_LOCK_KIND(expand4, )
2988 
2989 #undef expand1
2990 #undef expand2
2991 #undef expand3
2992 #undef expand4
2993 
2994 // Jump tables for the indirect lock functions
2995 // Only fill in the odd entries, that avoids the need to shift out the low bit
2996 
2997 // init functions
2998 #define expand(l, op) 0, __kmp_init_direct_lock,
2999 void (*__kmp_direct_init[])(kmp_dyna_lock_t *, kmp_dyna_lockseq_t) = {
3000  __kmp_init_indirect_lock, 0, KMP_FOREACH_D_LOCK(expand, init)};
3001 #undef expand
3002 
3003 // destroy functions
3004 #define expand(l, op) 0, (void (*)(kmp_dyna_lock_t *))__kmp_##op##_##l##_lock,
3005 static void (*direct_destroy[])(kmp_dyna_lock_t *) = {
3006  __kmp_destroy_indirect_lock, 0, KMP_FOREACH_D_LOCK(expand, destroy)};
3007 #undef expand
3008 #define expand(l, op) \
3009  0, (void (*)(kmp_dyna_lock_t *))__kmp_destroy_##l##_lock_with_checks,
3010 static void (*direct_destroy_check[])(kmp_dyna_lock_t *) = {
3011  __kmp_destroy_indirect_lock, 0, KMP_FOREACH_D_LOCK(expand, destroy)};
3012 #undef expand
3013 
3014 // set/acquire functions
3015 #define expand(l, op) \
3016  0, (int (*)(kmp_dyna_lock_t *, kmp_int32))__kmp_##op##_##l##_lock,
3017 static int (*direct_set[])(kmp_dyna_lock_t *, kmp_int32) = {
3018  __kmp_set_indirect_lock, 0, KMP_FOREACH_D_LOCK(expand, acquire)};
3019 #undef expand
3020 #define expand(l, op) \
3021  0, (int (*)(kmp_dyna_lock_t *, kmp_int32))__kmp_##op##_##l##_lock_with_checks,
3022 static int (*direct_set_check[])(kmp_dyna_lock_t *, kmp_int32) = {
3023  __kmp_set_indirect_lock_with_checks, 0,
3024  KMP_FOREACH_D_LOCK(expand, acquire)};
3025 #undef expand
3026 
3027 // unset/release and test functions
3028 #define expand(l, op) \
3029  0, (int (*)(kmp_dyna_lock_t *, kmp_int32))__kmp_##op##_##l##_lock,
3030 static int (*direct_unset[])(kmp_dyna_lock_t *, kmp_int32) = {
3031  __kmp_unset_indirect_lock, 0, KMP_FOREACH_D_LOCK(expand, release)};
3032 static int (*direct_test[])(kmp_dyna_lock_t *, kmp_int32) = {
3033  __kmp_test_indirect_lock, 0, KMP_FOREACH_D_LOCK(expand, test)};
3034 #undef expand
3035 #define expand(l, op) \
3036  0, (int (*)(kmp_dyna_lock_t *, kmp_int32))__kmp_##op##_##l##_lock_with_checks,
3037 static int (*direct_unset_check[])(kmp_dyna_lock_t *, kmp_int32) = {
3038  __kmp_unset_indirect_lock_with_checks, 0,
3039  KMP_FOREACH_D_LOCK(expand, release)};
3040 static int (*direct_test_check[])(kmp_dyna_lock_t *, kmp_int32) = {
3041  __kmp_test_indirect_lock_with_checks, 0, KMP_FOREACH_D_LOCK(expand, test)};
3042 #undef expand
3043 
3044 // Exposes only one set of jump tables (*lock or *lock_with_checks).
3045 void (**__kmp_direct_destroy)(kmp_dyna_lock_t *) = 0;
3046 int (**__kmp_direct_set)(kmp_dyna_lock_t *, kmp_int32) = 0;
3047 int (**__kmp_direct_unset)(kmp_dyna_lock_t *, kmp_int32) = 0;
3048 int (**__kmp_direct_test)(kmp_dyna_lock_t *, kmp_int32) = 0;
3049 
3050 // Jump tables for the indirect lock functions
3051 #define expand(l, op) (void (*)(kmp_user_lock_p)) __kmp_##op##_##l##_##lock,
3052 void (*__kmp_indirect_init[])(kmp_user_lock_p) = {
3053  KMP_FOREACH_I_LOCK(expand, init)};
3054 #undef expand
3055 
3056 #define expand(l, op) (void (*)(kmp_user_lock_p)) __kmp_##op##_##l##_##lock,
3057 static void (*indirect_destroy[])(kmp_user_lock_p) = {
3058  KMP_FOREACH_I_LOCK(expand, destroy)};
3059 #undef expand
3060 #define expand(l, op) \
3061  (void (*)(kmp_user_lock_p)) __kmp_##op##_##l##_##lock_with_checks,
3062 static void (*indirect_destroy_check[])(kmp_user_lock_p) = {
3063  KMP_FOREACH_I_LOCK(expand, destroy)};
3064 #undef expand
3065 
3066 // set/acquire functions
3067 #define expand(l, op) \
3068  (int (*)(kmp_user_lock_p, kmp_int32)) __kmp_##op##_##l##_##lock,
3069 static int (*indirect_set[])(kmp_user_lock_p,
3070  kmp_int32) = {KMP_FOREACH_I_LOCK(expand, acquire)};
3071 #undef expand
3072 #define expand(l, op) \
3073  (int (*)(kmp_user_lock_p, kmp_int32)) __kmp_##op##_##l##_##lock_with_checks,
3074 static int (*indirect_set_check[])(kmp_user_lock_p, kmp_int32) = {
3075  KMP_FOREACH_I_LOCK(expand, acquire)};
3076 #undef expand
3077 
3078 // unset/release and test functions
3079 #define expand(l, op) \
3080  (int (*)(kmp_user_lock_p, kmp_int32)) __kmp_##op##_##l##_##lock,
3081 static int (*indirect_unset[])(kmp_user_lock_p, kmp_int32) = {
3082  KMP_FOREACH_I_LOCK(expand, release)};
3083 static int (*indirect_test[])(kmp_user_lock_p,
3084  kmp_int32) = {KMP_FOREACH_I_LOCK(expand, test)};
3085 #undef expand
3086 #define expand(l, op) \
3087  (int (*)(kmp_user_lock_p, kmp_int32)) __kmp_##op##_##l##_##lock_with_checks,
3088 static int (*indirect_unset_check[])(kmp_user_lock_p, kmp_int32) = {
3089  KMP_FOREACH_I_LOCK(expand, release)};
3090 static int (*indirect_test_check[])(kmp_user_lock_p, kmp_int32) = {
3091  KMP_FOREACH_I_LOCK(expand, test)};
3092 #undef expand
3093 
3094 // Exposes only one jump tables (*lock or *lock_with_checks).
3095 void (**__kmp_indirect_destroy)(kmp_user_lock_p) = 0;
3096 int (**__kmp_indirect_set)(kmp_user_lock_p, kmp_int32) = 0;
3097 int (**__kmp_indirect_unset)(kmp_user_lock_p, kmp_int32) = 0;
3098 int (**__kmp_indirect_test)(kmp_user_lock_p, kmp_int32) = 0;
3099 
3100 // Lock index table.
3101 kmp_indirect_lock_table_t __kmp_i_lock_table;
3102 
3103 // Size of indirect locks.
3104 static kmp_uint32 __kmp_indirect_lock_size[KMP_NUM_I_LOCKS] = {0};
3105 
3106 // Jump tables for lock accessor/modifier.
3107 void (*__kmp_indirect_set_location[KMP_NUM_I_LOCKS])(kmp_user_lock_p,
3108  const ident_t *) = {0};
3109 void (*__kmp_indirect_set_flags[KMP_NUM_I_LOCKS])(kmp_user_lock_p,
3110  kmp_lock_flags_t) = {0};
3111 const ident_t *(*__kmp_indirect_get_location[KMP_NUM_I_LOCKS])(
3112  kmp_user_lock_p) = {0};
3113 kmp_lock_flags_t (*__kmp_indirect_get_flags[KMP_NUM_I_LOCKS])(
3114  kmp_user_lock_p) = {0};
3115 
3116 // Use different lock pools for different lock types.
3117 static kmp_indirect_lock_t *__kmp_indirect_lock_pool[KMP_NUM_I_LOCKS] = {0};
3118 
3119 // User lock allocator for dynamically dispatched indirect locks. Every entry of
3120 // the indirect lock table holds the address and type of the allocated indirect
3121 // lock (kmp_indirect_lock_t), and the size of the table doubles when it is
3122 // full. A destroyed indirect lock object is returned to the reusable pool of
3123 // locks, unique to each lock type.
3124 kmp_indirect_lock_t *__kmp_allocate_indirect_lock(void **user_lock,
3125  kmp_int32 gtid,
3126  kmp_indirect_locktag_t tag) {
3127  kmp_indirect_lock_t *lck;
3128  kmp_lock_index_t idx;
3129 
3130  __kmp_acquire_lock(&__kmp_global_lock, gtid);
3131 
3132  if (__kmp_indirect_lock_pool[tag] != NULL) {
3133  // Reuse the allocated and destroyed lock object
3134  lck = __kmp_indirect_lock_pool[tag];
3135  if (OMP_LOCK_T_SIZE < sizeof(void *))
3136  idx = lck->lock->pool.index;
3137  __kmp_indirect_lock_pool[tag] = (kmp_indirect_lock_t *)lck->lock->pool.next;
3138  KA_TRACE(20, ("__kmp_allocate_indirect_lock: reusing an existing lock %p\n",
3139  lck));
3140  } else {
3141  idx = __kmp_i_lock_table.next;
3142  // Check capacity and double the size if it is full
3143  if (idx == __kmp_i_lock_table.size) {
3144  // Double up the space for block pointers
3145  int row = __kmp_i_lock_table.size / KMP_I_LOCK_CHUNK;
3146  kmp_indirect_lock_t **new_table = (kmp_indirect_lock_t **)__kmp_allocate(
3147  2 * row * sizeof(kmp_indirect_lock_t *));
3148  KMP_MEMCPY(new_table, __kmp_i_lock_table.table,
3149  row * sizeof(kmp_indirect_lock_t *));
3150  kmp_indirect_lock_t **old_table = __kmp_i_lock_table.table;
3151  __kmp_i_lock_table.table = new_table;
3152  __kmp_free(old_table);
3153  // Allocate new objects in the new blocks
3154  for (int i = row; i < 2 * row; ++i)
3155  *(__kmp_i_lock_table.table + i) = (kmp_indirect_lock_t *)__kmp_allocate(
3156  KMP_I_LOCK_CHUNK * sizeof(kmp_indirect_lock_t));
3157  __kmp_i_lock_table.size = 2 * idx;
3158  }
3159  __kmp_i_lock_table.next++;
3160  lck = KMP_GET_I_LOCK(idx);
3161  // Allocate a new base lock object
3162  lck->lock = (kmp_user_lock_p)__kmp_allocate(__kmp_indirect_lock_size[tag]);
3163  KA_TRACE(20,
3164  ("__kmp_allocate_indirect_lock: allocated a new lock %p\n", lck));
3165  }
3166 
3167  __kmp_release_lock(&__kmp_global_lock, gtid);
3168 
3169  lck->type = tag;
3170 
3171  if (OMP_LOCK_T_SIZE < sizeof(void *)) {
3172  *((kmp_lock_index_t *)user_lock) = idx
3173  << 1; // indirect lock word must be even
3174  } else {
3175  *((kmp_indirect_lock_t **)user_lock) = lck;
3176  }
3177 
3178  return lck;
3179 }
3180 
3181 // User lock lookup for dynamically dispatched locks.
3182 static __forceinline kmp_indirect_lock_t *
3183 __kmp_lookup_indirect_lock(void **user_lock, const char *func) {
3184  if (__kmp_env_consistency_check) {
3185  kmp_indirect_lock_t *lck = NULL;
3186  if (user_lock == NULL) {
3187  KMP_FATAL(LockIsUninitialized, func);
3188  }
3189  if (OMP_LOCK_T_SIZE < sizeof(void *)) {
3190  kmp_lock_index_t idx = KMP_EXTRACT_I_INDEX(user_lock);
3191  if (idx >= __kmp_i_lock_table.size) {
3192  KMP_FATAL(LockIsUninitialized, func);
3193  }
3194  lck = KMP_GET_I_LOCK(idx);
3195  } else {
3196  lck = *((kmp_indirect_lock_t **)user_lock);
3197  }
3198  if (lck == NULL) {
3199  KMP_FATAL(LockIsUninitialized, func);
3200  }
3201  return lck;
3202  } else {
3203  if (OMP_LOCK_T_SIZE < sizeof(void *)) {
3204  return KMP_GET_I_LOCK(KMP_EXTRACT_I_INDEX(user_lock));
3205  } else {
3206  return *((kmp_indirect_lock_t **)user_lock);
3207  }
3208  }
3209 }
3210 
3211 static void __kmp_init_indirect_lock(kmp_dyna_lock_t *lock,
3212  kmp_dyna_lockseq_t seq) {
3213 #if KMP_USE_ADAPTIVE_LOCKS
3214  if (seq == lockseq_adaptive && !__kmp_cpuinfo.rtm) {
3215  KMP_WARNING(AdaptiveNotSupported, "kmp_lockseq_t", "adaptive");
3216  seq = lockseq_queuing;
3217  }
3218 #endif
3219 #if KMP_USE_TSX
3220  if (seq == lockseq_rtm_queuing && !__kmp_cpuinfo.rtm) {
3221  seq = lockseq_queuing;
3222  }
3223 #endif
3224  kmp_indirect_locktag_t tag = KMP_GET_I_TAG(seq);
3225  kmp_indirect_lock_t *l =
3226  __kmp_allocate_indirect_lock((void **)lock, __kmp_entry_gtid(), tag);
3227  KMP_I_LOCK_FUNC(l, init)(l->lock);
3228  KA_TRACE(
3229  20, ("__kmp_init_indirect_lock: initialized indirect lock with type#%d\n",
3230  seq));
3231 }
3232 
3233 static void __kmp_destroy_indirect_lock(kmp_dyna_lock_t *lock) {
3234  kmp_uint32 gtid = __kmp_entry_gtid();
3235  kmp_indirect_lock_t *l =
3236  __kmp_lookup_indirect_lock((void **)lock, "omp_destroy_lock");
3237  KMP_I_LOCK_FUNC(l, destroy)(l->lock);
3238  kmp_indirect_locktag_t tag = l->type;
3239 
3240  __kmp_acquire_lock(&__kmp_global_lock, gtid);
3241 
3242  // Use the base lock's space to keep the pool chain.
3243  l->lock->pool.next = (kmp_user_lock_p)__kmp_indirect_lock_pool[tag];
3244  if (OMP_LOCK_T_SIZE < sizeof(void *)) {
3245  l->lock->pool.index = KMP_EXTRACT_I_INDEX(lock);
3246  }
3247  __kmp_indirect_lock_pool[tag] = l;
3248 
3249  __kmp_release_lock(&__kmp_global_lock, gtid);
3250 }
3251 
3252 static int __kmp_set_indirect_lock(kmp_dyna_lock_t *lock, kmp_int32 gtid) {
3253  kmp_indirect_lock_t *l = KMP_LOOKUP_I_LOCK(lock);
3254  return KMP_I_LOCK_FUNC(l, set)(l->lock, gtid);
3255 }
3256 
3257 static int __kmp_unset_indirect_lock(kmp_dyna_lock_t *lock, kmp_int32 gtid) {
3258  kmp_indirect_lock_t *l = KMP_LOOKUP_I_LOCK(lock);
3259  return KMP_I_LOCK_FUNC(l, unset)(l->lock, gtid);
3260 }
3261 
3262 static int __kmp_test_indirect_lock(kmp_dyna_lock_t *lock, kmp_int32 gtid) {
3263  kmp_indirect_lock_t *l = KMP_LOOKUP_I_LOCK(lock);
3264  return KMP_I_LOCK_FUNC(l, test)(l->lock, gtid);
3265 }
3266 
3267 static int __kmp_set_indirect_lock_with_checks(kmp_dyna_lock_t *lock,
3268  kmp_int32 gtid) {
3269  kmp_indirect_lock_t *l =
3270  __kmp_lookup_indirect_lock((void **)lock, "omp_set_lock");
3271  return KMP_I_LOCK_FUNC(l, set)(l->lock, gtid);
3272 }
3273 
3274 static int __kmp_unset_indirect_lock_with_checks(kmp_dyna_lock_t *lock,
3275  kmp_int32 gtid) {
3276  kmp_indirect_lock_t *l =
3277  __kmp_lookup_indirect_lock((void **)lock, "omp_unset_lock");
3278  return KMP_I_LOCK_FUNC(l, unset)(l->lock, gtid);
3279 }
3280 
3281 static int __kmp_test_indirect_lock_with_checks(kmp_dyna_lock_t *lock,
3282  kmp_int32 gtid) {
3283  kmp_indirect_lock_t *l =
3284  __kmp_lookup_indirect_lock((void **)lock, "omp_test_lock");
3285  return KMP_I_LOCK_FUNC(l, test)(l->lock, gtid);
3286 }
3287 
3288 kmp_dyna_lockseq_t __kmp_user_lock_seq = lockseq_queuing;
3289 
3290 // This is used only in kmp_error.cpp when consistency checking is on.
3291 kmp_int32 __kmp_get_user_lock_owner(kmp_user_lock_p lck, kmp_uint32 seq) {
3292  switch (seq) {
3293  case lockseq_tas:
3294  case lockseq_nested_tas:
3295  return __kmp_get_tas_lock_owner((kmp_tas_lock_t *)lck);
3296 #if KMP_USE_FUTEX
3297  case lockseq_futex:
3298  case lockseq_nested_futex:
3299  return __kmp_get_futex_lock_owner((kmp_futex_lock_t *)lck);
3300 #endif
3301  case lockseq_ticket:
3302  case lockseq_nested_ticket:
3303  return __kmp_get_ticket_lock_owner((kmp_ticket_lock_t *)lck);
3304  case lockseq_queuing:
3305  case lockseq_nested_queuing:
3306 #if KMP_USE_ADAPTIVE_LOCKS
3307  case lockseq_adaptive:
3308 #endif
3309  return __kmp_get_queuing_lock_owner((kmp_queuing_lock_t *)lck);
3310  case lockseq_drdpa:
3311  case lockseq_nested_drdpa:
3312  return __kmp_get_drdpa_lock_owner((kmp_drdpa_lock_t *)lck);
3313  default:
3314  return 0;
3315  }
3316 }
3317 
3318 // Initializes data for dynamic user locks.
3319 void __kmp_init_dynamic_user_locks() {
3320  // Initialize jump table for the lock functions
3321  if (__kmp_env_consistency_check) {
3322  __kmp_direct_set = direct_set_check;
3323  __kmp_direct_unset = direct_unset_check;
3324  __kmp_direct_test = direct_test_check;
3325  __kmp_direct_destroy = direct_destroy_check;
3326  __kmp_indirect_set = indirect_set_check;
3327  __kmp_indirect_unset = indirect_unset_check;
3328  __kmp_indirect_test = indirect_test_check;
3329  __kmp_indirect_destroy = indirect_destroy_check;
3330  } else {
3331  __kmp_direct_set = direct_set;
3332  __kmp_direct_unset = direct_unset;
3333  __kmp_direct_test = direct_test;
3334  __kmp_direct_destroy = direct_destroy;
3335  __kmp_indirect_set = indirect_set;
3336  __kmp_indirect_unset = indirect_unset;
3337  __kmp_indirect_test = indirect_test;
3338  __kmp_indirect_destroy = indirect_destroy;
3339  }
3340  // If the user locks have already been initialized, then return. Allow the
3341  // switch between different KMP_CONSISTENCY_CHECK values, but do not allocate
3342  // new lock tables if they have already been allocated.
3343  if (__kmp_init_user_locks)
3344  return;
3345 
3346  // Initialize lock index table
3347  __kmp_i_lock_table.size = KMP_I_LOCK_CHUNK;
3348  __kmp_i_lock_table.table =
3349  (kmp_indirect_lock_t **)__kmp_allocate(sizeof(kmp_indirect_lock_t *));
3350  *(__kmp_i_lock_table.table) = (kmp_indirect_lock_t *)__kmp_allocate(
3351  KMP_I_LOCK_CHUNK * sizeof(kmp_indirect_lock_t));
3352  __kmp_i_lock_table.next = 0;
3353 
3354  // Indirect lock size
3355  __kmp_indirect_lock_size[locktag_ticket] = sizeof(kmp_ticket_lock_t);
3356  __kmp_indirect_lock_size[locktag_queuing] = sizeof(kmp_queuing_lock_t);
3357 #if KMP_USE_ADAPTIVE_LOCKS
3358  __kmp_indirect_lock_size[locktag_adaptive] = sizeof(kmp_adaptive_lock_t);
3359 #endif
3360  __kmp_indirect_lock_size[locktag_drdpa] = sizeof(kmp_drdpa_lock_t);
3361 #if KMP_USE_TSX
3362  __kmp_indirect_lock_size[locktag_rtm_queuing] = sizeof(kmp_queuing_lock_t);
3363 #endif
3364  __kmp_indirect_lock_size[locktag_nested_tas] = sizeof(kmp_tas_lock_t);
3365 #if KMP_USE_FUTEX
3366  __kmp_indirect_lock_size[locktag_nested_futex] = sizeof(kmp_futex_lock_t);
3367 #endif
3368  __kmp_indirect_lock_size[locktag_nested_ticket] = sizeof(kmp_ticket_lock_t);
3369  __kmp_indirect_lock_size[locktag_nested_queuing] = sizeof(kmp_queuing_lock_t);
3370  __kmp_indirect_lock_size[locktag_nested_drdpa] = sizeof(kmp_drdpa_lock_t);
3371 
3372 // Initialize lock accessor/modifier
3373 #define fill_jumps(table, expand, sep) \
3374  { \
3375  table[locktag##sep##ticket] = expand(ticket); \
3376  table[locktag##sep##queuing] = expand(queuing); \
3377  table[locktag##sep##drdpa] = expand(drdpa); \
3378  }
3379 
3380 #if KMP_USE_ADAPTIVE_LOCKS
3381 #define fill_table(table, expand) \
3382  { \
3383  fill_jumps(table, expand, _); \
3384  table[locktag_adaptive] = expand(queuing); \
3385  fill_jumps(table, expand, _nested_); \
3386  }
3387 #else
3388 #define fill_table(table, expand) \
3389  { \
3390  fill_jumps(table, expand, _); \
3391  fill_jumps(table, expand, _nested_); \
3392  }
3393 #endif // KMP_USE_ADAPTIVE_LOCKS
3394 
3395 #define expand(l) \
3396  (void (*)(kmp_user_lock_p, const ident_t *)) __kmp_set_##l##_lock_location
3397  fill_table(__kmp_indirect_set_location, expand);
3398 #undef expand
3399 #define expand(l) \
3400  (void (*)(kmp_user_lock_p, kmp_lock_flags_t)) __kmp_set_##l##_lock_flags
3401  fill_table(__kmp_indirect_set_flags, expand);
3402 #undef expand
3403 #define expand(l) \
3404  (const ident_t *(*)(kmp_user_lock_p)) __kmp_get_##l##_lock_location
3405  fill_table(__kmp_indirect_get_location, expand);
3406 #undef expand
3407 #define expand(l) \
3408  (kmp_lock_flags_t(*)(kmp_user_lock_p)) __kmp_get_##l##_lock_flags
3409  fill_table(__kmp_indirect_get_flags, expand);
3410 #undef expand
3411 
3412  __kmp_init_user_locks = TRUE;
3413 }
3414 
3415 // Clean up the lock table.
3416 void __kmp_cleanup_indirect_user_locks() {
3417  kmp_lock_index_t i;
3418  int k;
3419 
3420  // Clean up locks in the pools first (they were already destroyed before going
3421  // into the pools).
3422  for (k = 0; k < KMP_NUM_I_LOCKS; ++k) {
3423  kmp_indirect_lock_t *l = __kmp_indirect_lock_pool[k];
3424  while (l != NULL) {
3425  kmp_indirect_lock_t *ll = l;
3426  l = (kmp_indirect_lock_t *)l->lock->pool.next;
3427  KA_TRACE(20, ("__kmp_cleanup_indirect_user_locks: freeing %p from pool\n",
3428  ll));
3429  __kmp_free(ll->lock);
3430  ll->lock = NULL;
3431  }
3432  __kmp_indirect_lock_pool[k] = NULL;
3433  }
3434  // Clean up the remaining undestroyed locks.
3435  for (i = 0; i < __kmp_i_lock_table.next; i++) {
3436  kmp_indirect_lock_t *l = KMP_GET_I_LOCK(i);
3437  if (l->lock != NULL) {
3438  // Locks not destroyed explicitly need to be destroyed here.
3439  KMP_I_LOCK_FUNC(l, destroy)(l->lock);
3440  KA_TRACE(
3441  20,
3442  ("__kmp_cleanup_indirect_user_locks: destroy/freeing %p from table\n",
3443  l));
3444  __kmp_free(l->lock);
3445  }
3446  }
3447  // Free the table
3448  for (i = 0; i < __kmp_i_lock_table.size / KMP_I_LOCK_CHUNK; i++)
3449  __kmp_free(__kmp_i_lock_table.table[i]);
3450  __kmp_free(__kmp_i_lock_table.table);
3451 
3452  __kmp_init_user_locks = FALSE;
3453 }
3454 
3455 enum kmp_lock_kind __kmp_user_lock_kind = lk_default;
3456 int __kmp_num_locks_in_block = 1; // FIXME - tune this value
3457 
3458 #else // KMP_USE_DYNAMIC_LOCK
3459 
3460 static void __kmp_init_tas_lock_with_checks(kmp_tas_lock_t *lck) {
3461  __kmp_init_tas_lock(lck);
3462 }
3463 
3464 static void __kmp_init_nested_tas_lock_with_checks(kmp_tas_lock_t *lck) {
3465  __kmp_init_nested_tas_lock(lck);
3466 }
3467 
3468 #if KMP_USE_FUTEX
3469 static void __kmp_init_futex_lock_with_checks(kmp_futex_lock_t *lck) {
3470  __kmp_init_futex_lock(lck);
3471 }
3472 
3473 static void __kmp_init_nested_futex_lock_with_checks(kmp_futex_lock_t *lck) {
3474  __kmp_init_nested_futex_lock(lck);
3475 }
3476 #endif
3477 
3478 static int __kmp_is_ticket_lock_initialized(kmp_ticket_lock_t *lck) {
3479  return lck == lck->lk.self;
3480 }
3481 
3482 static void __kmp_init_ticket_lock_with_checks(kmp_ticket_lock_t *lck) {
3483  __kmp_init_ticket_lock(lck);
3484 }
3485 
3486 static void __kmp_init_nested_ticket_lock_with_checks(kmp_ticket_lock_t *lck) {
3487  __kmp_init_nested_ticket_lock(lck);
3488 }
3489 
3490 static int __kmp_is_queuing_lock_initialized(kmp_queuing_lock_t *lck) {
3491  return lck == lck->lk.initialized;
3492 }
3493 
3494 static void __kmp_init_queuing_lock_with_checks(kmp_queuing_lock_t *lck) {
3495  __kmp_init_queuing_lock(lck);
3496 }
3497 
3498 static void
3499 __kmp_init_nested_queuing_lock_with_checks(kmp_queuing_lock_t *lck) {
3500  __kmp_init_nested_queuing_lock(lck);
3501 }
3502 
3503 #if KMP_USE_ADAPTIVE_LOCKS
3504 static void __kmp_init_adaptive_lock_with_checks(kmp_adaptive_lock_t *lck) {
3505  __kmp_init_adaptive_lock(lck);
3506 }
3507 #endif
3508 
3509 static int __kmp_is_drdpa_lock_initialized(kmp_drdpa_lock_t *lck) {
3510  return lck == lck->lk.initialized;
3511 }
3512 
3513 static void __kmp_init_drdpa_lock_with_checks(kmp_drdpa_lock_t *lck) {
3514  __kmp_init_drdpa_lock(lck);
3515 }
3516 
3517 static void __kmp_init_nested_drdpa_lock_with_checks(kmp_drdpa_lock_t *lck) {
3518  __kmp_init_nested_drdpa_lock(lck);
3519 }
3520 
3521 /* user locks
3522  * They are implemented as a table of function pointers which are set to the
3523  * lock functions of the appropriate kind, once that has been determined. */
3524 
3525 enum kmp_lock_kind __kmp_user_lock_kind = lk_default;
3526 
3527 size_t __kmp_base_user_lock_size = 0;
3528 size_t __kmp_user_lock_size = 0;
3529 
3530 kmp_int32 (*__kmp_get_user_lock_owner_)(kmp_user_lock_p lck) = NULL;
3531 int (*__kmp_acquire_user_lock_with_checks_)(kmp_user_lock_p lck,
3532  kmp_int32 gtid) = NULL;
3533 
3534 int (*__kmp_test_user_lock_with_checks_)(kmp_user_lock_p lck,
3535  kmp_int32 gtid) = NULL;
3536 int (*__kmp_release_user_lock_with_checks_)(kmp_user_lock_p lck,
3537  kmp_int32 gtid) = NULL;
3538 void (*__kmp_init_user_lock_with_checks_)(kmp_user_lock_p lck) = NULL;
3539 void (*__kmp_destroy_user_lock_)(kmp_user_lock_p lck) = NULL;
3540 void (*__kmp_destroy_user_lock_with_checks_)(kmp_user_lock_p lck) = NULL;
3541 int (*__kmp_acquire_nested_user_lock_with_checks_)(kmp_user_lock_p lck,
3542  kmp_int32 gtid) = NULL;
3543 
3544 int (*__kmp_test_nested_user_lock_with_checks_)(kmp_user_lock_p lck,
3545  kmp_int32 gtid) = NULL;
3546 int (*__kmp_release_nested_user_lock_with_checks_)(kmp_user_lock_p lck,
3547  kmp_int32 gtid) = NULL;
3548 void (*__kmp_init_nested_user_lock_with_checks_)(kmp_user_lock_p lck) = NULL;
3549 void (*__kmp_destroy_nested_user_lock_with_checks_)(kmp_user_lock_p lck) = NULL;
3550 
3551 int (*__kmp_is_user_lock_initialized_)(kmp_user_lock_p lck) = NULL;
3552 const ident_t *(*__kmp_get_user_lock_location_)(kmp_user_lock_p lck) = NULL;
3553 void (*__kmp_set_user_lock_location_)(kmp_user_lock_p lck,
3554  const ident_t *loc) = NULL;
3555 kmp_lock_flags_t (*__kmp_get_user_lock_flags_)(kmp_user_lock_p lck) = NULL;
3556 void (*__kmp_set_user_lock_flags_)(kmp_user_lock_p lck,
3557  kmp_lock_flags_t flags) = NULL;
3558 
3559 void __kmp_set_user_lock_vptrs(kmp_lock_kind_t user_lock_kind) {
3560  switch (user_lock_kind) {
3561  case lk_default:
3562  default:
3563  KMP_ASSERT(0);
3564 
3565  case lk_tas: {
3566  __kmp_base_user_lock_size = sizeof(kmp_base_tas_lock_t);
3567  __kmp_user_lock_size = sizeof(kmp_tas_lock_t);
3568 
3569  __kmp_get_user_lock_owner_ =
3570  (kmp_int32(*)(kmp_user_lock_p))(&__kmp_get_tas_lock_owner);
3571 
3572  if (__kmp_env_consistency_check) {
3573  KMP_BIND_USER_LOCK_WITH_CHECKS(tas);
3574  KMP_BIND_NESTED_USER_LOCK_WITH_CHECKS(tas);
3575  } else {
3576  KMP_BIND_USER_LOCK(tas);
3577  KMP_BIND_NESTED_USER_LOCK(tas);
3578  }
3579 
3580  __kmp_destroy_user_lock_ =
3581  (void (*)(kmp_user_lock_p))(&__kmp_destroy_tas_lock);
3582 
3583  __kmp_is_user_lock_initialized_ = (int (*)(kmp_user_lock_p))NULL;
3584 
3585  __kmp_get_user_lock_location_ = (const ident_t *(*)(kmp_user_lock_p))NULL;
3586 
3587  __kmp_set_user_lock_location_ =
3588  (void (*)(kmp_user_lock_p, const ident_t *))NULL;
3589 
3590  __kmp_get_user_lock_flags_ = (kmp_lock_flags_t(*)(kmp_user_lock_p))NULL;
3591 
3592  __kmp_set_user_lock_flags_ =
3593  (void (*)(kmp_user_lock_p, kmp_lock_flags_t))NULL;
3594  } break;
3595 
3596 #if KMP_USE_FUTEX
3597 
3598  case lk_futex: {
3599  __kmp_base_user_lock_size = sizeof(kmp_base_futex_lock_t);
3600  __kmp_user_lock_size = sizeof(kmp_futex_lock_t);
3601 
3602  __kmp_get_user_lock_owner_ =
3603  (kmp_int32(*)(kmp_user_lock_p))(&__kmp_get_futex_lock_owner);
3604 
3605  if (__kmp_env_consistency_check) {
3606  KMP_BIND_USER_LOCK_WITH_CHECKS(futex);
3607  KMP_BIND_NESTED_USER_LOCK_WITH_CHECKS(futex);
3608  } else {
3609  KMP_BIND_USER_LOCK(futex);
3610  KMP_BIND_NESTED_USER_LOCK(futex);
3611  }
3612 
3613  __kmp_destroy_user_lock_ =
3614  (void (*)(kmp_user_lock_p))(&__kmp_destroy_futex_lock);
3615 
3616  __kmp_is_user_lock_initialized_ = (int (*)(kmp_user_lock_p))NULL;
3617 
3618  __kmp_get_user_lock_location_ = (const ident_t *(*)(kmp_user_lock_p))NULL;
3619 
3620  __kmp_set_user_lock_location_ =
3621  (void (*)(kmp_user_lock_p, const ident_t *))NULL;
3622 
3623  __kmp_get_user_lock_flags_ = (kmp_lock_flags_t(*)(kmp_user_lock_p))NULL;
3624 
3625  __kmp_set_user_lock_flags_ =
3626  (void (*)(kmp_user_lock_p, kmp_lock_flags_t))NULL;
3627  } break;
3628 
3629 #endif // KMP_USE_FUTEX
3630 
3631  case lk_ticket: {
3632  __kmp_base_user_lock_size = sizeof(kmp_base_ticket_lock_t);
3633  __kmp_user_lock_size = sizeof(kmp_ticket_lock_t);
3634 
3635  __kmp_get_user_lock_owner_ =
3636  (kmp_int32(*)(kmp_user_lock_p))(&__kmp_get_ticket_lock_owner);
3637 
3638  if (__kmp_env_consistency_check) {
3639  KMP_BIND_USER_LOCK_WITH_CHECKS(ticket);
3640  KMP_BIND_NESTED_USER_LOCK_WITH_CHECKS(ticket);
3641  } else {
3642  KMP_BIND_USER_LOCK(ticket);
3643  KMP_BIND_NESTED_USER_LOCK(ticket);
3644  }
3645 
3646  __kmp_destroy_user_lock_ =
3647  (void (*)(kmp_user_lock_p))(&__kmp_destroy_ticket_lock);
3648 
3649  __kmp_is_user_lock_initialized_ =
3650  (int (*)(kmp_user_lock_p))(&__kmp_is_ticket_lock_initialized);
3651 
3652  __kmp_get_user_lock_location_ =
3653  (const ident_t *(*)(kmp_user_lock_p))(&__kmp_get_ticket_lock_location);
3654 
3655  __kmp_set_user_lock_location_ = (void (*)(
3656  kmp_user_lock_p, const ident_t *))(&__kmp_set_ticket_lock_location);
3657 
3658  __kmp_get_user_lock_flags_ =
3659  (kmp_lock_flags_t(*)(kmp_user_lock_p))(&__kmp_get_ticket_lock_flags);
3660 
3661  __kmp_set_user_lock_flags_ = (void (*)(kmp_user_lock_p, kmp_lock_flags_t))(
3662  &__kmp_set_ticket_lock_flags);
3663  } break;
3664 
3665  case lk_queuing: {
3666  __kmp_base_user_lock_size = sizeof(kmp_base_queuing_lock_t);
3667  __kmp_user_lock_size = sizeof(kmp_queuing_lock_t);
3668 
3669  __kmp_get_user_lock_owner_ =
3670  (kmp_int32(*)(kmp_user_lock_p))(&__kmp_get_queuing_lock_owner);
3671 
3672  if (__kmp_env_consistency_check) {
3673  KMP_BIND_USER_LOCK_WITH_CHECKS(queuing);
3674  KMP_BIND_NESTED_USER_LOCK_WITH_CHECKS(queuing);
3675  } else {
3676  KMP_BIND_USER_LOCK(queuing);
3677  KMP_BIND_NESTED_USER_LOCK(queuing);
3678  }
3679 
3680  __kmp_destroy_user_lock_ =
3681  (void (*)(kmp_user_lock_p))(&__kmp_destroy_queuing_lock);
3682 
3683  __kmp_is_user_lock_initialized_ =
3684  (int (*)(kmp_user_lock_p))(&__kmp_is_queuing_lock_initialized);
3685 
3686  __kmp_get_user_lock_location_ =
3687  (const ident_t *(*)(kmp_user_lock_p))(&__kmp_get_queuing_lock_location);
3688 
3689  __kmp_set_user_lock_location_ = (void (*)(
3690  kmp_user_lock_p, const ident_t *))(&__kmp_set_queuing_lock_location);
3691 
3692  __kmp_get_user_lock_flags_ =
3693  (kmp_lock_flags_t(*)(kmp_user_lock_p))(&__kmp_get_queuing_lock_flags);
3694 
3695  __kmp_set_user_lock_flags_ = (void (*)(kmp_user_lock_p, kmp_lock_flags_t))(
3696  &__kmp_set_queuing_lock_flags);
3697  } break;
3698 
3699 #if KMP_USE_ADAPTIVE_LOCKS
3700  case lk_adaptive: {
3701  __kmp_base_user_lock_size = sizeof(kmp_base_adaptive_lock_t);
3702  __kmp_user_lock_size = sizeof(kmp_adaptive_lock_t);
3703 
3704  __kmp_get_user_lock_owner_ =
3705  (kmp_int32(*)(kmp_user_lock_p))(&__kmp_get_queuing_lock_owner);
3706 
3707  if (__kmp_env_consistency_check) {
3708  KMP_BIND_USER_LOCK_WITH_CHECKS(adaptive);
3709  } else {
3710  KMP_BIND_USER_LOCK(adaptive);
3711  }
3712 
3713  __kmp_destroy_user_lock_ =
3714  (void (*)(kmp_user_lock_p))(&__kmp_destroy_adaptive_lock);
3715 
3716  __kmp_is_user_lock_initialized_ =
3717  (int (*)(kmp_user_lock_p))(&__kmp_is_queuing_lock_initialized);
3718 
3719  __kmp_get_user_lock_location_ =
3720  (const ident_t *(*)(kmp_user_lock_p))(&__kmp_get_queuing_lock_location);
3721 
3722  __kmp_set_user_lock_location_ = (void (*)(
3723  kmp_user_lock_p, const ident_t *))(&__kmp_set_queuing_lock_location);
3724 
3725  __kmp_get_user_lock_flags_ =
3726  (kmp_lock_flags_t(*)(kmp_user_lock_p))(&__kmp_get_queuing_lock_flags);
3727 
3728  __kmp_set_user_lock_flags_ = (void (*)(kmp_user_lock_p, kmp_lock_flags_t))(
3729  &__kmp_set_queuing_lock_flags);
3730 
3731  } break;
3732 #endif // KMP_USE_ADAPTIVE_LOCKS
3733 
3734  case lk_drdpa: {
3735  __kmp_base_user_lock_size = sizeof(kmp_base_drdpa_lock_t);
3736  __kmp_user_lock_size = sizeof(kmp_drdpa_lock_t);
3737 
3738  __kmp_get_user_lock_owner_ =
3739  (kmp_int32(*)(kmp_user_lock_p))(&__kmp_get_drdpa_lock_owner);
3740 
3741  if (__kmp_env_consistency_check) {
3742  KMP_BIND_USER_LOCK_WITH_CHECKS(drdpa);
3743  KMP_BIND_NESTED_USER_LOCK_WITH_CHECKS(drdpa);
3744  } else {
3745  KMP_BIND_USER_LOCK(drdpa);
3746  KMP_BIND_NESTED_USER_LOCK(drdpa);
3747  }
3748 
3749  __kmp_destroy_user_lock_ =
3750  (void (*)(kmp_user_lock_p))(&__kmp_destroy_drdpa_lock);
3751 
3752  __kmp_is_user_lock_initialized_ =
3753  (int (*)(kmp_user_lock_p))(&__kmp_is_drdpa_lock_initialized);
3754 
3755  __kmp_get_user_lock_location_ =
3756  (const ident_t *(*)(kmp_user_lock_p))(&__kmp_get_drdpa_lock_location);
3757 
3758  __kmp_set_user_lock_location_ = (void (*)(
3759  kmp_user_lock_p, const ident_t *))(&__kmp_set_drdpa_lock_location);
3760 
3761  __kmp_get_user_lock_flags_ =
3762  (kmp_lock_flags_t(*)(kmp_user_lock_p))(&__kmp_get_drdpa_lock_flags);
3763 
3764  __kmp_set_user_lock_flags_ = (void (*)(kmp_user_lock_p, kmp_lock_flags_t))(
3765  &__kmp_set_drdpa_lock_flags);
3766  } break;
3767  }
3768 }
3769 
3770 // ----------------------------------------------------------------------------
3771 // User lock table & lock allocation
3772 
3773 kmp_lock_table_t __kmp_user_lock_table = {1, 0, NULL};
3774 kmp_user_lock_p __kmp_lock_pool = NULL;
3775 
3776 // Lock block-allocation support.
3777 kmp_block_of_locks *__kmp_lock_blocks = NULL;
3778 int __kmp_num_locks_in_block = 1; // FIXME - tune this value
3779 
3780 static kmp_lock_index_t __kmp_lock_table_insert(kmp_user_lock_p lck) {
3781  // Assume that kmp_global_lock is held upon entry/exit.
3782  kmp_lock_index_t index;
3783  if (__kmp_user_lock_table.used >= __kmp_user_lock_table.allocated) {
3784  kmp_lock_index_t size;
3785  kmp_user_lock_p *table;
3786  // Reallocate lock table.
3787  if (__kmp_user_lock_table.allocated == 0) {
3788  size = 1024;
3789  } else {
3790  size = __kmp_user_lock_table.allocated * 2;
3791  }
3792  table = (kmp_user_lock_p *)__kmp_allocate(sizeof(kmp_user_lock_p) * size);
3793  KMP_MEMCPY(table + 1, __kmp_user_lock_table.table + 1,
3794  sizeof(kmp_user_lock_p) * (__kmp_user_lock_table.used - 1));
3795  table[0] = (kmp_user_lock_p)__kmp_user_lock_table.table;
3796  // We cannot free the previous table now, since it may be in use by other
3797  // threads. So save the pointer to the previous table in in the first
3798  // element of the new table. All the tables will be organized into a list,
3799  // and could be freed when library shutting down.
3800  __kmp_user_lock_table.table = table;
3801  __kmp_user_lock_table.allocated = size;
3802  }
3803  KMP_DEBUG_ASSERT(__kmp_user_lock_table.used <
3804  __kmp_user_lock_table.allocated);
3805  index = __kmp_user_lock_table.used;
3806  __kmp_user_lock_table.table[index] = lck;
3807  ++__kmp_user_lock_table.used;
3808  return index;
3809 }
3810 
3811 static kmp_user_lock_p __kmp_lock_block_allocate() {
3812  // Assume that kmp_global_lock is held upon entry/exit.
3813  static int last_index = 0;
3814  if ((last_index >= __kmp_num_locks_in_block) || (__kmp_lock_blocks == NULL)) {
3815  // Restart the index.
3816  last_index = 0;
3817  // Need to allocate a new block.
3818  KMP_DEBUG_ASSERT(__kmp_user_lock_size > 0);
3819  size_t space_for_locks = __kmp_user_lock_size * __kmp_num_locks_in_block;
3820  char *buffer =
3821  (char *)__kmp_allocate(space_for_locks + sizeof(kmp_block_of_locks));
3822  // Set up the new block.
3823  kmp_block_of_locks *new_block =
3824  (kmp_block_of_locks *)(&buffer[space_for_locks]);
3825  new_block->next_block = __kmp_lock_blocks;
3826  new_block->locks = (void *)buffer;
3827  // Publish the new block.
3828  KMP_MB();
3829  __kmp_lock_blocks = new_block;
3830  }
3831  kmp_user_lock_p ret = (kmp_user_lock_p)(&(
3832  ((char *)(__kmp_lock_blocks->locks))[last_index * __kmp_user_lock_size]));
3833  last_index++;
3834  return ret;
3835 }
3836 
3837 // Get memory for a lock. It may be freshly allocated memory or reused memory
3838 // from lock pool.
3839 kmp_user_lock_p __kmp_user_lock_allocate(void **user_lock, kmp_int32 gtid,
3840  kmp_lock_flags_t flags) {
3841  kmp_user_lock_p lck;
3842  kmp_lock_index_t index;
3843  KMP_DEBUG_ASSERT(user_lock);
3844 
3845  __kmp_acquire_lock(&__kmp_global_lock, gtid);
3846 
3847  if (__kmp_lock_pool == NULL) {
3848  // Lock pool is empty. Allocate new memory.
3849 
3850  // ANNOTATION: Found no good way to express the syncronisation
3851  // between allocation and usage, so ignore the allocation
3852  ANNOTATE_IGNORE_WRITES_BEGIN();
3853  if (__kmp_num_locks_in_block <= 1) { // Tune this cutoff point.
3854  lck = (kmp_user_lock_p)__kmp_allocate(__kmp_user_lock_size);
3855  } else {
3856  lck = __kmp_lock_block_allocate();
3857  }
3858  ANNOTATE_IGNORE_WRITES_END();
3859 
3860  // Insert lock in the table so that it can be freed in __kmp_cleanup,
3861  // and debugger has info on all allocated locks.
3862  index = __kmp_lock_table_insert(lck);
3863  } else {
3864  // Pick up lock from pool.
3865  lck = __kmp_lock_pool;
3866  index = __kmp_lock_pool->pool.index;
3867  __kmp_lock_pool = __kmp_lock_pool->pool.next;
3868  }
3869 
3870  // We could potentially differentiate between nested and regular locks
3871  // here, and do the lock table lookup for regular locks only.
3872  if (OMP_LOCK_T_SIZE < sizeof(void *)) {
3873  *((kmp_lock_index_t *)user_lock) = index;
3874  } else {
3875  *((kmp_user_lock_p *)user_lock) = lck;
3876  }
3877 
3878  // mark the lock if it is critical section lock.
3879  __kmp_set_user_lock_flags(lck, flags);
3880 
3881  __kmp_release_lock(&__kmp_global_lock, gtid); // AC: TODO move this line upper
3882 
3883  return lck;
3884 }
3885 
3886 // Put lock's memory to pool for reusing.
3887 void __kmp_user_lock_free(void **user_lock, kmp_int32 gtid,
3888  kmp_user_lock_p lck) {
3889  KMP_DEBUG_ASSERT(user_lock != NULL);
3890  KMP_DEBUG_ASSERT(lck != NULL);
3891 
3892  __kmp_acquire_lock(&__kmp_global_lock, gtid);
3893 
3894  lck->pool.next = __kmp_lock_pool;
3895  __kmp_lock_pool = lck;
3896  if (OMP_LOCK_T_SIZE < sizeof(void *)) {
3897  kmp_lock_index_t index = *((kmp_lock_index_t *)user_lock);
3898  KMP_DEBUG_ASSERT(0 < index && index <= __kmp_user_lock_table.used);
3899  lck->pool.index = index;
3900  }
3901 
3902  __kmp_release_lock(&__kmp_global_lock, gtid);
3903 }
3904 
3905 kmp_user_lock_p __kmp_lookup_user_lock(void **user_lock, char const *func) {
3906  kmp_user_lock_p lck = NULL;
3907 
3908  if (__kmp_env_consistency_check) {
3909  if (user_lock == NULL) {
3910  KMP_FATAL(LockIsUninitialized, func);
3911  }
3912  }
3913 
3914  if (OMP_LOCK_T_SIZE < sizeof(void *)) {
3915  kmp_lock_index_t index = *((kmp_lock_index_t *)user_lock);
3916  if (__kmp_env_consistency_check) {
3917  if (!(0 < index && index < __kmp_user_lock_table.used)) {
3918  KMP_FATAL(LockIsUninitialized, func);
3919  }
3920  }
3921  KMP_DEBUG_ASSERT(0 < index && index < __kmp_user_lock_table.used);
3922  KMP_DEBUG_ASSERT(__kmp_user_lock_size > 0);
3923  lck = __kmp_user_lock_table.table[index];
3924  } else {
3925  lck = *((kmp_user_lock_p *)user_lock);
3926  }
3927 
3928  if (__kmp_env_consistency_check) {
3929  if (lck == NULL) {
3930  KMP_FATAL(LockIsUninitialized, func);
3931  }
3932  }
3933 
3934  return lck;
3935 }
3936 
3937 void __kmp_cleanup_user_locks(void) {
3938  // Reset lock pool. Don't worry about lock in the pool--we will free them when
3939  // iterating through lock table (it includes all the locks, dead or alive).
3940  __kmp_lock_pool = NULL;
3941 
3942 #define IS_CRITICAL(lck) \
3943  ((__kmp_get_user_lock_flags_ != NULL) && \
3944  ((*__kmp_get_user_lock_flags_)(lck)&kmp_lf_critical_section))
3945 
3946  // Loop through lock table, free all locks.
3947  // Do not free item [0], it is reserved for lock tables list.
3948  //
3949  // FIXME - we are iterating through a list of (pointers to) objects of type
3950  // union kmp_user_lock, but we have no way of knowing whether the base type is
3951  // currently "pool" or whatever the global user lock type is.
3952  //
3953  // We are relying on the fact that for all of the user lock types
3954  // (except "tas"), the first field in the lock struct is the "initialized"
3955  // field, which is set to the address of the lock object itself when
3956  // the lock is initialized. When the union is of type "pool", the
3957  // first field is a pointer to the next object in the free list, which
3958  // will not be the same address as the object itself.
3959  //
3960  // This means that the check (*__kmp_is_user_lock_initialized_)(lck) will fail
3961  // for "pool" objects on the free list. This must happen as the "location"
3962  // field of real user locks overlaps the "index" field of "pool" objects.
3963  //
3964  // It would be better to run through the free list, and remove all "pool"
3965  // objects from the lock table before executing this loop. However,
3966  // "pool" objects do not always have their index field set (only on
3967  // lin_32e), and I don't want to search the lock table for the address
3968  // of every "pool" object on the free list.
3969  while (__kmp_user_lock_table.used > 1) {
3970  const ident *loc;
3971 
3972  // reduce __kmp_user_lock_table.used before freeing the lock,
3973  // so that state of locks is consistent
3974  kmp_user_lock_p lck =
3975  __kmp_user_lock_table.table[--__kmp_user_lock_table.used];
3976 
3977  if ((__kmp_is_user_lock_initialized_ != NULL) &&
3978  (*__kmp_is_user_lock_initialized_)(lck)) {
3979  // Issue a warning if: KMP_CONSISTENCY_CHECK AND lock is initialized AND
3980  // it is NOT a critical section (user is not responsible for destroying
3981  // criticals) AND we know source location to report.
3982  if (__kmp_env_consistency_check && (!IS_CRITICAL(lck)) &&
3983  ((loc = __kmp_get_user_lock_location(lck)) != NULL) &&
3984  (loc->psource != NULL)) {
3985  kmp_str_loc_t str_loc = __kmp_str_loc_init(loc->psource, false);
3986  KMP_WARNING(CnsLockNotDestroyed, str_loc.file, str_loc.line);
3987  __kmp_str_loc_free(&str_loc);
3988  }
3989 
3990 #ifdef KMP_DEBUG
3991  if (IS_CRITICAL(lck)) {
3992  KA_TRACE(
3993  20,
3994  ("__kmp_cleanup_user_locks: free critical section lock %p (%p)\n",
3995  lck, *(void **)lck));
3996  } else {
3997  KA_TRACE(20, ("__kmp_cleanup_user_locks: free lock %p (%p)\n", lck,
3998  *(void **)lck));
3999  }
4000 #endif // KMP_DEBUG
4001 
4002  // Cleanup internal lock dynamic resources (for drdpa locks particularly).
4003  __kmp_destroy_user_lock(lck);
4004  }
4005 
4006  // Free the lock if block allocation of locks is not used.
4007  if (__kmp_lock_blocks == NULL) {
4008  __kmp_free(lck);
4009  }
4010  }
4011 
4012 #undef IS_CRITICAL
4013 
4014  // delete lock table(s).
4015  kmp_user_lock_p *table_ptr = __kmp_user_lock_table.table;
4016  __kmp_user_lock_table.table = NULL;
4017  __kmp_user_lock_table.allocated = 0;
4018 
4019  while (table_ptr != NULL) {
4020  // In the first element we saved the pointer to the previous
4021  // (smaller) lock table.
4022  kmp_user_lock_p *next = (kmp_user_lock_p *)(table_ptr[0]);
4023  __kmp_free(table_ptr);
4024  table_ptr = next;
4025  }
4026 
4027  // Free buffers allocated for blocks of locks.
4028  kmp_block_of_locks_t *block_ptr = __kmp_lock_blocks;
4029  __kmp_lock_blocks = NULL;
4030 
4031  while (block_ptr != NULL) {
4032  kmp_block_of_locks_t *next = block_ptr->next_block;
4033  __kmp_free(block_ptr->locks);
4034  // *block_ptr itself was allocated at the end of the locks vector.
4035  block_ptr = next;
4036  }
4037 
4038  TCW_4(__kmp_init_user_locks, FALSE);
4039 }
4040 
4041 #endif // KMP_USE_DYNAMIC_LOCK
void set_stdout()
Definition: kmp.h:4189
void open(const char *filename, const char *mode, const char *env_var=nullptr)
Definition: kmp.h:4172
Definition: kmp.h:229
char const * psource
Definition: kmp.h:239