xref: /freebsd/contrib/llvm-project/openmp/runtime/src/kmp_lock.h (revision ec0ea6efa1ad229d75c394c1a9b9cac33af2b1d3)
1 /*
2  * kmp_lock.h -- lock header file
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 #ifndef KMP_LOCK_H
14 #define KMP_LOCK_H
15 
16 #include <limits.h> // CHAR_BIT
17 #include <stddef.h> // offsetof
18 
19 #include "kmp_debug.h"
20 #include "kmp_os.h"
21 
22 #ifdef __cplusplus
23 #include <atomic>
24 
25 extern "C" {
26 #endif // __cplusplus
27 
28 // ----------------------------------------------------------------------------
29 // Have to copy these definitions from kmp.h because kmp.h cannot be included
30 // due to circular dependencies.  Will undef these at end of file.
31 
32 #define KMP_PAD(type, sz)                                                      \
33   (sizeof(type) + (sz - ((sizeof(type) - 1) % (sz)) - 1))
34 #define KMP_GTID_DNE (-2)
35 
36 // Forward declaration of ident and ident_t
37 
38 struct ident;
39 typedef struct ident ident_t;
40 
41 // End of copied code.
42 // ----------------------------------------------------------------------------
43 
44 // We need to know the size of the area we can assume that the compiler(s)
45 // allocated for objects of type omp_lock_t and omp_nest_lock_t.  The Intel
46 // compiler always allocates a pointer-sized area, as does visual studio.
47 //
48 // gcc however, only allocates 4 bytes for regular locks, even on 64-bit
49 // intel archs.  It allocates at least 8 bytes for nested lock (more on
50 // recent versions), but we are bounded by the pointer-sized chunks that
51 // the Intel compiler allocates.
52 
53 #if KMP_OS_LINUX && defined(KMP_GOMP_COMPAT)
54 #define OMP_LOCK_T_SIZE sizeof(int)
55 #define OMP_NEST_LOCK_T_SIZE sizeof(void *)
56 #else
57 #define OMP_LOCK_T_SIZE sizeof(void *)
58 #define OMP_NEST_LOCK_T_SIZE sizeof(void *)
59 #endif
60 
61 // The Intel compiler allocates a 32-byte chunk for a critical section.
62 // Both gcc and visual studio only allocate enough space for a pointer.
63 // Sometimes we know that the space was allocated by the Intel compiler.
64 #define OMP_CRITICAL_SIZE sizeof(void *)
65 #define INTEL_CRITICAL_SIZE 32
66 
67 // lock flags
68 typedef kmp_uint32 kmp_lock_flags_t;
69 
70 #define kmp_lf_critical_section 1
71 
72 // When a lock table is used, the indices are of kmp_lock_index_t
73 typedef kmp_uint32 kmp_lock_index_t;
74 
75 // When memory allocated for locks are on the lock pool (free list),
76 // it is treated as structs of this type.
77 struct kmp_lock_pool {
78   union kmp_user_lock *next;
79   kmp_lock_index_t index;
80 };
81 
82 typedef struct kmp_lock_pool kmp_lock_pool_t;
83 
84 extern void __kmp_validate_locks(void);
85 
86 // ----------------------------------------------------------------------------
87 //  There are 5 lock implementations:
88 //       1. Test and set locks.
89 //       2. futex locks (Linux* OS on x86 and
90 //          Intel(R) Many Integrated Core Architecture)
91 //       3. Ticket (Lamport bakery) locks.
92 //       4. Queuing locks (with separate spin fields).
93 //       5. DRPA (Dynamically Reconfigurable Distributed Polling Area) locks
94 //
95 //   and 3 lock purposes:
96 //       1. Bootstrap locks -- Used for a few locks available at library
97 //       startup-shutdown time.
98 //          These do not require non-negative global thread ID's.
99 //       2. Internal RTL locks -- Used everywhere else in the RTL
100 //       3. User locks (includes critical sections)
101 // ----------------------------------------------------------------------------
102 
103 // ============================================================================
104 // Lock implementations.
105 //
106 // Test and set locks.
107 //
108 // Non-nested test and set locks differ from the other lock kinds (except
109 // futex) in that we use the memory allocated by the compiler for the lock,
110 // rather than a pointer to it.
111 //
112 // On lin32, lin_32e, and win_32, the space allocated may be as small as 4
113 // bytes, so we have to use a lock table for nested locks, and avoid accessing
114 // the depth_locked field for non-nested locks.
115 //
116 // Information normally available to the tools, such as lock location, lock
117 // usage (normal lock vs. critical section), etc. is not available with test and
118 // set locks.
119 // ----------------------------------------------------------------------------
120 
121 struct kmp_base_tas_lock {
122   // KMP_LOCK_FREE(tas) => unlocked; locked: (gtid+1) of owning thread
123   std::atomic<kmp_int32> poll;
124   kmp_int32 depth_locked; // depth locked, for nested locks only
125 };
126 
127 typedef struct kmp_base_tas_lock kmp_base_tas_lock_t;
128 
129 union kmp_tas_lock {
130   kmp_base_tas_lock_t lk;
131   kmp_lock_pool_t pool; // make certain struct is large enough
132   double lk_align; // use worst case alignment; no cache line padding
133 };
134 
135 typedef union kmp_tas_lock kmp_tas_lock_t;
136 
137 // Static initializer for test and set lock variables. Usage:
138 //    kmp_tas_lock_t xlock = KMP_TAS_LOCK_INITIALIZER( xlock );
139 #define KMP_TAS_LOCK_INITIALIZER(lock)                                         \
140   {                                                                            \
141     { ATOMIC_VAR_INIT(KMP_LOCK_FREE(tas)), 0 }                                 \
142   }
143 
144 extern int __kmp_acquire_tas_lock(kmp_tas_lock_t *lck, kmp_int32 gtid);
145 extern int __kmp_test_tas_lock(kmp_tas_lock_t *lck, kmp_int32 gtid);
146 extern int __kmp_release_tas_lock(kmp_tas_lock_t *lck, kmp_int32 gtid);
147 extern void __kmp_init_tas_lock(kmp_tas_lock_t *lck);
148 extern void __kmp_destroy_tas_lock(kmp_tas_lock_t *lck);
149 
150 extern int __kmp_acquire_nested_tas_lock(kmp_tas_lock_t *lck, kmp_int32 gtid);
151 extern int __kmp_test_nested_tas_lock(kmp_tas_lock_t *lck, kmp_int32 gtid);
152 extern int __kmp_release_nested_tas_lock(kmp_tas_lock_t *lck, kmp_int32 gtid);
153 extern void __kmp_init_nested_tas_lock(kmp_tas_lock_t *lck);
154 extern void __kmp_destroy_nested_tas_lock(kmp_tas_lock_t *lck);
155 
156 #define KMP_LOCK_RELEASED 1
157 #define KMP_LOCK_STILL_HELD 0
158 #define KMP_LOCK_ACQUIRED_FIRST 1
159 #define KMP_LOCK_ACQUIRED_NEXT 0
160 #ifndef KMP_USE_FUTEX
161 #define KMP_USE_FUTEX                                                          \
162   (KMP_OS_LINUX &&                                                             \
163    (KMP_ARCH_X86 || KMP_ARCH_X86_64 || KMP_ARCH_ARM || KMP_ARCH_AARCH64))
164 #endif
165 #if KMP_USE_FUTEX
166 
167 // ----------------------------------------------------------------------------
168 // futex locks.  futex locks are only available on Linux* OS.
169 //
170 // Like non-nested test and set lock, non-nested futex locks use the memory
171 // allocated by the compiler for the lock, rather than a pointer to it.
172 //
173 // Information normally available to the tools, such as lock location, lock
174 // usage (normal lock vs. critical section), etc. is not available with test and
175 // set locks. With non-nested futex locks, the lock owner is not even available.
176 // ----------------------------------------------------------------------------
177 
178 struct kmp_base_futex_lock {
179   volatile kmp_int32 poll; // KMP_LOCK_FREE(futex) => unlocked
180   // 2*(gtid+1) of owning thread, 0 if unlocked
181   // locked: (gtid+1) of owning thread
182   kmp_int32 depth_locked; // depth locked, for nested locks only
183 };
184 
185 typedef struct kmp_base_futex_lock kmp_base_futex_lock_t;
186 
187 union kmp_futex_lock {
188   kmp_base_futex_lock_t lk;
189   kmp_lock_pool_t pool; // make certain struct is large enough
190   double lk_align; // use worst case alignment
191   // no cache line padding
192 };
193 
194 typedef union kmp_futex_lock kmp_futex_lock_t;
195 
196 // Static initializer for futex lock variables. Usage:
197 //    kmp_futex_lock_t xlock = KMP_FUTEX_LOCK_INITIALIZER( xlock );
198 #define KMP_FUTEX_LOCK_INITIALIZER(lock)                                       \
199   {                                                                            \
200     { KMP_LOCK_FREE(futex), 0 }                                                \
201   }
202 
203 extern int __kmp_acquire_futex_lock(kmp_futex_lock_t *lck, kmp_int32 gtid);
204 extern int __kmp_test_futex_lock(kmp_futex_lock_t *lck, kmp_int32 gtid);
205 extern int __kmp_release_futex_lock(kmp_futex_lock_t *lck, kmp_int32 gtid);
206 extern void __kmp_init_futex_lock(kmp_futex_lock_t *lck);
207 extern void __kmp_destroy_futex_lock(kmp_futex_lock_t *lck);
208 
209 extern int __kmp_acquire_nested_futex_lock(kmp_futex_lock_t *lck,
210                                            kmp_int32 gtid);
211 extern int __kmp_test_nested_futex_lock(kmp_futex_lock_t *lck, kmp_int32 gtid);
212 extern int __kmp_release_nested_futex_lock(kmp_futex_lock_t *lck,
213                                            kmp_int32 gtid);
214 extern void __kmp_init_nested_futex_lock(kmp_futex_lock_t *lck);
215 extern void __kmp_destroy_nested_futex_lock(kmp_futex_lock_t *lck);
216 
217 #endif // KMP_USE_FUTEX
218 
219 // ----------------------------------------------------------------------------
220 // Ticket locks.
221 
222 #ifdef __cplusplus
223 
224 #ifdef _MSC_VER
225 // MSVC won't allow use of std::atomic<> in a union since it has non-trivial
226 // copy constructor.
227 
228 struct kmp_base_ticket_lock {
229   // `initialized' must be the first entry in the lock data structure!
230   std::atomic_bool initialized;
231   volatile union kmp_ticket_lock *self; // points to the lock union
232   ident_t const *location; // Source code location of omp_init_lock().
233   std::atomic_uint
234       next_ticket; // ticket number to give to next thread which acquires
235   std::atomic_uint now_serving; // ticket number for thread which holds the lock
236   std::atomic_int owner_id; // (gtid+1) of owning thread, 0 if unlocked
237   std::atomic_int depth_locked; // depth locked, for nested locks only
238   kmp_lock_flags_t flags; // lock specifics, e.g. critical section lock
239 };
240 #else
241 struct kmp_base_ticket_lock {
242   // `initialized' must be the first entry in the lock data structure!
243   std::atomic<bool> initialized;
244   volatile union kmp_ticket_lock *self; // points to the lock union
245   ident_t const *location; // Source code location of omp_init_lock().
246   std::atomic<unsigned>
247       next_ticket; // ticket number to give to next thread which acquires
248   std::atomic<unsigned>
249       now_serving; // ticket number for thread which holds the lock
250   std::atomic<int> owner_id; // (gtid+1) of owning thread, 0 if unlocked
251   std::atomic<int> depth_locked; // depth locked, for nested locks only
252   kmp_lock_flags_t flags; // lock specifics, e.g. critical section lock
253 };
254 #endif
255 
256 #else // __cplusplus
257 
258 struct kmp_base_ticket_lock;
259 
260 #endif // !__cplusplus
261 
262 typedef struct kmp_base_ticket_lock kmp_base_ticket_lock_t;
263 
264 union KMP_ALIGN_CACHE kmp_ticket_lock {
265   kmp_base_ticket_lock_t
266       lk; // This field must be first to allow static initializing.
267   kmp_lock_pool_t pool;
268   double lk_align; // use worst case alignment
269   char lk_pad[KMP_PAD(kmp_base_ticket_lock_t, CACHE_LINE)];
270 };
271 
272 typedef union kmp_ticket_lock kmp_ticket_lock_t;
273 
274 // Static initializer for simple ticket lock variables. Usage:
275 //    kmp_ticket_lock_t xlock = KMP_TICKET_LOCK_INITIALIZER( xlock );
276 // Note the macro argument. It is important to make var properly initialized.
277 #define KMP_TICKET_LOCK_INITIALIZER(lock)                                      \
278   {                                                                            \
279     {                                                                          \
280       ATOMIC_VAR_INIT(true)                                                    \
281       , &(lock), NULL, ATOMIC_VAR_INIT(0U), ATOMIC_VAR_INIT(0U),               \
282           ATOMIC_VAR_INIT(0), ATOMIC_VAR_INIT(-1)                              \
283     }                                                                          \
284   }
285 
286 extern int __kmp_acquire_ticket_lock(kmp_ticket_lock_t *lck, kmp_int32 gtid);
287 extern int __kmp_test_ticket_lock(kmp_ticket_lock_t *lck, kmp_int32 gtid);
288 extern int __kmp_test_ticket_lock_with_cheks(kmp_ticket_lock_t *lck,
289                                              kmp_int32 gtid);
290 extern int __kmp_release_ticket_lock(kmp_ticket_lock_t *lck, kmp_int32 gtid);
291 extern void __kmp_init_ticket_lock(kmp_ticket_lock_t *lck);
292 extern void __kmp_destroy_ticket_lock(kmp_ticket_lock_t *lck);
293 
294 extern int __kmp_acquire_nested_ticket_lock(kmp_ticket_lock_t *lck,
295                                             kmp_int32 gtid);
296 extern int __kmp_test_nested_ticket_lock(kmp_ticket_lock_t *lck,
297                                          kmp_int32 gtid);
298 extern int __kmp_release_nested_ticket_lock(kmp_ticket_lock_t *lck,
299                                             kmp_int32 gtid);
300 extern void __kmp_init_nested_ticket_lock(kmp_ticket_lock_t *lck);
301 extern void __kmp_destroy_nested_ticket_lock(kmp_ticket_lock_t *lck);
302 
303 // ----------------------------------------------------------------------------
304 // Queuing locks.
305 
306 #if KMP_USE_ADAPTIVE_LOCKS
307 
308 struct kmp_adaptive_lock_info;
309 
310 typedef struct kmp_adaptive_lock_info kmp_adaptive_lock_info_t;
311 
312 #if KMP_DEBUG_ADAPTIVE_LOCKS
313 
314 struct kmp_adaptive_lock_statistics {
315   /* So we can get stats from locks that haven't been destroyed. */
316   kmp_adaptive_lock_info_t *next;
317   kmp_adaptive_lock_info_t *prev;
318 
319   /* Other statistics */
320   kmp_uint32 successfulSpeculations;
321   kmp_uint32 hardFailedSpeculations;
322   kmp_uint32 softFailedSpeculations;
323   kmp_uint32 nonSpeculativeAcquires;
324   kmp_uint32 nonSpeculativeAcquireAttempts;
325   kmp_uint32 lemmingYields;
326 };
327 
328 typedef struct kmp_adaptive_lock_statistics kmp_adaptive_lock_statistics_t;
329 
330 extern void __kmp_print_speculative_stats();
331 extern void __kmp_init_speculative_stats();
332 
333 #endif // KMP_DEBUG_ADAPTIVE_LOCKS
334 
335 struct kmp_adaptive_lock_info {
336   /* Values used for adaptivity.
337      Although these are accessed from multiple threads we don't access them
338      atomically, because if we miss updates it probably doesn't matter much. (It
339      just affects our decision about whether to try speculation on the lock). */
340   kmp_uint32 volatile badness;
341   kmp_uint32 volatile acquire_attempts;
342   /* Parameters of the lock. */
343   kmp_uint32 max_badness;
344   kmp_uint32 max_soft_retries;
345 
346 #if KMP_DEBUG_ADAPTIVE_LOCKS
347   kmp_adaptive_lock_statistics_t volatile stats;
348 #endif
349 };
350 
351 #endif // KMP_USE_ADAPTIVE_LOCKS
352 
353 struct kmp_base_queuing_lock {
354 
355   //  `initialized' must be the first entry in the lock data structure!
356   volatile union kmp_queuing_lock
357       *initialized; // Points to the lock union if in initialized state.
358 
359   ident_t const *location; // Source code location of omp_init_lock().
360 
361   KMP_ALIGN(8) // tail_id  must be 8-byte aligned!
362 
363   volatile kmp_int32
364       tail_id; // (gtid+1) of thread at tail of wait queue, 0 if empty
365   // Must be no padding here since head/tail used in 8-byte CAS
366   volatile kmp_int32
367       head_id; // (gtid+1) of thread at head of wait queue, 0 if empty
368   // Decl order assumes little endian
369   // bakery-style lock
370   volatile kmp_uint32
371       next_ticket; // ticket number to give to next thread which acquires
372   volatile kmp_uint32
373       now_serving; // ticket number for thread which holds the lock
374   volatile kmp_int32 owner_id; // (gtid+1) of owning thread, 0 if unlocked
375   kmp_int32 depth_locked; // depth locked, for nested locks only
376 
377   kmp_lock_flags_t flags; // lock specifics, e.g. critical section lock
378 };
379 
380 typedef struct kmp_base_queuing_lock kmp_base_queuing_lock_t;
381 
382 KMP_BUILD_ASSERT(offsetof(kmp_base_queuing_lock_t, tail_id) % 8 == 0);
383 
384 union KMP_ALIGN_CACHE kmp_queuing_lock {
385   kmp_base_queuing_lock_t
386       lk; // This field must be first to allow static initializing.
387   kmp_lock_pool_t pool;
388   double lk_align; // use worst case alignment
389   char lk_pad[KMP_PAD(kmp_base_queuing_lock_t, CACHE_LINE)];
390 };
391 
392 typedef union kmp_queuing_lock kmp_queuing_lock_t;
393 
394 extern int __kmp_acquire_queuing_lock(kmp_queuing_lock_t *lck, kmp_int32 gtid);
395 extern int __kmp_test_queuing_lock(kmp_queuing_lock_t *lck, kmp_int32 gtid);
396 extern int __kmp_release_queuing_lock(kmp_queuing_lock_t *lck, kmp_int32 gtid);
397 extern void __kmp_init_queuing_lock(kmp_queuing_lock_t *lck);
398 extern void __kmp_destroy_queuing_lock(kmp_queuing_lock_t *lck);
399 
400 extern int __kmp_acquire_nested_queuing_lock(kmp_queuing_lock_t *lck,
401                                              kmp_int32 gtid);
402 extern int __kmp_test_nested_queuing_lock(kmp_queuing_lock_t *lck,
403                                           kmp_int32 gtid);
404 extern int __kmp_release_nested_queuing_lock(kmp_queuing_lock_t *lck,
405                                              kmp_int32 gtid);
406 extern void __kmp_init_nested_queuing_lock(kmp_queuing_lock_t *lck);
407 extern void __kmp_destroy_nested_queuing_lock(kmp_queuing_lock_t *lck);
408 
409 #if KMP_USE_ADAPTIVE_LOCKS
410 
411 // ----------------------------------------------------------------------------
412 // Adaptive locks.
413 struct kmp_base_adaptive_lock {
414   kmp_base_queuing_lock qlk;
415   KMP_ALIGN(CACHE_LINE)
416   kmp_adaptive_lock_info_t
417       adaptive; // Information for the speculative adaptive lock
418 };
419 
420 typedef struct kmp_base_adaptive_lock kmp_base_adaptive_lock_t;
421 
422 union KMP_ALIGN_CACHE kmp_adaptive_lock {
423   kmp_base_adaptive_lock_t lk;
424   kmp_lock_pool_t pool;
425   double lk_align;
426   char lk_pad[KMP_PAD(kmp_base_adaptive_lock_t, CACHE_LINE)];
427 };
428 typedef union kmp_adaptive_lock kmp_adaptive_lock_t;
429 
430 #define GET_QLK_PTR(l) ((kmp_queuing_lock_t *)&(l)->lk.qlk)
431 
432 #endif // KMP_USE_ADAPTIVE_LOCKS
433 
434 // ----------------------------------------------------------------------------
435 // DRDPA ticket locks.
436 struct kmp_base_drdpa_lock {
437   // All of the fields on the first cache line are only written when
438   // initializing or reconfiguring the lock.  These are relatively rare
439   // operations, so data from the first cache line will usually stay resident in
440   // the cache of each thread trying to acquire the lock.
441   //
442   // initialized must be the first entry in the lock data structure!
443   KMP_ALIGN_CACHE
444 
445   volatile union kmp_drdpa_lock
446       *initialized; // points to the lock union if in initialized state
447   ident_t const *location; // Source code location of omp_init_lock().
448   std::atomic<std::atomic<kmp_uint64> *> polls;
449   std::atomic<kmp_uint64> mask; // is 2**num_polls-1 for mod op
450   kmp_uint64 cleanup_ticket; // thread with cleanup ticket
451   std::atomic<kmp_uint64> *old_polls; // will deallocate old_polls
452   kmp_uint32 num_polls; // must be power of 2
453 
454   // next_ticket it needs to exist in a separate cache line, as it is
455   // invalidated every time a thread takes a new ticket.
456   KMP_ALIGN_CACHE
457 
458   std::atomic<kmp_uint64> next_ticket;
459 
460   // now_serving is used to store our ticket value while we hold the lock. It
461   // has a slightly different meaning in the DRDPA ticket locks (where it is
462   // written by the acquiring thread) than it does in the simple ticket locks
463   // (where it is written by the releasing thread).
464   //
465   // Since now_serving is only read and written in the critical section,
466   // it is non-volatile, but it needs to exist on a separate cache line,
467   // as it is invalidated at every lock acquire.
468   //
469   // Likewise, the vars used for nested locks (owner_id and depth_locked) are
470   // only written by the thread owning the lock, so they are put in this cache
471   // line.  owner_id is read by other threads, so it must be declared volatile.
472   KMP_ALIGN_CACHE
473   kmp_uint64 now_serving; // doesn't have to be volatile
474   volatile kmp_uint32 owner_id; // (gtid+1) of owning thread, 0 if unlocked
475   kmp_int32 depth_locked; // depth locked
476   kmp_lock_flags_t flags; // lock specifics, e.g. critical section lock
477 };
478 
479 typedef struct kmp_base_drdpa_lock kmp_base_drdpa_lock_t;
480 
481 union KMP_ALIGN_CACHE kmp_drdpa_lock {
482   kmp_base_drdpa_lock_t
483       lk; // This field must be first to allow static initializing. */
484   kmp_lock_pool_t pool;
485   double lk_align; // use worst case alignment
486   char lk_pad[KMP_PAD(kmp_base_drdpa_lock_t, CACHE_LINE)];
487 };
488 
489 typedef union kmp_drdpa_lock kmp_drdpa_lock_t;
490 
491 extern int __kmp_acquire_drdpa_lock(kmp_drdpa_lock_t *lck, kmp_int32 gtid);
492 extern int __kmp_test_drdpa_lock(kmp_drdpa_lock_t *lck, kmp_int32 gtid);
493 extern int __kmp_release_drdpa_lock(kmp_drdpa_lock_t *lck, kmp_int32 gtid);
494 extern void __kmp_init_drdpa_lock(kmp_drdpa_lock_t *lck);
495 extern void __kmp_destroy_drdpa_lock(kmp_drdpa_lock_t *lck);
496 
497 extern int __kmp_acquire_nested_drdpa_lock(kmp_drdpa_lock_t *lck,
498                                            kmp_int32 gtid);
499 extern int __kmp_test_nested_drdpa_lock(kmp_drdpa_lock_t *lck, kmp_int32 gtid);
500 extern int __kmp_release_nested_drdpa_lock(kmp_drdpa_lock_t *lck,
501                                            kmp_int32 gtid);
502 extern void __kmp_init_nested_drdpa_lock(kmp_drdpa_lock_t *lck);
503 extern void __kmp_destroy_nested_drdpa_lock(kmp_drdpa_lock_t *lck);
504 
505 // ============================================================================
506 // Lock purposes.
507 // ============================================================================
508 
509 // Bootstrap locks.
510 //
511 // Bootstrap locks -- very few locks used at library initialization time.
512 // Bootstrap locks are currently implemented as ticket locks.
513 // They could also be implemented as test and set lock, but cannot be
514 // implemented with other lock kinds as they require gtids which are not
515 // available at initialization time.
516 
517 typedef kmp_ticket_lock_t kmp_bootstrap_lock_t;
518 
519 #define KMP_BOOTSTRAP_LOCK_INITIALIZER(lock) KMP_TICKET_LOCK_INITIALIZER((lock))
520 #define KMP_BOOTSTRAP_LOCK_INIT(lock)                                          \
521   kmp_bootstrap_lock_t lock = KMP_TICKET_LOCK_INITIALIZER(lock)
522 
523 static inline int __kmp_acquire_bootstrap_lock(kmp_bootstrap_lock_t *lck) {
524   return __kmp_acquire_ticket_lock(lck, KMP_GTID_DNE);
525 }
526 
527 static inline int __kmp_test_bootstrap_lock(kmp_bootstrap_lock_t *lck) {
528   return __kmp_test_ticket_lock(lck, KMP_GTID_DNE);
529 }
530 
531 static inline void __kmp_release_bootstrap_lock(kmp_bootstrap_lock_t *lck) {
532   __kmp_release_ticket_lock(lck, KMP_GTID_DNE);
533 }
534 
535 static inline void __kmp_init_bootstrap_lock(kmp_bootstrap_lock_t *lck) {
536   __kmp_init_ticket_lock(lck);
537 }
538 
539 static inline void __kmp_destroy_bootstrap_lock(kmp_bootstrap_lock_t *lck) {
540   __kmp_destroy_ticket_lock(lck);
541 }
542 
543 // Internal RTL locks.
544 //
545 // Internal RTL locks are also implemented as ticket locks, for now.
546 //
547 // FIXME - We should go through and figure out which lock kind works best for
548 // each internal lock, and use the type declaration and function calls for
549 // that explicit lock kind (and get rid of this section).
550 
551 typedef kmp_ticket_lock_t kmp_lock_t;
552 
553 #define KMP_LOCK_INIT(lock) kmp_lock_t lock = KMP_TICKET_LOCK_INITIALIZER(lock)
554 
555 static inline int __kmp_acquire_lock(kmp_lock_t *lck, kmp_int32 gtid) {
556   return __kmp_acquire_ticket_lock(lck, gtid);
557 }
558 
559 static inline int __kmp_test_lock(kmp_lock_t *lck, kmp_int32 gtid) {
560   return __kmp_test_ticket_lock(lck, gtid);
561 }
562 
563 static inline void __kmp_release_lock(kmp_lock_t *lck, kmp_int32 gtid) {
564   __kmp_release_ticket_lock(lck, gtid);
565 }
566 
567 static inline void __kmp_init_lock(kmp_lock_t *lck) {
568   __kmp_init_ticket_lock(lck);
569 }
570 
571 static inline void __kmp_destroy_lock(kmp_lock_t *lck) {
572   __kmp_destroy_ticket_lock(lck);
573 }
574 
575 // User locks.
576 //
577 // Do not allocate objects of type union kmp_user_lock!!! This will waste space
578 // unless __kmp_user_lock_kind == lk_drdpa. Instead, check the value of
579 // __kmp_user_lock_kind and allocate objects of the type of the appropriate
580 // union member, and cast their addresses to kmp_user_lock_p.
581 
582 enum kmp_lock_kind {
583   lk_default = 0,
584   lk_tas,
585 #if KMP_USE_FUTEX
586   lk_futex,
587 #endif
588 #if KMP_USE_DYNAMIC_LOCK && KMP_USE_TSX
589   lk_hle,
590   lk_rtm_queuing,
591   lk_rtm_spin,
592 #endif
593   lk_ticket,
594   lk_queuing,
595   lk_drdpa,
596 #if KMP_USE_ADAPTIVE_LOCKS
597   lk_adaptive
598 #endif // KMP_USE_ADAPTIVE_LOCKS
599 };
600 
601 typedef enum kmp_lock_kind kmp_lock_kind_t;
602 
603 extern kmp_lock_kind_t __kmp_user_lock_kind;
604 
605 union kmp_user_lock {
606   kmp_tas_lock_t tas;
607 #if KMP_USE_FUTEX
608   kmp_futex_lock_t futex;
609 #endif
610   kmp_ticket_lock_t ticket;
611   kmp_queuing_lock_t queuing;
612   kmp_drdpa_lock_t drdpa;
613 #if KMP_USE_ADAPTIVE_LOCKS
614   kmp_adaptive_lock_t adaptive;
615 #endif // KMP_USE_ADAPTIVE_LOCKS
616   kmp_lock_pool_t pool;
617 };
618 
619 typedef union kmp_user_lock *kmp_user_lock_p;
620 
621 #if !KMP_USE_DYNAMIC_LOCK
622 
623 extern size_t __kmp_base_user_lock_size;
624 extern size_t __kmp_user_lock_size;
625 
626 extern kmp_int32 (*__kmp_get_user_lock_owner_)(kmp_user_lock_p lck);
627 
628 static inline kmp_int32 __kmp_get_user_lock_owner(kmp_user_lock_p lck) {
629   KMP_DEBUG_ASSERT(__kmp_get_user_lock_owner_ != NULL);
630   return (*__kmp_get_user_lock_owner_)(lck);
631 }
632 
633 extern int (*__kmp_acquire_user_lock_with_checks_)(kmp_user_lock_p lck,
634                                                    kmp_int32 gtid);
635 
636 #if KMP_OS_LINUX &&                                                            \
637     (KMP_ARCH_X86 || KMP_ARCH_X86_64 || KMP_ARCH_ARM || KMP_ARCH_AARCH64)
638 
639 #define __kmp_acquire_user_lock_with_checks(lck, gtid)                         \
640   if (__kmp_user_lock_kind == lk_tas) {                                        \
641     if (__kmp_env_consistency_check) {                                         \
642       char const *const func = "omp_set_lock";                                 \
643       if ((sizeof(kmp_tas_lock_t) <= OMP_LOCK_T_SIZE) &&                       \
644           lck->tas.lk.depth_locked != -1) {                                    \
645         KMP_FATAL(LockNestableUsedAsSimple, func);                             \
646       }                                                                        \
647       if ((gtid >= 0) && (lck->tas.lk.poll - 1 == gtid)) {                     \
648         KMP_FATAL(LockIsAlreadyOwned, func);                                   \
649       }                                                                        \
650     }                                                                          \
651     if (lck->tas.lk.poll != 0 ||                                               \
652         !__kmp_atomic_compare_store_acq(&lck->tas.lk.poll, 0, gtid + 1)) {     \
653       kmp_uint32 spins;                                                        \
654       KMP_FSYNC_PREPARE(lck);                                                  \
655       KMP_INIT_YIELD(spins);                                                   \
656       do {                                                                     \
657         KMP_YIELD_OVERSUB_ELSE_SPIN(spins);                                    \
658       } while (lck->tas.lk.poll != 0 || !__kmp_atomic_compare_store_acq(       \
659                                             &lck->tas.lk.poll, 0, gtid + 1));  \
660     }                                                                          \
661     KMP_FSYNC_ACQUIRED(lck);                                                   \
662   } else {                                                                     \
663     KMP_DEBUG_ASSERT(__kmp_acquire_user_lock_with_checks_ != NULL);            \
664     (*__kmp_acquire_user_lock_with_checks_)(lck, gtid);                        \
665   }
666 
667 #else
668 static inline int __kmp_acquire_user_lock_with_checks(kmp_user_lock_p lck,
669                                                       kmp_int32 gtid) {
670   KMP_DEBUG_ASSERT(__kmp_acquire_user_lock_with_checks_ != NULL);
671   return (*__kmp_acquire_user_lock_with_checks_)(lck, gtid);
672 }
673 #endif
674 
675 extern int (*__kmp_test_user_lock_with_checks_)(kmp_user_lock_p lck,
676                                                 kmp_int32 gtid);
677 
678 #if KMP_OS_LINUX &&                                                            \
679     (KMP_ARCH_X86 || KMP_ARCH_X86_64 || KMP_ARCH_ARM || KMP_ARCH_AARCH64)
680 
681 #include "kmp_i18n.h" /* AC: KMP_FATAL definition */
682 extern int __kmp_env_consistency_check; /* AC: copy from kmp.h here */
683 static inline int __kmp_test_user_lock_with_checks(kmp_user_lock_p lck,
684                                                    kmp_int32 gtid) {
685   if (__kmp_user_lock_kind == lk_tas) {
686     if (__kmp_env_consistency_check) {
687       char const *const func = "omp_test_lock";
688       if ((sizeof(kmp_tas_lock_t) <= OMP_LOCK_T_SIZE) &&
689           lck->tas.lk.depth_locked != -1) {
690         KMP_FATAL(LockNestableUsedAsSimple, func);
691       }
692     }
693     return ((lck->tas.lk.poll == 0) &&
694             __kmp_atomic_compare_store_acq(&lck->tas.lk.poll, 0, gtid + 1));
695   } else {
696     KMP_DEBUG_ASSERT(__kmp_test_user_lock_with_checks_ != NULL);
697     return (*__kmp_test_user_lock_with_checks_)(lck, gtid);
698   }
699 }
700 #else
701 static inline int __kmp_test_user_lock_with_checks(kmp_user_lock_p lck,
702                                                    kmp_int32 gtid) {
703   KMP_DEBUG_ASSERT(__kmp_test_user_lock_with_checks_ != NULL);
704   return (*__kmp_test_user_lock_with_checks_)(lck, gtid);
705 }
706 #endif
707 
708 extern int (*__kmp_release_user_lock_with_checks_)(kmp_user_lock_p lck,
709                                                    kmp_int32 gtid);
710 
711 static inline void __kmp_release_user_lock_with_checks(kmp_user_lock_p lck,
712                                                        kmp_int32 gtid) {
713   KMP_DEBUG_ASSERT(__kmp_release_user_lock_with_checks_ != NULL);
714   (*__kmp_release_user_lock_with_checks_)(lck, gtid);
715 }
716 
717 extern void (*__kmp_init_user_lock_with_checks_)(kmp_user_lock_p lck);
718 
719 static inline void __kmp_init_user_lock_with_checks(kmp_user_lock_p lck) {
720   KMP_DEBUG_ASSERT(__kmp_init_user_lock_with_checks_ != NULL);
721   (*__kmp_init_user_lock_with_checks_)(lck);
722 }
723 
724 // We need a non-checking version of destroy lock for when the RTL is
725 // doing the cleanup as it can't always tell if the lock is nested or not.
726 extern void (*__kmp_destroy_user_lock_)(kmp_user_lock_p lck);
727 
728 static inline void __kmp_destroy_user_lock(kmp_user_lock_p lck) {
729   KMP_DEBUG_ASSERT(__kmp_destroy_user_lock_ != NULL);
730   (*__kmp_destroy_user_lock_)(lck);
731 }
732 
733 extern void (*__kmp_destroy_user_lock_with_checks_)(kmp_user_lock_p lck);
734 
735 static inline void __kmp_destroy_user_lock_with_checks(kmp_user_lock_p lck) {
736   KMP_DEBUG_ASSERT(__kmp_destroy_user_lock_with_checks_ != NULL);
737   (*__kmp_destroy_user_lock_with_checks_)(lck);
738 }
739 
740 extern int (*__kmp_acquire_nested_user_lock_with_checks_)(kmp_user_lock_p lck,
741                                                           kmp_int32 gtid);
742 
743 #if KMP_OS_LINUX && (KMP_ARCH_X86 || KMP_ARCH_X86_64)
744 
745 #define __kmp_acquire_nested_user_lock_with_checks(lck, gtid, depth)           \
746   if (__kmp_user_lock_kind == lk_tas) {                                        \
747     if (__kmp_env_consistency_check) {                                         \
748       char const *const func = "omp_set_nest_lock";                            \
749       if ((sizeof(kmp_tas_lock_t) <= OMP_NEST_LOCK_T_SIZE) &&                  \
750           lck->tas.lk.depth_locked == -1) {                                    \
751         KMP_FATAL(LockSimpleUsedAsNestable, func);                             \
752       }                                                                        \
753     }                                                                          \
754     if (lck->tas.lk.poll - 1 == gtid) {                                        \
755       lck->tas.lk.depth_locked += 1;                                           \
756       *depth = KMP_LOCK_ACQUIRED_NEXT;                                         \
757     } else {                                                                   \
758       if ((lck->tas.lk.poll != 0) ||                                           \
759           !__kmp_atomic_compare_store_acq(&lck->tas.lk.poll, 0, gtid + 1)) {   \
760         kmp_uint32 spins;                                                      \
761         KMP_FSYNC_PREPARE(lck);                                                \
762         KMP_INIT_YIELD(spins);                                                 \
763         do {                                                                   \
764           KMP_YIELD_OVERSUB_ELSE_SPIN(spins);                                  \
765         } while (                                                              \
766             (lck->tas.lk.poll != 0) ||                                         \
767             !__kmp_atomic_compare_store_acq(&lck->tas.lk.poll, 0, gtid + 1));  \
768       }                                                                        \
769       lck->tas.lk.depth_locked = 1;                                            \
770       *depth = KMP_LOCK_ACQUIRED_FIRST;                                        \
771     }                                                                          \
772     KMP_FSYNC_ACQUIRED(lck);                                                   \
773   } else {                                                                     \
774     KMP_DEBUG_ASSERT(__kmp_acquire_nested_user_lock_with_checks_ != NULL);     \
775     *depth = (*__kmp_acquire_nested_user_lock_with_checks_)(lck, gtid);        \
776   }
777 
778 #else
779 static inline void
780 __kmp_acquire_nested_user_lock_with_checks(kmp_user_lock_p lck, kmp_int32 gtid,
781                                            int *depth) {
782   KMP_DEBUG_ASSERT(__kmp_acquire_nested_user_lock_with_checks_ != NULL);
783   *depth = (*__kmp_acquire_nested_user_lock_with_checks_)(lck, gtid);
784 }
785 #endif
786 
787 extern int (*__kmp_test_nested_user_lock_with_checks_)(kmp_user_lock_p lck,
788                                                        kmp_int32 gtid);
789 
790 #if KMP_OS_LINUX && (KMP_ARCH_X86 || KMP_ARCH_X86_64)
791 static inline int __kmp_test_nested_user_lock_with_checks(kmp_user_lock_p lck,
792                                                           kmp_int32 gtid) {
793   if (__kmp_user_lock_kind == lk_tas) {
794     int retval;
795     if (__kmp_env_consistency_check) {
796       char const *const func = "omp_test_nest_lock";
797       if ((sizeof(kmp_tas_lock_t) <= OMP_NEST_LOCK_T_SIZE) &&
798           lck->tas.lk.depth_locked == -1) {
799         KMP_FATAL(LockSimpleUsedAsNestable, func);
800       }
801     }
802     KMP_DEBUG_ASSERT(gtid >= 0);
803     if (lck->tas.lk.poll - 1 ==
804         gtid) { /* __kmp_get_tas_lock_owner( lck ) == gtid */
805       return ++lck->tas.lk.depth_locked; /* same owner, depth increased */
806     }
807     retval = ((lck->tas.lk.poll == 0) &&
808               __kmp_atomic_compare_store_acq(&lck->tas.lk.poll, 0, gtid + 1));
809     if (retval) {
810       KMP_MB();
811       lck->tas.lk.depth_locked = 1;
812     }
813     return retval;
814   } else {
815     KMP_DEBUG_ASSERT(__kmp_test_nested_user_lock_with_checks_ != NULL);
816     return (*__kmp_test_nested_user_lock_with_checks_)(lck, gtid);
817   }
818 }
819 #else
820 static inline int __kmp_test_nested_user_lock_with_checks(kmp_user_lock_p lck,
821                                                           kmp_int32 gtid) {
822   KMP_DEBUG_ASSERT(__kmp_test_nested_user_lock_with_checks_ != NULL);
823   return (*__kmp_test_nested_user_lock_with_checks_)(lck, gtid);
824 }
825 #endif
826 
827 extern int (*__kmp_release_nested_user_lock_with_checks_)(kmp_user_lock_p lck,
828                                                           kmp_int32 gtid);
829 
830 static inline int
831 __kmp_release_nested_user_lock_with_checks(kmp_user_lock_p lck,
832                                            kmp_int32 gtid) {
833   KMP_DEBUG_ASSERT(__kmp_release_nested_user_lock_with_checks_ != NULL);
834   return (*__kmp_release_nested_user_lock_with_checks_)(lck, gtid);
835 }
836 
837 extern void (*__kmp_init_nested_user_lock_with_checks_)(kmp_user_lock_p lck);
838 
839 static inline void
840 __kmp_init_nested_user_lock_with_checks(kmp_user_lock_p lck) {
841   KMP_DEBUG_ASSERT(__kmp_init_nested_user_lock_with_checks_ != NULL);
842   (*__kmp_init_nested_user_lock_with_checks_)(lck);
843 }
844 
845 extern void (*__kmp_destroy_nested_user_lock_with_checks_)(kmp_user_lock_p lck);
846 
847 static inline void
848 __kmp_destroy_nested_user_lock_with_checks(kmp_user_lock_p lck) {
849   KMP_DEBUG_ASSERT(__kmp_destroy_nested_user_lock_with_checks_ != NULL);
850   (*__kmp_destroy_nested_user_lock_with_checks_)(lck);
851 }
852 
853 // user lock functions which do not necessarily exist for all lock kinds.
854 //
855 // The "set" functions usually have wrapper routines that check for a NULL set
856 // function pointer and call it if non-NULL.
857 //
858 // In some cases, it makes sense to have a "get" wrapper function check for a
859 // NULL get function pointer and return NULL / invalid value / error code if
860 // the function pointer is NULL.
861 //
862 // In other cases, the calling code really should differentiate between an
863 // unimplemented function and one that is implemented but returning NULL /
864 // invalid value.  If this is the case, no get function wrapper exists.
865 
866 extern int (*__kmp_is_user_lock_initialized_)(kmp_user_lock_p lck);
867 
868 // no set function; fields set during local allocation
869 
870 extern const ident_t *(*__kmp_get_user_lock_location_)(kmp_user_lock_p lck);
871 
872 static inline const ident_t *__kmp_get_user_lock_location(kmp_user_lock_p lck) {
873   if (__kmp_get_user_lock_location_ != NULL) {
874     return (*__kmp_get_user_lock_location_)(lck);
875   } else {
876     return NULL;
877   }
878 }
879 
880 extern void (*__kmp_set_user_lock_location_)(kmp_user_lock_p lck,
881                                              const ident_t *loc);
882 
883 static inline void __kmp_set_user_lock_location(kmp_user_lock_p lck,
884                                                 const ident_t *loc) {
885   if (__kmp_set_user_lock_location_ != NULL) {
886     (*__kmp_set_user_lock_location_)(lck, loc);
887   }
888 }
889 
890 extern kmp_lock_flags_t (*__kmp_get_user_lock_flags_)(kmp_user_lock_p lck);
891 
892 extern void (*__kmp_set_user_lock_flags_)(kmp_user_lock_p lck,
893                                           kmp_lock_flags_t flags);
894 
895 static inline void __kmp_set_user_lock_flags(kmp_user_lock_p lck,
896                                              kmp_lock_flags_t flags) {
897   if (__kmp_set_user_lock_flags_ != NULL) {
898     (*__kmp_set_user_lock_flags_)(lck, flags);
899   }
900 }
901 
902 // The function which sets up all of the vtbl pointers for kmp_user_lock_t.
903 extern void __kmp_set_user_lock_vptrs(kmp_lock_kind_t user_lock_kind);
904 
905 // Macros for binding user lock functions.
906 #define KMP_BIND_USER_LOCK_TEMPLATE(nest, kind, suffix)                        \
907   {                                                                            \
908     __kmp_acquire##nest##user_lock_with_checks_ = (int (*)(                    \
909         kmp_user_lock_p, kmp_int32))__kmp_acquire##nest##kind##_##suffix;      \
910     __kmp_release##nest##user_lock_with_checks_ = (int (*)(                    \
911         kmp_user_lock_p, kmp_int32))__kmp_release##nest##kind##_##suffix;      \
912     __kmp_test##nest##user_lock_with_checks_ = (int (*)(                       \
913         kmp_user_lock_p, kmp_int32))__kmp_test##nest##kind##_##suffix;         \
914     __kmp_init##nest##user_lock_with_checks_ =                                 \
915         (void (*)(kmp_user_lock_p))__kmp_init##nest##kind##_##suffix;          \
916     __kmp_destroy##nest##user_lock_with_checks_ =                              \
917         (void (*)(kmp_user_lock_p))__kmp_destroy##nest##kind##_##suffix;       \
918   }
919 
920 #define KMP_BIND_USER_LOCK(kind) KMP_BIND_USER_LOCK_TEMPLATE(_, kind, lock)
921 #define KMP_BIND_USER_LOCK_WITH_CHECKS(kind)                                   \
922   KMP_BIND_USER_LOCK_TEMPLATE(_, kind, lock_with_checks)
923 #define KMP_BIND_NESTED_USER_LOCK(kind)                                        \
924   KMP_BIND_USER_LOCK_TEMPLATE(_nested_, kind, lock)
925 #define KMP_BIND_NESTED_USER_LOCK_WITH_CHECKS(kind)                            \
926   KMP_BIND_USER_LOCK_TEMPLATE(_nested_, kind, lock_with_checks)
927 
928 // User lock table & lock allocation
929 /* On 64-bit Linux* OS (and OS X*) GNU compiler allocates only 4 bytems memory
930    for lock variable, which is not enough to store a pointer, so we have to use
931    lock indexes instead of pointers and maintain lock table to map indexes to
932    pointers.
933 
934 
935    Note: The first element of the table is not a pointer to lock! It is a
936    pointer to previously allocated table (or NULL if it is the first table).
937 
938    Usage:
939 
940    if ( OMP_LOCK_T_SIZE < sizeof( <lock> ) ) { // or OMP_NEST_LOCK_T_SIZE
941      Lock table is fully utilized. User locks are indexes, so table is used on
942      user lock operation.
943      Note: it may be the case (lin_32) that we don't need to use a lock
944      table for regular locks, but do need the table for nested locks.
945    }
946    else {
947      Lock table initialized but not actually used.
948    }
949 */
950 
951 struct kmp_lock_table {
952   kmp_lock_index_t used; // Number of used elements
953   kmp_lock_index_t allocated; // Number of allocated elements
954   kmp_user_lock_p *table; // Lock table.
955 };
956 
957 typedef struct kmp_lock_table kmp_lock_table_t;
958 
959 extern kmp_lock_table_t __kmp_user_lock_table;
960 extern kmp_user_lock_p __kmp_lock_pool;
961 
962 struct kmp_block_of_locks {
963   struct kmp_block_of_locks *next_block;
964   void *locks;
965 };
966 
967 typedef struct kmp_block_of_locks kmp_block_of_locks_t;
968 
969 extern kmp_block_of_locks_t *__kmp_lock_blocks;
970 extern int __kmp_num_locks_in_block;
971 
972 extern kmp_user_lock_p __kmp_user_lock_allocate(void **user_lock,
973                                                 kmp_int32 gtid,
974                                                 kmp_lock_flags_t flags);
975 extern void __kmp_user_lock_free(void **user_lock, kmp_int32 gtid,
976                                  kmp_user_lock_p lck);
977 extern kmp_user_lock_p __kmp_lookup_user_lock(void **user_lock,
978                                               char const *func);
979 extern void __kmp_cleanup_user_locks();
980 
981 #define KMP_CHECK_USER_LOCK_INIT()                                             \
982   {                                                                            \
983     if (!TCR_4(__kmp_init_user_locks)) {                                       \
984       __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);                         \
985       if (!TCR_4(__kmp_init_user_locks)) {                                     \
986         TCW_4(__kmp_init_user_locks, TRUE);                                    \
987       }                                                                        \
988       __kmp_release_bootstrap_lock(&__kmp_initz_lock);                         \
989     }                                                                          \
990   }
991 
992 #endif // KMP_USE_DYNAMIC_LOCK
993 
994 #undef KMP_PAD
995 #undef KMP_GTID_DNE
996 
997 #if KMP_USE_DYNAMIC_LOCK
998 // KMP_USE_DYNAMIC_LOCK enables dynamic dispatch of lock functions without
999 // breaking the current compatibility. Essential functionality of this new code
1000 // is dynamic dispatch, but it also implements (or enables implementation of)
1001 // hinted user lock and critical section which will be part of OMP 4.5 soon.
1002 //
1003 // Lock type can be decided at creation time (i.e., lock initialization), and
1004 // subsequent lock function call on the created lock object requires type
1005 // extraction and call through jump table using the extracted type. This type
1006 // information is stored in two different ways depending on the size of the lock
1007 // object, and we differentiate lock types by this size requirement - direct and
1008 // indirect locks.
1009 //
1010 // Direct locks:
1011 // A direct lock object fits into the space created by the compiler for an
1012 // omp_lock_t object, and TAS/Futex lock falls into this category. We use low
1013 // one byte of the lock object as the storage for the lock type, and appropriate
1014 // bit operation is required to access the data meaningful to the lock
1015 // algorithms. Also, to differentiate direct lock from indirect lock, 1 is
1016 // written to LSB of the lock object. The newly introduced "hle" lock is also a
1017 // direct lock.
1018 //
1019 // Indirect locks:
1020 // An indirect lock object requires more space than the compiler-generated
1021 // space, and it should be allocated from heap. Depending on the size of the
1022 // compiler-generated space for the lock (i.e., size of omp_lock_t), this
1023 // omp_lock_t object stores either the address of the heap-allocated indirect
1024 // lock (void * fits in the object) or an index to the indirect lock table entry
1025 // that holds the address. Ticket/Queuing/DRDPA/Adaptive lock falls into this
1026 // category, and the newly introduced "rtm" lock is also an indirect lock which
1027 // was implemented on top of the Queuing lock. When the omp_lock_t object holds
1028 // an index (not lock address), 0 is written to LSB to differentiate the lock
1029 // from a direct lock, and the remaining part is the actual index to the
1030 // indirect lock table.
1031 
1032 #include <stdint.h> // for uintptr_t
1033 
1034 // Shortcuts
1035 #define KMP_USE_INLINED_TAS                                                    \
1036   (KMP_OS_LINUX && (KMP_ARCH_X86 || KMP_ARCH_X86_64 || KMP_ARCH_ARM)) && 1
1037 #define KMP_USE_INLINED_FUTEX KMP_USE_FUTEX && 0
1038 
1039 // List of lock definitions; all nested locks are indirect locks.
1040 // hle lock is xchg lock prefixed with XACQUIRE/XRELEASE.
1041 // All nested locks are indirect lock types.
1042 #if KMP_USE_TSX
1043 #if KMP_USE_FUTEX
1044 #define KMP_FOREACH_D_LOCK(m, a) m(tas, a) m(futex, a) m(hle, a) m(rtm_spin, a)
1045 #define KMP_FOREACH_I_LOCK(m, a)                                               \
1046   m(ticket, a) m(queuing, a) m(adaptive, a) m(drdpa, a) m(rtm_queuing, a)      \
1047       m(nested_tas, a) m(nested_futex, a) m(nested_ticket, a)                  \
1048           m(nested_queuing, a) m(nested_drdpa, a)
1049 #else
1050 #define KMP_FOREACH_D_LOCK(m, a) m(tas, a) m(hle, a) m(rtm_spin, a)
1051 #define KMP_FOREACH_I_LOCK(m, a)                                               \
1052   m(ticket, a) m(queuing, a) m(adaptive, a) m(drdpa, a) m(rtm_queuing, a)      \
1053       m(nested_tas, a) m(nested_ticket, a) m(nested_queuing, a)                \
1054           m(nested_drdpa, a)
1055 #endif // KMP_USE_FUTEX
1056 #define KMP_LAST_D_LOCK lockseq_rtm_spin
1057 #else
1058 #if KMP_USE_FUTEX
1059 #define KMP_FOREACH_D_LOCK(m, a) m(tas, a) m(futex, a)
1060 #define KMP_FOREACH_I_LOCK(m, a)                                               \
1061   m(ticket, a) m(queuing, a) m(drdpa, a) m(nested_tas, a) m(nested_futex, a)   \
1062       m(nested_ticket, a) m(nested_queuing, a) m(nested_drdpa, a)
1063 #define KMP_LAST_D_LOCK lockseq_futex
1064 #else
1065 #define KMP_FOREACH_D_LOCK(m, a) m(tas, a)
1066 #define KMP_FOREACH_I_LOCK(m, a)                                               \
1067   m(ticket, a) m(queuing, a) m(drdpa, a) m(nested_tas, a) m(nested_ticket, a)  \
1068       m(nested_queuing, a) m(nested_drdpa, a)
1069 #define KMP_LAST_D_LOCK lockseq_tas
1070 #endif // KMP_USE_FUTEX
1071 #endif // KMP_USE_TSX
1072 
1073 // Information used in dynamic dispatch
1074 #define KMP_LOCK_SHIFT                                                         \
1075   8 // number of low bits to be used as tag for direct locks
1076 #define KMP_FIRST_D_LOCK lockseq_tas
1077 #define KMP_FIRST_I_LOCK lockseq_ticket
1078 #define KMP_LAST_I_LOCK lockseq_nested_drdpa
1079 #define KMP_NUM_I_LOCKS                                                        \
1080   (locktag_nested_drdpa + 1) // number of indirect lock types
1081 
1082 // Base type for dynamic locks.
1083 typedef kmp_uint32 kmp_dyna_lock_t;
1084 
1085 // Lock sequence that enumerates all lock kinds. Always make this enumeration
1086 // consistent with kmp_lockseq_t in the include directory.
1087 typedef enum {
1088   lockseq_indirect = 0,
1089 #define expand_seq(l, a) lockseq_##l,
1090   KMP_FOREACH_D_LOCK(expand_seq, 0) KMP_FOREACH_I_LOCK(expand_seq, 0)
1091 #undef expand_seq
1092 } kmp_dyna_lockseq_t;
1093 
1094 // Enumerates indirect lock tags.
1095 typedef enum {
1096 #define expand_tag(l, a) locktag_##l,
1097   KMP_FOREACH_I_LOCK(expand_tag, 0)
1098 #undef expand_tag
1099 } kmp_indirect_locktag_t;
1100 
1101 // Utility macros that extract information from lock sequences.
1102 #define KMP_IS_D_LOCK(seq)                                                     \
1103   ((seq) >= KMP_FIRST_D_LOCK && (seq) <= KMP_LAST_D_LOCK)
1104 #define KMP_IS_I_LOCK(seq)                                                     \
1105   ((seq) >= KMP_FIRST_I_LOCK && (seq) <= KMP_LAST_I_LOCK)
1106 #define KMP_GET_I_TAG(seq) (kmp_indirect_locktag_t)((seq)-KMP_FIRST_I_LOCK)
1107 #define KMP_GET_D_TAG(seq) ((seq) << 1 | 1)
1108 
1109 // Enumerates direct lock tags starting from indirect tag.
1110 typedef enum {
1111 #define expand_tag(l, a) locktag_##l = KMP_GET_D_TAG(lockseq_##l),
1112   KMP_FOREACH_D_LOCK(expand_tag, 0)
1113 #undef expand_tag
1114 } kmp_direct_locktag_t;
1115 
1116 // Indirect lock type
1117 typedef struct {
1118   kmp_user_lock_p lock;
1119   kmp_indirect_locktag_t type;
1120 } kmp_indirect_lock_t;
1121 
1122 // Function tables for direct locks. Set/unset/test differentiate functions
1123 // with/without consistency checking.
1124 extern void (*__kmp_direct_init[])(kmp_dyna_lock_t *, kmp_dyna_lockseq_t);
1125 extern void (**__kmp_direct_destroy)(kmp_dyna_lock_t *);
1126 extern int (**__kmp_direct_set)(kmp_dyna_lock_t *, kmp_int32);
1127 extern int (**__kmp_direct_unset)(kmp_dyna_lock_t *, kmp_int32);
1128 extern int (**__kmp_direct_test)(kmp_dyna_lock_t *, kmp_int32);
1129 
1130 // Function tables for indirect locks. Set/unset/test differentiate functions
1131 // with/without consistency checking.
1132 extern void (*__kmp_indirect_init[])(kmp_user_lock_p);
1133 extern void (**__kmp_indirect_destroy)(kmp_user_lock_p);
1134 extern int (**__kmp_indirect_set)(kmp_user_lock_p, kmp_int32);
1135 extern int (**__kmp_indirect_unset)(kmp_user_lock_p, kmp_int32);
1136 extern int (**__kmp_indirect_test)(kmp_user_lock_p, kmp_int32);
1137 
1138 // Extracts direct lock tag from a user lock pointer
1139 #define KMP_EXTRACT_D_TAG(l)                                                   \
1140   (*((kmp_dyna_lock_t *)(l)) & ((1 << KMP_LOCK_SHIFT) - 1) &                   \
1141    -(*((kmp_dyna_lock_t *)(l)) & 1))
1142 
1143 // Extracts indirect lock index from a user lock pointer
1144 #define KMP_EXTRACT_I_INDEX(l) (*(kmp_lock_index_t *)(l) >> 1)
1145 
1146 // Returns function pointer to the direct lock function with l (kmp_dyna_lock_t
1147 // *) and op (operation type).
1148 #define KMP_D_LOCK_FUNC(l, op) __kmp_direct_##op[KMP_EXTRACT_D_TAG(l)]
1149 
1150 // Returns function pointer to the indirect lock function with l
1151 // (kmp_indirect_lock_t *) and op (operation type).
1152 #define KMP_I_LOCK_FUNC(l, op)                                                 \
1153   __kmp_indirect_##op[((kmp_indirect_lock_t *)(l))->type]
1154 
1155 // Initializes a direct lock with the given lock pointer and lock sequence.
1156 #define KMP_INIT_D_LOCK(l, seq)                                                \
1157   __kmp_direct_init[KMP_GET_D_TAG(seq)]((kmp_dyna_lock_t *)l, seq)
1158 
1159 // Initializes an indirect lock with the given lock pointer and lock sequence.
1160 #define KMP_INIT_I_LOCK(l, seq)                                                \
1161   __kmp_direct_init[0]((kmp_dyna_lock_t *)(l), seq)
1162 
1163 // Returns "free" lock value for the given lock type.
1164 #define KMP_LOCK_FREE(type) (locktag_##type)
1165 
1166 // Returns "busy" lock value for the given lock teyp.
1167 #define KMP_LOCK_BUSY(v, type) ((v) << KMP_LOCK_SHIFT | locktag_##type)
1168 
1169 // Returns lock value after removing (shifting) lock tag.
1170 #define KMP_LOCK_STRIP(v) ((v) >> KMP_LOCK_SHIFT)
1171 
1172 // Initializes global states and data structures for managing dynamic user
1173 // locks.
1174 extern void __kmp_init_dynamic_user_locks();
1175 
1176 // Allocates and returns an indirect lock with the given indirect lock tag.
1177 extern kmp_indirect_lock_t *
1178 __kmp_allocate_indirect_lock(void **, kmp_int32, kmp_indirect_locktag_t);
1179 
1180 // Cleans up global states and data structures for managing dynamic user locks.
1181 extern void __kmp_cleanup_indirect_user_locks();
1182 
1183 // Default user lock sequence when not using hinted locks.
1184 extern kmp_dyna_lockseq_t __kmp_user_lock_seq;
1185 
1186 // Jump table for "set lock location", available only for indirect locks.
1187 extern void (*__kmp_indirect_set_location[KMP_NUM_I_LOCKS])(kmp_user_lock_p,
1188                                                             const ident_t *);
1189 #define KMP_SET_I_LOCK_LOCATION(lck, loc)                                      \
1190   {                                                                            \
1191     if (__kmp_indirect_set_location[(lck)->type] != NULL)                      \
1192       __kmp_indirect_set_location[(lck)->type]((lck)->lock, loc);              \
1193   }
1194 
1195 // Jump table for "set lock flags", available only for indirect locks.
1196 extern void (*__kmp_indirect_set_flags[KMP_NUM_I_LOCKS])(kmp_user_lock_p,
1197                                                          kmp_lock_flags_t);
1198 #define KMP_SET_I_LOCK_FLAGS(lck, flag)                                        \
1199   {                                                                            \
1200     if (__kmp_indirect_set_flags[(lck)->type] != NULL)                         \
1201       __kmp_indirect_set_flags[(lck)->type]((lck)->lock, flag);                \
1202   }
1203 
1204 // Jump table for "get lock location", available only for indirect locks.
1205 extern const ident_t *(*__kmp_indirect_get_location[KMP_NUM_I_LOCKS])(
1206     kmp_user_lock_p);
1207 #define KMP_GET_I_LOCK_LOCATION(lck)                                           \
1208   (__kmp_indirect_get_location[(lck)->type] != NULL                            \
1209        ? __kmp_indirect_get_location[(lck)->type]((lck)->lock)                 \
1210        : NULL)
1211 
1212 // Jump table for "get lock flags", available only for indirect locks.
1213 extern kmp_lock_flags_t (*__kmp_indirect_get_flags[KMP_NUM_I_LOCKS])(
1214     kmp_user_lock_p);
1215 #define KMP_GET_I_LOCK_FLAGS(lck)                                              \
1216   (__kmp_indirect_get_flags[(lck)->type] != NULL                               \
1217        ? __kmp_indirect_get_flags[(lck)->type]((lck)->lock)                    \
1218        : NULL)
1219 
1220 #define KMP_I_LOCK_CHUNK                                                       \
1221   1024 // number of kmp_indirect_lock_t objects to be allocated together
1222 
1223 // Lock table for indirect locks.
1224 typedef struct kmp_indirect_lock_table {
1225   kmp_indirect_lock_t **table; // blocks of indirect locks allocated
1226   kmp_lock_index_t size; // size of the indirect lock table
1227   kmp_lock_index_t next; // index to the next lock to be allocated
1228 } kmp_indirect_lock_table_t;
1229 
1230 extern kmp_indirect_lock_table_t __kmp_i_lock_table;
1231 
1232 // Returns the indirect lock associated with the given index.
1233 #define KMP_GET_I_LOCK(index)                                                  \
1234   (*(__kmp_i_lock_table.table + (index) / KMP_I_LOCK_CHUNK) +                  \
1235    (index) % KMP_I_LOCK_CHUNK)
1236 
1237 // Number of locks in a lock block, which is fixed to "1" now.
1238 // TODO: No lock block implementation now. If we do support, we need to manage
1239 // lock block data structure for each indirect lock type.
1240 extern int __kmp_num_locks_in_block;
1241 
1242 // Fast lock table lookup without consistency checking
1243 #define KMP_LOOKUP_I_LOCK(l)                                                   \
1244   ((OMP_LOCK_T_SIZE < sizeof(void *)) ? KMP_GET_I_LOCK(KMP_EXTRACT_I_INDEX(l)) \
1245                                       : *((kmp_indirect_lock_t **)(l)))
1246 
1247 // Used once in kmp_error.cpp
1248 extern kmp_int32 __kmp_get_user_lock_owner(kmp_user_lock_p, kmp_uint32);
1249 
1250 #else // KMP_USE_DYNAMIC_LOCK
1251 
1252 #define KMP_LOCK_BUSY(v, type) (v)
1253 #define KMP_LOCK_FREE(type) 0
1254 #define KMP_LOCK_STRIP(v) (v)
1255 
1256 #endif // KMP_USE_DYNAMIC_LOCK
1257 
1258 // data structure for using backoff within spin locks.
1259 typedef struct {
1260   kmp_uint32 step; // current step
1261   kmp_uint32 max_backoff; // upper bound of outer delay loop
1262   kmp_uint32 min_tick; // size of inner delay loop in ticks (machine-dependent)
1263 } kmp_backoff_t;
1264 
1265 // Runtime's default backoff parameters
1266 extern kmp_backoff_t __kmp_spin_backoff_params;
1267 
1268 // Backoff function
1269 extern void __kmp_spin_backoff(kmp_backoff_t *);
1270 
1271 #ifdef __cplusplus
1272 } // extern "C"
1273 #endif // __cplusplus
1274 
1275 #endif /* KMP_LOCK_H */
1276