LLVM OpenMP* Runtime Library
kmp_gsupport.cpp
1 /*
2  * kmp_gsupport.cpp
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 "kmp.h"
14 #include "kmp_atomic.h"
15 
16 #if OMPT_SUPPORT
17 #include "ompt-specific.h"
18 #endif
19 
20 #ifdef __cplusplus
21 extern "C" {
22 #endif // __cplusplus
23 
24 #define MKLOC(loc, routine) \
25  static ident_t(loc) = {0, KMP_IDENT_KMPC, 0, 0, ";unknown;unknown;0;0;;"};
26 
27 #include "kmp_ftn_os.h"
28 
29 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_BARRIER)(void) {
30  int gtid = __kmp_entry_gtid();
31  MKLOC(loc, "GOMP_barrier");
32  KA_TRACE(20, ("GOMP_barrier: T#%d\n", gtid));
33 #if OMPT_SUPPORT && OMPT_OPTIONAL
34  ompt_frame_t *ompt_frame;
35  if (ompt_enabled.enabled) {
36  __ompt_get_task_info_internal(0, NULL, NULL, &ompt_frame, NULL, NULL);
37  ompt_frame->enter_frame.ptr = OMPT_GET_FRAME_ADDRESS(0);
38  OMPT_STORE_RETURN_ADDRESS(gtid);
39  }
40 #endif
41  __kmpc_barrier(&loc, gtid);
42 #if OMPT_SUPPORT && OMPT_OPTIONAL
43  if (ompt_enabled.enabled) {
44  ompt_frame->enter_frame = ompt_data_none;
45  }
46 #endif
47 }
48 
49 // Mutual exclusion
50 
51 // The symbol that icc/ifort generates for unnamed for unnamed critical sections
52 // - .gomp_critical_user_ - is defined using .comm in any objects reference it.
53 // We can't reference it directly here in C code, as the symbol contains a ".".
54 //
55 // The RTL contains an assembly language definition of .gomp_critical_user_
56 // with another symbol __kmp_unnamed_critical_addr initialized with it's
57 // address.
58 extern kmp_critical_name *__kmp_unnamed_critical_addr;
59 
60 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_CRITICAL_START)(void) {
61  int gtid = __kmp_entry_gtid();
62  MKLOC(loc, "GOMP_critical_start");
63  KA_TRACE(20, ("GOMP_critical_start: T#%d\n", gtid));
64 #if OMPT_SUPPORT && OMPT_OPTIONAL
65  OMPT_STORE_RETURN_ADDRESS(gtid);
66 #endif
67  __kmpc_critical(&loc, gtid, __kmp_unnamed_critical_addr);
68 }
69 
70 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_CRITICAL_END)(void) {
71  int gtid = __kmp_get_gtid();
72  MKLOC(loc, "GOMP_critical_end");
73  KA_TRACE(20, ("GOMP_critical_end: T#%d\n", gtid));
74 #if OMPT_SUPPORT && OMPT_OPTIONAL
75  OMPT_STORE_RETURN_ADDRESS(gtid);
76 #endif
77  __kmpc_end_critical(&loc, gtid, __kmp_unnamed_critical_addr);
78 }
79 
80 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_CRITICAL_NAME_START)(void **pptr) {
81  int gtid = __kmp_entry_gtid();
82  MKLOC(loc, "GOMP_critical_name_start");
83  KA_TRACE(20, ("GOMP_critical_name_start: T#%d\n", gtid));
84  __kmpc_critical(&loc, gtid, (kmp_critical_name *)pptr);
85 }
86 
87 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_CRITICAL_NAME_END)(void **pptr) {
88  int gtid = __kmp_get_gtid();
89  MKLOC(loc, "GOMP_critical_name_end");
90  KA_TRACE(20, ("GOMP_critical_name_end: T#%d\n", gtid));
91  __kmpc_end_critical(&loc, gtid, (kmp_critical_name *)pptr);
92 }
93 
94 // The Gnu codegen tries to use locked operations to perform atomic updates
95 // inline. If it can't, then it calls GOMP_atomic_start() before performing
96 // the update and GOMP_atomic_end() afterward, regardless of the data type.
97 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_ATOMIC_START)(void) {
98  int gtid = __kmp_entry_gtid();
99  KA_TRACE(20, ("GOMP_atomic_start: T#%d\n", gtid));
100 
101 #if OMPT_SUPPORT
102  __ompt_thread_assign_wait_id(0);
103 #endif
104 
105  __kmp_acquire_atomic_lock(&__kmp_atomic_lock, gtid);
106 }
107 
108 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_ATOMIC_END)(void) {
109  int gtid = __kmp_get_gtid();
110  KA_TRACE(20, ("GOMP_atomic_end: T#%d\n", gtid));
111  __kmp_release_atomic_lock(&__kmp_atomic_lock, gtid);
112 }
113 
114 int KMP_EXPAND_NAME(KMP_API_NAME_GOMP_SINGLE_START)(void) {
115  int gtid = __kmp_entry_gtid();
116  MKLOC(loc, "GOMP_single_start");
117  KA_TRACE(20, ("GOMP_single_start: T#%d\n", gtid));
118 
119  if (!TCR_4(__kmp_init_parallel))
120  __kmp_parallel_initialize();
121  __kmp_resume_if_soft_paused();
122 
123  // 3rd parameter == FALSE prevents kmp_enter_single from pushing a
124  // workshare when USE_CHECKS is defined. We need to avoid the push,
125  // as there is no corresponding GOMP_single_end() call.
126  kmp_int32 rc = __kmp_enter_single(gtid, &loc, FALSE);
127 
128 #if OMPT_SUPPORT && OMPT_OPTIONAL
129  kmp_info_t *this_thr = __kmp_threads[gtid];
130  kmp_team_t *team = this_thr->th.th_team;
131  int tid = __kmp_tid_from_gtid(gtid);
132 
133  if (ompt_enabled.enabled) {
134  if (rc) {
135  if (ompt_enabled.ompt_callback_work) {
136  ompt_callbacks.ompt_callback(ompt_callback_work)(
137  ompt_work_single_executor, ompt_scope_begin,
138  &(team->t.ompt_team_info.parallel_data),
139  &(team->t.t_implicit_task_taskdata[tid].ompt_task_info.task_data),
140  1, OMPT_GET_RETURN_ADDRESS(0));
141  }
142  } else {
143  if (ompt_enabled.ompt_callback_work) {
144  ompt_callbacks.ompt_callback(ompt_callback_work)(
145  ompt_work_single_other, ompt_scope_begin,
146  &(team->t.ompt_team_info.parallel_data),
147  &(team->t.t_implicit_task_taskdata[tid].ompt_task_info.task_data),
148  1, OMPT_GET_RETURN_ADDRESS(0));
149  ompt_callbacks.ompt_callback(ompt_callback_work)(
150  ompt_work_single_other, ompt_scope_end,
151  &(team->t.ompt_team_info.parallel_data),
152  &(team->t.t_implicit_task_taskdata[tid].ompt_task_info.task_data),
153  1, OMPT_GET_RETURN_ADDRESS(0));
154  }
155  }
156  }
157 #endif
158 
159  return rc;
160 }
161 
162 void *KMP_EXPAND_NAME(KMP_API_NAME_GOMP_SINGLE_COPY_START)(void) {
163  void *retval;
164  int gtid = __kmp_entry_gtid();
165  MKLOC(loc, "GOMP_single_copy_start");
166  KA_TRACE(20, ("GOMP_single_copy_start: T#%d\n", gtid));
167 
168  if (!TCR_4(__kmp_init_parallel))
169  __kmp_parallel_initialize();
170  __kmp_resume_if_soft_paused();
171 
172  // If this is the first thread to enter, return NULL. The generated code will
173  // then call GOMP_single_copy_end() for this thread only, with the
174  // copyprivate data pointer as an argument.
175  if (__kmp_enter_single(gtid, &loc, FALSE))
176  return NULL;
177 
178 // Wait for the first thread to set the copyprivate data pointer,
179 // and for all other threads to reach this point.
180 
181 #if OMPT_SUPPORT && OMPT_OPTIONAL
182  ompt_frame_t *ompt_frame;
183  if (ompt_enabled.enabled) {
184  __ompt_get_task_info_internal(0, NULL, NULL, &ompt_frame, NULL, NULL);
185  ompt_frame->enter_frame.ptr = OMPT_GET_FRAME_ADDRESS(0);
186  OMPT_STORE_RETURN_ADDRESS(gtid);
187  }
188 #endif
189  __kmp_barrier(bs_plain_barrier, gtid, FALSE, 0, NULL, NULL);
190 
191  // Retrieve the value of the copyprivate data point, and wait for all
192  // threads to do likewise, then return.
193  retval = __kmp_team_from_gtid(gtid)->t.t_copypriv_data;
194 #if OMPT_SUPPORT && OMPT_OPTIONAL
195  if (ompt_enabled.enabled) {
196  OMPT_STORE_RETURN_ADDRESS(gtid);
197  }
198 #endif
199  __kmp_barrier(bs_plain_barrier, gtid, FALSE, 0, NULL, NULL);
200 #if OMPT_SUPPORT && OMPT_OPTIONAL
201  if (ompt_enabled.enabled) {
202  ompt_frame->enter_frame = ompt_data_none;
203  }
204 #endif
205  return retval;
206 }
207 
208 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_SINGLE_COPY_END)(void *data) {
209  int gtid = __kmp_get_gtid();
210  KA_TRACE(20, ("GOMP_single_copy_end: T#%d\n", gtid));
211 
212  // Set the copyprivate data pointer fo the team, then hit the barrier so that
213  // the other threads will continue on and read it. Hit another barrier before
214  // continuing, so that the know that the copyprivate data pointer has been
215  // propagated to all threads before trying to reuse the t_copypriv_data field.
216  __kmp_team_from_gtid(gtid)->t.t_copypriv_data = data;
217 #if OMPT_SUPPORT && OMPT_OPTIONAL
218  ompt_frame_t *ompt_frame;
219  if (ompt_enabled.enabled) {
220  __ompt_get_task_info_internal(0, NULL, NULL, &ompt_frame, NULL, NULL);
221  ompt_frame->enter_frame.ptr = OMPT_GET_FRAME_ADDRESS(0);
222  OMPT_STORE_RETURN_ADDRESS(gtid);
223  }
224 #endif
225  __kmp_barrier(bs_plain_barrier, gtid, FALSE, 0, NULL, NULL);
226 #if OMPT_SUPPORT && OMPT_OPTIONAL
227  if (ompt_enabled.enabled) {
228  OMPT_STORE_RETURN_ADDRESS(gtid);
229  }
230 #endif
231  __kmp_barrier(bs_plain_barrier, gtid, FALSE, 0, NULL, NULL);
232 #if OMPT_SUPPORT && OMPT_OPTIONAL
233  if (ompt_enabled.enabled) {
234  ompt_frame->enter_frame = ompt_data_none;
235  }
236 #endif
237 }
238 
239 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_ORDERED_START)(void) {
240  int gtid = __kmp_entry_gtid();
241  MKLOC(loc, "GOMP_ordered_start");
242  KA_TRACE(20, ("GOMP_ordered_start: T#%d\n", gtid));
243 #if OMPT_SUPPORT && OMPT_OPTIONAL
244  OMPT_STORE_RETURN_ADDRESS(gtid);
245 #endif
246  __kmpc_ordered(&loc, gtid);
247 }
248 
249 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_ORDERED_END)(void) {
250  int gtid = __kmp_get_gtid();
251  MKLOC(loc, "GOMP_ordered_end");
252  KA_TRACE(20, ("GOMP_ordered_start: T#%d\n", gtid));
253 #if OMPT_SUPPORT && OMPT_OPTIONAL
254  OMPT_STORE_RETURN_ADDRESS(gtid);
255 #endif
256  __kmpc_end_ordered(&loc, gtid);
257 }
258 
259 // Dispatch macro defs
260 //
261 // They come in two flavors: 64-bit unsigned, and either 32-bit signed
262 // (IA-32 architecture) or 64-bit signed (Intel(R) 64).
263 
264 #if KMP_ARCH_X86 || KMP_ARCH_ARM || KMP_ARCH_MIPS
265 #define KMP_DISPATCH_INIT __kmp_aux_dispatch_init_4
266 #define KMP_DISPATCH_FINI_CHUNK __kmp_aux_dispatch_fini_chunk_4
267 #define KMP_DISPATCH_NEXT __kmpc_dispatch_next_4
268 #else
269 #define KMP_DISPATCH_INIT __kmp_aux_dispatch_init_8
270 #define KMP_DISPATCH_FINI_CHUNK __kmp_aux_dispatch_fini_chunk_8
271 #define KMP_DISPATCH_NEXT __kmpc_dispatch_next_8
272 #endif /* KMP_ARCH_X86 */
273 
274 #define KMP_DISPATCH_INIT_ULL __kmp_aux_dispatch_init_8u
275 #define KMP_DISPATCH_FINI_CHUNK_ULL __kmp_aux_dispatch_fini_chunk_8u
276 #define KMP_DISPATCH_NEXT_ULL __kmpc_dispatch_next_8u
277 
278 // The parallel contruct
279 
280 #ifndef KMP_DEBUG
281 static
282 #endif /* KMP_DEBUG */
283  void
284  __kmp_GOMP_microtask_wrapper(int *gtid, int *npr, void (*task)(void *),
285  void *data) {
286 #if OMPT_SUPPORT
287  kmp_info_t *thr;
288  ompt_frame_t *ompt_frame;
289  ompt_state_t enclosing_state;
290 
291  if (ompt_enabled.enabled) {
292  // get pointer to thread data structure
293  thr = __kmp_threads[*gtid];
294 
295  // save enclosing task state; set current state for task
296  enclosing_state = thr->th.ompt_thread_info.state;
297  thr->th.ompt_thread_info.state = ompt_state_work_parallel;
298 
299  // set task frame
300  __ompt_get_task_info_internal(0, NULL, NULL, &ompt_frame, NULL, NULL);
301  ompt_frame->exit_frame.ptr = OMPT_GET_FRAME_ADDRESS(0);
302  }
303 #endif
304 
305  task(data);
306 
307 #if OMPT_SUPPORT
308  if (ompt_enabled.enabled) {
309  // clear task frame
310  ompt_frame->exit_frame = ompt_data_none;
311 
312  // restore enclosing state
313  thr->th.ompt_thread_info.state = enclosing_state;
314  }
315 #endif
316 }
317 
318 #ifndef KMP_DEBUG
319 static
320 #endif /* KMP_DEBUG */
321  void
322  __kmp_GOMP_parallel_microtask_wrapper(int *gtid, int *npr,
323  void (*task)(void *), void *data,
324  unsigned num_threads, ident_t *loc,
325  enum sched_type schedule, long start,
326  long end, long incr,
327  long chunk_size) {
328  // Intialize the loop worksharing construct.
329 
330  KMP_DISPATCH_INIT(loc, *gtid, schedule, start, end, incr, chunk_size,
331  schedule != kmp_sch_static);
332 
333 #if OMPT_SUPPORT
334  kmp_info_t *thr;
335  ompt_frame_t *ompt_frame;
336  ompt_state_t enclosing_state;
337 
338  if (ompt_enabled.enabled) {
339  thr = __kmp_threads[*gtid];
340  // save enclosing task state; set current state for task
341  enclosing_state = thr->th.ompt_thread_info.state;
342  thr->th.ompt_thread_info.state = ompt_state_work_parallel;
343 
344  // set task frame
345  __ompt_get_task_info_internal(0, NULL, NULL, &ompt_frame, NULL, NULL);
346  ompt_frame->exit_frame.ptr = OMPT_GET_FRAME_ADDRESS(0);
347  }
348 #endif
349 
350  // Now invoke the microtask.
351  task(data);
352 
353 #if OMPT_SUPPORT
354  if (ompt_enabled.enabled) {
355  // clear task frame
356  ompt_frame->exit_frame = ompt_data_none;
357 
358  // reset enclosing state
359  thr->th.ompt_thread_info.state = enclosing_state;
360  }
361 #endif
362 }
363 
364 #ifndef KMP_DEBUG
365 static
366 #endif /* KMP_DEBUG */
367  void
368  __kmp_GOMP_fork_call(ident_t *loc, int gtid, void (*unwrapped_task)(void *),
369  microtask_t wrapper, int argc, ...) {
370  int rc;
371  kmp_info_t *thr = __kmp_threads[gtid];
372  kmp_team_t *team = thr->th.th_team;
373  int tid = __kmp_tid_from_gtid(gtid);
374 
375  va_list ap;
376  va_start(ap, argc);
377 
378  rc = __kmp_fork_call(loc, gtid, fork_context_gnu, argc, wrapper,
379  __kmp_invoke_task_func,
380 #if (KMP_ARCH_X86_64 || KMP_ARCH_ARM || KMP_ARCH_AARCH64) && KMP_OS_LINUX
381  &ap
382 #else
383  ap
384 #endif
385  );
386 
387  va_end(ap);
388 
389  if (rc) {
390  __kmp_run_before_invoked_task(gtid, tid, thr, team);
391  }
392 
393 #if OMPT_SUPPORT
394  int ompt_team_size;
395  if (ompt_enabled.enabled) {
396  ompt_team_info_t *team_info = __ompt_get_teaminfo(0, NULL);
397  ompt_task_info_t *task_info = __ompt_get_task_info_object(0);
398 
399  // implicit task callback
400  if (ompt_enabled.ompt_callback_implicit_task) {
401  ompt_team_size = __kmp_team_from_gtid(gtid)->t.t_nproc;
402  ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
403  ompt_scope_begin, &(team_info->parallel_data),
404  &(task_info->task_data), ompt_team_size, __kmp_tid_from_gtid(gtid), ompt_task_implicit); // TODO: Can this be ompt_task_initial?
405  task_info->thread_num = __kmp_tid_from_gtid(gtid);
406  }
407  thr->th.ompt_thread_info.state = ompt_state_work_parallel;
408  }
409 #endif
410 }
411 
412 static void __kmp_GOMP_serialized_parallel(ident_t *loc, kmp_int32 gtid,
413  void (*task)(void *)) {
414 #if OMPT_SUPPORT
415  OMPT_STORE_RETURN_ADDRESS(gtid);
416 #endif
417  __kmp_serialized_parallel(loc, gtid);
418 }
419 
420 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_START)(void (*task)(void *),
421  void *data,
422  unsigned num_threads) {
423  int gtid = __kmp_entry_gtid();
424 
425 #if OMPT_SUPPORT
426  ompt_frame_t *parent_frame, *frame;
427 
428  if (ompt_enabled.enabled) {
429  __ompt_get_task_info_internal(0, NULL, NULL, &parent_frame, NULL, NULL);
430  parent_frame->enter_frame.ptr = OMPT_GET_FRAME_ADDRESS(0);
431  OMPT_STORE_RETURN_ADDRESS(gtid);
432  }
433 #endif
434 
435  MKLOC(loc, "GOMP_parallel_start");
436  KA_TRACE(20, ("GOMP_parallel_start: T#%d\n", gtid));
437 
438  if (__kmpc_ok_to_fork(&loc) && (num_threads != 1)) {
439  if (num_threads != 0) {
440  __kmp_push_num_threads(&loc, gtid, num_threads);
441  }
442  __kmp_GOMP_fork_call(&loc, gtid, task,
443  (microtask_t)__kmp_GOMP_microtask_wrapper, 2, task,
444  data);
445  } else {
446  __kmp_GOMP_serialized_parallel(&loc, gtid, task);
447  }
448 
449 #if OMPT_SUPPORT
450  if (ompt_enabled.enabled) {
451  __ompt_get_task_info_internal(0, NULL, NULL, &frame, NULL, NULL);
452  frame->exit_frame.ptr = OMPT_GET_FRAME_ADDRESS(0);
453  }
454 #endif
455 }
456 
457 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_END)(void) {
458  int gtid = __kmp_get_gtid();
459  kmp_info_t *thr;
460 
461  thr = __kmp_threads[gtid];
462 
463  MKLOC(loc, "GOMP_parallel_end");
464  KA_TRACE(20, ("GOMP_parallel_end: T#%d\n", gtid));
465 
466  if (!thr->th.th_team->t.t_serialized) {
467  __kmp_run_after_invoked_task(gtid, __kmp_tid_from_gtid(gtid), thr,
468  thr->th.th_team);
469 
470 #if OMPT_SUPPORT
471  if (ompt_enabled.enabled) {
472  // Implicit task is finished here, in the barrier we might schedule
473  // deferred tasks,
474  // these don't see the implicit task on the stack
475  OMPT_CUR_TASK_INFO(thr)->frame.exit_frame = ompt_data_none;
476  }
477 #endif
478 
479  __kmp_join_call(&loc, gtid
480 #if OMPT_SUPPORT
481  ,
482  fork_context_gnu
483 #endif
484  );
485  } else {
486  __kmpc_end_serialized_parallel(&loc, gtid);
487  }
488 }
489 
490 // Loop worksharing constructs
491 
492 // The Gnu codegen passes in an exclusive upper bound for the overall range,
493 // but the libguide dispatch code expects an inclusive upper bound, hence the
494 // "end - incr" 5th argument to KMP_DISPATCH_INIT (and the " ub - str" 11th
495 // argument to __kmp_GOMP_fork_call).
496 //
497 // Conversely, KMP_DISPATCH_NEXT returns and inclusive upper bound in *p_ub,
498 // but the Gnu codegen expects an excluside upper bound, so the adjustment
499 // "*p_ub += stride" compenstates for the discrepancy.
500 //
501 // Correction: the gnu codegen always adjusts the upper bound by +-1, not the
502 // stride value. We adjust the dispatch parameters accordingly (by +-1), but
503 // we still adjust p_ub by the actual stride value.
504 //
505 // The "runtime" versions do not take a chunk_sz parameter.
506 //
507 // The profile lib cannot support construct checking of unordered loops that
508 // are predetermined by the compiler to be statically scheduled, as the gcc
509 // codegen will not always emit calls to GOMP_loop_static_next() to get the
510 // next iteration. Instead, it emits inline code to call omp_get_thread_num()
511 // num and calculate the iteration space using the result. It doesn't do this
512 // with ordered static loop, so they can be checked.
513 
514 #if OMPT_SUPPORT
515 #define IF_OMPT_SUPPORT(code) code
516 #else
517 #define IF_OMPT_SUPPORT(code)
518 #endif
519 
520 #define LOOP_START(func, schedule) \
521  int func(long lb, long ub, long str, long chunk_sz, long *p_lb, \
522  long *p_ub) { \
523  int status; \
524  long stride; \
525  int gtid = __kmp_entry_gtid(); \
526  MKLOC(loc, KMP_STR(func)); \
527  KA_TRACE( \
528  20, \
529  (KMP_STR( \
530  func) ": T#%d, lb 0x%lx, ub 0x%lx, str 0x%lx, chunk_sz 0x%lx\n", \
531  gtid, lb, ub, str, chunk_sz)); \
532  \
533  if ((str > 0) ? (lb < ub) : (lb > ub)) { \
534  IF_OMPT_SUPPORT(OMPT_STORE_RETURN_ADDRESS(gtid);) \
535  KMP_DISPATCH_INIT(&loc, gtid, (schedule), lb, \
536  (str > 0) ? (ub - 1) : (ub + 1), str, chunk_sz, \
537  (schedule) != kmp_sch_static); \
538  IF_OMPT_SUPPORT(OMPT_STORE_RETURN_ADDRESS(gtid);) \
539  status = KMP_DISPATCH_NEXT(&loc, gtid, NULL, (kmp_int *)p_lb, \
540  (kmp_int *)p_ub, (kmp_int *)&stride); \
541  if (status) { \
542  KMP_DEBUG_ASSERT(stride == str); \
543  *p_ub += (str > 0) ? 1 : -1; \
544  } \
545  } else { \
546  status = 0; \
547  } \
548  \
549  KA_TRACE( \
550  20, \
551  (KMP_STR( \
552  func) " exit: T#%d, *p_lb 0x%lx, *p_ub 0x%lx, returning %d\n", \
553  gtid, *p_lb, *p_ub, status)); \
554  return status; \
555  }
556 
557 #define LOOP_RUNTIME_START(func, schedule) \
558  int func(long lb, long ub, long str, long *p_lb, long *p_ub) { \
559  int status; \
560  long stride; \
561  long chunk_sz = 0; \
562  int gtid = __kmp_entry_gtid(); \
563  MKLOC(loc, KMP_STR(func)); \
564  KA_TRACE( \
565  20, \
566  (KMP_STR(func) ": T#%d, lb 0x%lx, ub 0x%lx, str 0x%lx, chunk_sz %d\n", \
567  gtid, lb, ub, str, chunk_sz)); \
568  \
569  if ((str > 0) ? (lb < ub) : (lb > ub)) { \
570  IF_OMPT_SUPPORT(OMPT_STORE_RETURN_ADDRESS(gtid);) \
571  KMP_DISPATCH_INIT(&loc, gtid, (schedule), lb, \
572  (str > 0) ? (ub - 1) : (ub + 1), str, chunk_sz, TRUE); \
573  IF_OMPT_SUPPORT(OMPT_STORE_RETURN_ADDRESS(gtid);) \
574  status = KMP_DISPATCH_NEXT(&loc, gtid, NULL, (kmp_int *)p_lb, \
575  (kmp_int *)p_ub, (kmp_int *)&stride); \
576  if (status) { \
577  KMP_DEBUG_ASSERT(stride == str); \
578  *p_ub += (str > 0) ? 1 : -1; \
579  } \
580  } else { \
581  status = 0; \
582  } \
583  \
584  KA_TRACE( \
585  20, \
586  (KMP_STR( \
587  func) " exit: T#%d, *p_lb 0x%lx, *p_ub 0x%lx, returning %d\n", \
588  gtid, *p_lb, *p_ub, status)); \
589  return status; \
590  }
591 
592 #define KMP_DOACROSS_FINI(status, gtid) \
593  if (!status && __kmp_threads[gtid]->th.th_dispatch->th_doacross_flags) { \
594  __kmpc_doacross_fini(NULL, gtid); \
595  }
596 
597 #define LOOP_NEXT(func, fini_code) \
598  int func(long *p_lb, long *p_ub) { \
599  int status; \
600  long stride; \
601  int gtid = __kmp_get_gtid(); \
602  MKLOC(loc, KMP_STR(func)); \
603  KA_TRACE(20, (KMP_STR(func) ": T#%d\n", gtid)); \
604  \
605  IF_OMPT_SUPPORT(OMPT_STORE_RETURN_ADDRESS(gtid);) \
606  fini_code status = KMP_DISPATCH_NEXT(&loc, gtid, NULL, (kmp_int *)p_lb, \
607  (kmp_int *)p_ub, (kmp_int *)&stride); \
608  if (status) { \
609  *p_ub += (stride > 0) ? 1 : -1; \
610  } \
611  KMP_DOACROSS_FINI(status, gtid) \
612  \
613  KA_TRACE( \
614  20, \
615  (KMP_STR(func) " exit: T#%d, *p_lb 0x%lx, *p_ub 0x%lx, stride 0x%lx, " \
616  "returning %d\n", \
617  gtid, *p_lb, *p_ub, stride, status)); \
618  return status; \
619  }
620 
621 LOOP_START(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_STATIC_START), kmp_sch_static)
622 LOOP_NEXT(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_STATIC_NEXT), {})
623 LOOP_START(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_DYNAMIC_START),
624  kmp_sch_dynamic_chunked)
625 LOOP_NEXT(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_DYNAMIC_NEXT), {})
626 LOOP_START(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_GUIDED_START),
628 LOOP_NEXT(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_GUIDED_NEXT), {})
629 LOOP_RUNTIME_START(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_RUNTIME_START),
630  kmp_sch_runtime)
631 LOOP_NEXT(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_RUNTIME_NEXT), {})
632 
633 LOOP_START(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ORDERED_STATIC_START),
635 LOOP_NEXT(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ORDERED_STATIC_NEXT),
636  { KMP_DISPATCH_FINI_CHUNK(&loc, gtid); })
637 LOOP_START(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ORDERED_DYNAMIC_START),
638  kmp_ord_dynamic_chunked)
639 LOOP_NEXT(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ORDERED_DYNAMIC_NEXT),
640  { KMP_DISPATCH_FINI_CHUNK(&loc, gtid); })
641 LOOP_START(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ORDERED_GUIDED_START),
642  kmp_ord_guided_chunked)
643 LOOP_NEXT(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ORDERED_GUIDED_NEXT),
644  { KMP_DISPATCH_FINI_CHUNK(&loc, gtid); })
645 LOOP_RUNTIME_START(
646  KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ORDERED_RUNTIME_START),
647  kmp_ord_runtime)
648 LOOP_NEXT(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ORDERED_RUNTIME_NEXT),
649  { KMP_DISPATCH_FINI_CHUNK(&loc, gtid); })
650 
651 #define LOOP_DOACROSS_START(func, schedule) \
652  bool func(unsigned ncounts, long *counts, long chunk_sz, long *p_lb, \
653  long *p_ub) { \
654  int status; \
655  long stride, lb, ub, str; \
656  int gtid = __kmp_entry_gtid(); \
657  struct kmp_dim *dims = \
658  (struct kmp_dim *)__kmp_allocate(sizeof(struct kmp_dim) * ncounts); \
659  MKLOC(loc, KMP_STR(func)); \
660  for (unsigned i = 0; i < ncounts; ++i) { \
661  dims[i].lo = 0; \
662  dims[i].up = counts[i] - 1; \
663  dims[i].st = 1; \
664  } \
665  __kmpc_doacross_init(&loc, gtid, (int)ncounts, dims); \
666  lb = 0; \
667  ub = counts[0]; \
668  str = 1; \
669  KA_TRACE(20, (KMP_STR(func) ": T#%d, ncounts %u, lb 0x%lx, ub 0x%lx, str " \
670  "0x%lx, chunk_sz " \
671  "0x%lx\n", \
672  gtid, ncounts, lb, ub, str, chunk_sz)); \
673  \
674  if ((str > 0) ? (lb < ub) : (lb > ub)) { \
675  KMP_DISPATCH_INIT(&loc, gtid, (schedule), lb, \
676  (str > 0) ? (ub - 1) : (ub + 1), str, chunk_sz, \
677  (schedule) != kmp_sch_static); \
678  status = KMP_DISPATCH_NEXT(&loc, gtid, NULL, (kmp_int *)p_lb, \
679  (kmp_int *)p_ub, (kmp_int *)&stride); \
680  if (status) { \
681  KMP_DEBUG_ASSERT(stride == str); \
682  *p_ub += (str > 0) ? 1 : -1; \
683  } \
684  } else { \
685  status = 0; \
686  } \
687  KMP_DOACROSS_FINI(status, gtid); \
688  \
689  KA_TRACE( \
690  20, \
691  (KMP_STR( \
692  func) " exit: T#%d, *p_lb 0x%lx, *p_ub 0x%lx, returning %d\n", \
693  gtid, *p_lb, *p_ub, status)); \
694  __kmp_free(dims); \
695  return status; \
696  }
697 
698 #define LOOP_DOACROSS_RUNTIME_START(func, schedule) \
699  int func(unsigned ncounts, long *counts, long *p_lb, long *p_ub) { \
700  int status; \
701  long stride, lb, ub, str; \
702  long chunk_sz = 0; \
703  int gtid = __kmp_entry_gtid(); \
704  struct kmp_dim *dims = \
705  (struct kmp_dim *)__kmp_allocate(sizeof(struct kmp_dim) * ncounts); \
706  MKLOC(loc, KMP_STR(func)); \
707  for (unsigned i = 0; i < ncounts; ++i) { \
708  dims[i].lo = 0; \
709  dims[i].up = counts[i] - 1; \
710  dims[i].st = 1; \
711  } \
712  __kmpc_doacross_init(&loc, gtid, (int)ncounts, dims); \
713  lb = 0; \
714  ub = counts[0]; \
715  str = 1; \
716  KA_TRACE( \
717  20, \
718  (KMP_STR(func) ": T#%d, lb 0x%lx, ub 0x%lx, str 0x%lx, chunk_sz %d\n", \
719  gtid, lb, ub, str, chunk_sz)); \
720  \
721  if ((str > 0) ? (lb < ub) : (lb > ub)) { \
722  KMP_DISPATCH_INIT(&loc, gtid, (schedule), lb, \
723  (str > 0) ? (ub - 1) : (ub + 1), str, chunk_sz, TRUE); \
724  status = KMP_DISPATCH_NEXT(&loc, gtid, NULL, (kmp_int *)p_lb, \
725  (kmp_int *)p_ub, (kmp_int *)&stride); \
726  if (status) { \
727  KMP_DEBUG_ASSERT(stride == str); \
728  *p_ub += (str > 0) ? 1 : -1; \
729  } \
730  } else { \
731  status = 0; \
732  } \
733  KMP_DOACROSS_FINI(status, gtid); \
734  \
735  KA_TRACE( \
736  20, \
737  (KMP_STR( \
738  func) " exit: T#%d, *p_lb 0x%lx, *p_ub 0x%lx, returning %d\n", \
739  gtid, *p_lb, *p_ub, status)); \
740  __kmp_free(dims); \
741  return status; \
742  }
743 
744 LOOP_DOACROSS_START(
745  KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_DOACROSS_STATIC_START),
747 LOOP_DOACROSS_START(
748  KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_DOACROSS_DYNAMIC_START),
749  kmp_sch_dynamic_chunked)
750 LOOP_DOACROSS_START(
751  KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_DOACROSS_GUIDED_START),
753 LOOP_DOACROSS_RUNTIME_START(
754  KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_DOACROSS_RUNTIME_START),
755  kmp_sch_runtime)
756 
757 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_END)(void) {
758  int gtid = __kmp_get_gtid();
759  KA_TRACE(20, ("GOMP_loop_end: T#%d\n", gtid))
760 
761 #if OMPT_SUPPORT && OMPT_OPTIONAL
762  ompt_frame_t *ompt_frame;
763  if (ompt_enabled.enabled) {
764  __ompt_get_task_info_internal(0, NULL, NULL, &ompt_frame, NULL, NULL);
765  ompt_frame->enter_frame.ptr = OMPT_GET_FRAME_ADDRESS(0);
766  OMPT_STORE_RETURN_ADDRESS(gtid);
767  }
768 #endif
769  __kmp_barrier(bs_plain_barrier, gtid, FALSE, 0, NULL, NULL);
770 #if OMPT_SUPPORT && OMPT_OPTIONAL
771  if (ompt_enabled.enabled) {
772  ompt_frame->enter_frame = ompt_data_none;
773  }
774 #endif
775 
776  KA_TRACE(20, ("GOMP_loop_end exit: T#%d\n", gtid))
777 }
778 
779 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_END_NOWAIT)(void) {
780  KA_TRACE(20, ("GOMP_loop_end_nowait: T#%d\n", __kmp_get_gtid()))
781 }
782 
783 // Unsigned long long loop worksharing constructs
784 //
785 // These are new with gcc 4.4
786 
787 #define LOOP_START_ULL(func, schedule) \
788  int func(int up, unsigned long long lb, unsigned long long ub, \
789  unsigned long long str, unsigned long long chunk_sz, \
790  unsigned long long *p_lb, unsigned long long *p_ub) { \
791  int status; \
792  long long str2 = up ? ((long long)str) : -((long long)str); \
793  long long stride; \
794  int gtid = __kmp_entry_gtid(); \
795  MKLOC(loc, KMP_STR(func)); \
796  \
797  KA_TRACE(20, (KMP_STR(func) ": T#%d, up %d, lb 0x%llx, ub 0x%llx, str " \
798  "0x%llx, chunk_sz 0x%llx\n", \
799  gtid, up, lb, ub, str, chunk_sz)); \
800  \
801  if ((str > 0) ? (lb < ub) : (lb > ub)) { \
802  KMP_DISPATCH_INIT_ULL(&loc, gtid, (schedule), lb, \
803  (str2 > 0) ? (ub - 1) : (ub + 1), str2, chunk_sz, \
804  (schedule) != kmp_sch_static); \
805  status = \
806  KMP_DISPATCH_NEXT_ULL(&loc, gtid, NULL, (kmp_uint64 *)p_lb, \
807  (kmp_uint64 *)p_ub, (kmp_int64 *)&stride); \
808  if (status) { \
809  KMP_DEBUG_ASSERT(stride == str2); \
810  *p_ub += (str > 0) ? 1 : -1; \
811  } \
812  } else { \
813  status = 0; \
814  } \
815  \
816  KA_TRACE( \
817  20, \
818  (KMP_STR( \
819  func) " exit: T#%d, *p_lb 0x%llx, *p_ub 0x%llx, returning %d\n", \
820  gtid, *p_lb, *p_ub, status)); \
821  return status; \
822  }
823 
824 #define LOOP_RUNTIME_START_ULL(func, schedule) \
825  int func(int up, unsigned long long lb, unsigned long long ub, \
826  unsigned long long str, unsigned long long *p_lb, \
827  unsigned long long *p_ub) { \
828  int status; \
829  long long str2 = up ? ((long long)str) : -((long long)str); \
830  unsigned long long stride; \
831  unsigned long long chunk_sz = 0; \
832  int gtid = __kmp_entry_gtid(); \
833  MKLOC(loc, KMP_STR(func)); \
834  \
835  KA_TRACE(20, (KMP_STR(func) ": T#%d, up %d, lb 0x%llx, ub 0x%llx, str " \
836  "0x%llx, chunk_sz 0x%llx\n", \
837  gtid, up, lb, ub, str, chunk_sz)); \
838  \
839  if ((str > 0) ? (lb < ub) : (lb > ub)) { \
840  KMP_DISPATCH_INIT_ULL(&loc, gtid, (schedule), lb, \
841  (str2 > 0) ? (ub - 1) : (ub + 1), str2, chunk_sz, \
842  TRUE); \
843  status = \
844  KMP_DISPATCH_NEXT_ULL(&loc, gtid, NULL, (kmp_uint64 *)p_lb, \
845  (kmp_uint64 *)p_ub, (kmp_int64 *)&stride); \
846  if (status) { \
847  KMP_DEBUG_ASSERT((long long)stride == str2); \
848  *p_ub += (str > 0) ? 1 : -1; \
849  } \
850  } else { \
851  status = 0; \
852  } \
853  \
854  KA_TRACE( \
855  20, \
856  (KMP_STR( \
857  func) " exit: T#%d, *p_lb 0x%llx, *p_ub 0x%llx, returning %d\n", \
858  gtid, *p_lb, *p_ub, status)); \
859  return status; \
860  }
861 
862 #define LOOP_NEXT_ULL(func, fini_code) \
863  int func(unsigned long long *p_lb, unsigned long long *p_ub) { \
864  int status; \
865  long long stride; \
866  int gtid = __kmp_get_gtid(); \
867  MKLOC(loc, KMP_STR(func)); \
868  KA_TRACE(20, (KMP_STR(func) ": T#%d\n", gtid)); \
869  \
870  fini_code status = \
871  KMP_DISPATCH_NEXT_ULL(&loc, gtid, NULL, (kmp_uint64 *)p_lb, \
872  (kmp_uint64 *)p_ub, (kmp_int64 *)&stride); \
873  if (status) { \
874  *p_ub += (stride > 0) ? 1 : -1; \
875  } \
876  \
877  KA_TRACE( \
878  20, \
879  (KMP_STR( \
880  func) " exit: T#%d, *p_lb 0x%llx, *p_ub 0x%llx, stride 0x%llx, " \
881  "returning %d\n", \
882  gtid, *p_lb, *p_ub, stride, status)); \
883  return status; \
884  }
885 
886 LOOP_START_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_STATIC_START),
888 LOOP_NEXT_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_STATIC_NEXT), {})
889 LOOP_START_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_DYNAMIC_START),
890  kmp_sch_dynamic_chunked)
891 LOOP_NEXT_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_DYNAMIC_NEXT), {})
892 LOOP_START_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_GUIDED_START),
894 LOOP_NEXT_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_GUIDED_NEXT), {})
895 LOOP_RUNTIME_START_ULL(
896  KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_RUNTIME_START), kmp_sch_runtime)
897 LOOP_NEXT_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_RUNTIME_NEXT), {})
898 
899 LOOP_START_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_STATIC_START),
901 LOOP_NEXT_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_STATIC_NEXT),
902  { KMP_DISPATCH_FINI_CHUNK_ULL(&loc, gtid); })
903 LOOP_START_ULL(
904  KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_DYNAMIC_START),
905  kmp_ord_dynamic_chunked)
906 LOOP_NEXT_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_DYNAMIC_NEXT),
907  { KMP_DISPATCH_FINI_CHUNK_ULL(&loc, gtid); })
908 LOOP_START_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_GUIDED_START),
909  kmp_ord_guided_chunked)
910 LOOP_NEXT_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_GUIDED_NEXT),
911  { KMP_DISPATCH_FINI_CHUNK_ULL(&loc, gtid); })
912 LOOP_RUNTIME_START_ULL(
913  KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_RUNTIME_START),
914  kmp_ord_runtime)
915 LOOP_NEXT_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_RUNTIME_NEXT),
916  { KMP_DISPATCH_FINI_CHUNK_ULL(&loc, gtid); })
917 
918 #define LOOP_DOACROSS_START_ULL(func, schedule) \
919  int func(unsigned ncounts, unsigned long long *counts, \
920  unsigned long long chunk_sz, unsigned long long *p_lb, \
921  unsigned long long *p_ub) { \
922  int status; \
923  long long stride, str, lb, ub; \
924  int gtid = __kmp_entry_gtid(); \
925  struct kmp_dim *dims = \
926  (struct kmp_dim *)__kmp_allocate(sizeof(struct kmp_dim) * ncounts); \
927  MKLOC(loc, KMP_STR(func)); \
928  for (unsigned i = 0; i < ncounts; ++i) { \
929  dims[i].lo = 0; \
930  dims[i].up = counts[i] - 1; \
931  dims[i].st = 1; \
932  } \
933  __kmpc_doacross_init(&loc, gtid, (int)ncounts, dims); \
934  lb = 0; \
935  ub = counts[0]; \
936  str = 1; \
937  \
938  KA_TRACE(20, (KMP_STR(func) ": T#%d, lb 0x%llx, ub 0x%llx, str " \
939  "0x%llx, chunk_sz 0x%llx\n", \
940  gtid, lb, ub, str, chunk_sz)); \
941  \
942  if ((str > 0) ? (lb < ub) : (lb > ub)) { \
943  KMP_DISPATCH_INIT_ULL(&loc, gtid, (schedule), lb, \
944  (str > 0) ? (ub - 1) : (ub + 1), str, chunk_sz, \
945  (schedule) != kmp_sch_static); \
946  status = \
947  KMP_DISPATCH_NEXT_ULL(&loc, gtid, NULL, (kmp_uint64 *)p_lb, \
948  (kmp_uint64 *)p_ub, (kmp_int64 *)&stride); \
949  if (status) { \
950  KMP_DEBUG_ASSERT(stride == str); \
951  *p_ub += (str > 0) ? 1 : -1; \
952  } \
953  } else { \
954  status = 0; \
955  } \
956  KMP_DOACROSS_FINI(status, gtid); \
957  \
958  KA_TRACE( \
959  20, \
960  (KMP_STR( \
961  func) " exit: T#%d, *p_lb 0x%llx, *p_ub 0x%llx, returning %d\n", \
962  gtid, *p_lb, *p_ub, status)); \
963  __kmp_free(dims); \
964  return status; \
965  }
966 
967 #define LOOP_DOACROSS_RUNTIME_START_ULL(func, schedule) \
968  int func(unsigned ncounts, unsigned long long *counts, \
969  unsigned long long *p_lb, unsigned long long *p_ub) { \
970  int status; \
971  unsigned long long stride, str, lb, ub; \
972  unsigned long long chunk_sz = 0; \
973  int gtid = __kmp_entry_gtid(); \
974  struct kmp_dim *dims = \
975  (struct kmp_dim *)__kmp_allocate(sizeof(struct kmp_dim) * ncounts); \
976  MKLOC(loc, KMP_STR(func)); \
977  for (unsigned i = 0; i < ncounts; ++i) { \
978  dims[i].lo = 0; \
979  dims[i].up = counts[i] - 1; \
980  dims[i].st = 1; \
981  } \
982  __kmpc_doacross_init(&loc, gtid, (int)ncounts, dims); \
983  lb = 0; \
984  ub = counts[0]; \
985  str = 1; \
986  KA_TRACE(20, (KMP_STR(func) ": T#%d, lb 0x%llx, ub 0x%llx, str " \
987  "0x%llx, chunk_sz 0x%llx\n", \
988  gtid, lb, ub, str, chunk_sz)); \
989  \
990  if ((str > 0) ? (lb < ub) : (lb > ub)) { \
991  KMP_DISPATCH_INIT_ULL(&loc, gtid, (schedule), lb, \
992  (str > 0) ? (ub - 1) : (ub + 1), str, chunk_sz, \
993  TRUE); \
994  status = \
995  KMP_DISPATCH_NEXT_ULL(&loc, gtid, NULL, (kmp_uint64 *)p_lb, \
996  (kmp_uint64 *)p_ub, (kmp_int64 *)&stride); \
997  if (status) { \
998  KMP_DEBUG_ASSERT(stride == str); \
999  *p_ub += (str > 0) ? 1 : -1; \
1000  } \
1001  } else { \
1002  status = 0; \
1003  } \
1004  KMP_DOACROSS_FINI(status, gtid); \
1005  \
1006  KA_TRACE( \
1007  20, \
1008  (KMP_STR( \
1009  func) " exit: T#%d, *p_lb 0x%llx, *p_ub 0x%llx, returning %d\n", \
1010  gtid, *p_lb, *p_ub, status)); \
1011  __kmp_free(dims); \
1012  return status; \
1013  }
1014 
1015 LOOP_DOACROSS_START_ULL(
1016  KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_DOACROSS_STATIC_START),
1018 LOOP_DOACROSS_START_ULL(
1019  KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_DOACROSS_DYNAMIC_START),
1020  kmp_sch_dynamic_chunked)
1021 LOOP_DOACROSS_START_ULL(
1022  KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_DOACROSS_GUIDED_START),
1024 LOOP_DOACROSS_RUNTIME_START_ULL(
1025  KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_DOACROSS_RUNTIME_START),
1026  kmp_sch_runtime)
1027 
1028 // Combined parallel / loop worksharing constructs
1029 //
1030 // There are no ull versions (yet).
1031 
1032 #define PARALLEL_LOOP_START(func, schedule, ompt_pre, ompt_post) \
1033  void func(void (*task)(void *), void *data, unsigned num_threads, long lb, \
1034  long ub, long str, long chunk_sz) { \
1035  int gtid = __kmp_entry_gtid(); \
1036  MKLOC(loc, KMP_STR(func)); \
1037  KA_TRACE( \
1038  20, \
1039  (KMP_STR( \
1040  func) ": T#%d, lb 0x%lx, ub 0x%lx, str 0x%lx, chunk_sz 0x%lx\n", \
1041  gtid, lb, ub, str, chunk_sz)); \
1042  \
1043  ompt_pre(); \
1044  \
1045  if (__kmpc_ok_to_fork(&loc) && (num_threads != 1)) { \
1046  if (num_threads != 0) { \
1047  __kmp_push_num_threads(&loc, gtid, num_threads); \
1048  } \
1049  __kmp_GOMP_fork_call(&loc, gtid, task, \
1050  (microtask_t)__kmp_GOMP_parallel_microtask_wrapper, \
1051  9, task, data, num_threads, &loc, (schedule), lb, \
1052  (str > 0) ? (ub - 1) : (ub + 1), str, chunk_sz); \
1053  IF_OMPT_SUPPORT(OMPT_STORE_RETURN_ADDRESS(gtid)); \
1054  } else { \
1055  __kmp_GOMP_serialized_parallel(&loc, gtid, task); \
1056  IF_OMPT_SUPPORT(OMPT_STORE_RETURN_ADDRESS(gtid)); \
1057  } \
1058  \
1059  KMP_DISPATCH_INIT(&loc, gtid, (schedule), lb, \
1060  (str > 0) ? (ub - 1) : (ub + 1), str, chunk_sz, \
1061  (schedule) != kmp_sch_static); \
1062  \
1063  ompt_post(); \
1064  \
1065  KA_TRACE(20, (KMP_STR(func) " exit: T#%d\n", gtid)); \
1066  }
1067 
1068 #if OMPT_SUPPORT && OMPT_OPTIONAL
1069 
1070 #define OMPT_LOOP_PRE() \
1071  ompt_frame_t *parent_frame; \
1072  if (ompt_enabled.enabled) { \
1073  __ompt_get_task_info_internal(0, NULL, NULL, &parent_frame, NULL, NULL); \
1074  parent_frame->enter_frame.ptr = OMPT_GET_FRAME_ADDRESS(0); \
1075  OMPT_STORE_RETURN_ADDRESS(gtid); \
1076  }
1077 
1078 #define OMPT_LOOP_POST() \
1079  if (ompt_enabled.enabled) { \
1080  parent_frame->enter_frame = ompt_data_none; \
1081  }
1082 
1083 #else
1084 
1085 #define OMPT_LOOP_PRE()
1086 
1087 #define OMPT_LOOP_POST()
1088 
1089 #endif
1090 
1091 PARALLEL_LOOP_START(
1092  KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_LOOP_STATIC_START),
1093  kmp_sch_static, OMPT_LOOP_PRE, OMPT_LOOP_POST)
1094 PARALLEL_LOOP_START(
1095  KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_LOOP_DYNAMIC_START),
1096  kmp_sch_dynamic_chunked, OMPT_LOOP_PRE, OMPT_LOOP_POST)
1097 PARALLEL_LOOP_START(
1098  KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_LOOP_GUIDED_START),
1099  kmp_sch_guided_chunked, OMPT_LOOP_PRE, OMPT_LOOP_POST)
1100 PARALLEL_LOOP_START(
1101  KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_LOOP_RUNTIME_START),
1102  kmp_sch_runtime, OMPT_LOOP_PRE, OMPT_LOOP_POST)
1103 
1104 // Tasking constructs
1105 
1106 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TASK)(void (*func)(void *), void *data,
1107  void (*copy_func)(void *, void *),
1108  long arg_size, long arg_align,
1109  bool if_cond, unsigned gomp_flags,
1110  void **depend) {
1111  MKLOC(loc, "GOMP_task");
1112  int gtid = __kmp_entry_gtid();
1113  kmp_int32 flags = 0;
1114  kmp_tasking_flags_t *input_flags = (kmp_tasking_flags_t *)&flags;
1115 
1116  KA_TRACE(20, ("GOMP_task: T#%d\n", gtid));
1117 
1118  // The low-order bit is the "untied" flag
1119  if (!(gomp_flags & 1)) {
1120  input_flags->tiedness = 1;
1121  }
1122  // The second low-order bit is the "final" flag
1123  if (gomp_flags & 2) {
1124  input_flags->final = 1;
1125  }
1126  input_flags->native = 1;
1127  // __kmp_task_alloc() sets up all other flags
1128 
1129  if (!if_cond) {
1130  arg_size = 0;
1131  }
1132 
1133  kmp_task_t *task = __kmp_task_alloc(
1134  &loc, gtid, input_flags, sizeof(kmp_task_t),
1135  arg_size ? arg_size + arg_align - 1 : 0, (kmp_routine_entry_t)func);
1136 
1137  if (arg_size > 0) {
1138  if (arg_align > 0) {
1139  task->shareds = (void *)((((size_t)task->shareds) + arg_align - 1) /
1140  arg_align * arg_align);
1141  }
1142  // else error??
1143 
1144  if (copy_func) {
1145  (*copy_func)(task->shareds, data);
1146  } else {
1147  KMP_MEMCPY(task->shareds, data, arg_size);
1148  }
1149  }
1150 
1151 #if OMPT_SUPPORT
1152  kmp_taskdata_t *current_task;
1153  if (ompt_enabled.enabled) {
1154  OMPT_STORE_RETURN_ADDRESS(gtid);
1155  current_task = __kmp_threads[gtid]->th.th_current_task;
1156  current_task->ompt_task_info.frame.enter_frame.ptr = OMPT_GET_FRAME_ADDRESS(0);
1157  }
1158 #endif
1159 
1160  if (if_cond) {
1161  if (gomp_flags & 8) {
1162  KMP_ASSERT(depend);
1163  const size_t ndeps = (kmp_intptr_t)depend[0];
1164  const size_t nout = (kmp_intptr_t)depend[1];
1165  kmp_depend_info_t dep_list[ndeps];
1166 
1167  for (size_t i = 0U; i < ndeps; i++) {
1168  dep_list[i].base_addr = (kmp_intptr_t)depend[2U + i];
1169  dep_list[i].len = 0U;
1170  dep_list[i].flags.in = 1;
1171  dep_list[i].flags.out = (i < nout);
1172  }
1173  __kmpc_omp_task_with_deps(&loc, gtid, task, ndeps, dep_list, 0, NULL);
1174  } else {
1175  __kmpc_omp_task(&loc, gtid, task);
1176  }
1177  } else {
1178 #if OMPT_SUPPORT
1179  ompt_thread_info_t oldInfo;
1180  kmp_info_t *thread;
1181  kmp_taskdata_t *taskdata;
1182  if (ompt_enabled.enabled) {
1183  // Store the threads states and restore them after the task
1184  thread = __kmp_threads[gtid];
1185  taskdata = KMP_TASK_TO_TASKDATA(task);
1186  oldInfo = thread->th.ompt_thread_info;
1187  thread->th.ompt_thread_info.wait_id = 0;
1188  thread->th.ompt_thread_info.state = ompt_state_work_parallel;
1189  taskdata->ompt_task_info.frame.exit_frame.ptr = OMPT_GET_FRAME_ADDRESS(0);
1190  OMPT_STORE_RETURN_ADDRESS(gtid);
1191  }
1192 #endif
1193 
1194  __kmpc_omp_task_begin_if0(&loc, gtid, task);
1195  func(data);
1196  __kmpc_omp_task_complete_if0(&loc, gtid, task);
1197 
1198 #if OMPT_SUPPORT
1199  if (ompt_enabled.enabled) {
1200  thread->th.ompt_thread_info = oldInfo;
1201  taskdata->ompt_task_info.frame.exit_frame = ompt_data_none;
1202  }
1203 #endif
1204  }
1205 #if OMPT_SUPPORT
1206  if (ompt_enabled.enabled) {
1207  current_task->ompt_task_info.frame.enter_frame = ompt_data_none;
1208  }
1209 #endif
1210 
1211  KA_TRACE(20, ("GOMP_task exit: T#%d\n", gtid));
1212 }
1213 
1214 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TASKWAIT)(void) {
1215  MKLOC(loc, "GOMP_taskwait");
1216  int gtid = __kmp_entry_gtid();
1217 
1218 #if OMPT_SUPPORT
1219  if (ompt_enabled.enabled)
1220  OMPT_STORE_RETURN_ADDRESS(gtid);
1221 #endif
1222 
1223  KA_TRACE(20, ("GOMP_taskwait: T#%d\n", gtid));
1224 
1225  __kmpc_omp_taskwait(&loc, gtid);
1226 
1227  KA_TRACE(20, ("GOMP_taskwait exit: T#%d\n", gtid));
1228 }
1229 
1230 // Sections worksharing constructs
1231 //
1232 // For the sections construct, we initialize a dynamically scheduled loop
1233 // worksharing construct with lb 1 and stride 1, and use the iteration #'s
1234 // that its returns as sections ids.
1235 //
1236 // There are no special entry points for ordered sections, so we always use
1237 // the dynamically scheduled workshare, even if the sections aren't ordered.
1238 
1239 unsigned KMP_EXPAND_NAME(KMP_API_NAME_GOMP_SECTIONS_START)(unsigned count) {
1240  int status;
1241  kmp_int lb, ub, stride;
1242  int gtid = __kmp_entry_gtid();
1243  MKLOC(loc, "GOMP_sections_start");
1244  KA_TRACE(20, ("GOMP_sections_start: T#%d\n", gtid));
1245 
1246  KMP_DISPATCH_INIT(&loc, gtid, kmp_nm_dynamic_chunked, 1, count, 1, 1, TRUE);
1247 
1248  status = KMP_DISPATCH_NEXT(&loc, gtid, NULL, &lb, &ub, &stride);
1249  if (status) {
1250  KMP_DEBUG_ASSERT(stride == 1);
1251  KMP_DEBUG_ASSERT(lb > 0);
1252  KMP_ASSERT(lb == ub);
1253  } else {
1254  lb = 0;
1255  }
1256 
1257  KA_TRACE(20, ("GOMP_sections_start exit: T#%d returning %u\n", gtid,
1258  (unsigned)lb));
1259  return (unsigned)lb;
1260 }
1261 
1262 unsigned KMP_EXPAND_NAME(KMP_API_NAME_GOMP_SECTIONS_NEXT)(void) {
1263  int status;
1264  kmp_int lb, ub, stride;
1265  int gtid = __kmp_get_gtid();
1266  MKLOC(loc, "GOMP_sections_next");
1267  KA_TRACE(20, ("GOMP_sections_next: T#%d\n", gtid));
1268 
1269 #if OMPT_SUPPORT
1270  OMPT_STORE_RETURN_ADDRESS(gtid);
1271 #endif
1272 
1273  status = KMP_DISPATCH_NEXT(&loc, gtid, NULL, &lb, &ub, &stride);
1274  if (status) {
1275  KMP_DEBUG_ASSERT(stride == 1);
1276  KMP_DEBUG_ASSERT(lb > 0);
1277  KMP_ASSERT(lb == ub);
1278  } else {
1279  lb = 0;
1280  }
1281 
1282  KA_TRACE(
1283  20, ("GOMP_sections_next exit: T#%d returning %u\n", gtid, (unsigned)lb));
1284  return (unsigned)lb;
1285 }
1286 
1287 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_SECTIONS_START)(
1288  void (*task)(void *), void *data, unsigned num_threads, unsigned count) {
1289  int gtid = __kmp_entry_gtid();
1290 
1291 #if OMPT_SUPPORT
1292  ompt_frame_t *parent_frame;
1293 
1294  if (ompt_enabled.enabled) {
1295  __ompt_get_task_info_internal(0, NULL, NULL, &parent_frame, NULL, NULL);
1296  parent_frame->enter_frame.ptr = OMPT_GET_FRAME_ADDRESS(0);
1297  OMPT_STORE_RETURN_ADDRESS(gtid);
1298  }
1299 #endif
1300 
1301  MKLOC(loc, "GOMP_parallel_sections_start");
1302  KA_TRACE(20, ("GOMP_parallel_sections_start: T#%d\n", gtid));
1303 
1304  if (__kmpc_ok_to_fork(&loc) && (num_threads != 1)) {
1305  if (num_threads != 0) {
1306  __kmp_push_num_threads(&loc, gtid, num_threads);
1307  }
1308  __kmp_GOMP_fork_call(&loc, gtid, task,
1309  (microtask_t)__kmp_GOMP_parallel_microtask_wrapper, 9,
1310  task, data, num_threads, &loc, kmp_nm_dynamic_chunked,
1311  (kmp_int)1, (kmp_int)count, (kmp_int)1, (kmp_int)1);
1312  } else {
1313  __kmp_GOMP_serialized_parallel(&loc, gtid, task);
1314  }
1315 
1316 #if OMPT_SUPPORT
1317  if (ompt_enabled.enabled) {
1318  parent_frame->enter_frame = ompt_data_none;
1319  }
1320 #endif
1321 
1322  KMP_DISPATCH_INIT(&loc, gtid, kmp_nm_dynamic_chunked, 1, count, 1, 1, TRUE);
1323 
1324  KA_TRACE(20, ("GOMP_parallel_sections_start exit: T#%d\n", gtid));
1325 }
1326 
1327 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_SECTIONS_END)(void) {
1328  int gtid = __kmp_get_gtid();
1329  KA_TRACE(20, ("GOMP_sections_end: T#%d\n", gtid))
1330 
1331 #if OMPT_SUPPORT
1332  ompt_frame_t *ompt_frame;
1333  if (ompt_enabled.enabled) {
1334  __ompt_get_task_info_internal(0, NULL, NULL, &ompt_frame, NULL, NULL);
1335  ompt_frame->enter_frame.ptr = OMPT_GET_FRAME_ADDRESS(0);
1336  OMPT_STORE_RETURN_ADDRESS(gtid);
1337  }
1338 #endif
1339  __kmp_barrier(bs_plain_barrier, gtid, FALSE, 0, NULL, NULL);
1340 #if OMPT_SUPPORT
1341  if (ompt_enabled.enabled) {
1342  ompt_frame->enter_frame = ompt_data_none;
1343  }
1344 #endif
1345 
1346  KA_TRACE(20, ("GOMP_sections_end exit: T#%d\n", gtid))
1347 }
1348 
1349 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_SECTIONS_END_NOWAIT)(void) {
1350  KA_TRACE(20, ("GOMP_sections_end_nowait: T#%d\n", __kmp_get_gtid()))
1351 }
1352 
1353 // libgomp has an empty function for GOMP_taskyield as of 2013-10-10
1354 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TASKYIELD)(void) {
1355  KA_TRACE(20, ("GOMP_taskyield: T#%d\n", __kmp_get_gtid()))
1356  return;
1357 }
1358 
1359 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL)(void (*task)(void *),
1360  void *data,
1361  unsigned num_threads,
1362  unsigned int flags) {
1363  int gtid = __kmp_entry_gtid();
1364  MKLOC(loc, "GOMP_parallel");
1365  KA_TRACE(20, ("GOMP_parallel: T#%d\n", gtid));
1366 
1367 #if OMPT_SUPPORT
1368  ompt_task_info_t *parent_task_info, *task_info;
1369  if (ompt_enabled.enabled) {
1370  parent_task_info = __ompt_get_task_info_object(0);
1371  parent_task_info->frame.enter_frame.ptr = OMPT_GET_FRAME_ADDRESS(0);
1372  OMPT_STORE_RETURN_ADDRESS(gtid);
1373  }
1374 #endif
1375  if (__kmpc_ok_to_fork(&loc) && (num_threads != 1)) {
1376  if (num_threads != 0) {
1377  __kmp_push_num_threads(&loc, gtid, num_threads);
1378  }
1379  if (flags != 0) {
1380  __kmp_push_proc_bind(&loc, gtid, (kmp_proc_bind_t)flags);
1381  }
1382  __kmp_GOMP_fork_call(&loc, gtid, task,
1383  (microtask_t)__kmp_GOMP_microtask_wrapper, 2, task,
1384  data);
1385  } else {
1386  __kmp_GOMP_serialized_parallel(&loc, gtid, task);
1387  }
1388 #if OMPT_SUPPORT
1389  if (ompt_enabled.enabled) {
1390  task_info = __ompt_get_task_info_object(0);
1391  task_info->frame.exit_frame.ptr = OMPT_GET_FRAME_ADDRESS(0);
1392  }
1393 #endif
1394  task(data);
1395 #if OMPT_SUPPORT
1396  if (ompt_enabled.enabled) {
1397  OMPT_STORE_RETURN_ADDRESS(gtid);
1398  }
1399 #endif
1400  KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_END)();
1401 #if OMPT_SUPPORT
1402  if (ompt_enabled.enabled) {
1403  task_info->frame.exit_frame = ompt_data_none;
1404  parent_task_info->frame.enter_frame = ompt_data_none;
1405  }
1406 #endif
1407 }
1408 
1409 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_SECTIONS)(void (*task)(void *),
1410  void *data,
1411  unsigned num_threads,
1412  unsigned count,
1413  unsigned flags) {
1414  int gtid = __kmp_entry_gtid();
1415  MKLOC(loc, "GOMP_parallel_sections");
1416  KA_TRACE(20, ("GOMP_parallel_sections: T#%d\n", gtid));
1417 
1418 #if OMPT_SUPPORT
1419  OMPT_STORE_RETURN_ADDRESS(gtid);
1420 #endif
1421 
1422  if (__kmpc_ok_to_fork(&loc) && (num_threads != 1)) {
1423  if (num_threads != 0) {
1424  __kmp_push_num_threads(&loc, gtid, num_threads);
1425  }
1426  if (flags != 0) {
1427  __kmp_push_proc_bind(&loc, gtid, (kmp_proc_bind_t)flags);
1428  }
1429  __kmp_GOMP_fork_call(&loc, gtid, task,
1430  (microtask_t)__kmp_GOMP_parallel_microtask_wrapper, 9,
1431  task, data, num_threads, &loc, kmp_nm_dynamic_chunked,
1432  (kmp_int)1, (kmp_int)count, (kmp_int)1, (kmp_int)1);
1433  } else {
1434  __kmp_GOMP_serialized_parallel(&loc, gtid, task);
1435  }
1436 
1437 #if OMPT_SUPPORT
1438  OMPT_STORE_RETURN_ADDRESS(gtid);
1439 #endif
1440 
1441  KMP_DISPATCH_INIT(&loc, gtid, kmp_nm_dynamic_chunked, 1, count, 1, 1, TRUE);
1442 
1443  task(data);
1444  KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_END)();
1445  KA_TRACE(20, ("GOMP_parallel_sections exit: T#%d\n", gtid));
1446 }
1447 
1448 #define PARALLEL_LOOP(func, schedule, ompt_pre, ompt_post) \
1449  void func(void (*task)(void *), void *data, unsigned num_threads, long lb, \
1450  long ub, long str, long chunk_sz, unsigned flags) { \
1451  int gtid = __kmp_entry_gtid(); \
1452  MKLOC(loc, KMP_STR(func)); \
1453  KA_TRACE( \
1454  20, \
1455  (KMP_STR( \
1456  func) ": T#%d, lb 0x%lx, ub 0x%lx, str 0x%lx, chunk_sz 0x%lx\n", \
1457  gtid, lb, ub, str, chunk_sz)); \
1458  \
1459  ompt_pre(); \
1460  if (__kmpc_ok_to_fork(&loc) && (num_threads != 1)) { \
1461  if (num_threads != 0) { \
1462  __kmp_push_num_threads(&loc, gtid, num_threads); \
1463  } \
1464  if (flags != 0) { \
1465  __kmp_push_proc_bind(&loc, gtid, (kmp_proc_bind_t)flags); \
1466  } \
1467  __kmp_GOMP_fork_call(&loc, gtid, task, \
1468  (microtask_t)__kmp_GOMP_parallel_microtask_wrapper, \
1469  9, task, data, num_threads, &loc, (schedule), lb, \
1470  (str > 0) ? (ub - 1) : (ub + 1), str, chunk_sz); \
1471  } else { \
1472  __kmp_GOMP_serialized_parallel(&loc, gtid, task); \
1473  } \
1474  \
1475  IF_OMPT_SUPPORT(OMPT_STORE_RETURN_ADDRESS(gtid);) \
1476  KMP_DISPATCH_INIT(&loc, gtid, (schedule), lb, \
1477  (str > 0) ? (ub - 1) : (ub + 1), str, chunk_sz, \
1478  (schedule) != kmp_sch_static); \
1479  task(data); \
1480  KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_END)(); \
1481  ompt_post(); \
1482  \
1483  KA_TRACE(20, (KMP_STR(func) " exit: T#%d\n", gtid)); \
1484  }
1485 
1486 PARALLEL_LOOP(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_LOOP_STATIC),
1487  kmp_sch_static, OMPT_LOOP_PRE, OMPT_LOOP_POST)
1488 PARALLEL_LOOP(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_LOOP_DYNAMIC),
1489  kmp_sch_dynamic_chunked, OMPT_LOOP_PRE, OMPT_LOOP_POST)
1490 PARALLEL_LOOP(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_LOOP_GUIDED),
1491  kmp_sch_guided_chunked, OMPT_LOOP_PRE, OMPT_LOOP_POST)
1492 PARALLEL_LOOP(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_LOOP_RUNTIME),
1493  kmp_sch_runtime, OMPT_LOOP_PRE, OMPT_LOOP_POST)
1494 
1495 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TASKGROUP_START)(void) {
1496  int gtid = __kmp_entry_gtid();
1497  MKLOC(loc, "GOMP_taskgroup_start");
1498  KA_TRACE(20, ("GOMP_taskgroup_start: T#%d\n", gtid));
1499 
1500 #if OMPT_SUPPORT
1501  if (ompt_enabled.enabled)
1502  OMPT_STORE_RETURN_ADDRESS(gtid);
1503 #endif
1504 
1505  __kmpc_taskgroup(&loc, gtid);
1506 
1507  return;
1508 }
1509 
1510 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TASKGROUP_END)(void) {
1511  int gtid = __kmp_get_gtid();
1512  MKLOC(loc, "GOMP_taskgroup_end");
1513  KA_TRACE(20, ("GOMP_taskgroup_end: T#%d\n", gtid));
1514 
1515 #if OMPT_SUPPORT
1516  if (ompt_enabled.enabled)
1517  OMPT_STORE_RETURN_ADDRESS(gtid);
1518 #endif
1519 
1520  __kmpc_end_taskgroup(&loc, gtid);
1521 
1522  return;
1523 }
1524 
1525 static kmp_int32 __kmp_gomp_to_omp_cancellation_kind(int gomp_kind) {
1526  kmp_int32 cncl_kind = 0;
1527  switch (gomp_kind) {
1528  case 1:
1529  cncl_kind = cancel_parallel;
1530  break;
1531  case 2:
1532  cncl_kind = cancel_loop;
1533  break;
1534  case 4:
1535  cncl_kind = cancel_sections;
1536  break;
1537  case 8:
1538  cncl_kind = cancel_taskgroup;
1539  break;
1540  }
1541  return cncl_kind;
1542 }
1543 
1544 // Return true if cancellation should take place, false otherwise
1545 bool KMP_EXPAND_NAME(KMP_API_NAME_GOMP_CANCELLATION_POINT)(int which) {
1546  int gtid = __kmp_get_gtid();
1547  MKLOC(loc, "GOMP_cancellation_point");
1548  KA_TRACE(20, ("GOMP_cancellation_point: T#%d which:%d\n", gtid, which));
1549  kmp_int32 cncl_kind = __kmp_gomp_to_omp_cancellation_kind(which);
1550  return __kmpc_cancellationpoint(&loc, gtid, cncl_kind);
1551 }
1552 
1553 // Return true if cancellation should take place, false otherwise
1554 bool KMP_EXPAND_NAME(KMP_API_NAME_GOMP_CANCEL)(int which, bool do_cancel) {
1555  int gtid = __kmp_get_gtid();
1556  MKLOC(loc, "GOMP_cancel");
1557  KA_TRACE(20, ("GOMP_cancel: T#%d which:%d do_cancel:%d\n", gtid, which,
1558  (int)do_cancel));
1559  kmp_int32 cncl_kind = __kmp_gomp_to_omp_cancellation_kind(which);
1560 
1561  if (do_cancel == FALSE) {
1562  return __kmpc_cancellationpoint(&loc, gtid, cncl_kind);
1563  } else {
1564  return __kmpc_cancel(&loc, gtid, cncl_kind);
1565  }
1566 }
1567 
1568 // Return true if cancellation should take place, false otherwise
1569 bool KMP_EXPAND_NAME(KMP_API_NAME_GOMP_BARRIER_CANCEL)(void) {
1570  int gtid = __kmp_get_gtid();
1571  KA_TRACE(20, ("GOMP_barrier_cancel: T#%d\n", gtid));
1572  return __kmp_barrier_gomp_cancel(gtid);
1573 }
1574 
1575 // Return true if cancellation should take place, false otherwise
1576 bool KMP_EXPAND_NAME(KMP_API_NAME_GOMP_SECTIONS_END_CANCEL)(void) {
1577  int gtid = __kmp_get_gtid();
1578  KA_TRACE(20, ("GOMP_sections_end_cancel: T#%d\n", gtid));
1579  return __kmp_barrier_gomp_cancel(gtid);
1580 }
1581 
1582 // Return true if cancellation should take place, false otherwise
1583 bool KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_END_CANCEL)(void) {
1584  int gtid = __kmp_get_gtid();
1585  KA_TRACE(20, ("GOMP_loop_end_cancel: T#%d\n", gtid));
1586  return __kmp_barrier_gomp_cancel(gtid);
1587 }
1588 
1589 // All target functions are empty as of 2014-05-29
1590 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TARGET)(int device, void (*fn)(void *),
1591  const void *openmp_target,
1592  size_t mapnum, void **hostaddrs,
1593  size_t *sizes,
1594  unsigned char *kinds) {
1595  return;
1596 }
1597 
1598 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TARGET_DATA)(
1599  int device, const void *openmp_target, size_t mapnum, void **hostaddrs,
1600  size_t *sizes, unsigned char *kinds) {
1601  return;
1602 }
1603 
1604 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TARGET_END_DATA)(void) { return; }
1605 
1606 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TARGET_UPDATE)(
1607  int device, const void *openmp_target, size_t mapnum, void **hostaddrs,
1608  size_t *sizes, unsigned char *kinds) {
1609  return;
1610 }
1611 
1612 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TEAMS)(unsigned int num_teams,
1613  unsigned int thread_limit) {
1614  return;
1615 }
1616 
1617 // Task duplication function which copies src to dest (both are
1618 // preallocated task structures)
1619 static void __kmp_gomp_task_dup(kmp_task_t *dest, kmp_task_t *src,
1620  kmp_int32 last_private) {
1621  kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(src);
1622  if (taskdata->td_copy_func) {
1623  (taskdata->td_copy_func)(dest->shareds, src->shareds);
1624  }
1625 }
1626 
1627 #ifdef __cplusplus
1628 } // extern "C"
1629 #endif
1630 
1631 template <typename T>
1632 void __GOMP_taskloop(void (*func)(void *), void *data,
1633  void (*copy_func)(void *, void *), long arg_size,
1634  long arg_align, unsigned gomp_flags,
1635  unsigned long num_tasks, int priority, T start, T end,
1636  T step) {
1637  typedef void (*p_task_dup_t)(kmp_task_t *, kmp_task_t *, kmp_int32);
1638  MKLOC(loc, "GOMP_taskloop");
1639  int sched;
1640  T *loop_bounds;
1641  int gtid = __kmp_entry_gtid();
1642  kmp_int32 flags = 0;
1643  int if_val = gomp_flags & (1u << 10);
1644  int nogroup = gomp_flags & (1u << 11);
1645  int up = gomp_flags & (1u << 8);
1646  p_task_dup_t task_dup = NULL;
1647  kmp_tasking_flags_t *input_flags = (kmp_tasking_flags_t *)&flags;
1648 #ifdef KMP_DEBUG
1649  {
1650  char *buff;
1651  buff = __kmp_str_format(
1652  "GOMP_taskloop: T#%%d: func:%%p data:%%p copy_func:%%p "
1653  "arg_size:%%ld arg_align:%%ld gomp_flags:0x%%x num_tasks:%%lu "
1654  "priority:%%d start:%%%s end:%%%s step:%%%s\n",
1655  traits_t<T>::spec, traits_t<T>::spec, traits_t<T>::spec);
1656  KA_TRACE(20, (buff, gtid, func, data, copy_func, arg_size, arg_align,
1657  gomp_flags, num_tasks, priority, start, end, step));
1658  __kmp_str_free(&buff);
1659  }
1660 #endif
1661  KMP_ASSERT((size_t)arg_size >= 2 * sizeof(T));
1662  KMP_ASSERT(arg_align > 0);
1663  // The low-order bit is the "untied" flag
1664  if (!(gomp_flags & 1)) {
1665  input_flags->tiedness = 1;
1666  }
1667  // The second low-order bit is the "final" flag
1668  if (gomp_flags & 2) {
1669  input_flags->final = 1;
1670  }
1671  // Negative step flag
1672  if (!up) {
1673  // If step is flagged as negative, but isn't properly sign extended
1674  // Then manually sign extend it. Could be a short, int, char embedded
1675  // in a long. So cannot assume any cast.
1676  if (step > 0) {
1677  for (int i = sizeof(T) * CHAR_BIT - 1; i >= 0L; --i) {
1678  // break at the first 1 bit
1679  if (step & ((T)1 << i))
1680  break;
1681  step |= ((T)1 << i);
1682  }
1683  }
1684  }
1685  input_flags->native = 1;
1686  // Figure out if none/grainsize/num_tasks clause specified
1687  if (num_tasks > 0) {
1688  if (gomp_flags & (1u << 9))
1689  sched = 1; // grainsize specified
1690  else
1691  sched = 2; // num_tasks specified
1692  // neither grainsize nor num_tasks specified
1693  } else {
1694  sched = 0;
1695  }
1696 
1697  // __kmp_task_alloc() sets up all other flags
1698  kmp_task_t *task =
1699  __kmp_task_alloc(&loc, gtid, input_flags, sizeof(kmp_task_t),
1700  arg_size + arg_align - 1, (kmp_routine_entry_t)func);
1701  kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task);
1702  taskdata->td_copy_func = copy_func;
1703  taskdata->td_size_loop_bounds = sizeof(T);
1704 
1705  // re-align shareds if needed and setup firstprivate copy constructors
1706  // through the task_dup mechanism
1707  task->shareds = (void *)((((size_t)task->shareds) + arg_align - 1) /
1708  arg_align * arg_align);
1709  if (copy_func) {
1710  task_dup = __kmp_gomp_task_dup;
1711  }
1712  KMP_MEMCPY(task->shareds, data, arg_size);
1713 
1714  loop_bounds = (T *)task->shareds;
1715  loop_bounds[0] = start;
1716  loop_bounds[1] = end + (up ? -1 : 1);
1717  __kmpc_taskloop(&loc, gtid, task, if_val, (kmp_uint64 *)&(loop_bounds[0]),
1718  (kmp_uint64 *)&(loop_bounds[1]), (kmp_int64)step, nogroup,
1719  sched, (kmp_uint64)num_tasks, (void *)task_dup);
1720 }
1721 
1722 // 4 byte version of GOMP_doacross_post
1723 // This verison needs to create a temporary array which converts 4 byte
1724 // integers into 8 byte integeres
1725 template <typename T, bool need_conversion = (sizeof(long) == 4)>
1726 void __kmp_GOMP_doacross_post(T *count);
1727 
1728 template <> void __kmp_GOMP_doacross_post<long, true>(long *count) {
1729  int gtid = __kmp_entry_gtid();
1730  kmp_info_t *th = __kmp_threads[gtid];
1731  MKLOC(loc, "GOMP_doacross_post");
1732  kmp_int64 num_dims = th->th.th_dispatch->th_doacross_info[0];
1733  kmp_int64 *vec =
1734  (kmp_int64 *)__kmp_thread_malloc(th, sizeof(kmp_int64) * num_dims);
1735  for (kmp_int64 i = 0; i < num_dims; ++i) {
1736  vec[i] = (kmp_int64)count[i];
1737  }
1738  __kmpc_doacross_post(&loc, gtid, vec);
1739  __kmp_thread_free(th, vec);
1740 }
1741 
1742 // 8 byte versions of GOMP_doacross_post
1743 // This version can just pass in the count array directly instead of creating
1744 // a temporary array
1745 template <> void __kmp_GOMP_doacross_post<long, false>(long *count) {
1746  int gtid = __kmp_entry_gtid();
1747  MKLOC(loc, "GOMP_doacross_post");
1748  __kmpc_doacross_post(&loc, gtid, RCAST(kmp_int64 *, count));
1749 }
1750 
1751 template <typename T> void __kmp_GOMP_doacross_wait(T first, va_list args) {
1752  int gtid = __kmp_entry_gtid();
1753  kmp_info_t *th = __kmp_threads[gtid];
1754  MKLOC(loc, "GOMP_doacross_wait");
1755  kmp_int64 num_dims = th->th.th_dispatch->th_doacross_info[0];
1756  kmp_int64 *vec =
1757  (kmp_int64 *)__kmp_thread_malloc(th, sizeof(kmp_int64) * num_dims);
1758  vec[0] = (kmp_int64)first;
1759  for (kmp_int64 i = 1; i < num_dims; ++i) {
1760  T item = va_arg(args, T);
1761  vec[i] = (kmp_int64)item;
1762  }
1763  __kmpc_doacross_wait(&loc, gtid, vec);
1764  __kmp_thread_free(th, vec);
1765  return;
1766 }
1767 
1768 #ifdef __cplusplus
1769 extern "C" {
1770 #endif // __cplusplus
1771 
1772 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TASKLOOP)(
1773  void (*func)(void *), void *data, void (*copy_func)(void *, void *),
1774  long arg_size, long arg_align, unsigned gomp_flags, unsigned long num_tasks,
1775  int priority, long start, long end, long step) {
1776  __GOMP_taskloop<long>(func, data, copy_func, arg_size, arg_align, gomp_flags,
1777  num_tasks, priority, start, end, step);
1778 }
1779 
1780 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TASKLOOP_ULL)(
1781  void (*func)(void *), void *data, void (*copy_func)(void *, void *),
1782  long arg_size, long arg_align, unsigned gomp_flags, unsigned long num_tasks,
1783  int priority, unsigned long long start, unsigned long long end,
1784  unsigned long long step) {
1785  __GOMP_taskloop<unsigned long long>(func, data, copy_func, arg_size,
1786  arg_align, gomp_flags, num_tasks,
1787  priority, start, end, step);
1788 }
1789 
1790 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_DOACROSS_POST)(long *count) {
1791  __kmp_GOMP_doacross_post(count);
1792 }
1793 
1794 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_DOACROSS_WAIT)(long first, ...) {
1795  va_list args;
1796  va_start(args, first);
1797  __kmp_GOMP_doacross_wait<long>(first, args);
1798  va_end(args);
1799 }
1800 
1801 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_DOACROSS_ULL_POST)(
1802  unsigned long long *count) {
1803  int gtid = __kmp_entry_gtid();
1804  MKLOC(loc, "GOMP_doacross_ull_post");
1805  __kmpc_doacross_post(&loc, gtid, RCAST(kmp_int64 *, count));
1806 }
1807 
1808 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_DOACROSS_ULL_WAIT)(
1809  unsigned long long first, ...) {
1810  va_list args;
1811  va_start(args, first);
1812  __kmp_GOMP_doacross_wait<unsigned long long>(first, args);
1813  va_end(args);
1814 }
1815 
1816 /* The following sections of code create aliases for the GOMP_* functions, then
1817  create versioned symbols using the assembler directive .symver. This is only
1818  pertinent for ELF .so library. The KMP_VERSION_SYMBOL macro is defined in
1819  kmp_os.h */
1820 
1821 #ifdef KMP_USE_VERSION_SYMBOLS
1822 // GOMP_1.0 versioned symbols
1823 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_ATOMIC_END, 10, "GOMP_1.0");
1824 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_ATOMIC_START, 10, "GOMP_1.0");
1825 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_BARRIER, 10, "GOMP_1.0");
1826 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_CRITICAL_END, 10, "GOMP_1.0");
1827 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_CRITICAL_NAME_END, 10, "GOMP_1.0");
1828 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_CRITICAL_NAME_START, 10, "GOMP_1.0");
1829 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_CRITICAL_START, 10, "GOMP_1.0");
1830 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_DYNAMIC_NEXT, 10, "GOMP_1.0");
1831 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_DYNAMIC_START, 10, "GOMP_1.0");
1832 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_END, 10, "GOMP_1.0");
1833 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_END_NOWAIT, 10, "GOMP_1.0");
1834 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_GUIDED_NEXT, 10, "GOMP_1.0");
1835 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_GUIDED_START, 10, "GOMP_1.0");
1836 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ORDERED_DYNAMIC_NEXT, 10, "GOMP_1.0");
1837 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ORDERED_DYNAMIC_START, 10,
1838  "GOMP_1.0");
1839 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ORDERED_GUIDED_NEXT, 10, "GOMP_1.0");
1840 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ORDERED_GUIDED_START, 10, "GOMP_1.0");
1841 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ORDERED_RUNTIME_NEXT, 10, "GOMP_1.0");
1842 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ORDERED_RUNTIME_START, 10,
1843  "GOMP_1.0");
1844 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ORDERED_STATIC_NEXT, 10, "GOMP_1.0");
1845 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ORDERED_STATIC_START, 10, "GOMP_1.0");
1846 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_RUNTIME_NEXT, 10, "GOMP_1.0");
1847 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_RUNTIME_START, 10, "GOMP_1.0");
1848 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_STATIC_NEXT, 10, "GOMP_1.0");
1849 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_STATIC_START, 10, "GOMP_1.0");
1850 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_ORDERED_END, 10, "GOMP_1.0");
1851 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_ORDERED_START, 10, "GOMP_1.0");
1852 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_END, 10, "GOMP_1.0");
1853 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_LOOP_DYNAMIC_START, 10,
1854  "GOMP_1.0");
1855 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_LOOP_GUIDED_START, 10,
1856  "GOMP_1.0");
1857 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_LOOP_RUNTIME_START, 10,
1858  "GOMP_1.0");
1859 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_LOOP_STATIC_START, 10,
1860  "GOMP_1.0");
1861 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_SECTIONS_START, 10, "GOMP_1.0");
1862 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_START, 10, "GOMP_1.0");
1863 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_SECTIONS_END, 10, "GOMP_1.0");
1864 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_SECTIONS_END_NOWAIT, 10, "GOMP_1.0");
1865 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_SECTIONS_NEXT, 10, "GOMP_1.0");
1866 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_SECTIONS_START, 10, "GOMP_1.0");
1867 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_SINGLE_COPY_END, 10, "GOMP_1.0");
1868 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_SINGLE_COPY_START, 10, "GOMP_1.0");
1869 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_SINGLE_START, 10, "GOMP_1.0");
1870 
1871 // GOMP_2.0 versioned symbols
1872 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TASK, 20, "GOMP_2.0");
1873 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TASKWAIT, 20, "GOMP_2.0");
1874 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_DYNAMIC_NEXT, 20, "GOMP_2.0");
1875 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_DYNAMIC_START, 20, "GOMP_2.0");
1876 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_GUIDED_NEXT, 20, "GOMP_2.0");
1877 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_GUIDED_START, 20, "GOMP_2.0");
1878 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_DYNAMIC_NEXT, 20,
1879  "GOMP_2.0");
1880 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_DYNAMIC_START, 20,
1881  "GOMP_2.0");
1882 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_GUIDED_NEXT, 20,
1883  "GOMP_2.0");
1884 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_GUIDED_START, 20,
1885  "GOMP_2.0");
1886 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_RUNTIME_NEXT, 20,
1887  "GOMP_2.0");
1888 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_RUNTIME_START, 20,
1889  "GOMP_2.0");
1890 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_STATIC_NEXT, 20,
1891  "GOMP_2.0");
1892 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_STATIC_START, 20,
1893  "GOMP_2.0");
1894 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_RUNTIME_NEXT, 20, "GOMP_2.0");
1895 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_RUNTIME_START, 20, "GOMP_2.0");
1896 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_STATIC_NEXT, 20, "GOMP_2.0");
1897 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_STATIC_START, 20, "GOMP_2.0");
1898 
1899 // GOMP_3.0 versioned symbols
1900 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TASKYIELD, 30, "GOMP_3.0");
1901 
1902 // GOMP_4.0 versioned symbols
1903 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL, 40, "GOMP_4.0");
1904 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_SECTIONS, 40, "GOMP_4.0");
1905 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_LOOP_DYNAMIC, 40, "GOMP_4.0");
1906 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_LOOP_GUIDED, 40, "GOMP_4.0");
1907 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_LOOP_RUNTIME, 40, "GOMP_4.0");
1908 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_LOOP_STATIC, 40, "GOMP_4.0");
1909 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TASKGROUP_START, 40, "GOMP_4.0");
1910 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TASKGROUP_END, 40, "GOMP_4.0");
1911 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_BARRIER_CANCEL, 40, "GOMP_4.0");
1912 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_CANCEL, 40, "GOMP_4.0");
1913 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_CANCELLATION_POINT, 40, "GOMP_4.0");
1914 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_END_CANCEL, 40, "GOMP_4.0");
1915 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_SECTIONS_END_CANCEL, 40, "GOMP_4.0");
1916 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TARGET, 40, "GOMP_4.0");
1917 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TARGET_DATA, 40, "GOMP_4.0");
1918 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TARGET_END_DATA, 40, "GOMP_4.0");
1919 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TARGET_UPDATE, 40, "GOMP_4.0");
1920 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TEAMS, 40, "GOMP_4.0");
1921 
1922 // GOMP_4.5 versioned symbols
1923 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TASKLOOP, 45, "GOMP_4.5");
1924 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TASKLOOP_ULL, 45, "GOMP_4.5");
1925 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_DOACROSS_POST, 45, "GOMP_4.5");
1926 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_DOACROSS_WAIT, 45, "GOMP_4.5");
1927 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_DOACROSS_STATIC_START, 45,
1928  "GOMP_4.5");
1929 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_DOACROSS_DYNAMIC_START, 45,
1930  "GOMP_4.5");
1931 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_DOACROSS_GUIDED_START, 45,
1932  "GOMP_4.5");
1933 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_DOACROSS_RUNTIME_START, 45,
1934  "GOMP_4.5");
1935 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_DOACROSS_ULL_POST, 45, "GOMP_4.5");
1936 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_DOACROSS_ULL_WAIT, 45, "GOMP_4.5");
1937 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_DOACROSS_STATIC_START, 45,
1938  "GOMP_4.5");
1939 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_DOACROSS_DYNAMIC_START, 45,
1940  "GOMP_4.5");
1941 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_DOACROSS_GUIDED_START, 45,
1942  "GOMP_4.5");
1943 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_DOACROSS_RUNTIME_START, 45,
1944  "GOMP_4.5");
1945 
1946 #endif // KMP_USE_VERSION_SYMBOLS
1947 
1948 #ifdef __cplusplus
1949 } // extern "C"
1950 #endif // __cplusplus
void __kmpc_taskloop(ident_t *loc, int gtid, kmp_task_t *task, int if_val, kmp_uint64 *lb, kmp_uint64 *ub, kmp_int64 st, int nogroup, int sched, kmp_uint64 grainsize, void *task_dup)
KMP_EXPORT void __kmpc_end_ordered(ident_t *, kmp_int32 global_tid)
KMP_EXPORT void __kmpc_end_serialized_parallel(ident_t *, kmp_int32 global_tid)
KMP_EXPORT kmp_int32 __kmpc_omp_task_with_deps(ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *new_task, kmp_int32 ndeps, kmp_depend_info_t *dep_list, kmp_int32 ndeps_noalias, kmp_depend_info_t *noalias_dep_list)
KMP_EXPORT void __kmpc_ordered(ident_t *, kmp_int32 global_tid)
sched_type
Definition: kmp.h:336
KMP_EXPORT void __kmpc_critical(ident_t *, kmp_int32 global_tid, kmp_critical_name *)
Definition: kmp.h:222
KMP_EXPORT kmp_int32 __kmpc_ok_to_fork(ident_t *)
KMP_EXPORT void __kmpc_barrier(ident_t *, kmp_int32 global_tid)
KMP_EXPORT void __kmpc_end_critical(ident_t *, kmp_int32 global_tid, kmp_critical_name *)