xref: /freebsd/contrib/llvm-project/openmp/runtime/src/z_Linux_util.cpp (revision 770cf0a5f02dc8983a89c6568d741fbc25baa999)
1 /*
2  * z_Linux_util.cpp -- platform specific routines.
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_affinity.h"
15 #include "kmp_i18n.h"
16 #include "kmp_io.h"
17 #include "kmp_itt.h"
18 #include "kmp_lock.h"
19 #include "kmp_stats.h"
20 #include "kmp_str.h"
21 #include "kmp_wait_release.h"
22 #include "kmp_wrapper_getpid.h"
23 
24 #if !KMP_OS_DRAGONFLY && !KMP_OS_FREEBSD && !KMP_OS_NETBSD && !KMP_OS_OPENBSD
25 #include <alloca.h>
26 #endif
27 #include <math.h> // HUGE_VAL.
28 #if KMP_OS_LINUX
29 #include <semaphore.h>
30 #endif // KMP_OS_LINUX
31 #include <sys/resource.h>
32 #if KMP_OS_AIX
33 #include <sys/ldr.h>
34 #include <libperfstat.h>
35 #elif !KMP_OS_HAIKU
36 #include <sys/syscall.h>
37 #endif
38 #include <sys/time.h>
39 #include <sys/times.h>
40 #include <unistd.h>
41 
42 #if KMP_OS_LINUX
43 #include <sys/sysinfo.h>
44 #if KMP_USE_FUTEX
45 // We should really include <futex.h>, but that causes compatibility problems on
46 // different Linux* OS distributions that either require that you include (or
47 // break when you try to include) <pci/types.h>. Since all we need is the two
48 // macros below (which are part of the kernel ABI, so can't change) we just
49 // define the constants here and don't include <futex.h>
50 #ifndef FUTEX_WAIT
51 #define FUTEX_WAIT 0
52 #endif
53 #ifndef FUTEX_WAKE
54 #define FUTEX_WAKE 1
55 #endif
56 #endif
57 #elif KMP_OS_DARWIN
58 #include <mach/mach.h>
59 #include <sys/sysctl.h>
60 #elif KMP_OS_DRAGONFLY || KMP_OS_FREEBSD
61 #include <sys/types.h>
62 #include <sys/sysctl.h>
63 #include <sys/user.h>
64 #include <pthread_np.h>
65 #if KMP_OS_DRAGONFLY
66 #include <kvm.h>
67 #endif
68 #elif KMP_OS_NETBSD || KMP_OS_OPENBSD
69 #include <sys/types.h>
70 #include <sys/sysctl.h>
71 #if KMP_OS_NETBSD
72 #include <sched.h>
73 #endif
74 #if KMP_OS_OPENBSD
75 #include <pthread_np.h>
76 #endif
77 #elif KMP_OS_SOLARIS
78 #include <procfs.h>
79 #include <thread.h>
80 #include <sys/loadavg.h>
81 #endif
82 
83 #include <ctype.h>
84 #include <dirent.h>
85 #include <fcntl.h>
86 
87 struct kmp_sys_timer {
88   struct timespec start;
89 };
90 
91 #ifndef TIMEVAL_TO_TIMESPEC
92 // Convert timeval to timespec.
93 #define TIMEVAL_TO_TIMESPEC(tv, ts)                                            \
94   do {                                                                         \
95     (ts)->tv_sec = (tv)->tv_sec;                                               \
96     (ts)->tv_nsec = (tv)->tv_usec * 1000;                                      \
97   } while (0)
98 #endif
99 
100 // Convert timespec to nanoseconds.
101 #define TS2NS(timespec)                                                        \
102   (((timespec).tv_sec * (long int)1e9) + (timespec).tv_nsec)
103 
104 static struct kmp_sys_timer __kmp_sys_timer_data;
105 
106 #if KMP_HANDLE_SIGNALS
107 typedef void (*sig_func_t)(int);
108 STATIC_EFI2_WORKAROUND struct sigaction __kmp_sighldrs[NSIG];
109 static sigset_t __kmp_sigset;
110 #endif
111 
112 static int __kmp_init_runtime = FALSE;
113 
114 static int __kmp_fork_count = 0;
115 
116 static pthread_condattr_t __kmp_suspend_cond_attr;
117 static pthread_mutexattr_t __kmp_suspend_mutex_attr;
118 
119 static kmp_cond_align_t __kmp_wait_cv;
120 static kmp_mutex_align_t __kmp_wait_mx;
121 
122 kmp_uint64 __kmp_ticks_per_msec = 1000000;
123 kmp_uint64 __kmp_ticks_per_usec = 1000;
124 
125 #ifdef DEBUG_SUSPEND
126 static void __kmp_print_cond(char *buffer, kmp_cond_align_t *cond) {
127   KMP_SNPRINTF(buffer, 128, "(cond (lock (%ld, %d)), (descr (%p)))",
128                cond->c_cond.__c_lock.__status, cond->c_cond.__c_lock.__spinlock,
129                cond->c_cond.__c_waiting);
130 }
131 #endif
132 
133 #if ((KMP_OS_LINUX || KMP_OS_FREEBSD || KMP_OS_NETBSD || KMP_OS_DRAGONFLY ||   \
134       KMP_OS_AIX) &&                                                           \
135      KMP_AFFINITY_SUPPORTED)
136 
137 /* Affinity support */
138 
139 void __kmp_affinity_bind_thread(int which) {
140   KMP_ASSERT2(KMP_AFFINITY_CAPABLE(),
141               "Illegal set affinity operation when not capable");
142 
143   kmp_affin_mask_t *mask;
144   KMP_CPU_ALLOC_ON_STACK(mask);
145   KMP_CPU_ZERO(mask);
146   KMP_CPU_SET(which, mask);
147   __kmp_set_system_affinity(mask, TRUE);
148   KMP_CPU_FREE_FROM_STACK(mask);
149 }
150 
151 #if KMP_OS_AIX
152 void __kmp_affinity_determine_capable(const char *env_var) {
153   // All versions of AIX support bindprocessor().
154 
155   size_t mask_size = __kmp_xproc / CHAR_BIT;
156   // Round up to byte boundary.
157   if (__kmp_xproc % CHAR_BIT)
158     ++mask_size;
159 
160   // Round up to the mask_size_type boundary.
161   if (mask_size % sizeof(__kmp_affin_mask_size))
162     mask_size += sizeof(__kmp_affin_mask_size) -
163                  mask_size % sizeof(__kmp_affin_mask_size);
164   KMP_AFFINITY_ENABLE(mask_size);
165   KA_TRACE(10,
166            ("__kmp_affinity_determine_capable: "
167             "AIX OS affinity interface bindprocessor functional (mask size = "
168             "%" KMP_SIZE_T_SPEC ").\n",
169             __kmp_affin_mask_size));
170 }
171 
172 #else // !KMP_OS_AIX
173 
174 /* Determine if we can access affinity functionality on this version of
175  * Linux* OS by checking __NR_sched_{get,set}affinity system calls, and set
176  * __kmp_affin_mask_size to the appropriate value (0 means not capable). */
177 void __kmp_affinity_determine_capable(const char *env_var) {
178   // Check and see if the OS supports thread affinity.
179 
180 #if KMP_OS_LINUX
181 #define KMP_CPU_SET_SIZE_LIMIT (1024 * 1024)
182 #define KMP_CPU_SET_TRY_SIZE CACHE_LINE
183 #elif KMP_OS_FREEBSD || KMP_OS_DRAGONFLY
184 #define KMP_CPU_SET_SIZE_LIMIT (sizeof(cpuset_t))
185 #elif KMP_OS_NETBSD
186 #define KMP_CPU_SET_SIZE_LIMIT (256)
187 #endif
188 
189   int verbose = __kmp_affinity.flags.verbose;
190   int warnings = __kmp_affinity.flags.warnings;
191   enum affinity_type type = __kmp_affinity.type;
192 
193 #if KMP_OS_LINUX
194   long gCode;
195   unsigned char *buf;
196   buf = (unsigned char *)KMP_INTERNAL_MALLOC(KMP_CPU_SET_SIZE_LIMIT);
197 
198   // If the syscall returns a suggestion for the size,
199   // then we don't have to search for an appropriate size.
200   gCode = syscall(__NR_sched_getaffinity, 0, KMP_CPU_SET_TRY_SIZE, buf);
201   KA_TRACE(30, ("__kmp_affinity_determine_capable: "
202                 "initial getaffinity call returned %ld errno = %d\n",
203                 gCode, errno));
204 
205   if (gCode < 0 && errno != EINVAL) {
206     // System call not supported
207     if (verbose ||
208         (warnings && (type != affinity_none) && (type != affinity_default) &&
209          (type != affinity_disabled))) {
210       int error = errno;
211       kmp_msg_t err_code = KMP_ERR(error);
212       __kmp_msg(kmp_ms_warning, KMP_MSG(GetAffSysCallNotSupported, env_var),
213                 err_code, __kmp_msg_null);
214       if (__kmp_generate_warnings == kmp_warnings_off) {
215         __kmp_str_free(&err_code.str);
216       }
217     }
218     KMP_AFFINITY_DISABLE();
219     KMP_INTERNAL_FREE(buf);
220     return;
221   } else if (gCode > 0) {
222     // The optimal situation: the OS returns the size of the buffer it expects.
223     KMP_AFFINITY_ENABLE(gCode);
224     KA_TRACE(10, ("__kmp_affinity_determine_capable: "
225                   "affinity supported (mask size %d)\n",
226                   (int)__kmp_affin_mask_size));
227     KMP_INTERNAL_FREE(buf);
228     return;
229   }
230 
231   // Call the getaffinity system call repeatedly with increasing set sizes
232   // until we succeed, or reach an upper bound on the search.
233   KA_TRACE(30, ("__kmp_affinity_determine_capable: "
234                 "searching for proper set size\n"));
235   int size;
236   for (size = 1; size <= KMP_CPU_SET_SIZE_LIMIT; size *= 2) {
237     gCode = syscall(__NR_sched_getaffinity, 0, size, buf);
238     KA_TRACE(30, ("__kmp_affinity_determine_capable: "
239                   "getaffinity for mask size %ld returned %ld errno = %d\n",
240                   size, gCode, errno));
241 
242     if (gCode < 0) {
243       if (errno == ENOSYS) {
244         // We shouldn't get here
245         KA_TRACE(30, ("__kmp_affinity_determine_capable: "
246                       "inconsistent OS call behavior: errno == ENOSYS for mask "
247                       "size %d\n",
248                       size));
249         if (verbose ||
250             (warnings && (type != affinity_none) &&
251              (type != affinity_default) && (type != affinity_disabled))) {
252           int error = errno;
253           kmp_msg_t err_code = KMP_ERR(error);
254           __kmp_msg(kmp_ms_warning, KMP_MSG(GetAffSysCallNotSupported, env_var),
255                     err_code, __kmp_msg_null);
256           if (__kmp_generate_warnings == kmp_warnings_off) {
257             __kmp_str_free(&err_code.str);
258           }
259         }
260         KMP_AFFINITY_DISABLE();
261         KMP_INTERNAL_FREE(buf);
262         return;
263       }
264       continue;
265     }
266 
267     KMP_AFFINITY_ENABLE(gCode);
268     KA_TRACE(10, ("__kmp_affinity_determine_capable: "
269                   "affinity supported (mask size %d)\n",
270                   (int)__kmp_affin_mask_size));
271     KMP_INTERNAL_FREE(buf);
272     return;
273   }
274 #elif KMP_OS_FREEBSD || KMP_OS_NETBSD || KMP_OS_DRAGONFLY
275   long gCode;
276   unsigned char *buf;
277   buf = (unsigned char *)KMP_INTERNAL_MALLOC(KMP_CPU_SET_SIZE_LIMIT);
278   gCode = pthread_getaffinity_np(pthread_self(), KMP_CPU_SET_SIZE_LIMIT,
279                                  reinterpret_cast<cpuset_t *>(buf));
280   KA_TRACE(30, ("__kmp_affinity_determine_capable: "
281                 "initial getaffinity call returned %d errno = %d\n",
282                 gCode, errno));
283   if (gCode == 0) {
284     KMP_AFFINITY_ENABLE(KMP_CPU_SET_SIZE_LIMIT);
285     KA_TRACE(10, ("__kmp_affinity_determine_capable: "
286                   "affinity supported (mask size %d)\n",
287                   (int)__kmp_affin_mask_size));
288     KMP_INTERNAL_FREE(buf);
289     return;
290   }
291 #endif
292   KMP_INTERNAL_FREE(buf);
293 
294   // Affinity is not supported
295   KMP_AFFINITY_DISABLE();
296   KA_TRACE(10, ("__kmp_affinity_determine_capable: "
297                 "cannot determine mask size - affinity not supported\n"));
298   if (verbose || (warnings && (type != affinity_none) &&
299                   (type != affinity_default) && (type != affinity_disabled))) {
300     KMP_WARNING(AffCantGetMaskSize, env_var);
301   }
302 }
303 #endif // KMP_OS_AIX
304 #endif // (KMP_OS_LINUX || KMP_OS_FREEBSD || KMP_OS_NETBSD ||                  \
305            KMP_OS_DRAGONFLY || KMP_OS_AIX) && KMP_AFFINITY_SUPPORTED
306 
307 #if KMP_USE_FUTEX
308 
309 int __kmp_futex_determine_capable() {
310   int loc = 0;
311   long rc = syscall(__NR_futex, &loc, FUTEX_WAKE, 1, NULL, NULL, 0);
312   int retval = (rc == 0) || (errno != ENOSYS);
313 
314   KA_TRACE(10,
315            ("__kmp_futex_determine_capable: rc = %d errno = %d\n", rc, errno));
316   KA_TRACE(10, ("__kmp_futex_determine_capable: futex syscall%s supported\n",
317                 retval ? "" : " not"));
318 
319   return retval;
320 }
321 
322 #endif // KMP_USE_FUTEX
323 
324 #if (KMP_ARCH_X86 || KMP_ARCH_X86_64 || KMP_ARCH_WASM) && (!KMP_ASM_INTRINS)
325 /* Only 32-bit "add-exchange" instruction on IA-32 architecture causes us to
326    use compare_and_store for these routines */
327 
328 kmp_int8 __kmp_test_then_or8(volatile kmp_int8 *p, kmp_int8 d) {
329   kmp_int8 old_value, new_value;
330 
331   old_value = TCR_1(*p);
332   new_value = old_value | d;
333 
334   while (!KMP_COMPARE_AND_STORE_REL8(p, old_value, new_value)) {
335     KMP_CPU_PAUSE();
336     old_value = TCR_1(*p);
337     new_value = old_value | d;
338   }
339   return old_value;
340 }
341 
342 kmp_int8 __kmp_test_then_and8(volatile kmp_int8 *p, kmp_int8 d) {
343   kmp_int8 old_value, new_value;
344 
345   old_value = TCR_1(*p);
346   new_value = old_value & d;
347 
348   while (!KMP_COMPARE_AND_STORE_REL8(p, old_value, new_value)) {
349     KMP_CPU_PAUSE();
350     old_value = TCR_1(*p);
351     new_value = old_value & d;
352   }
353   return old_value;
354 }
355 
356 kmp_uint32 __kmp_test_then_or32(volatile kmp_uint32 *p, kmp_uint32 d) {
357   kmp_uint32 old_value, new_value;
358 
359   old_value = TCR_4(*p);
360   new_value = old_value | d;
361 
362   while (!KMP_COMPARE_AND_STORE_REL32(p, old_value, new_value)) {
363     KMP_CPU_PAUSE();
364     old_value = TCR_4(*p);
365     new_value = old_value | d;
366   }
367   return old_value;
368 }
369 
370 kmp_uint32 __kmp_test_then_and32(volatile kmp_uint32 *p, kmp_uint32 d) {
371   kmp_uint32 old_value, new_value;
372 
373   old_value = TCR_4(*p);
374   new_value = old_value & d;
375 
376   while (!KMP_COMPARE_AND_STORE_REL32(p, old_value, new_value)) {
377     KMP_CPU_PAUSE();
378     old_value = TCR_4(*p);
379     new_value = old_value & d;
380   }
381   return old_value;
382 }
383 
384 #if KMP_ARCH_X86 || KMP_ARCH_WASM
385 kmp_int8 __kmp_test_then_add8(volatile kmp_int8 *p, kmp_int8 d) {
386   kmp_int8 old_value, new_value;
387 
388   old_value = TCR_1(*p);
389   new_value = old_value + d;
390 
391   while (!KMP_COMPARE_AND_STORE_REL8(p, old_value, new_value)) {
392     KMP_CPU_PAUSE();
393     old_value = TCR_1(*p);
394     new_value = old_value + d;
395   }
396   return old_value;
397 }
398 
399 kmp_int64 __kmp_test_then_add64(volatile kmp_int64 *p, kmp_int64 d) {
400   kmp_int64 old_value, new_value;
401 
402   old_value = TCR_8(*p);
403   new_value = old_value + d;
404 
405   while (!KMP_COMPARE_AND_STORE_REL64(p, old_value, new_value)) {
406     KMP_CPU_PAUSE();
407     old_value = TCR_8(*p);
408     new_value = old_value + d;
409   }
410   return old_value;
411 }
412 #endif /* KMP_ARCH_X86 */
413 
414 kmp_uint64 __kmp_test_then_or64(volatile kmp_uint64 *p, kmp_uint64 d) {
415   kmp_uint64 old_value, new_value;
416 
417   old_value = TCR_8(*p);
418   new_value = old_value | d;
419   while (!KMP_COMPARE_AND_STORE_REL64(p, old_value, new_value)) {
420     KMP_CPU_PAUSE();
421     old_value = TCR_8(*p);
422     new_value = old_value | d;
423   }
424   return old_value;
425 }
426 
427 kmp_uint64 __kmp_test_then_and64(volatile kmp_uint64 *p, kmp_uint64 d) {
428   kmp_uint64 old_value, new_value;
429 
430   old_value = TCR_8(*p);
431   new_value = old_value & d;
432   while (!KMP_COMPARE_AND_STORE_REL64(p, old_value, new_value)) {
433     KMP_CPU_PAUSE();
434     old_value = TCR_8(*p);
435     new_value = old_value & d;
436   }
437   return old_value;
438 }
439 
440 #endif /* (KMP_ARCH_X86 || KMP_ARCH_X86_64) && (! KMP_ASM_INTRINS) */
441 
442 void __kmp_terminate_thread(int gtid) {
443   int status;
444   kmp_info_t *th = __kmp_threads[gtid];
445 
446   if (!th)
447     return;
448 
449 #ifdef KMP_CANCEL_THREADS
450   KA_TRACE(10, ("__kmp_terminate_thread: kill (%d)\n", gtid));
451   status = pthread_cancel(th->th.th_info.ds.ds_thread);
452   if (status != 0 && status != ESRCH) {
453     __kmp_fatal(KMP_MSG(CantTerminateWorkerThread), KMP_ERR(status),
454                 __kmp_msg_null);
455   }
456 #endif
457   KMP_YIELD(TRUE);
458 } //
459 
460 /* Set thread stack info.
461    If values are unreasonable, assume call failed and use incremental stack
462    refinement method instead. Returns TRUE if the stack parameters could be
463    determined exactly, FALSE if incremental refinement is necessary. */
464 static kmp_int32 __kmp_set_stack_info(int gtid, kmp_info_t *th) {
465   int stack_data;
466 #if KMP_OS_LINUX || KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD ||     \
467     KMP_OS_HAIKU || KMP_OS_HURD || KMP_OS_SOLARIS || KMP_OS_AIX
468   int status;
469   size_t size = 0;
470   void *addr = 0;
471 
472   /* Always do incremental stack refinement for ubermaster threads since the
473      initial thread stack range can be reduced by sibling thread creation so
474      pthread_attr_getstack may cause thread gtid aliasing */
475   if (!KMP_UBER_GTID(gtid)) {
476 
477 #if KMP_OS_SOLARIS
478     stack_t s;
479     if ((status = thr_stksegment(&s)) < 0) {
480       KMP_CHECK_SYSFAIL("thr_stksegment", status);
481     }
482 
483     addr = s.ss_sp;
484     size = s.ss_size;
485     KA_TRACE(60, ("__kmp_set_stack_info: T#%d thr_stksegment returned size:"
486                   " %lu, low addr: %p\n",
487                   gtid, size, addr));
488 #else
489     pthread_attr_t attr;
490     /* Fetch the real thread attributes */
491     status = pthread_attr_init(&attr);
492     KMP_CHECK_SYSFAIL("pthread_attr_init", status);
493 #if KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD
494     status = pthread_attr_get_np(pthread_self(), &attr);
495     KMP_CHECK_SYSFAIL("pthread_attr_get_np", status);
496 #else
497     status = pthread_getattr_np(pthread_self(), &attr);
498     KMP_CHECK_SYSFAIL("pthread_getattr_np", status);
499 #endif
500     status = pthread_attr_getstack(&attr, &addr, &size);
501     KMP_CHECK_SYSFAIL("pthread_attr_getstack", status);
502     KA_TRACE(60,
503              ("__kmp_set_stack_info: T#%d pthread_attr_getstack returned size:"
504               " %lu, low addr: %p\n",
505               gtid, size, addr));
506     status = pthread_attr_destroy(&attr);
507     KMP_CHECK_SYSFAIL("pthread_attr_destroy", status);
508 #endif
509   }
510 
511   if (size != 0 && addr != 0) { // was stack parameter determination successful?
512     /* Store the correct base and size */
513     TCW_PTR(th->th.th_info.ds.ds_stackbase, (((char *)addr) + size));
514     TCW_PTR(th->th.th_info.ds.ds_stacksize, size);
515     TCW_4(th->th.th_info.ds.ds_stackgrow, FALSE);
516     return TRUE;
517   }
518 #endif /* KMP_OS_LINUX || KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD  \
519           || KMP_OS_HAIKU || KMP_OS_HURD || KMP_OS_SOLARIS */
520   /* Use incremental refinement starting from initial conservative estimate */
521   TCW_PTR(th->th.th_info.ds.ds_stacksize, 0);
522   TCW_PTR(th->th.th_info.ds.ds_stackbase, &stack_data);
523   TCW_4(th->th.th_info.ds.ds_stackgrow, TRUE);
524   return FALSE;
525 }
526 
527 static void *__kmp_launch_worker(void *thr) {
528   int status, old_type, old_state;
529 #ifdef KMP_BLOCK_SIGNALS
530   sigset_t new_set, old_set;
531 #endif /* KMP_BLOCK_SIGNALS */
532   void *exit_val;
533 #if KMP_OS_LINUX || KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD ||     \
534     KMP_OS_OPENBSD || KMP_OS_HAIKU || KMP_OS_HURD || KMP_OS_SOLARIS ||         \
535     KMP_OS_AIX
536   void *volatile padding = 0;
537 #endif
538   int gtid;
539 
540   gtid = ((kmp_info_t *)thr)->th.th_info.ds.ds_gtid;
541   __kmp_gtid_set_specific(gtid);
542 #ifdef KMP_TDATA_GTID
543   __kmp_gtid = gtid;
544 #endif
545 #if KMP_STATS_ENABLED
546   // set thread local index to point to thread-specific stats
547   __kmp_stats_thread_ptr = ((kmp_info_t *)thr)->th.th_stats;
548   __kmp_stats_thread_ptr->startLife();
549   KMP_SET_THREAD_STATE(IDLE);
550   KMP_INIT_PARTITIONED_TIMERS(OMP_idle);
551 #endif
552 
553 #if USE_ITT_BUILD
554   __kmp_itt_thread_name(gtid);
555 #endif /* USE_ITT_BUILD */
556 
557 #if KMP_AFFINITY_SUPPORTED
558   __kmp_affinity_bind_init_mask(gtid);
559 #endif
560 
561 #ifdef KMP_CANCEL_THREADS
562   status = pthread_setcanceltype(PTHREAD_CANCEL_ASYNCHRONOUS, &old_type);
563   KMP_CHECK_SYSFAIL("pthread_setcanceltype", status);
564   // josh todo: isn't PTHREAD_CANCEL_ENABLE default for newly-created threads?
565   status = pthread_setcancelstate(PTHREAD_CANCEL_ENABLE, &old_state);
566   KMP_CHECK_SYSFAIL("pthread_setcancelstate", status);
567 #endif
568 
569 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
570   // Set FP control regs to be a copy of the parallel initialization thread's.
571   __kmp_clear_x87_fpu_status_word();
572   __kmp_load_x87_fpu_control_word(&__kmp_init_x87_fpu_control_word);
573   __kmp_load_mxcsr(&__kmp_init_mxcsr);
574 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
575 
576 #ifdef KMP_BLOCK_SIGNALS
577   status = sigfillset(&new_set);
578   KMP_CHECK_SYSFAIL_ERRNO("sigfillset", status);
579   status = pthread_sigmask(SIG_BLOCK, &new_set, &old_set);
580   KMP_CHECK_SYSFAIL("pthread_sigmask", status);
581 #endif /* KMP_BLOCK_SIGNALS */
582 
583 #if KMP_OS_LINUX || KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD ||     \
584     KMP_OS_OPENBSD || KMP_OS_HAIKU || KMP_OS_HURD || KMP_OS_SOLARIS ||         \
585     KMP_OS_AIX
586   if (__kmp_stkoffset > 0 && gtid > 0) {
587     padding = KMP_ALLOCA(gtid * __kmp_stkoffset);
588     (void)padding;
589   }
590 #endif
591 
592   KMP_MB();
593   __kmp_set_stack_info(gtid, (kmp_info_t *)thr);
594 
595   __kmp_check_stack_overlap((kmp_info_t *)thr);
596 
597   exit_val = __kmp_launch_thread((kmp_info_t *)thr);
598 
599 #ifdef KMP_BLOCK_SIGNALS
600   status = pthread_sigmask(SIG_SETMASK, &old_set, NULL);
601   KMP_CHECK_SYSFAIL("pthread_sigmask", status);
602 #endif /* KMP_BLOCK_SIGNALS */
603 
604   return exit_val;
605 }
606 
607 #if KMP_USE_MONITOR
608 /* The monitor thread controls all of the threads in the complex */
609 
610 static void *__kmp_launch_monitor(void *thr) {
611   int status, old_type, old_state;
612 #ifdef KMP_BLOCK_SIGNALS
613   sigset_t new_set;
614 #endif /* KMP_BLOCK_SIGNALS */
615   struct timespec interval;
616 
617   KMP_MB(); /* Flush all pending memory write invalidates.  */
618 
619   KA_TRACE(10, ("__kmp_launch_monitor: #1 launched\n"));
620 
621   /* register us as the monitor thread */
622   __kmp_gtid_set_specific(KMP_GTID_MONITOR);
623 #ifdef KMP_TDATA_GTID
624   __kmp_gtid = KMP_GTID_MONITOR;
625 #endif
626 
627   KMP_MB();
628 
629 #if USE_ITT_BUILD
630   // Instruct Intel(R) Threading Tools to ignore monitor thread.
631   __kmp_itt_thread_ignore();
632 #endif /* USE_ITT_BUILD */
633 
634   __kmp_set_stack_info(((kmp_info_t *)thr)->th.th_info.ds.ds_gtid,
635                        (kmp_info_t *)thr);
636 
637   __kmp_check_stack_overlap((kmp_info_t *)thr);
638 
639 #ifdef KMP_CANCEL_THREADS
640   status = pthread_setcanceltype(PTHREAD_CANCEL_ASYNCHRONOUS, &old_type);
641   KMP_CHECK_SYSFAIL("pthread_setcanceltype", status);
642   // josh todo: isn't PTHREAD_CANCEL_ENABLE default for newly-created threads?
643   status = pthread_setcancelstate(PTHREAD_CANCEL_ENABLE, &old_state);
644   KMP_CHECK_SYSFAIL("pthread_setcancelstate", status);
645 #endif
646 
647 #if KMP_REAL_TIME_FIX
648   // This is a potential fix which allows application with real-time scheduling
649   // policy work. However, decision about the fix is not made yet, so it is
650   // disabled by default.
651   { // Are program started with real-time scheduling policy?
652     int sched = sched_getscheduler(0);
653     if (sched == SCHED_FIFO || sched == SCHED_RR) {
654       // Yes, we are a part of real-time application. Try to increase the
655       // priority of the monitor.
656       struct sched_param param;
657       int max_priority = sched_get_priority_max(sched);
658       int rc;
659       KMP_WARNING(RealTimeSchedNotSupported);
660       sched_getparam(0, &param);
661       if (param.sched_priority < max_priority) {
662         param.sched_priority += 1;
663         rc = sched_setscheduler(0, sched, &param);
664         if (rc != 0) {
665           int error = errno;
666           kmp_msg_t err_code = KMP_ERR(error);
667           __kmp_msg(kmp_ms_warning, KMP_MSG(CantChangeMonitorPriority),
668                     err_code, KMP_MSG(MonitorWillStarve), __kmp_msg_null);
669           if (__kmp_generate_warnings == kmp_warnings_off) {
670             __kmp_str_free(&err_code.str);
671           }
672         }
673       } else {
674         // We cannot abort here, because number of CPUs may be enough for all
675         // the threads, including the monitor thread, so application could
676         // potentially work...
677         __kmp_msg(kmp_ms_warning, KMP_MSG(RunningAtMaxPriority),
678                   KMP_MSG(MonitorWillStarve), KMP_HNT(RunningAtMaxPriority),
679                   __kmp_msg_null);
680       }
681     }
682     // AC: free thread that waits for monitor started
683     TCW_4(__kmp_global.g.g_time.dt.t_value, 0);
684   }
685 #endif // KMP_REAL_TIME_FIX
686 
687   KMP_MB(); /* Flush all pending memory write invalidates.  */
688 
689   if (__kmp_monitor_wakeups == 1) {
690     interval.tv_sec = 1;
691     interval.tv_nsec = 0;
692   } else {
693     interval.tv_sec = 0;
694     interval.tv_nsec = (KMP_NSEC_PER_SEC / __kmp_monitor_wakeups);
695   }
696 
697   KA_TRACE(10, ("__kmp_launch_monitor: #2 monitor\n"));
698 
699   while (!TCR_4(__kmp_global.g.g_done)) {
700     struct timespec now;
701     struct timeval tval;
702 
703     /*  This thread monitors the state of the system */
704 
705     KA_TRACE(15, ("__kmp_launch_monitor: update\n"));
706 
707     status = gettimeofday(&tval, NULL);
708     KMP_CHECK_SYSFAIL_ERRNO("gettimeofday", status);
709     TIMEVAL_TO_TIMESPEC(&tval, &now);
710 
711     now.tv_sec += interval.tv_sec;
712     now.tv_nsec += interval.tv_nsec;
713 
714     if (now.tv_nsec >= KMP_NSEC_PER_SEC) {
715       now.tv_sec += 1;
716       now.tv_nsec -= KMP_NSEC_PER_SEC;
717     }
718 
719     status = pthread_mutex_lock(&__kmp_wait_mx.m_mutex);
720     KMP_CHECK_SYSFAIL("pthread_mutex_lock", status);
721     // AC: the monitor should not fall asleep if g_done has been set
722     if (!TCR_4(__kmp_global.g.g_done)) { // check once more under mutex
723       status = pthread_cond_timedwait(&__kmp_wait_cv.c_cond,
724                                       &__kmp_wait_mx.m_mutex, &now);
725       if (status != 0) {
726         if (status != ETIMEDOUT && status != EINTR) {
727           KMP_SYSFAIL("pthread_cond_timedwait", status);
728         }
729       }
730     }
731     status = pthread_mutex_unlock(&__kmp_wait_mx.m_mutex);
732     KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status);
733 
734     TCW_4(__kmp_global.g.g_time.dt.t_value,
735           TCR_4(__kmp_global.g.g_time.dt.t_value) + 1);
736 
737     KMP_MB(); /* Flush all pending memory write invalidates.  */
738   }
739 
740   KA_TRACE(10, ("__kmp_launch_monitor: #3 cleanup\n"));
741 
742 #ifdef KMP_BLOCK_SIGNALS
743   status = sigfillset(&new_set);
744   KMP_CHECK_SYSFAIL_ERRNO("sigfillset", status);
745   status = pthread_sigmask(SIG_UNBLOCK, &new_set, NULL);
746   KMP_CHECK_SYSFAIL("pthread_sigmask", status);
747 #endif /* KMP_BLOCK_SIGNALS */
748 
749   KA_TRACE(10, ("__kmp_launch_monitor: #4 finished\n"));
750 
751   if (__kmp_global.g.g_abort != 0) {
752     /* now we need to terminate the worker threads  */
753     /* the value of t_abort is the signal we caught */
754 
755     int gtid;
756 
757     KA_TRACE(10, ("__kmp_launch_monitor: #5 terminate sig=%d\n",
758                   __kmp_global.g.g_abort));
759 
760     /* terminate the OpenMP worker threads */
761     /* TODO this is not valid for sibling threads!!
762      * the uber master might not be 0 anymore.. */
763     for (gtid = 1; gtid < __kmp_threads_capacity; ++gtid)
764       __kmp_terminate_thread(gtid);
765 
766     __kmp_cleanup();
767 
768     KA_TRACE(10, ("__kmp_launch_monitor: #6 raise sig=%d\n",
769                   __kmp_global.g.g_abort));
770 
771     if (__kmp_global.g.g_abort > 0)
772       raise(__kmp_global.g.g_abort);
773   }
774 
775   KA_TRACE(10, ("__kmp_launch_monitor: #7 exit\n"));
776 
777   return thr;
778 }
779 #endif // KMP_USE_MONITOR
780 
781 void __kmp_create_worker(int gtid, kmp_info_t *th, size_t stack_size) {
782   pthread_t handle;
783   pthread_attr_t thread_attr;
784   int status;
785 
786   th->th.th_info.ds.ds_gtid = gtid;
787 
788 #if KMP_STATS_ENABLED
789   // sets up worker thread stats
790   __kmp_acquire_tas_lock(&__kmp_stats_lock, gtid);
791 
792   // th->th.th_stats is used to transfer thread-specific stats-pointer to
793   // __kmp_launch_worker. So when thread is created (goes into
794   // __kmp_launch_worker) it will set its thread local pointer to
795   // th->th.th_stats
796   if (!KMP_UBER_GTID(gtid)) {
797     th->th.th_stats = __kmp_stats_list->push_back(gtid);
798   } else {
799     // For root threads, __kmp_stats_thread_ptr is set in __kmp_register_root(),
800     // so set the th->th.th_stats field to it.
801     th->th.th_stats = __kmp_stats_thread_ptr;
802   }
803   __kmp_release_tas_lock(&__kmp_stats_lock, gtid);
804 
805 #endif // KMP_STATS_ENABLED
806 
807   if (KMP_UBER_GTID(gtid)) {
808     KA_TRACE(10, ("__kmp_create_worker: uber thread (%d)\n", gtid));
809     th->th.th_info.ds.ds_thread = pthread_self();
810     __kmp_set_stack_info(gtid, th);
811     __kmp_check_stack_overlap(th);
812     return;
813   }
814 
815   KA_TRACE(10, ("__kmp_create_worker: try to create thread (%d)\n", gtid));
816 
817   KMP_MB(); /* Flush all pending memory write invalidates.  */
818 
819 #ifdef KMP_THREAD_ATTR
820   status = pthread_attr_init(&thread_attr);
821   if (status != 0) {
822     __kmp_fatal(KMP_MSG(CantInitThreadAttrs), KMP_ERR(status), __kmp_msg_null);
823   }
824   status = pthread_attr_setdetachstate(&thread_attr, PTHREAD_CREATE_JOINABLE);
825   if (status != 0) {
826     __kmp_fatal(KMP_MSG(CantSetWorkerState), KMP_ERR(status), __kmp_msg_null);
827   }
828 
829   /* Set stack size for this thread now.
830      The multiple of 2 is there because on some machines, requesting an unusual
831      stacksize causes the thread to have an offset before the dummy alloca()
832      takes place to create the offset.  Since we want the user to have a
833      sufficient stacksize AND support a stack offset, we alloca() twice the
834      offset so that the upcoming alloca() does not eliminate any premade offset,
835      and also gives the user the stack space they requested for all threads */
836   stack_size += gtid * __kmp_stkoffset * 2;
837 
838   KA_TRACE(10, ("__kmp_create_worker: T#%d, default stacksize = %lu bytes, "
839                 "__kmp_stksize = %lu bytes, final stacksize = %lu bytes\n",
840                 gtid, KMP_DEFAULT_STKSIZE, __kmp_stksize, stack_size));
841 
842 #ifdef _POSIX_THREAD_ATTR_STACKSIZE
843   status = pthread_attr_setstacksize(&thread_attr, stack_size);
844 #ifdef KMP_BACKUP_STKSIZE
845   if (status != 0) {
846     if (!__kmp_env_stksize) {
847       stack_size = KMP_BACKUP_STKSIZE + gtid * __kmp_stkoffset;
848       __kmp_stksize = KMP_BACKUP_STKSIZE;
849       KA_TRACE(10, ("__kmp_create_worker: T#%d, default stacksize = %lu bytes, "
850                     "__kmp_stksize = %lu bytes, (backup) final stacksize = %lu "
851                     "bytes\n",
852                     gtid, KMP_DEFAULT_STKSIZE, __kmp_stksize, stack_size));
853       status = pthread_attr_setstacksize(&thread_attr, stack_size);
854     }
855   }
856 #endif /* KMP_BACKUP_STKSIZE */
857   if (status != 0) {
858     __kmp_fatal(KMP_MSG(CantSetWorkerStackSize, stack_size), KMP_ERR(status),
859                 KMP_HNT(ChangeWorkerStackSize), __kmp_msg_null);
860   }
861 #endif /* _POSIX_THREAD_ATTR_STACKSIZE */
862 
863 #endif /* KMP_THREAD_ATTR */
864 
865   status =
866       pthread_create(&handle, &thread_attr, __kmp_launch_worker, (void *)th);
867   if (status != 0 || !handle) { // ??? Why do we check handle??
868 #ifdef _POSIX_THREAD_ATTR_STACKSIZE
869     if (status == EINVAL) {
870       __kmp_fatal(KMP_MSG(CantSetWorkerStackSize, stack_size), KMP_ERR(status),
871                   KMP_HNT(IncreaseWorkerStackSize), __kmp_msg_null);
872     }
873     if (status == ENOMEM) {
874       __kmp_fatal(KMP_MSG(CantSetWorkerStackSize, stack_size), KMP_ERR(status),
875                   KMP_HNT(DecreaseWorkerStackSize), __kmp_msg_null);
876     }
877 #endif /* _POSIX_THREAD_ATTR_STACKSIZE */
878     if (status == EAGAIN) {
879       __kmp_fatal(KMP_MSG(NoResourcesForWorkerThread), KMP_ERR(status),
880                   KMP_HNT(Decrease_NUM_THREADS), __kmp_msg_null);
881     }
882     KMP_SYSFAIL("pthread_create", status);
883   }
884 
885   // Rename worker threads for improved debuggability
886   if (!KMP_UBER_GTID(gtid)) {
887 #if defined(LIBOMP_HAVE_PTHREAD_SET_NAME_NP)
888     pthread_set_name_np(handle, "openmp_worker");
889 #elif defined(LIBOMP_HAVE_PTHREAD_SETNAME_NP) && !KMP_OS_DARWIN
890 #if KMP_OS_NETBSD
891     pthread_setname_np(handle, "%s", const_cast<char *>("openmp_worker"));
892 #else
893     pthread_setname_np(handle, "openmp_worker");
894 #endif
895 #endif
896   }
897 
898   th->th.th_info.ds.ds_thread = handle;
899 
900 #ifdef KMP_THREAD_ATTR
901   status = pthread_attr_destroy(&thread_attr);
902   if (status) {
903     kmp_msg_t err_code = KMP_ERR(status);
904     __kmp_msg(kmp_ms_warning, KMP_MSG(CantDestroyThreadAttrs), err_code,
905               __kmp_msg_null);
906     if (__kmp_generate_warnings == kmp_warnings_off) {
907       __kmp_str_free(&err_code.str);
908     }
909   }
910 #endif /* KMP_THREAD_ATTR */
911 
912   KMP_MB(); /* Flush all pending memory write invalidates.  */
913 
914   KA_TRACE(10, ("__kmp_create_worker: done creating thread (%d)\n", gtid));
915 
916 } // __kmp_create_worker
917 
918 #if KMP_USE_MONITOR
919 void __kmp_create_monitor(kmp_info_t *th) {
920   pthread_t handle;
921   pthread_attr_t thread_attr;
922   size_t size;
923   int status;
924   int auto_adj_size = FALSE;
925 
926   if (__kmp_dflt_blocktime == KMP_MAX_BLOCKTIME) {
927     // We don't need monitor thread in case of MAX_BLOCKTIME
928     KA_TRACE(10, ("__kmp_create_monitor: skipping monitor thread because of "
929                   "MAX blocktime\n"));
930     th->th.th_info.ds.ds_tid = 0; // this makes reap_monitor no-op
931     th->th.th_info.ds.ds_gtid = 0;
932     return;
933   }
934   KA_TRACE(10, ("__kmp_create_monitor: try to create monitor\n"));
935 
936   KMP_MB(); /* Flush all pending memory write invalidates.  */
937 
938   th->th.th_info.ds.ds_tid = KMP_GTID_MONITOR;
939   th->th.th_info.ds.ds_gtid = KMP_GTID_MONITOR;
940 #if KMP_REAL_TIME_FIX
941   TCW_4(__kmp_global.g.g_time.dt.t_value,
942         -1); // Will use it for synchronization a bit later.
943 #else
944   TCW_4(__kmp_global.g.g_time.dt.t_value, 0);
945 #endif // KMP_REAL_TIME_FIX
946 
947 #ifdef KMP_THREAD_ATTR
948   if (__kmp_monitor_stksize == 0) {
949     __kmp_monitor_stksize = KMP_DEFAULT_MONITOR_STKSIZE;
950     auto_adj_size = TRUE;
951   }
952   status = pthread_attr_init(&thread_attr);
953   if (status != 0) {
954     __kmp_fatal(KMP_MSG(CantInitThreadAttrs), KMP_ERR(status), __kmp_msg_null);
955   }
956   status = pthread_attr_setdetachstate(&thread_attr, PTHREAD_CREATE_JOINABLE);
957   if (status != 0) {
958     __kmp_fatal(KMP_MSG(CantSetMonitorState), KMP_ERR(status), __kmp_msg_null);
959   }
960 
961 #ifdef _POSIX_THREAD_ATTR_STACKSIZE
962   status = pthread_attr_getstacksize(&thread_attr, &size);
963   KMP_CHECK_SYSFAIL("pthread_attr_getstacksize", status);
964 #else
965   size = __kmp_sys_min_stksize;
966 #endif /* _POSIX_THREAD_ATTR_STACKSIZE */
967 #endif /* KMP_THREAD_ATTR */
968 
969   if (__kmp_monitor_stksize == 0) {
970     __kmp_monitor_stksize = KMP_DEFAULT_MONITOR_STKSIZE;
971   }
972   if (__kmp_monitor_stksize < __kmp_sys_min_stksize) {
973     __kmp_monitor_stksize = __kmp_sys_min_stksize;
974   }
975 
976   KA_TRACE(10, ("__kmp_create_monitor: default stacksize = %lu bytes,"
977                 "requested stacksize = %lu bytes\n",
978                 size, __kmp_monitor_stksize));
979 
980 retry:
981 
982 /* Set stack size for this thread now. */
983 #ifdef _POSIX_THREAD_ATTR_STACKSIZE
984   KA_TRACE(10, ("__kmp_create_monitor: setting stacksize = %lu bytes,",
985                 __kmp_monitor_stksize));
986   status = pthread_attr_setstacksize(&thread_attr, __kmp_monitor_stksize);
987   if (status != 0) {
988     if (auto_adj_size) {
989       __kmp_monitor_stksize *= 2;
990       goto retry;
991     }
992     kmp_msg_t err_code = KMP_ERR(status);
993     __kmp_msg(kmp_ms_warning, // should this be fatal?  BB
994               KMP_MSG(CantSetMonitorStackSize, (long int)__kmp_monitor_stksize),
995               err_code, KMP_HNT(ChangeMonitorStackSize), __kmp_msg_null);
996     if (__kmp_generate_warnings == kmp_warnings_off) {
997       __kmp_str_free(&err_code.str);
998     }
999   }
1000 #endif /* _POSIX_THREAD_ATTR_STACKSIZE */
1001 
1002   status =
1003       pthread_create(&handle, &thread_attr, __kmp_launch_monitor, (void *)th);
1004 
1005   if (status != 0) {
1006 #ifdef _POSIX_THREAD_ATTR_STACKSIZE
1007     if (status == EINVAL) {
1008       if (auto_adj_size && (__kmp_monitor_stksize < (size_t)0x40000000)) {
1009         __kmp_monitor_stksize *= 2;
1010         goto retry;
1011       }
1012       __kmp_fatal(KMP_MSG(CantSetMonitorStackSize, __kmp_monitor_stksize),
1013                   KMP_ERR(status), KMP_HNT(IncreaseMonitorStackSize),
1014                   __kmp_msg_null);
1015     }
1016     if (status == ENOMEM) {
1017       __kmp_fatal(KMP_MSG(CantSetMonitorStackSize, __kmp_monitor_stksize),
1018                   KMP_ERR(status), KMP_HNT(DecreaseMonitorStackSize),
1019                   __kmp_msg_null);
1020     }
1021 #endif /* _POSIX_THREAD_ATTR_STACKSIZE */
1022     if (status == EAGAIN) {
1023       __kmp_fatal(KMP_MSG(NoResourcesForMonitorThread), KMP_ERR(status),
1024                   KMP_HNT(DecreaseNumberOfThreadsInUse), __kmp_msg_null);
1025     }
1026     KMP_SYSFAIL("pthread_create", status);
1027   }
1028 
1029   th->th.th_info.ds.ds_thread = handle;
1030 
1031 #if KMP_REAL_TIME_FIX
1032   // Wait for the monitor thread is really started and set its *priority*.
1033   KMP_DEBUG_ASSERT(sizeof(kmp_uint32) ==
1034                    sizeof(__kmp_global.g.g_time.dt.t_value));
1035   __kmp_wait_4((kmp_uint32 volatile *)&__kmp_global.g.g_time.dt.t_value, -1,
1036                &__kmp_neq_4, NULL);
1037 #endif // KMP_REAL_TIME_FIX
1038 
1039 #ifdef KMP_THREAD_ATTR
1040   status = pthread_attr_destroy(&thread_attr);
1041   if (status != 0) {
1042     kmp_msg_t err_code = KMP_ERR(status);
1043     __kmp_msg(kmp_ms_warning, KMP_MSG(CantDestroyThreadAttrs), err_code,
1044               __kmp_msg_null);
1045     if (__kmp_generate_warnings == kmp_warnings_off) {
1046       __kmp_str_free(&err_code.str);
1047     }
1048   }
1049 #endif
1050 
1051   KMP_MB(); /* Flush all pending memory write invalidates.  */
1052 
1053   KA_TRACE(10, ("__kmp_create_monitor: monitor created %#.8lx\n",
1054                 th->th.th_info.ds.ds_thread));
1055 
1056 } // __kmp_create_monitor
1057 #endif // KMP_USE_MONITOR
1058 
1059 void __kmp_exit_thread(int exit_status) {
1060 #if KMP_OS_WASI
1061 // TODO: the wasm32-wasi-threads target does not yet support pthread_exit.
1062 #else
1063   pthread_exit((void *)(intptr_t)exit_status);
1064 #endif
1065 } // __kmp_exit_thread
1066 
1067 #if KMP_USE_MONITOR
1068 void __kmp_resume_monitor();
1069 
1070 extern "C" void __kmp_reap_monitor(kmp_info_t *th) {
1071   int status;
1072   void *exit_val;
1073 
1074   KA_TRACE(10, ("__kmp_reap_monitor: try to reap monitor thread with handle"
1075                 " %#.8lx\n",
1076                 th->th.th_info.ds.ds_thread));
1077 
1078   // If monitor has been created, its tid and gtid should be KMP_GTID_MONITOR.
1079   // If both tid and gtid are 0, it means the monitor did not ever start.
1080   // If both tid and gtid are KMP_GTID_DNE, the monitor has been shut down.
1081   KMP_DEBUG_ASSERT(th->th.th_info.ds.ds_tid == th->th.th_info.ds.ds_gtid);
1082   if (th->th.th_info.ds.ds_gtid != KMP_GTID_MONITOR) {
1083     KA_TRACE(10, ("__kmp_reap_monitor: monitor did not start, returning\n"));
1084     return;
1085   }
1086 
1087   KMP_MB(); /* Flush all pending memory write invalidates.  */
1088 
1089   /* First, check to see whether the monitor thread exists to wake it up. This
1090      is to avoid performance problem when the monitor sleeps during
1091      blocktime-size interval */
1092 
1093   status = pthread_kill(th->th.th_info.ds.ds_thread, 0);
1094   if (status != ESRCH) {
1095     __kmp_resume_monitor(); // Wake up the monitor thread
1096   }
1097   KA_TRACE(10, ("__kmp_reap_monitor: try to join with monitor\n"));
1098   status = pthread_join(th->th.th_info.ds.ds_thread, &exit_val);
1099   if (exit_val != th) {
1100     __kmp_fatal(KMP_MSG(ReapMonitorError), KMP_ERR(status), __kmp_msg_null);
1101   }
1102 
1103   th->th.th_info.ds.ds_tid = KMP_GTID_DNE;
1104   th->th.th_info.ds.ds_gtid = KMP_GTID_DNE;
1105 
1106   KA_TRACE(10, ("__kmp_reap_monitor: done reaping monitor thread with handle"
1107                 " %#.8lx\n",
1108                 th->th.th_info.ds.ds_thread));
1109 
1110   KMP_MB(); /* Flush all pending memory write invalidates.  */
1111 }
1112 #else
1113 // Empty symbol to export (see exports_so.txt) when
1114 // monitor thread feature is disabled
1115 extern "C" void __kmp_reap_monitor(kmp_info_t *th) { (void)th; }
1116 #endif // KMP_USE_MONITOR
1117 
1118 void __kmp_reap_worker(kmp_info_t *th) {
1119   int status;
1120   void *exit_val;
1121 
1122   KMP_MB(); /* Flush all pending memory write invalidates.  */
1123 
1124   KA_TRACE(
1125       10, ("__kmp_reap_worker: try to reap T#%d\n", th->th.th_info.ds.ds_gtid));
1126 
1127   status = pthread_join(th->th.th_info.ds.ds_thread, &exit_val);
1128 #ifdef KMP_DEBUG
1129   /* Don't expose these to the user until we understand when they trigger */
1130   if (status != 0) {
1131     __kmp_fatal(KMP_MSG(ReapWorkerError), KMP_ERR(status), __kmp_msg_null);
1132   }
1133   if (exit_val != th) {
1134     KA_TRACE(10, ("__kmp_reap_worker: worker T#%d did not reap properly, "
1135                   "exit_val = %p\n",
1136                   th->th.th_info.ds.ds_gtid, exit_val));
1137   }
1138 #else
1139   (void)status; // unused variable
1140 #endif /* KMP_DEBUG */
1141 
1142   KA_TRACE(10, ("__kmp_reap_worker: done reaping T#%d\n",
1143                 th->th.th_info.ds.ds_gtid));
1144 
1145   KMP_MB(); /* Flush all pending memory write invalidates.  */
1146 }
1147 
1148 #if KMP_HANDLE_SIGNALS
1149 
1150 static void __kmp_null_handler(int signo) {
1151   //  Do nothing, for doing SIG_IGN-type actions.
1152 } // __kmp_null_handler
1153 
1154 static void __kmp_team_handler(int signo) {
1155   if (__kmp_global.g.g_abort == 0) {
1156 /* Stage 1 signal handler, let's shut down all of the threads */
1157 #ifdef KMP_DEBUG
1158     __kmp_debug_printf("__kmp_team_handler: caught signal = %d\n", signo);
1159 #endif
1160     switch (signo) {
1161     case SIGHUP:
1162     case SIGINT:
1163     case SIGQUIT:
1164     case SIGILL:
1165     case SIGABRT:
1166     case SIGFPE:
1167     case SIGBUS:
1168     case SIGSEGV:
1169 #ifdef SIGSYS
1170     case SIGSYS:
1171 #endif
1172     case SIGTERM:
1173       if (__kmp_debug_buf) {
1174         __kmp_dump_debug_buffer();
1175       }
1176       __kmp_unregister_library(); // cleanup shared memory
1177       KMP_MB(); // Flush all pending memory write invalidates.
1178       TCW_4(__kmp_global.g.g_abort, signo);
1179       KMP_MB(); // Flush all pending memory write invalidates.
1180       TCW_4(__kmp_global.g.g_done, TRUE);
1181       KMP_MB(); // Flush all pending memory write invalidates.
1182       break;
1183     default:
1184 #ifdef KMP_DEBUG
1185       __kmp_debug_printf("__kmp_team_handler: unknown signal type");
1186 #endif
1187       break;
1188     }
1189   }
1190 } // __kmp_team_handler
1191 
1192 static void __kmp_sigaction(int signum, const struct sigaction *act,
1193                             struct sigaction *oldact) {
1194   int rc = sigaction(signum, act, oldact);
1195   KMP_CHECK_SYSFAIL_ERRNO("sigaction", rc);
1196 }
1197 
1198 static void __kmp_install_one_handler(int sig, sig_func_t handler_func,
1199                                       int parallel_init) {
1200   KMP_MB(); // Flush all pending memory write invalidates.
1201   KB_TRACE(60,
1202            ("__kmp_install_one_handler( %d, ..., %d )\n", sig, parallel_init));
1203   if (parallel_init) {
1204     struct sigaction new_action;
1205     struct sigaction old_action;
1206     new_action.sa_handler = handler_func;
1207     new_action.sa_flags = 0;
1208     sigfillset(&new_action.sa_mask);
1209     __kmp_sigaction(sig, &new_action, &old_action);
1210     if (old_action.sa_handler == __kmp_sighldrs[sig].sa_handler) {
1211       sigaddset(&__kmp_sigset, sig);
1212     } else {
1213       // Restore/keep user's handler if one previously installed.
1214       __kmp_sigaction(sig, &old_action, NULL);
1215     }
1216   } else {
1217     // Save initial/system signal handlers to see if user handlers installed.
1218     __kmp_sigaction(sig, NULL, &__kmp_sighldrs[sig]);
1219   }
1220   KMP_MB(); // Flush all pending memory write invalidates.
1221 } // __kmp_install_one_handler
1222 
1223 static void __kmp_remove_one_handler(int sig) {
1224   KB_TRACE(60, ("__kmp_remove_one_handler( %d )\n", sig));
1225   if (sigismember(&__kmp_sigset, sig)) {
1226     struct sigaction old;
1227     KMP_MB(); // Flush all pending memory write invalidates.
1228     __kmp_sigaction(sig, &__kmp_sighldrs[sig], &old);
1229     if ((old.sa_handler != __kmp_team_handler) &&
1230         (old.sa_handler != __kmp_null_handler)) {
1231       // Restore the users signal handler.
1232       KB_TRACE(10, ("__kmp_remove_one_handler: oops, not our handler, "
1233                     "restoring: sig=%d\n",
1234                     sig));
1235       __kmp_sigaction(sig, &old, NULL);
1236     }
1237     sigdelset(&__kmp_sigset, sig);
1238     KMP_MB(); // Flush all pending memory write invalidates.
1239   }
1240 } // __kmp_remove_one_handler
1241 
1242 void __kmp_install_signals(int parallel_init) {
1243   KB_TRACE(10, ("__kmp_install_signals( %d )\n", parallel_init));
1244   if (__kmp_handle_signals || !parallel_init) {
1245     // If ! parallel_init, we do not install handlers, just save original
1246     // handlers. Let us do it even __handle_signals is 0.
1247     sigemptyset(&__kmp_sigset);
1248     __kmp_install_one_handler(SIGHUP, __kmp_team_handler, parallel_init);
1249     __kmp_install_one_handler(SIGINT, __kmp_team_handler, parallel_init);
1250     __kmp_install_one_handler(SIGQUIT, __kmp_team_handler, parallel_init);
1251     __kmp_install_one_handler(SIGILL, __kmp_team_handler, parallel_init);
1252     __kmp_install_one_handler(SIGABRT, __kmp_team_handler, parallel_init);
1253     __kmp_install_one_handler(SIGFPE, __kmp_team_handler, parallel_init);
1254     __kmp_install_one_handler(SIGBUS, __kmp_team_handler, parallel_init);
1255     __kmp_install_one_handler(SIGSEGV, __kmp_team_handler, parallel_init);
1256 #ifdef SIGSYS
1257     __kmp_install_one_handler(SIGSYS, __kmp_team_handler, parallel_init);
1258 #endif // SIGSYS
1259     __kmp_install_one_handler(SIGTERM, __kmp_team_handler, parallel_init);
1260 #ifdef SIGPIPE
1261     __kmp_install_one_handler(SIGPIPE, __kmp_team_handler, parallel_init);
1262 #endif // SIGPIPE
1263   }
1264 } // __kmp_install_signals
1265 
1266 void __kmp_remove_signals(void) {
1267   int sig;
1268   KB_TRACE(10, ("__kmp_remove_signals()\n"));
1269   for (sig = 1; sig < NSIG; ++sig) {
1270     __kmp_remove_one_handler(sig);
1271   }
1272 } // __kmp_remove_signals
1273 
1274 #endif // KMP_HANDLE_SIGNALS
1275 
1276 void __kmp_enable(int new_state) {
1277 #ifdef KMP_CANCEL_THREADS
1278   int status, old_state;
1279   status = pthread_setcancelstate(new_state, &old_state);
1280   KMP_CHECK_SYSFAIL("pthread_setcancelstate", status);
1281   KMP_DEBUG_ASSERT(old_state == PTHREAD_CANCEL_DISABLE);
1282 #endif
1283 }
1284 
1285 void __kmp_disable(int *old_state) {
1286 #ifdef KMP_CANCEL_THREADS
1287   int status;
1288   status = pthread_setcancelstate(PTHREAD_CANCEL_DISABLE, old_state);
1289   KMP_CHECK_SYSFAIL("pthread_setcancelstate", status);
1290 #endif
1291 }
1292 
1293 static void __kmp_atfork_prepare(void) {
1294   __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);
1295   __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock);
1296 }
1297 
1298 static void __kmp_atfork_parent(void) {
1299   __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
1300   __kmp_release_bootstrap_lock(&__kmp_initz_lock);
1301 }
1302 
1303 /* Reset the library so execution in the child starts "all over again" with
1304    clean data structures in initial states.  Don't worry about freeing memory
1305    allocated by parent, just abandon it to be safe. */
1306 static void __kmp_atfork_child(void) {
1307   __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
1308   __kmp_release_bootstrap_lock(&__kmp_initz_lock);
1309   /* TODO make sure this is done right for nested/sibling */
1310   // ATT:  Memory leaks are here? TODO: Check it and fix.
1311   /* KMP_ASSERT( 0 ); */
1312 
1313   ++__kmp_fork_count;
1314 
1315 #if KMP_AFFINITY_SUPPORTED
1316 #if KMP_OS_LINUX || KMP_OS_FREEBSD || KMP_OS_NETBSD || KMP_OS_DRAGONFLY ||     \
1317     KMP_OS_AIX
1318   // reset the affinity in the child to the initial thread
1319   // affinity in the parent
1320   kmp_set_thread_affinity_mask_initial();
1321 #endif
1322   // Set default not to bind threads tightly in the child (we're expecting
1323   // over-subscription after the fork and this can improve things for
1324   // scripting languages that use OpenMP inside process-parallel code).
1325   if (__kmp_nested_proc_bind.bind_types != NULL) {
1326     __kmp_nested_proc_bind.bind_types[0] = proc_bind_false;
1327   }
1328   for (kmp_affinity_t *affinity : __kmp_affinities)
1329     *affinity = KMP_AFFINITY_INIT(affinity->env_var);
1330   __kmp_affin_fullMask = nullptr;
1331   __kmp_affin_origMask = nullptr;
1332   __kmp_topology = nullptr;
1333 #endif // KMP_AFFINITY_SUPPORTED
1334 
1335 #if KMP_USE_MONITOR
1336   __kmp_init_monitor = 0;
1337 #endif
1338   __kmp_init_parallel = FALSE;
1339   __kmp_init_middle = FALSE;
1340   __kmp_init_serial = FALSE;
1341   TCW_4(__kmp_init_gtid, FALSE);
1342   __kmp_init_common = FALSE;
1343 
1344   TCW_4(__kmp_init_user_locks, FALSE);
1345 #if !KMP_USE_DYNAMIC_LOCK
1346   __kmp_user_lock_table.used = 1;
1347   __kmp_user_lock_table.allocated = 0;
1348   __kmp_user_lock_table.table = NULL;
1349   __kmp_lock_blocks = NULL;
1350 #endif
1351 
1352   __kmp_all_nth = 0;
1353   TCW_4(__kmp_nth, 0);
1354 
1355   __kmp_thread_pool = NULL;
1356   __kmp_thread_pool_insert_pt = NULL;
1357   __kmp_team_pool = NULL;
1358 
1359   /* Must actually zero all the *cache arguments passed to __kmpc_threadprivate
1360      here so threadprivate doesn't use stale data */
1361   KA_TRACE(10, ("__kmp_atfork_child: checking cache address list %p\n",
1362                 __kmp_threadpriv_cache_list));
1363 
1364   while (__kmp_threadpriv_cache_list != NULL) {
1365 
1366     if (*__kmp_threadpriv_cache_list->addr != NULL) {
1367       KC_TRACE(50, ("__kmp_atfork_child: zeroing cache at address %p\n",
1368                     &(*__kmp_threadpriv_cache_list->addr)));
1369 
1370       *__kmp_threadpriv_cache_list->addr = NULL;
1371     }
1372     __kmp_threadpriv_cache_list = __kmp_threadpriv_cache_list->next;
1373   }
1374 
1375   __kmp_init_runtime = FALSE;
1376 
1377   /* reset statically initialized locks */
1378   __kmp_init_bootstrap_lock(&__kmp_initz_lock);
1379   __kmp_init_bootstrap_lock(&__kmp_stdio_lock);
1380   __kmp_init_bootstrap_lock(&__kmp_console_lock);
1381   __kmp_init_bootstrap_lock(&__kmp_task_team_lock);
1382 
1383 #if USE_ITT_BUILD
1384   __kmp_itt_reset(); // reset ITT's global state
1385 #endif /* USE_ITT_BUILD */
1386 
1387   {
1388     // Child process often get terminated without any use of OpenMP. That might
1389     // cause mapped shared memory file to be left unattended. Thus we postpone
1390     // library registration till middle initialization in the child process.
1391     __kmp_need_register_serial = FALSE;
1392     __kmp_serial_initialize();
1393   }
1394 
1395   /* This is necessary to make sure no stale data is left around */
1396   /* AC: customers complain that we use unsafe routines in the atfork
1397      handler. Mathworks: dlsym() is unsafe. We call dlsym and dlopen
1398      in dynamic_link when check the presence of shared tbbmalloc library.
1399      Suggestion is to make the library initialization lazier, similar
1400      to what done for __kmpc_begin(). */
1401   // TODO: synchronize all static initializations with regular library
1402   //       startup; look at kmp_global.cpp and etc.
1403   //__kmp_internal_begin ();
1404 }
1405 
1406 void __kmp_register_atfork(void) {
1407   if (__kmp_need_register_atfork) {
1408 #if !KMP_OS_WASI
1409     int status = pthread_atfork(__kmp_atfork_prepare, __kmp_atfork_parent,
1410                                 __kmp_atfork_child);
1411     KMP_CHECK_SYSFAIL("pthread_atfork", status);
1412 #endif
1413     __kmp_need_register_atfork = FALSE;
1414   }
1415 }
1416 
1417 void __kmp_suspend_initialize(void) {
1418   int status;
1419   status = pthread_mutexattr_init(&__kmp_suspend_mutex_attr);
1420   KMP_CHECK_SYSFAIL("pthread_mutexattr_init", status);
1421   status = pthread_condattr_init(&__kmp_suspend_cond_attr);
1422   KMP_CHECK_SYSFAIL("pthread_condattr_init", status);
1423 }
1424 
1425 void __kmp_suspend_initialize_thread(kmp_info_t *th) {
1426   int old_value = KMP_ATOMIC_LD_RLX(&th->th.th_suspend_init_count);
1427   int new_value = __kmp_fork_count + 1;
1428   // Return if already initialized
1429   if (old_value == new_value)
1430     return;
1431   // Wait, then return if being initialized
1432   if (old_value == -1 || !__kmp_atomic_compare_store(
1433                              &th->th.th_suspend_init_count, old_value, -1)) {
1434     while (KMP_ATOMIC_LD_ACQ(&th->th.th_suspend_init_count) != new_value) {
1435       KMP_CPU_PAUSE();
1436     }
1437   } else {
1438     // Claim to be the initializer and do initializations
1439     int status;
1440     status = pthread_cond_init(&th->th.th_suspend_cv.c_cond,
1441                                &__kmp_suspend_cond_attr);
1442     KMP_CHECK_SYSFAIL("pthread_cond_init", status);
1443     status = pthread_mutex_init(&th->th.th_suspend_mx.m_mutex,
1444                                 &__kmp_suspend_mutex_attr);
1445     KMP_CHECK_SYSFAIL("pthread_mutex_init", status);
1446     KMP_ATOMIC_ST_REL(&th->th.th_suspend_init_count, new_value);
1447   }
1448 }
1449 
1450 void __kmp_suspend_uninitialize_thread(kmp_info_t *th) {
1451   if (KMP_ATOMIC_LD_ACQ(&th->th.th_suspend_init_count) > __kmp_fork_count) {
1452     /* this means we have initialize the suspension pthread objects for this
1453        thread in this instance of the process */
1454     int status;
1455 
1456     status = pthread_cond_destroy(&th->th.th_suspend_cv.c_cond);
1457     if (status != 0 && status != EBUSY) {
1458       KMP_SYSFAIL("pthread_cond_destroy", status);
1459     }
1460     status = pthread_mutex_destroy(&th->th.th_suspend_mx.m_mutex);
1461     if (status != 0 && status != EBUSY) {
1462       KMP_SYSFAIL("pthread_mutex_destroy", status);
1463     }
1464     --th->th.th_suspend_init_count;
1465     KMP_DEBUG_ASSERT(KMP_ATOMIC_LD_RLX(&th->th.th_suspend_init_count) ==
1466                      __kmp_fork_count);
1467   }
1468 }
1469 
1470 // return true if lock obtained, false otherwise
1471 int __kmp_try_suspend_mx(kmp_info_t *th) {
1472   return (pthread_mutex_trylock(&th->th.th_suspend_mx.m_mutex) == 0);
1473 }
1474 
1475 void __kmp_lock_suspend_mx(kmp_info_t *th) {
1476   int status = pthread_mutex_lock(&th->th.th_suspend_mx.m_mutex);
1477   KMP_CHECK_SYSFAIL("pthread_mutex_lock", status);
1478 }
1479 
1480 void __kmp_unlock_suspend_mx(kmp_info_t *th) {
1481   int status = pthread_mutex_unlock(&th->th.th_suspend_mx.m_mutex);
1482   KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status);
1483 }
1484 
1485 /* This routine puts the calling thread to sleep after setting the
1486    sleep bit for the indicated flag variable to true. */
1487 template <class C>
1488 static inline void __kmp_suspend_template(int th_gtid, C *flag) {
1489   KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(USER_suspend);
1490   kmp_info_t *th = __kmp_threads[th_gtid];
1491   int status;
1492   typename C::flag_t old_spin;
1493 
1494   KF_TRACE(30, ("__kmp_suspend_template: T#%d enter for flag = %p\n", th_gtid,
1495                 flag->get()));
1496 
1497   __kmp_suspend_initialize_thread(th);
1498 
1499   __kmp_lock_suspend_mx(th);
1500 
1501   KF_TRACE(10, ("__kmp_suspend_template: T#%d setting sleep bit for spin(%p)\n",
1502                 th_gtid, flag->get()));
1503 
1504   /* TODO: shouldn't this use release semantics to ensure that
1505      __kmp_suspend_initialize_thread gets called first? */
1506   old_spin = flag->set_sleeping();
1507   TCW_PTR(th->th.th_sleep_loc, (void *)flag);
1508   th->th.th_sleep_loc_type = flag->get_type();
1509   if (__kmp_dflt_blocktime == KMP_MAX_BLOCKTIME &&
1510       __kmp_pause_status != kmp_soft_paused) {
1511     flag->unset_sleeping();
1512     TCW_PTR(th->th.th_sleep_loc, NULL);
1513     th->th.th_sleep_loc_type = flag_unset;
1514     __kmp_unlock_suspend_mx(th);
1515     return;
1516   }
1517   KF_TRACE(5, ("__kmp_suspend_template: T#%d set sleep bit for spin(%p)==%x,"
1518                " was %x\n",
1519                th_gtid, flag->get(), flag->load(), old_spin));
1520 
1521   if (flag->done_check_val(old_spin) || flag->done_check()) {
1522     flag->unset_sleeping();
1523     TCW_PTR(th->th.th_sleep_loc, NULL);
1524     th->th.th_sleep_loc_type = flag_unset;
1525     KF_TRACE(5, ("__kmp_suspend_template: T#%d false alarm, reset sleep bit "
1526                  "for spin(%p)\n",
1527                  th_gtid, flag->get()));
1528   } else {
1529     /* Encapsulate in a loop as the documentation states that this may
1530        "with low probability" return when the condition variable has
1531        not been signaled or broadcast */
1532     int deactivated = FALSE;
1533 
1534     while (flag->is_sleeping()) {
1535 #ifdef DEBUG_SUSPEND
1536       char buffer[128];
1537       __kmp_suspend_count++;
1538       __kmp_print_cond(buffer, &th->th.th_suspend_cv);
1539       __kmp_printf("__kmp_suspend_template: suspending T#%d: %s\n", th_gtid,
1540                    buffer);
1541 #endif
1542       // Mark the thread as no longer active (only in the first iteration of the
1543       // loop).
1544       if (!deactivated) {
1545         th->th.th_active = FALSE;
1546         if (th->th.th_active_in_pool) {
1547           th->th.th_active_in_pool = FALSE;
1548           KMP_ATOMIC_DEC(&__kmp_thread_pool_active_nth);
1549           KMP_DEBUG_ASSERT(TCR_4(__kmp_thread_pool_active_nth) >= 0);
1550         }
1551         deactivated = TRUE;
1552       }
1553 
1554       KMP_DEBUG_ASSERT(th->th.th_sleep_loc);
1555       KMP_DEBUG_ASSERT(flag->get_type() == th->th.th_sleep_loc_type);
1556 
1557 #if USE_SUSPEND_TIMEOUT
1558       struct timespec now;
1559       struct timeval tval;
1560       int msecs;
1561 
1562       status = gettimeofday(&tval, NULL);
1563       KMP_CHECK_SYSFAIL_ERRNO("gettimeofday", status);
1564       TIMEVAL_TO_TIMESPEC(&tval, &now);
1565 
1566       msecs = (4 * __kmp_dflt_blocktime) + 200;
1567       now.tv_sec += msecs / 1000;
1568       now.tv_nsec += (msecs % 1000) * 1000;
1569 
1570       KF_TRACE(15, ("__kmp_suspend_template: T#%d about to perform "
1571                     "pthread_cond_timedwait\n",
1572                     th_gtid));
1573       status = pthread_cond_timedwait(&th->th.th_suspend_cv.c_cond,
1574                                       &th->th.th_suspend_mx.m_mutex, &now);
1575 #else
1576       KF_TRACE(15, ("__kmp_suspend_template: T#%d about to perform"
1577                     " pthread_cond_wait\n",
1578                     th_gtid));
1579       status = pthread_cond_wait(&th->th.th_suspend_cv.c_cond,
1580                                  &th->th.th_suspend_mx.m_mutex);
1581 #endif // USE_SUSPEND_TIMEOUT
1582 
1583       if ((status != 0) && (status != EINTR) && (status != ETIMEDOUT)) {
1584         KMP_SYSFAIL("pthread_cond_wait", status);
1585       }
1586 
1587       KMP_DEBUG_ASSERT(flag->get_type() == flag->get_ptr_type());
1588 
1589       if (!flag->is_sleeping() &&
1590           ((status == EINTR) || (status == ETIMEDOUT))) {
1591         // if interrupt or timeout, and thread is no longer sleeping, we need to
1592         // make sure sleep_loc gets reset; however, this shouldn't be needed if
1593         // we woke up with resume
1594         flag->unset_sleeping();
1595         TCW_PTR(th->th.th_sleep_loc, NULL);
1596         th->th.th_sleep_loc_type = flag_unset;
1597       }
1598 #ifdef KMP_DEBUG
1599       if (status == ETIMEDOUT) {
1600         if (flag->is_sleeping()) {
1601           KF_TRACE(100,
1602                    ("__kmp_suspend_template: T#%d timeout wakeup\n", th_gtid));
1603         } else {
1604           KF_TRACE(2, ("__kmp_suspend_template: T#%d timeout wakeup, sleep bit "
1605                        "not set!\n",
1606                        th_gtid));
1607           TCW_PTR(th->th.th_sleep_loc, NULL);
1608           th->th.th_sleep_loc_type = flag_unset;
1609         }
1610       } else if (flag->is_sleeping()) {
1611         KF_TRACE(100,
1612                  ("__kmp_suspend_template: T#%d spurious wakeup\n", th_gtid));
1613       }
1614 #endif
1615     } // while
1616 
1617     // Mark the thread as active again (if it was previous marked as inactive)
1618     if (deactivated) {
1619       th->th.th_active = TRUE;
1620       if (TCR_4(th->th.th_in_pool)) {
1621         KMP_ATOMIC_INC(&__kmp_thread_pool_active_nth);
1622         th->th.th_active_in_pool = TRUE;
1623       }
1624     }
1625   }
1626   // We may have had the loop variable set before entering the loop body;
1627   // so we need to reset sleep_loc.
1628   TCW_PTR(th->th.th_sleep_loc, NULL);
1629   th->th.th_sleep_loc_type = flag_unset;
1630 
1631   KMP_DEBUG_ASSERT(!flag->is_sleeping());
1632   KMP_DEBUG_ASSERT(!th->th.th_sleep_loc);
1633 #ifdef DEBUG_SUSPEND
1634   {
1635     char buffer[128];
1636     __kmp_print_cond(buffer, &th->th.th_suspend_cv);
1637     __kmp_printf("__kmp_suspend_template: T#%d has awakened: %s\n", th_gtid,
1638                  buffer);
1639   }
1640 #endif
1641 
1642   __kmp_unlock_suspend_mx(th);
1643   KF_TRACE(30, ("__kmp_suspend_template: T#%d exit\n", th_gtid));
1644 }
1645 
1646 template <bool C, bool S>
1647 void __kmp_suspend_32(int th_gtid, kmp_flag_32<C, S> *flag) {
1648   __kmp_suspend_template(th_gtid, flag);
1649 }
1650 template <bool C, bool S>
1651 void __kmp_suspend_64(int th_gtid, kmp_flag_64<C, S> *flag) {
1652   __kmp_suspend_template(th_gtid, flag);
1653 }
1654 template <bool C, bool S>
1655 void __kmp_atomic_suspend_64(int th_gtid, kmp_atomic_flag_64<C, S> *flag) {
1656   __kmp_suspend_template(th_gtid, flag);
1657 }
1658 void __kmp_suspend_oncore(int th_gtid, kmp_flag_oncore *flag) {
1659   __kmp_suspend_template(th_gtid, flag);
1660 }
1661 
1662 template void __kmp_suspend_32<false, false>(int, kmp_flag_32<false, false> *);
1663 template void __kmp_suspend_64<false, true>(int, kmp_flag_64<false, true> *);
1664 template void __kmp_suspend_64<true, false>(int, kmp_flag_64<true, false> *);
1665 template void
1666 __kmp_atomic_suspend_64<false, true>(int, kmp_atomic_flag_64<false, true> *);
1667 template void
1668 __kmp_atomic_suspend_64<true, false>(int, kmp_atomic_flag_64<true, false> *);
1669 
1670 /* This routine signals the thread specified by target_gtid to wake up
1671    after setting the sleep bit indicated by the flag argument to FALSE.
1672    The target thread must already have called __kmp_suspend_template() */
1673 template <class C>
1674 static inline void __kmp_resume_template(int target_gtid, C *flag) {
1675   KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(USER_resume);
1676   kmp_info_t *th = __kmp_threads[target_gtid];
1677   int status;
1678 
1679 #ifdef KMP_DEBUG
1680   int gtid = TCR_4(__kmp_init_gtid) ? __kmp_get_gtid() : -1;
1681 #endif
1682 
1683   KF_TRACE(30, ("__kmp_resume_template: T#%d wants to wakeup T#%d enter\n",
1684                 gtid, target_gtid));
1685   KMP_DEBUG_ASSERT(gtid != target_gtid);
1686 
1687   __kmp_suspend_initialize_thread(th);
1688 
1689   __kmp_lock_suspend_mx(th);
1690 
1691   if (!flag || flag != th->th.th_sleep_loc) {
1692     // coming from __kmp_null_resume_wrapper, or thread is now sleeping on a
1693     // different location; wake up at new location
1694     flag = (C *)CCAST(void *, th->th.th_sleep_loc);
1695   }
1696 
1697   // First, check if the flag is null or its type has changed. If so, someone
1698   // else woke it up.
1699   if (!flag) { // Thread doesn't appear to be sleeping on anything
1700     KF_TRACE(5, ("__kmp_resume_template: T#%d exiting, thread T#%d already "
1701                  "awake: flag(%p)\n",
1702                  gtid, target_gtid, (void *)NULL));
1703     __kmp_unlock_suspend_mx(th);
1704     return;
1705   } else if (flag->get_type() != th->th.th_sleep_loc_type) {
1706     // Flag type does not appear to match this function template; possibly the
1707     // thread is sleeping on something else. Try null resume again.
1708     KF_TRACE(
1709         5,
1710         ("__kmp_resume_template: T#%d retrying, thread T#%d Mismatch flag(%p), "
1711          "spin(%p) type=%d ptr_type=%d\n",
1712          gtid, target_gtid, flag, flag->get(), flag->get_type(),
1713          th->th.th_sleep_loc_type));
1714     __kmp_unlock_suspend_mx(th);
1715     __kmp_null_resume_wrapper(th);
1716     return;
1717   } else { // if multiple threads are sleeping, flag should be internally
1718     // referring to a specific thread here
1719     if (!flag->is_sleeping()) {
1720       KF_TRACE(5, ("__kmp_resume_template: T#%d exiting, thread T#%d already "
1721                    "awake: flag(%p): %u\n",
1722                    gtid, target_gtid, flag->get(), (unsigned int)flag->load()));
1723       __kmp_unlock_suspend_mx(th);
1724       return;
1725     }
1726   }
1727   KMP_DEBUG_ASSERT(flag);
1728   flag->unset_sleeping();
1729   TCW_PTR(th->th.th_sleep_loc, NULL);
1730   th->th.th_sleep_loc_type = flag_unset;
1731 
1732   KF_TRACE(5, ("__kmp_resume_template: T#%d about to wakeup T#%d, reset "
1733                "sleep bit for flag's loc(%p): %u\n",
1734                gtid, target_gtid, flag->get(), (unsigned int)flag->load()));
1735 
1736 #ifdef DEBUG_SUSPEND
1737   {
1738     char buffer[128];
1739     __kmp_print_cond(buffer, &th->th.th_suspend_cv);
1740     __kmp_printf("__kmp_resume_template: T#%d resuming T#%d: %s\n", gtid,
1741                  target_gtid, buffer);
1742   }
1743 #endif
1744   status = pthread_cond_signal(&th->th.th_suspend_cv.c_cond);
1745   KMP_CHECK_SYSFAIL("pthread_cond_signal", status);
1746   __kmp_unlock_suspend_mx(th);
1747   KF_TRACE(30, ("__kmp_resume_template: T#%d exiting after signaling wake up"
1748                 " for T#%d\n",
1749                 gtid, target_gtid));
1750 }
1751 
1752 template <bool C, bool S>
1753 void __kmp_resume_32(int target_gtid, kmp_flag_32<C, S> *flag) {
1754   __kmp_resume_template(target_gtid, flag);
1755 }
1756 template <bool C, bool S>
1757 void __kmp_resume_64(int target_gtid, kmp_flag_64<C, S> *flag) {
1758   __kmp_resume_template(target_gtid, flag);
1759 }
1760 template <bool C, bool S>
1761 void __kmp_atomic_resume_64(int target_gtid, kmp_atomic_flag_64<C, S> *flag) {
1762   __kmp_resume_template(target_gtid, flag);
1763 }
1764 void __kmp_resume_oncore(int target_gtid, kmp_flag_oncore *flag) {
1765   __kmp_resume_template(target_gtid, flag);
1766 }
1767 
1768 template void __kmp_resume_32<false, true>(int, kmp_flag_32<false, true> *);
1769 template void __kmp_resume_32<false, false>(int, kmp_flag_32<false, false> *);
1770 template void __kmp_resume_64<false, true>(int, kmp_flag_64<false, true> *);
1771 template void
1772 __kmp_atomic_resume_64<false, true>(int, kmp_atomic_flag_64<false, true> *);
1773 
1774 #if KMP_USE_MONITOR
1775 void __kmp_resume_monitor() {
1776   KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(USER_resume);
1777   int status;
1778 #ifdef KMP_DEBUG
1779   int gtid = TCR_4(__kmp_init_gtid) ? __kmp_get_gtid() : -1;
1780   KF_TRACE(30, ("__kmp_resume_monitor: T#%d wants to wakeup T#%d enter\n", gtid,
1781                 KMP_GTID_MONITOR));
1782   KMP_DEBUG_ASSERT(gtid != KMP_GTID_MONITOR);
1783 #endif
1784   status = pthread_mutex_lock(&__kmp_wait_mx.m_mutex);
1785   KMP_CHECK_SYSFAIL("pthread_mutex_lock", status);
1786 #ifdef DEBUG_SUSPEND
1787   {
1788     char buffer[128];
1789     __kmp_print_cond(buffer, &__kmp_wait_cv.c_cond);
1790     __kmp_printf("__kmp_resume_monitor: T#%d resuming T#%d: %s\n", gtid,
1791                  KMP_GTID_MONITOR, buffer);
1792   }
1793 #endif
1794   status = pthread_cond_signal(&__kmp_wait_cv.c_cond);
1795   KMP_CHECK_SYSFAIL("pthread_cond_signal", status);
1796   status = pthread_mutex_unlock(&__kmp_wait_mx.m_mutex);
1797   KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status);
1798   KF_TRACE(30, ("__kmp_resume_monitor: T#%d exiting after signaling wake up"
1799                 " for T#%d\n",
1800                 gtid, KMP_GTID_MONITOR));
1801 }
1802 #endif // KMP_USE_MONITOR
1803 
1804 void __kmp_yield() { sched_yield(); }
1805 
1806 void __kmp_gtid_set_specific(int gtid) {
1807   if (__kmp_init_gtid) {
1808     int status;
1809     status = pthread_setspecific(__kmp_gtid_threadprivate_key,
1810                                  (void *)(intptr_t)(gtid + 1));
1811     KMP_CHECK_SYSFAIL("pthread_setspecific", status);
1812   } else {
1813     KA_TRACE(50, ("__kmp_gtid_set_specific: runtime shutdown, returning\n"));
1814   }
1815 }
1816 
1817 int __kmp_gtid_get_specific() {
1818   int gtid;
1819   if (!__kmp_init_gtid) {
1820     KA_TRACE(50, ("__kmp_gtid_get_specific: runtime shutdown, returning "
1821                   "KMP_GTID_SHUTDOWN\n"));
1822     return KMP_GTID_SHUTDOWN;
1823   }
1824   gtid = (int)(size_t)pthread_getspecific(__kmp_gtid_threadprivate_key);
1825   if (gtid == 0) {
1826     gtid = KMP_GTID_DNE;
1827   } else {
1828     gtid--;
1829   }
1830   KA_TRACE(50, ("__kmp_gtid_get_specific: key:%d gtid:%d\n",
1831                 __kmp_gtid_threadprivate_key, gtid));
1832   return gtid;
1833 }
1834 
1835 double __kmp_read_cpu_time(void) {
1836   /*clock_t   t;*/
1837   struct tms buffer;
1838 
1839   /*t =*/times(&buffer);
1840 
1841   return (double)(buffer.tms_utime + buffer.tms_cutime) /
1842          (double)CLOCKS_PER_SEC;
1843 }
1844 
1845 int __kmp_read_system_info(struct kmp_sys_info *info) {
1846   int status;
1847   struct rusage r_usage;
1848 
1849   memset(info, 0, sizeof(*info));
1850 
1851   status = getrusage(RUSAGE_SELF, &r_usage);
1852   KMP_CHECK_SYSFAIL_ERRNO("getrusage", status);
1853 
1854 #if !KMP_OS_WASI
1855   // The maximum resident set size utilized (in kilobytes)
1856   info->maxrss = r_usage.ru_maxrss;
1857   // The number of page faults serviced without any I/O
1858   info->minflt = r_usage.ru_minflt;
1859   // The number of page faults serviced that required I/O
1860   info->majflt = r_usage.ru_majflt;
1861   // The number of times a process was "swapped" out of memory
1862   info->nswap = r_usage.ru_nswap;
1863   // The number of times the file system had to perform input
1864   info->inblock = r_usage.ru_inblock;
1865   // The number of times the file system had to perform output
1866   info->oublock = r_usage.ru_oublock;
1867   // The number of times a context switch was voluntarily
1868   info->nvcsw = r_usage.ru_nvcsw;
1869   // The number of times a context switch was forced
1870   info->nivcsw = r_usage.ru_nivcsw;
1871 #endif
1872 
1873   return (status != 0);
1874 }
1875 
1876 void __kmp_read_system_time(double *delta) {
1877   double t_ns;
1878   struct timeval tval;
1879   struct timespec stop;
1880   int status;
1881 
1882   status = gettimeofday(&tval, NULL);
1883   KMP_CHECK_SYSFAIL_ERRNO("gettimeofday", status);
1884   TIMEVAL_TO_TIMESPEC(&tval, &stop);
1885   t_ns = (double)(TS2NS(stop) - TS2NS(__kmp_sys_timer_data.start));
1886   *delta = (t_ns * 1e-9);
1887 }
1888 
1889 void __kmp_clear_system_time(void) {
1890   struct timeval tval;
1891   int status;
1892   status = gettimeofday(&tval, NULL);
1893   KMP_CHECK_SYSFAIL_ERRNO("gettimeofday", status);
1894   TIMEVAL_TO_TIMESPEC(&tval, &__kmp_sys_timer_data.start);
1895 }
1896 
1897 static int __kmp_get_xproc(void) {
1898 
1899   int r = 0;
1900 
1901 #if KMP_OS_LINUX
1902 
1903   __kmp_type_convert(sysconf(_SC_NPROCESSORS_CONF), &(r));
1904 
1905 #elif KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD || KMP_OS_OPENBSD || \
1906     KMP_OS_HAIKU || KMP_OS_HURD || KMP_OS_SOLARIS || KMP_OS_WASI || KMP_OS_AIX
1907 
1908   __kmp_type_convert(sysconf(_SC_NPROCESSORS_ONLN), &(r));
1909 
1910 #elif KMP_OS_DARWIN
1911 
1912   size_t len = sizeof(r);
1913   sysctlbyname("hw.logicalcpu", &r, &len, NULL, 0);
1914 
1915 #else
1916 
1917 #error "Unknown or unsupported OS."
1918 
1919 #endif
1920 
1921   return r > 0 ? r : 2; /* guess value of 2 if OS told us 0 */
1922 
1923 } // __kmp_get_xproc
1924 
1925 int __kmp_read_from_file(char const *path, char const *format, ...) {
1926   int result;
1927   va_list args;
1928 
1929   va_start(args, format);
1930   FILE *f = fopen(path, "rb");
1931   if (f == NULL) {
1932     va_end(args);
1933     return 0;
1934   }
1935   result = vfscanf(f, format, args);
1936   fclose(f);
1937   va_end(args);
1938 
1939   return result;
1940 }
1941 
1942 void __kmp_runtime_initialize(void) {
1943   int status;
1944   pthread_mutexattr_t mutex_attr;
1945   pthread_condattr_t cond_attr;
1946 
1947   if (__kmp_init_runtime) {
1948     return;
1949   }
1950 
1951 #if (KMP_ARCH_X86 || KMP_ARCH_X86_64)
1952   if (!__kmp_cpuinfo.initialized) {
1953     __kmp_query_cpuid(&__kmp_cpuinfo);
1954   }
1955 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
1956 
1957   __kmp_xproc = __kmp_get_xproc();
1958 
1959 #if !KMP_32_BIT_ARCH
1960   struct rlimit rlim;
1961   // read stack size of calling thread, save it as default for worker threads;
1962   // this should be done before reading environment variables
1963   status = getrlimit(RLIMIT_STACK, &rlim);
1964   if (status == 0) { // success?
1965     __kmp_stksize = rlim.rlim_cur;
1966     __kmp_check_stksize(&__kmp_stksize); // check value and adjust if needed
1967   }
1968 #endif /* KMP_32_BIT_ARCH */
1969 
1970   if (sysconf(_SC_THREADS)) {
1971 
1972     /* Query the maximum number of threads */
1973     __kmp_type_convert(sysconf(_SC_THREAD_THREADS_MAX), &(__kmp_sys_max_nth));
1974 #ifdef __ve__
1975     if (__kmp_sys_max_nth == -1) {
1976       // VE's pthread supports only up to 64 threads per a VE process.
1977       // So we use that KMP_MAX_NTH (predefined as 64) here.
1978       __kmp_sys_max_nth = KMP_MAX_NTH;
1979     }
1980 #else
1981     if (__kmp_sys_max_nth == -1) {
1982       /* Unlimited threads for NPTL */
1983       __kmp_sys_max_nth = INT_MAX;
1984     } else if (__kmp_sys_max_nth <= 1) {
1985       /* Can't tell, just use PTHREAD_THREADS_MAX */
1986       __kmp_sys_max_nth = KMP_MAX_NTH;
1987     }
1988 #endif
1989 
1990     /* Query the minimum stack size */
1991     __kmp_sys_min_stksize = sysconf(_SC_THREAD_STACK_MIN);
1992     if (__kmp_sys_min_stksize <= 1) {
1993       __kmp_sys_min_stksize = KMP_MIN_STKSIZE;
1994     }
1995   }
1996 
1997   /* Set up minimum number of threads to switch to TLS gtid */
1998   __kmp_tls_gtid_min = KMP_TLS_GTID_MIN;
1999 
2000   status = pthread_key_create(&__kmp_gtid_threadprivate_key,
2001                               __kmp_internal_end_dest);
2002   KMP_CHECK_SYSFAIL("pthread_key_create", status);
2003   status = pthread_mutexattr_init(&mutex_attr);
2004   KMP_CHECK_SYSFAIL("pthread_mutexattr_init", status);
2005   status = pthread_mutex_init(&__kmp_wait_mx.m_mutex, &mutex_attr);
2006   KMP_CHECK_SYSFAIL("pthread_mutex_init", status);
2007   status = pthread_mutexattr_destroy(&mutex_attr);
2008   KMP_CHECK_SYSFAIL("pthread_mutexattr_destroy", status);
2009   status = pthread_condattr_init(&cond_attr);
2010   KMP_CHECK_SYSFAIL("pthread_condattr_init", status);
2011   status = pthread_cond_init(&__kmp_wait_cv.c_cond, &cond_attr);
2012   KMP_CHECK_SYSFAIL("pthread_cond_init", status);
2013   status = pthread_condattr_destroy(&cond_attr);
2014   KMP_CHECK_SYSFAIL("pthread_condattr_destroy", status);
2015 #if USE_ITT_BUILD
2016   __kmp_itt_initialize();
2017 #endif /* USE_ITT_BUILD */
2018 
2019   __kmp_init_runtime = TRUE;
2020 }
2021 
2022 void __kmp_runtime_destroy(void) {
2023   int status;
2024 
2025   if (!__kmp_init_runtime) {
2026     return; // Nothing to do.
2027   }
2028 
2029 #if USE_ITT_BUILD
2030   __kmp_itt_destroy();
2031 #endif /* USE_ITT_BUILD */
2032 
2033   status = pthread_key_delete(__kmp_gtid_threadprivate_key);
2034   KMP_CHECK_SYSFAIL("pthread_key_delete", status);
2035 
2036   status = pthread_mutex_destroy(&__kmp_wait_mx.m_mutex);
2037   if (status != 0 && status != EBUSY) {
2038     KMP_SYSFAIL("pthread_mutex_destroy", status);
2039   }
2040   status = pthread_cond_destroy(&__kmp_wait_cv.c_cond);
2041   if (status != 0 && status != EBUSY) {
2042     KMP_SYSFAIL("pthread_cond_destroy", status);
2043   }
2044 #if KMP_AFFINITY_SUPPORTED
2045   __kmp_affinity_uninitialize();
2046 #endif
2047 
2048   __kmp_init_runtime = FALSE;
2049 }
2050 
2051 /* Put the thread to sleep for a time period */
2052 /* NOTE: not currently used anywhere */
2053 void __kmp_thread_sleep(int millis) { sleep((millis + 500) / 1000); }
2054 
2055 /* Calculate the elapsed wall clock time for the user */
2056 void __kmp_elapsed(double *t) {
2057   int status;
2058 #ifdef FIX_SGI_CLOCK
2059   struct timespec ts;
2060 
2061   status = clock_gettime(CLOCK_PROCESS_CPUTIME_ID, &ts);
2062   KMP_CHECK_SYSFAIL_ERRNO("clock_gettime", status);
2063   *t =
2064       (double)ts.tv_nsec * (1.0 / (double)KMP_NSEC_PER_SEC) + (double)ts.tv_sec;
2065 #else
2066   struct timeval tv;
2067 
2068   status = gettimeofday(&tv, NULL);
2069   KMP_CHECK_SYSFAIL_ERRNO("gettimeofday", status);
2070   *t =
2071       (double)tv.tv_usec * (1.0 / (double)KMP_USEC_PER_SEC) + (double)tv.tv_sec;
2072 #endif
2073 }
2074 
2075 /* Calculate the elapsed wall clock tick for the user */
2076 void __kmp_elapsed_tick(double *t) { *t = 1 / (double)CLOCKS_PER_SEC; }
2077 
2078 /* Return the current time stamp in nsec */
2079 kmp_uint64 __kmp_now_nsec() {
2080   struct timeval t;
2081   gettimeofday(&t, NULL);
2082   kmp_uint64 nsec = (kmp_uint64)KMP_NSEC_PER_SEC * (kmp_uint64)t.tv_sec +
2083                     (kmp_uint64)1000 * (kmp_uint64)t.tv_usec;
2084   return nsec;
2085 }
2086 
2087 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
2088 /* Measure clock ticks per millisecond */
2089 void __kmp_initialize_system_tick() {
2090   kmp_uint64 now, nsec2, diff;
2091   kmp_uint64 delay = 1000000; // ~450 usec on most machines.
2092   kmp_uint64 nsec = __kmp_now_nsec();
2093   kmp_uint64 goal = __kmp_hardware_timestamp() + delay;
2094   while ((now = __kmp_hardware_timestamp()) < goal)
2095     ;
2096   nsec2 = __kmp_now_nsec();
2097   diff = nsec2 - nsec;
2098   if (diff > 0) {
2099     double tpus = 1000.0 * (double)(delay + (now - goal)) / (double)diff;
2100     if (tpus > 0.0) {
2101       __kmp_ticks_per_msec = (kmp_uint64)(tpus * 1000.0);
2102       __kmp_ticks_per_usec = (kmp_uint64)tpus;
2103     }
2104   }
2105 }
2106 #endif
2107 
2108 /* Determine whether the given address is mapped into the current address
2109    space. */
2110 
2111 int __kmp_is_address_mapped(void *addr) {
2112 
2113   int found = 0;
2114   int rc;
2115 
2116 #if KMP_OS_LINUX || KMP_OS_HURD
2117 
2118   /* On GNUish OSes, read the /proc/<pid>/maps pseudo-file to get all the
2119      address ranges mapped into the address space. */
2120 
2121   char *name = __kmp_str_format("/proc/%d/maps", getpid());
2122   FILE *file = NULL;
2123 
2124   file = fopen(name, "r");
2125   KMP_ASSERT(file != NULL);
2126 
2127   for (;;) {
2128 
2129     void *beginning = NULL;
2130     void *ending = NULL;
2131     char perms[5];
2132 
2133     rc = fscanf(file, "%p-%p %4s %*[^\n]\n", &beginning, &ending, perms);
2134     if (rc == EOF) {
2135       break;
2136     }
2137     KMP_ASSERT(rc == 3 &&
2138                KMP_STRLEN(perms) == 4); // Make sure all fields are read.
2139 
2140     // Ending address is not included in the region, but beginning is.
2141     if ((addr >= beginning) && (addr < ending)) {
2142       perms[2] = 0; // 3th and 4th character does not matter.
2143       if (strcmp(perms, "rw") == 0) {
2144         // Memory we are looking for should be readable and writable.
2145         found = 1;
2146       }
2147       break;
2148     }
2149   }
2150 
2151   // Free resources.
2152   fclose(file);
2153   KMP_INTERNAL_FREE(name);
2154 #elif KMP_OS_FREEBSD
2155   char *buf;
2156   size_t lstsz;
2157   int mib[] = {CTL_KERN, KERN_PROC, KERN_PROC_VMMAP, getpid()};
2158   rc = sysctl(mib, 4, NULL, &lstsz, NULL, 0);
2159   if (rc < 0)
2160     return 0;
2161   // We pass from number of vm entry's semantic
2162   // to size of whole entry map list.
2163   lstsz = lstsz * 4 / 3;
2164   buf = reinterpret_cast<char *>(KMP_INTERNAL_MALLOC(lstsz));
2165   rc = sysctl(mib, 4, buf, &lstsz, NULL, 0);
2166   if (rc < 0) {
2167     KMP_INTERNAL_FREE(buf);
2168     return 0;
2169   }
2170 
2171   char *lw = buf;
2172   char *up = buf + lstsz;
2173 
2174   while (lw < up) {
2175     struct kinfo_vmentry *cur = reinterpret_cast<struct kinfo_vmentry *>(lw);
2176     size_t cursz = cur->kve_structsize;
2177     if (cursz == 0)
2178       break;
2179     void *start = reinterpret_cast<void *>(cur->kve_start);
2180     void *end = reinterpret_cast<void *>(cur->kve_end);
2181     // Readable/Writable addresses within current map entry
2182     if ((addr >= start) && (addr < end)) {
2183       if ((cur->kve_protection & KVME_PROT_READ) != 0 &&
2184           (cur->kve_protection & KVME_PROT_WRITE) != 0) {
2185         found = 1;
2186         break;
2187       }
2188     }
2189     lw += cursz;
2190   }
2191   KMP_INTERNAL_FREE(buf);
2192 #elif KMP_OS_DRAGONFLY
2193   char err[_POSIX2_LINE_MAX];
2194   kinfo_proc *proc;
2195   vmspace sp;
2196   vm_map *cur;
2197   vm_map_entry entry, *c;
2198   struct proc p;
2199   kvm_t *fd;
2200   uintptr_t uaddr;
2201   int num;
2202 
2203   fd = kvm_openfiles(nullptr, nullptr, nullptr, O_RDONLY, err);
2204   if (!fd) {
2205     return 0;
2206   }
2207 
2208   proc = kvm_getprocs(fd, KERN_PROC_PID, getpid(), &num);
2209 
2210   if (kvm_read(fd, static_cast<uintptr_t>(proc->kp_paddr), &p, sizeof(p)) !=
2211           sizeof(p) ||
2212       kvm_read(fd, reinterpret_cast<uintptr_t>(p.p_vmspace), &sp, sizeof(sp)) !=
2213           sizeof(sp)) {
2214     kvm_close(fd);
2215     return 0;
2216   }
2217 
2218   (void)rc;
2219   cur = &sp.vm_map;
2220   uaddr = reinterpret_cast<uintptr_t>(addr);
2221   for (c = kvm_vm_map_entry_first(fd, cur, &entry); c;
2222        c = kvm_vm_map_entry_next(fd, c, &entry)) {
2223     if ((uaddr >= entry.ba.start) && (uaddr <= entry.ba.end)) {
2224       if ((entry.protection & VM_PROT_READ) != 0 &&
2225           (entry.protection & VM_PROT_WRITE) != 0) {
2226         found = 1;
2227         break;
2228       }
2229     }
2230   }
2231 
2232   kvm_close(fd);
2233 #elif KMP_OS_SOLARIS
2234   prxmap_t *cur, *map;
2235   void *buf;
2236   uintptr_t uaddr;
2237   ssize_t rd;
2238   int fd;
2239   pid_t pid = getpid();
2240   char *name = __kmp_str_format("/proc/%d/xmap", pid);
2241   fd = open(name, O_RDONLY);
2242   if (fd == -1) {
2243     KMP_INTERNAL_FREE(name);
2244     return 0;
2245   }
2246 
2247   size_t sz = (1 << 20);
2248   buf = KMP_INTERNAL_MALLOC(sz);
2249 
2250   while (sz > 0 && (rd = pread(fd, buf, sz, 0)) == sz) {
2251     void *newbuf;
2252     sz <<= 1;
2253     newbuf = KMP_INTERNAL_REALLOC(buf, sz);
2254     buf = newbuf;
2255   }
2256 
2257   map = reinterpret_cast<prxmap_t *>(buf);
2258   uaddr = reinterpret_cast<uintptr_t>(addr);
2259 
2260   for (cur = map; rd > 0; cur++, rd = -sizeof(*map)) {
2261     if (uaddr >= cur->pr_vaddr && uaddr < cur->pr_vaddr) {
2262       if ((cur->pr_mflags & MA_READ) != 0 && (cur->pr_mflags & MA_WRITE) != 0) {
2263         found = 1;
2264         break;
2265       }
2266     }
2267   }
2268 
2269   KMP_INTERNAL_FREE(map);
2270   close(fd);
2271   KMP_INTERNAL_FREE(name);
2272 #elif KMP_OS_DARWIN
2273 
2274   /* On OS X*, /proc pseudo filesystem is not available. Try to read memory
2275      using vm interface. */
2276 
2277   int buffer;
2278   vm_size_t count;
2279   rc = vm_read_overwrite(
2280       mach_task_self(), // Task to read memory of.
2281       (vm_address_t)(addr), // Address to read from.
2282       1, // Number of bytes to be read.
2283       (vm_address_t)(&buffer), // Address of buffer to save read bytes in.
2284       &count // Address of var to save number of read bytes in.
2285   );
2286   if (rc == 0) {
2287     // Memory successfully read.
2288     found = 1;
2289   }
2290 
2291 #elif KMP_OS_NETBSD
2292 
2293   int mib[5];
2294   mib[0] = CTL_VM;
2295   mib[1] = VM_PROC;
2296   mib[2] = VM_PROC_MAP;
2297   mib[3] = getpid();
2298   mib[4] = sizeof(struct kinfo_vmentry);
2299 
2300   size_t size;
2301   rc = sysctl(mib, __arraycount(mib), NULL, &size, NULL, 0);
2302   KMP_ASSERT(!rc);
2303   KMP_ASSERT(size);
2304 
2305   size = size * 4 / 3;
2306   struct kinfo_vmentry *kiv = (struct kinfo_vmentry *)KMP_INTERNAL_MALLOC(size);
2307   KMP_ASSERT(kiv);
2308 
2309   rc = sysctl(mib, __arraycount(mib), kiv, &size, NULL, 0);
2310   KMP_ASSERT(!rc);
2311   KMP_ASSERT(size);
2312 
2313   for (size_t i = 0; i < size; i++) {
2314     if (kiv[i].kve_start >= (uint64_t)addr &&
2315         kiv[i].kve_end <= (uint64_t)addr) {
2316       found = 1;
2317       break;
2318     }
2319   }
2320   KMP_INTERNAL_FREE(kiv);
2321 #elif KMP_OS_OPENBSD
2322 
2323   int mib[3];
2324   mib[0] = CTL_KERN;
2325   mib[1] = KERN_PROC_VMMAP;
2326   mib[2] = getpid();
2327 
2328   size_t size;
2329   uint64_t end;
2330   rc = sysctl(mib, 3, NULL, &size, NULL, 0);
2331   KMP_ASSERT(!rc);
2332   KMP_ASSERT(size);
2333   end = size;
2334 
2335   struct kinfo_vmentry kiv = {.kve_start = 0};
2336 
2337   while ((rc = sysctl(mib, 3, &kiv, &size, NULL, 0)) == 0) {
2338     KMP_ASSERT(size);
2339     if (kiv.kve_end == end)
2340       break;
2341 
2342     if (kiv.kve_start >= (uint64_t)addr && kiv.kve_end <= (uint64_t)addr) {
2343       found = 1;
2344       break;
2345     }
2346     kiv.kve_start += 1;
2347   }
2348 #elif KMP_OS_WASI
2349   found = (int)addr < (__builtin_wasm_memory_size(0) * PAGESIZE);
2350 #elif KMP_OS_AIX
2351 
2352   uint32_t loadQueryBufSize = 4096u; // Default loadquery buffer size.
2353   char *loadQueryBuf;
2354 
2355   for (;;) {
2356     loadQueryBuf = (char *)KMP_INTERNAL_MALLOC(loadQueryBufSize);
2357     if (loadQueryBuf == NULL) {
2358       return 0;
2359     }
2360 
2361     rc = loadquery(L_GETXINFO | L_IGNOREUNLOAD, loadQueryBuf, loadQueryBufSize);
2362     if (rc < 0) {
2363       KMP_INTERNAL_FREE(loadQueryBuf);
2364       if (errno != ENOMEM) {
2365         return 0;
2366       }
2367       // errno == ENOMEM; double the size.
2368       loadQueryBufSize <<= 1;
2369       continue;
2370     }
2371     // Obtained the load info successfully.
2372     break;
2373   }
2374 
2375   struct ld_xinfo *curLdInfo = (struct ld_xinfo *)loadQueryBuf;
2376 
2377   // Loop through the load info to find if there is a match.
2378   for (;;) {
2379     uintptr_t curDataStart = (uintptr_t)curLdInfo->ldinfo_dataorg;
2380     uintptr_t curDataEnd = curDataStart + curLdInfo->ldinfo_datasize;
2381 
2382     // The data segment is readable and writable.
2383     if (curDataStart <= (uintptr_t)addr && (uintptr_t)addr < curDataEnd) {
2384       found = 1;
2385       break;
2386     }
2387     if (curLdInfo->ldinfo_next == 0u) {
2388       // Reached the end of load info.
2389       break;
2390     }
2391     curLdInfo = (struct ld_xinfo *)((char *)curLdInfo + curLdInfo->ldinfo_next);
2392   }
2393   KMP_INTERNAL_FREE(loadQueryBuf);
2394 
2395 #elif KMP_OS_HAIKU
2396 
2397   found = 1;
2398 #else
2399 
2400 #error "Unknown or unsupported OS"
2401 
2402 #endif
2403 
2404   return found;
2405 
2406 } // __kmp_is_address_mapped
2407 
2408 #ifdef USE_LOAD_BALANCE
2409 
2410 #if KMP_OS_DARWIN || KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD ||    \
2411     KMP_OS_OPENBSD || KMP_OS_SOLARIS
2412 
2413 // The function returns the rounded value of the system load average
2414 // during given time interval which depends on the value of
2415 // __kmp_load_balance_interval variable (default is 60 sec, other values
2416 // may be 300 sec or 900 sec).
2417 // It returns -1 in case of error.
2418 int __kmp_get_load_balance(int max) {
2419   double averages[3];
2420   int ret_avg = 0;
2421 
2422   int res = getloadavg(averages, 3);
2423 
2424   // Check __kmp_load_balance_interval to determine which of averages to use.
2425   // getloadavg() may return the number of samples less than requested that is
2426   // less than 3.
2427   if (__kmp_load_balance_interval < 180 && (res >= 1)) {
2428     ret_avg = (int)averages[0]; // 1 min
2429   } else if ((__kmp_load_balance_interval >= 180 &&
2430               __kmp_load_balance_interval < 600) &&
2431              (res >= 2)) {
2432     ret_avg = (int)averages[1]; // 5 min
2433   } else if ((__kmp_load_balance_interval >= 600) && (res == 3)) {
2434     ret_avg = (int)averages[2]; // 15 min
2435   } else { // Error occurred
2436     return -1;
2437   }
2438 
2439   return ret_avg;
2440 }
2441 
2442 #elif KMP_OS_AIX
2443 
2444 // The function returns number of running (not sleeping) threads, or -1 in case
2445 // of error.
2446 int __kmp_get_load_balance(int max) {
2447 
2448   static int glb_running_threads = 0; // Saved count of the running threads for
2449                                       // the thread balance algorithm.
2450   static double glb_call_time = 0; // Thread balance algorithm call time.
2451   int running_threads = 0; // Number of running threads in the system.
2452 
2453   double call_time = 0.0;
2454 
2455   __kmp_elapsed(&call_time);
2456 
2457   if (glb_call_time &&
2458       (call_time - glb_call_time < __kmp_load_balance_interval))
2459     return glb_running_threads;
2460 
2461   glb_call_time = call_time;
2462 
2463   if (max <= 0) {
2464     max = INT_MAX;
2465   }
2466 
2467   // Check how many perfstat_cpu_t structures are available.
2468   int logical_cpus = perfstat_cpu(NULL, NULL, sizeof(perfstat_cpu_t), 0);
2469   if (logical_cpus <= 0) {
2470     glb_call_time = -1;
2471     return -1;
2472   }
2473 
2474   perfstat_cpu_t *cpu_stat = (perfstat_cpu_t *)KMP_INTERNAL_MALLOC(
2475       logical_cpus * sizeof(perfstat_cpu_t));
2476   if (cpu_stat == NULL) {
2477     glb_call_time = -1;
2478     return -1;
2479   }
2480 
2481   // Set first CPU as the name of the first logical CPU for which the info is
2482   // desired.
2483   perfstat_id_t first_cpu_name;
2484   strcpy(first_cpu_name.name, FIRST_CPU);
2485 
2486   // Get the stat info of logical CPUs.
2487   int rc = perfstat_cpu(&first_cpu_name, cpu_stat, sizeof(perfstat_cpu_t),
2488                         logical_cpus);
2489   KMP_DEBUG_ASSERT(rc == logical_cpus);
2490   if (rc <= 0) {
2491     KMP_INTERNAL_FREE(cpu_stat);
2492     glb_call_time = -1;
2493     return -1;
2494   }
2495   for (int i = 0; i < logical_cpus; ++i) {
2496     running_threads += cpu_stat[i].runque;
2497     if (running_threads >= max)
2498       break;
2499   }
2500 
2501   // There _might_ be a timing hole where the thread executing this
2502   // code gets skipped in the load balance, and running_threads is 0.
2503   // Assert in the debug builds only!!!
2504   KMP_DEBUG_ASSERT(running_threads > 0);
2505   if (running_threads <= 0)
2506     running_threads = 1;
2507 
2508   KMP_INTERNAL_FREE(cpu_stat);
2509 
2510   glb_running_threads = running_threads;
2511 
2512   return running_threads;
2513 }
2514 
2515 #else // Linux* OS
2516 
2517 // The function returns number of running (not sleeping) threads, or -1 in case
2518 // of error. Error could be reported if Linux* OS kernel too old (without
2519 // "/proc" support). Counting running threads stops if max running threads
2520 // encountered.
2521 int __kmp_get_load_balance(int max) {
2522   static int permanent_error = 0;
2523   static int glb_running_threads = 0; // Saved count of the running threads for
2524   // the thread balance algorithm
2525   static double glb_call_time = 0; /* Thread balance algorithm call time */
2526 
2527   int running_threads = 0; // Number of running threads in the system.
2528 
2529   DIR *proc_dir = NULL; // Handle of "/proc/" directory.
2530   struct dirent *proc_entry = NULL;
2531 
2532   kmp_str_buf_t task_path; // "/proc/<pid>/task/<tid>/" path.
2533   DIR *task_dir = NULL; // Handle of "/proc/<pid>/task/<tid>/" directory.
2534   struct dirent *task_entry = NULL;
2535   int task_path_fixed_len;
2536 
2537   kmp_str_buf_t stat_path; // "/proc/<pid>/task/<tid>/stat" path.
2538   int stat_file = -1;
2539   int stat_path_fixed_len;
2540 
2541 #ifdef KMP_DEBUG
2542   int total_processes = 0; // Total number of processes in system.
2543 #endif
2544 
2545   double call_time = 0.0;
2546 
2547   __kmp_str_buf_init(&task_path);
2548   __kmp_str_buf_init(&stat_path);
2549 
2550   __kmp_elapsed(&call_time);
2551 
2552   if (glb_call_time &&
2553       (call_time - glb_call_time < __kmp_load_balance_interval)) {
2554     running_threads = glb_running_threads;
2555     goto finish;
2556   }
2557 
2558   glb_call_time = call_time;
2559 
2560   // Do not spend time on scanning "/proc/" if we have a permanent error.
2561   if (permanent_error) {
2562     running_threads = -1;
2563     goto finish;
2564   }
2565 
2566   if (max <= 0) {
2567     max = INT_MAX;
2568   }
2569 
2570   // Open "/proc/" directory.
2571   proc_dir = opendir("/proc");
2572   if (proc_dir == NULL) {
2573     // Cannot open "/proc/". Probably the kernel does not support it. Return an
2574     // error now and in subsequent calls.
2575     running_threads = -1;
2576     permanent_error = 1;
2577     goto finish;
2578   }
2579 
2580   // Initialize fixed part of task_path. This part will not change.
2581   __kmp_str_buf_cat(&task_path, "/proc/", 6);
2582   task_path_fixed_len = task_path.used; // Remember number of used characters.
2583 
2584   proc_entry = readdir(proc_dir);
2585   while (proc_entry != NULL) {
2586     // Proc entry is a directory and name starts with a digit. Assume it is a
2587     // process' directory.
2588     if (proc_entry->d_type == DT_DIR && isdigit(proc_entry->d_name[0])) {
2589 
2590 #ifdef KMP_DEBUG
2591       ++total_processes;
2592 #endif
2593       // Make sure init process is the very first in "/proc", so we can replace
2594       // strcmp( proc_entry->d_name, "1" ) == 0 with simpler total_processes ==
2595       // 1. We are going to check that total_processes == 1 => d_name == "1" is
2596       // true (where "=>" is implication). Since C++ does not have => operator,
2597       // let us replace it with its equivalent: a => b == ! a || b.
2598       KMP_DEBUG_ASSERT(total_processes != 1 ||
2599                        strcmp(proc_entry->d_name, "1") == 0);
2600 
2601       // Construct task_path.
2602       task_path.used = task_path_fixed_len; // Reset task_path to "/proc/".
2603       __kmp_str_buf_cat(&task_path, proc_entry->d_name,
2604                         KMP_STRLEN(proc_entry->d_name));
2605       __kmp_str_buf_cat(&task_path, "/task", 5);
2606 
2607       task_dir = opendir(task_path.str);
2608       if (task_dir == NULL) {
2609         // Process can finish between reading "/proc/" directory entry and
2610         // opening process' "task/" directory. So, in general case we should not
2611         // complain, but have to skip this process and read the next one. But on
2612         // systems with no "task/" support we will spend lot of time to scan
2613         // "/proc/" tree again and again without any benefit. "init" process
2614         // (its pid is 1) should exist always, so, if we cannot open
2615         // "/proc/1/task/" directory, it means "task/" is not supported by
2616         // kernel. Report an error now and in the future.
2617         if (strcmp(proc_entry->d_name, "1") == 0) {
2618           running_threads = -1;
2619           permanent_error = 1;
2620           goto finish;
2621         }
2622       } else {
2623         // Construct fixed part of stat file path.
2624         __kmp_str_buf_clear(&stat_path);
2625         __kmp_str_buf_cat(&stat_path, task_path.str, task_path.used);
2626         __kmp_str_buf_cat(&stat_path, "/", 1);
2627         stat_path_fixed_len = stat_path.used;
2628 
2629         task_entry = readdir(task_dir);
2630         while (task_entry != NULL) {
2631           // It is a directory and name starts with a digit.
2632           if (proc_entry->d_type == DT_DIR && isdigit(task_entry->d_name[0])) {
2633 
2634             // Construct complete stat file path. Easiest way would be:
2635             //  __kmp_str_buf_print( & stat_path, "%s/%s/stat", task_path.str,
2636             //  task_entry->d_name );
2637             // but seriae of __kmp_str_buf_cat works a bit faster.
2638             stat_path.used =
2639                 stat_path_fixed_len; // Reset stat path to its fixed part.
2640             __kmp_str_buf_cat(&stat_path, task_entry->d_name,
2641                               KMP_STRLEN(task_entry->d_name));
2642             __kmp_str_buf_cat(&stat_path, "/stat", 5);
2643 
2644             // Note: Low-level API (open/read/close) is used. High-level API
2645             // (fopen/fclose)  works ~ 30 % slower.
2646             stat_file = open(stat_path.str, O_RDONLY);
2647             if (stat_file == -1) {
2648               // We cannot report an error because task (thread) can terminate
2649               // just before reading this file.
2650             } else {
2651               /* Content of "stat" file looks like:
2652                  24285 (program) S ...
2653 
2654                  It is a single line (if program name does not include funny
2655                  symbols). First number is a thread id, then name of executable
2656                  file name in paretheses, then state of the thread. We need just
2657                  thread state.
2658 
2659                  Good news: Length of program name is 15 characters max. Longer
2660                  names are truncated.
2661 
2662                  Thus, we need rather short buffer: 15 chars for program name +
2663                  2 parenthesis, + 3 spaces + ~7 digits of pid = 37.
2664 
2665                  Bad news: Program name may contain special symbols like space,
2666                  closing parenthesis, or even new line. This makes parsing
2667                  "stat" file not 100 % reliable. In case of fanny program names
2668                  parsing may fail (report incorrect thread state).
2669 
2670                  Parsing "status" file looks more promissing (due to different
2671                  file structure and escaping special symbols) but reading and
2672                  parsing of "status" file works slower.
2673                   -- ln
2674               */
2675               char buffer[65];
2676               ssize_t len;
2677               len = read(stat_file, buffer, sizeof(buffer) - 1);
2678               if (len >= 0) {
2679                 buffer[len] = 0;
2680                 // Using scanf:
2681                 //     sscanf( buffer, "%*d (%*s) %c ", & state );
2682                 // looks very nice, but searching for a closing parenthesis
2683                 // works a bit faster.
2684                 char *close_parent = strstr(buffer, ") ");
2685                 if (close_parent != NULL) {
2686                   char state = *(close_parent + 2);
2687                   if (state == 'R') {
2688                     ++running_threads;
2689                     if (running_threads >= max) {
2690                       goto finish;
2691                     }
2692                   }
2693                 }
2694               }
2695               close(stat_file);
2696               stat_file = -1;
2697             }
2698           }
2699           task_entry = readdir(task_dir);
2700         }
2701         closedir(task_dir);
2702         task_dir = NULL;
2703       }
2704     }
2705     proc_entry = readdir(proc_dir);
2706   }
2707 
2708   // There _might_ be a timing hole where the thread executing this
2709   // code get skipped in the load balance, and running_threads is 0.
2710   // Assert in the debug builds only!!!
2711   KMP_DEBUG_ASSERT(running_threads > 0);
2712   if (running_threads <= 0) {
2713     running_threads = 1;
2714   }
2715 
2716 finish: // Clean up and exit.
2717   if (proc_dir != NULL) {
2718     closedir(proc_dir);
2719   }
2720   __kmp_str_buf_free(&task_path);
2721   if (task_dir != NULL) {
2722     closedir(task_dir);
2723   }
2724   __kmp_str_buf_free(&stat_path);
2725   if (stat_file != -1) {
2726     close(stat_file);
2727   }
2728 
2729   glb_running_threads = running_threads;
2730 
2731   return running_threads;
2732 
2733 } // __kmp_get_load_balance
2734 
2735 #endif // KMP_OS_DARWIN
2736 
2737 #endif // USE_LOAD_BALANCE
2738 
2739 #if !(KMP_ARCH_X86 || KMP_ARCH_X86_64 || KMP_MIC ||                            \
2740       ((KMP_OS_LINUX || KMP_OS_DARWIN) && KMP_ARCH_AARCH64) ||                 \
2741       KMP_ARCH_PPC64 || KMP_ARCH_RISCV64 || KMP_ARCH_LOONGARCH64 ||            \
2742       KMP_ARCH_ARM || KMP_ARCH_VE || KMP_ARCH_S390X || KMP_ARCH_PPC_XCOFF ||   \
2743       KMP_ARCH_AARCH64_32)
2744 
2745 // Because WebAssembly will use `call_indirect` to invoke the microtask and
2746 // WebAssembly indirect calls check that the called signature is a precise
2747 // match, we need to cast each microtask function pointer back from `void *` to
2748 // its original type.
2749 typedef void (*microtask_t0)(int *, int *);
2750 typedef void (*microtask_t1)(int *, int *, void *);
2751 typedef void (*microtask_t2)(int *, int *, void *, void *);
2752 typedef void (*microtask_t3)(int *, int *, void *, void *, void *);
2753 typedef void (*microtask_t4)(int *, int *, void *, void *, void *, void *);
2754 typedef void (*microtask_t5)(int *, int *, void *, void *, void *, void *,
2755                              void *);
2756 typedef void (*microtask_t6)(int *, int *, void *, void *, void *, void *,
2757                              void *, void *);
2758 typedef void (*microtask_t7)(int *, int *, void *, void *, void *, void *,
2759                              void *, void *, void *);
2760 typedef void (*microtask_t8)(int *, int *, void *, void *, void *, void *,
2761                              void *, void *, void *, void *);
2762 typedef void (*microtask_t9)(int *, int *, void *, void *, void *, void *,
2763                              void *, void *, void *, void *, void *);
2764 typedef void (*microtask_t10)(int *, int *, void *, void *, void *, void *,
2765                               void *, void *, void *, void *, void *, void *);
2766 typedef void (*microtask_t11)(int *, int *, void *, void *, void *, void *,
2767                               void *, void *, void *, void *, void *, void *,
2768                               void *);
2769 typedef void (*microtask_t12)(int *, int *, void *, void *, void *, void *,
2770                               void *, void *, void *, void *, void *, void *,
2771                               void *, void *);
2772 typedef void (*microtask_t13)(int *, int *, void *, void *, void *, void *,
2773                               void *, void *, void *, void *, void *, void *,
2774                               void *, void *, void *);
2775 typedef void (*microtask_t14)(int *, int *, void *, void *, void *, void *,
2776                               void *, void *, void *, void *, void *, void *,
2777                               void *, void *, void *, void *);
2778 typedef void (*microtask_t15)(int *, int *, void *, void *, void *, void *,
2779                               void *, void *, void *, void *, void *, void *,
2780                               void *, void *, void *, void *, void *);
2781 
2782 // we really only need the case with 1 argument, because CLANG always build
2783 // a struct of pointers to shared variables referenced in the outlined function
2784 int __kmp_invoke_microtask(microtask_t pkfn, int gtid, int tid, int argc,
2785                            void *p_argv[]
2786 #if OMPT_SUPPORT
2787                            ,
2788                            void **exit_frame_ptr
2789 #endif
2790 ) {
2791 #if OMPT_SUPPORT
2792   *exit_frame_ptr = OMPT_GET_FRAME_ADDRESS(0);
2793 #endif
2794 
2795   switch (argc) {
2796   default:
2797     fprintf(stderr, "Too many args to microtask: %d!\n", argc);
2798     fflush(stderr);
2799     exit(-1);
2800   case 0:
2801     (*(microtask_t0)pkfn)(&gtid, &tid);
2802     break;
2803   case 1:
2804     (*(microtask_t1)pkfn)(&gtid, &tid, p_argv[0]);
2805     break;
2806   case 2:
2807     (*(microtask_t2)pkfn)(&gtid, &tid, p_argv[0], p_argv[1]);
2808     break;
2809   case 3:
2810     (*(microtask_t3)pkfn)(&gtid, &tid, p_argv[0], p_argv[1], p_argv[2]);
2811     break;
2812   case 4:
2813     (*(microtask_t4)pkfn)(&gtid, &tid, p_argv[0], p_argv[1], p_argv[2],
2814                           p_argv[3]);
2815     break;
2816   case 5:
2817     (*(microtask_t5)pkfn)(&gtid, &tid, p_argv[0], p_argv[1], p_argv[2],
2818                           p_argv[3], p_argv[4]);
2819     break;
2820   case 6:
2821     (*(microtask_t6)pkfn)(&gtid, &tid, p_argv[0], p_argv[1], p_argv[2],
2822                           p_argv[3], p_argv[4], p_argv[5]);
2823     break;
2824   case 7:
2825     (*(microtask_t7)pkfn)(&gtid, &tid, p_argv[0], p_argv[1], p_argv[2],
2826                           p_argv[3], p_argv[4], p_argv[5], p_argv[6]);
2827     break;
2828   case 8:
2829     (*(microtask_t8)pkfn)(&gtid, &tid, p_argv[0], p_argv[1], p_argv[2],
2830                           p_argv[3], p_argv[4], p_argv[5], p_argv[6],
2831                           p_argv[7]);
2832     break;
2833   case 9:
2834     (*(microtask_t9)pkfn)(&gtid, &tid, p_argv[0], p_argv[1], p_argv[2],
2835                           p_argv[3], p_argv[4], p_argv[5], p_argv[6], p_argv[7],
2836                           p_argv[8]);
2837     break;
2838   case 10:
2839     (*(microtask_t10)pkfn)(&gtid, &tid, p_argv[0], p_argv[1], p_argv[2],
2840                            p_argv[3], p_argv[4], p_argv[5], p_argv[6],
2841                            p_argv[7], p_argv[8], p_argv[9]);
2842     break;
2843   case 11:
2844     (*(microtask_t11)pkfn)(&gtid, &tid, p_argv[0], p_argv[1], p_argv[2],
2845                            p_argv[3], p_argv[4], p_argv[5], p_argv[6],
2846                            p_argv[7], p_argv[8], p_argv[9], p_argv[10]);
2847     break;
2848   case 12:
2849     (*(microtask_t12)pkfn)(&gtid, &tid, p_argv[0], p_argv[1], p_argv[2],
2850                            p_argv[3], p_argv[4], p_argv[5], p_argv[6],
2851                            p_argv[7], p_argv[8], p_argv[9], p_argv[10],
2852                            p_argv[11]);
2853     break;
2854   case 13:
2855     (*(microtask_t13)pkfn)(&gtid, &tid, p_argv[0], p_argv[1], p_argv[2],
2856                            p_argv[3], p_argv[4], p_argv[5], p_argv[6],
2857                            p_argv[7], p_argv[8], p_argv[9], p_argv[10],
2858                            p_argv[11], p_argv[12]);
2859     break;
2860   case 14:
2861     (*(microtask_t14)pkfn)(&gtid, &tid, p_argv[0], p_argv[1], p_argv[2],
2862                            p_argv[3], p_argv[4], p_argv[5], p_argv[6],
2863                            p_argv[7], p_argv[8], p_argv[9], p_argv[10],
2864                            p_argv[11], p_argv[12], p_argv[13]);
2865     break;
2866   case 15:
2867     (*(microtask_t15)pkfn)(&gtid, &tid, p_argv[0], p_argv[1], p_argv[2],
2868                            p_argv[3], p_argv[4], p_argv[5], p_argv[6],
2869                            p_argv[7], p_argv[8], p_argv[9], p_argv[10],
2870                            p_argv[11], p_argv[12], p_argv[13], p_argv[14]);
2871     break;
2872   }
2873 
2874   return 1;
2875 }
2876 
2877 #endif
2878 
2879 #if KMP_OS_LINUX
2880 // Functions for hidden helper task
2881 namespace {
2882 // Condition variable for initializing hidden helper team
2883 pthread_cond_t hidden_helper_threads_initz_cond_var;
2884 pthread_mutex_t hidden_helper_threads_initz_lock;
2885 volatile int hidden_helper_initz_signaled = FALSE;
2886 
2887 // Condition variable for deinitializing hidden helper team
2888 pthread_cond_t hidden_helper_threads_deinitz_cond_var;
2889 pthread_mutex_t hidden_helper_threads_deinitz_lock;
2890 volatile int hidden_helper_deinitz_signaled = FALSE;
2891 
2892 // Condition variable for the wrapper function of main thread
2893 pthread_cond_t hidden_helper_main_thread_cond_var;
2894 pthread_mutex_t hidden_helper_main_thread_lock;
2895 volatile int hidden_helper_main_thread_signaled = FALSE;
2896 
2897 // Semaphore for worker threads. We don't use condition variable here in case
2898 // that when multiple signals are sent at the same time, only one thread might
2899 // be waken.
2900 sem_t hidden_helper_task_sem;
2901 } // namespace
2902 
2903 void __kmp_hidden_helper_worker_thread_wait() {
2904   int status = sem_wait(&hidden_helper_task_sem);
2905   KMP_CHECK_SYSFAIL("sem_wait", status);
2906 }
2907 
2908 void __kmp_do_initialize_hidden_helper_threads() {
2909   // Initialize condition variable
2910   int status =
2911       pthread_cond_init(&hidden_helper_threads_initz_cond_var, nullptr);
2912   KMP_CHECK_SYSFAIL("pthread_cond_init", status);
2913 
2914   status = pthread_cond_init(&hidden_helper_threads_deinitz_cond_var, nullptr);
2915   KMP_CHECK_SYSFAIL("pthread_cond_init", status);
2916 
2917   status = pthread_cond_init(&hidden_helper_main_thread_cond_var, nullptr);
2918   KMP_CHECK_SYSFAIL("pthread_cond_init", status);
2919 
2920   status = pthread_mutex_init(&hidden_helper_threads_initz_lock, nullptr);
2921   KMP_CHECK_SYSFAIL("pthread_mutex_init", status);
2922 
2923   status = pthread_mutex_init(&hidden_helper_threads_deinitz_lock, nullptr);
2924   KMP_CHECK_SYSFAIL("pthread_mutex_init", status);
2925 
2926   status = pthread_mutex_init(&hidden_helper_main_thread_lock, nullptr);
2927   KMP_CHECK_SYSFAIL("pthread_mutex_init", status);
2928 
2929   // Initialize the semaphore
2930   status = sem_init(&hidden_helper_task_sem, 0, 0);
2931   KMP_CHECK_SYSFAIL("sem_init", status);
2932 
2933   // Create a new thread to finish initialization
2934   pthread_t handle;
2935   status = pthread_create(
2936       &handle, nullptr,
2937       [](void *) -> void * {
2938         __kmp_hidden_helper_threads_initz_routine();
2939         return nullptr;
2940       },
2941       nullptr);
2942   KMP_CHECK_SYSFAIL("pthread_create", status);
2943 }
2944 
2945 void __kmp_hidden_helper_threads_initz_wait() {
2946   // Initial thread waits here for the completion of the initialization. The
2947   // condition variable will be notified by main thread of hidden helper teams.
2948   int status = pthread_mutex_lock(&hidden_helper_threads_initz_lock);
2949   KMP_CHECK_SYSFAIL("pthread_mutex_lock", status);
2950 
2951   if (!TCR_4(hidden_helper_initz_signaled)) {
2952     status = pthread_cond_wait(&hidden_helper_threads_initz_cond_var,
2953                                &hidden_helper_threads_initz_lock);
2954     KMP_CHECK_SYSFAIL("pthread_cond_wait", status);
2955   }
2956 
2957   status = pthread_mutex_unlock(&hidden_helper_threads_initz_lock);
2958   KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status);
2959 }
2960 
2961 void __kmp_hidden_helper_initz_release() {
2962   // After all initialization, reset __kmp_init_hidden_helper_threads to false.
2963   int status = pthread_mutex_lock(&hidden_helper_threads_initz_lock);
2964   KMP_CHECK_SYSFAIL("pthread_mutex_lock", status);
2965 
2966   status = pthread_cond_signal(&hidden_helper_threads_initz_cond_var);
2967   KMP_CHECK_SYSFAIL("pthread_cond_wait", status);
2968 
2969   TCW_SYNC_4(hidden_helper_initz_signaled, TRUE);
2970 
2971   status = pthread_mutex_unlock(&hidden_helper_threads_initz_lock);
2972   KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status);
2973 }
2974 
2975 void __kmp_hidden_helper_main_thread_wait() {
2976   // The main thread of hidden helper team will be blocked here. The
2977   // condition variable can only be signal in the destructor of RTL.
2978   int status = pthread_mutex_lock(&hidden_helper_main_thread_lock);
2979   KMP_CHECK_SYSFAIL("pthread_mutex_lock", status);
2980 
2981   if (!TCR_4(hidden_helper_main_thread_signaled)) {
2982     status = pthread_cond_wait(&hidden_helper_main_thread_cond_var,
2983                                &hidden_helper_main_thread_lock);
2984     KMP_CHECK_SYSFAIL("pthread_cond_wait", status);
2985   }
2986 
2987   status = pthread_mutex_unlock(&hidden_helper_main_thread_lock);
2988   KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status);
2989 }
2990 
2991 void __kmp_hidden_helper_main_thread_release() {
2992   // The initial thread of OpenMP RTL should call this function to wake up the
2993   // main thread of hidden helper team.
2994   int status = pthread_mutex_lock(&hidden_helper_main_thread_lock);
2995   KMP_CHECK_SYSFAIL("pthread_mutex_lock", status);
2996 
2997   status = pthread_cond_signal(&hidden_helper_main_thread_cond_var);
2998   KMP_CHECK_SYSFAIL("pthread_cond_signal", status);
2999 
3000   // The hidden helper team is done here
3001   TCW_SYNC_4(hidden_helper_main_thread_signaled, TRUE);
3002 
3003   status = pthread_mutex_unlock(&hidden_helper_main_thread_lock);
3004   KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status);
3005 }
3006 
3007 void __kmp_hidden_helper_worker_thread_signal() {
3008   int status = sem_post(&hidden_helper_task_sem);
3009   KMP_CHECK_SYSFAIL("sem_post", status);
3010 }
3011 
3012 void __kmp_hidden_helper_threads_deinitz_wait() {
3013   // Initial thread waits here for the completion of the deinitialization. The
3014   // condition variable will be notified by main thread of hidden helper teams.
3015   int status = pthread_mutex_lock(&hidden_helper_threads_deinitz_lock);
3016   KMP_CHECK_SYSFAIL("pthread_mutex_lock", status);
3017 
3018   if (!TCR_4(hidden_helper_deinitz_signaled)) {
3019     status = pthread_cond_wait(&hidden_helper_threads_deinitz_cond_var,
3020                                &hidden_helper_threads_deinitz_lock);
3021     KMP_CHECK_SYSFAIL("pthread_cond_wait", status);
3022   }
3023 
3024   status = pthread_mutex_unlock(&hidden_helper_threads_deinitz_lock);
3025   KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status);
3026 }
3027 
3028 void __kmp_hidden_helper_threads_deinitz_release() {
3029   int status = pthread_mutex_lock(&hidden_helper_threads_deinitz_lock);
3030   KMP_CHECK_SYSFAIL("pthread_mutex_lock", status);
3031 
3032   status = pthread_cond_signal(&hidden_helper_threads_deinitz_cond_var);
3033   KMP_CHECK_SYSFAIL("pthread_cond_wait", status);
3034 
3035   TCW_SYNC_4(hidden_helper_deinitz_signaled, TRUE);
3036 
3037   status = pthread_mutex_unlock(&hidden_helper_threads_deinitz_lock);
3038   KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status);
3039 }
3040 #else // KMP_OS_LINUX
3041 void __kmp_hidden_helper_worker_thread_wait() {
3042   KMP_ASSERT(0 && "Hidden helper task is not supported on this OS");
3043 }
3044 
3045 void __kmp_do_initialize_hidden_helper_threads() {
3046   KMP_ASSERT(0 && "Hidden helper task is not supported on this OS");
3047 }
3048 
3049 void __kmp_hidden_helper_threads_initz_wait() {
3050   KMP_ASSERT(0 && "Hidden helper task is not supported on this OS");
3051 }
3052 
3053 void __kmp_hidden_helper_initz_release() {
3054   KMP_ASSERT(0 && "Hidden helper task is not supported on this OS");
3055 }
3056 
3057 void __kmp_hidden_helper_main_thread_wait() {
3058   KMP_ASSERT(0 && "Hidden helper task is not supported on this OS");
3059 }
3060 
3061 void __kmp_hidden_helper_main_thread_release() {
3062   KMP_ASSERT(0 && "Hidden helper task is not supported on this OS");
3063 }
3064 
3065 void __kmp_hidden_helper_worker_thread_signal() {
3066   KMP_ASSERT(0 && "Hidden helper task is not supported on this OS");
3067 }
3068 
3069 void __kmp_hidden_helper_threads_deinitz_wait() {
3070   KMP_ASSERT(0 && "Hidden helper task is not supported on this OS");
3071 }
3072 
3073 void __kmp_hidden_helper_threads_deinitz_release() {
3074   KMP_ASSERT(0 && "Hidden helper task is not supported on this OS");
3075 }
3076 #endif // KMP_OS_LINUX
3077 
3078 bool __kmp_detect_shm() {
3079   DIR *dir = opendir("/dev/shm");
3080   if (dir) { // /dev/shm exists
3081     closedir(dir);
3082     return true;
3083   } else if (ENOENT == errno) { // /dev/shm does not exist
3084     return false;
3085   } else { // opendir() failed
3086     return false;
3087   }
3088 }
3089 
3090 bool __kmp_detect_tmp() {
3091   DIR *dir = opendir("/tmp");
3092   if (dir) { // /tmp exists
3093     closedir(dir);
3094     return true;
3095   } else if (ENOENT == errno) { // /tmp does not exist
3096     return false;
3097   } else { // opendir() failed
3098     return false;
3099   }
3100 }
3101 
3102 // end of file //
3103