1 /*
2  * kmp_barrier.cpp
3  */
4 
5 
6 //===----------------------------------------------------------------------===//
7 //
8 //                     The LLVM Compiler Infrastructure
9 //
10 // This file is dual licensed under the MIT and the University of Illinois Open
11 // Source Licenses. See LICENSE.txt for details.
12 //
13 //===----------------------------------------------------------------------===//
14 
15 
16 #include "kmp.h"
17 #include "kmp_wait_release.h"
18 #include "kmp_stats.h"
19 #include "kmp_itt.h"
20 #include "kmp_os.h"
21 
22 
23 #if KMP_MIC
24 #include <immintrin.h>
25 #define USE_NGO_STORES 1
26 #endif // KMP_MIC
27 
28 #if KMP_MIC && USE_NGO_STORES
29 // ICV copying
30 #define ngo_load(src)            __m512d Vt = _mm512_load_pd((void *)(src))
31 #define ngo_store_icvs(dst, src) _mm512_storenrngo_pd((void *)(dst), Vt)
32 #define ngo_store_go(dst, src)   _mm512_storenrngo_pd((void *)(dst), Vt)
33 #define ngo_sync()               __asm__ volatile ("lock; addl $0,0(%%rsp)" ::: "memory")
34 #else
35 #define ngo_load(src)            ((void)0)
36 #define ngo_store_icvs(dst, src) copy_icvs((dst), (src))
37 #define ngo_store_go(dst, src)   KMP_MEMCPY((dst), (src), CACHE_LINE)
38 #define ngo_sync()               ((void)0)
39 #endif /* KMP_MIC && USE_NGO_STORES */
40 
41 void __kmp_print_structure(void); // Forward declaration
42 
43 // ---------------------------- Barrier Algorithms ----------------------------
44 
45 // Linear Barrier
46 static void
47 __kmp_linear_barrier_gather(enum barrier_type bt, kmp_info_t *this_thr, int gtid, int tid,
48                             void (*reduce)(void *, void *)
49                             USE_ITT_BUILD_ARG(void * itt_sync_obj) )
50 {
51     KMP_TIME_DEVELOPER_BLOCK(KMP_linear_gather);
52     register kmp_team_t *team = this_thr->th.th_team;
53     register kmp_bstate_t *thr_bar = & this_thr->th.th_bar[bt].bb;
54     register kmp_info_t **other_threads = team->t.t_threads;
55 
56     KA_TRACE(20, ("__kmp_linear_barrier_gather: T#%d(%d:%d) enter for barrier type %d\n",
57                   gtid, team->t.t_id, tid, bt));
58     KMP_DEBUG_ASSERT(this_thr == other_threads[this_thr->th.th_info.ds.ds_tid]);
59 
60 #if USE_ITT_BUILD && USE_ITT_NOTIFY
61     // Barrier imbalance - save arrive time to the thread
62     if(__kmp_forkjoin_frames_mode == 3 || __kmp_forkjoin_frames_mode == 2) {
63         this_thr->th.th_bar_arrive_time = this_thr->th.th_bar_min_time = __itt_get_timestamp();
64     }
65 #endif
66     // We now perform a linear reduction to signal that all of the threads have arrived.
67     if (!KMP_MASTER_TID(tid)) {
68         KA_TRACE(20, ("__kmp_linear_barrier_gather: T#%d(%d:%d) releasing T#%d(%d:%d)"
69                       "arrived(%p): %llu => %llu\n", gtid, team->t.t_id, tid,
70                       __kmp_gtid_from_tid(0, team), team->t.t_id, 0, &thr_bar->b_arrived,
71                       thr_bar->b_arrived, thr_bar->b_arrived + KMP_BARRIER_STATE_BUMP));
72         // Mark arrival to master thread
73         /* After performing this write, a worker thread may not assume that the team is valid
74            any more - it could be deallocated by the master thread at any time. */
75         kmp_flag_64 flag(&thr_bar->b_arrived, other_threads[0]);
76         flag.release();
77     } else {
78         register kmp_balign_team_t *team_bar = &team->t.t_bar[bt];
79         register int nproc = this_thr->th.th_team_nproc;
80         register int i;
81         // Don't have to worry about sleep bit here or atomic since team setting
82         register kmp_uint64 new_state = team_bar->b_arrived + KMP_BARRIER_STATE_BUMP;
83 
84         // Collect all the worker team member threads.
85         for (i=1; i<nproc; ++i) {
86 #if KMP_CACHE_MANAGE
87             // Prefetch next thread's arrived count
88             if (i+1 < nproc)
89                 KMP_CACHE_PREFETCH(&other_threads[i+1]->th.th_bar[bt].bb.b_arrived);
90 #endif /* KMP_CACHE_MANAGE */
91             KA_TRACE(20, ("__kmp_linear_barrier_gather: T#%d(%d:%d) wait T#%d(%d:%d) "
92                           "arrived(%p) == %llu\n", gtid, team->t.t_id, tid,
93                             __kmp_gtid_from_tid(i, team), team->t.t_id, i,
94                             &other_threads[i]->th.th_bar[bt].bb.b_arrived, new_state));
95 
96             // Wait for worker thread to arrive
97             kmp_flag_64 flag(&other_threads[i]->th.th_bar[bt].bb.b_arrived, new_state);
98             flag.wait(this_thr, FALSE
99                       USE_ITT_BUILD_ARG(itt_sync_obj) );
100 #if USE_ITT_BUILD && USE_ITT_NOTIFY
101             // Barrier imbalance - write min of the thread time and the other thread time to the thread.
102             if (__kmp_forkjoin_frames_mode == 2) {
103                 this_thr->th.th_bar_min_time = KMP_MIN(this_thr->th.th_bar_min_time,
104                                                           other_threads[i]->th.th_bar_min_time);
105             }
106 #endif
107             if (reduce) {
108                 KA_TRACE(100, ("__kmp_linear_barrier_gather: T#%d(%d:%d) += T#%d(%d:%d)\n", gtid,
109                                team->t.t_id, tid, __kmp_gtid_from_tid(i, team), team->t.t_id, i));
110                 (*reduce)(this_thr->th.th_local.reduce_data,
111                           other_threads[i]->th.th_local.reduce_data);
112             }
113         }
114         // Don't have to worry about sleep bit here or atomic since team setting
115         team_bar->b_arrived = new_state;
116         KA_TRACE(20, ("__kmp_linear_barrier_gather: T#%d(%d:%d) set team %d arrived(%p) = %llu\n",
117                       gtid, team->t.t_id, tid, team->t.t_id, &team_bar->b_arrived, new_state));
118     }
119     KA_TRACE(20, ("__kmp_linear_barrier_gather: T#%d(%d:%d) exit for barrier type %d\n",
120                   gtid, team->t.t_id, tid, bt));
121 }
122 
123 static void
124 __kmp_linear_barrier_release(enum barrier_type bt, kmp_info_t *this_thr, int gtid, int tid,
125                              int propagate_icvs
126                              USE_ITT_BUILD_ARG(void *itt_sync_obj) )
127 {
128     KMP_TIME_DEVELOPER_BLOCK(KMP_linear_release);
129     register kmp_bstate_t *thr_bar = &this_thr->th.th_bar[bt].bb;
130     register kmp_team_t *team;
131 
132     if (KMP_MASTER_TID(tid)) {
133         register unsigned int i;
134         register kmp_uint32 nproc = this_thr->th.th_team_nproc;
135         register kmp_info_t **other_threads;
136 
137         team = __kmp_threads[gtid]->th.th_team;
138         KMP_DEBUG_ASSERT(team != NULL);
139         other_threads = team->t.t_threads;
140 
141         KA_TRACE(20, ("__kmp_linear_barrier_release: T#%d(%d:%d) master enter for barrier type %d\n",
142                       gtid, team->t.t_id, tid, bt));
143 
144         if (nproc > 1) {
145 #if KMP_BARRIER_ICV_PUSH
146             {
147                 KMP_TIME_DEVELOPER_BLOCK(USER_icv_copy);
148                 if (propagate_icvs) {
149                     ngo_load(&team->t.t_implicit_task_taskdata[0].td_icvs);
150                     for (i=1; i<nproc; ++i) {
151                         __kmp_init_implicit_task(team->t.t_ident, team->t.t_threads[i], team, i, FALSE);
152                         ngo_store_icvs(&team->t.t_implicit_task_taskdata[i].td_icvs,
153                                        &team->t.t_implicit_task_taskdata[0].td_icvs);
154                     }
155                     ngo_sync();
156                 }
157             }
158 #endif // KMP_BARRIER_ICV_PUSH
159 
160             // Now, release all of the worker threads
161             for (i=1; i<nproc; ++i) {
162 #if KMP_CACHE_MANAGE
163                 // Prefetch next thread's go flag
164                 if (i+1 < nproc)
165                     KMP_CACHE_PREFETCH(&other_threads[i+1]->th.th_bar[bt].bb.b_go);
166 #endif /* KMP_CACHE_MANAGE */
167                 KA_TRACE(20, ("__kmp_linear_barrier_release: T#%d(%d:%d) releasing T#%d(%d:%d) "
168                               "go(%p): %u => %u\n", gtid, team->t.t_id, tid,
169                               other_threads[i]->th.th_info.ds.ds_gtid, team->t.t_id, i,
170                               &other_threads[i]->th.th_bar[bt].bb.b_go,
171                               other_threads[i]->th.th_bar[bt].bb.b_go,
172                               other_threads[i]->th.th_bar[bt].bb.b_go + KMP_BARRIER_STATE_BUMP));
173                 kmp_flag_64 flag(&other_threads[i]->th.th_bar[bt].bb.b_go, other_threads[i]);
174                 flag.release();
175             }
176         }
177     } else { // Wait for the MASTER thread to release us
178         KA_TRACE(20, ("__kmp_linear_barrier_release: T#%d wait go(%p) == %u\n",
179                       gtid, &thr_bar->b_go, KMP_BARRIER_STATE_BUMP));
180         kmp_flag_64 flag(&thr_bar->b_go, KMP_BARRIER_STATE_BUMP);
181         flag.wait(this_thr, TRUE
182                   USE_ITT_BUILD_ARG(itt_sync_obj) );
183 #if USE_ITT_BUILD && USE_ITT_NOTIFY
184         if ((__itt_sync_create_ptr && itt_sync_obj == NULL) || KMP_ITT_DEBUG) {
185             // In a fork barrier; cannot get the object reliably (or ITTNOTIFY is disabled)
186             itt_sync_obj = __kmp_itt_barrier_object(gtid, bs_forkjoin_barrier, 0, -1);
187             // Cancel wait on previous parallel region...
188             __kmp_itt_task_starting(itt_sync_obj);
189 
190             if (bt == bs_forkjoin_barrier && TCR_4(__kmp_global.g.g_done))
191                 return;
192 
193             itt_sync_obj = __kmp_itt_barrier_object(gtid, bs_forkjoin_barrier);
194             if (itt_sync_obj != NULL)
195                 // Call prepare as early as possible for "new" barrier
196                 __kmp_itt_task_finished(itt_sync_obj);
197         } else
198 #endif /* USE_ITT_BUILD && USE_ITT_NOTIFY */
199         // Early exit for reaping threads releasing forkjoin barrier
200         if ( bt == bs_forkjoin_barrier && TCR_4(__kmp_global.g.g_done) )
201             return;
202         // The worker thread may now assume that the team is valid.
203 #ifdef KMP_DEBUG
204         tid = __kmp_tid_from_gtid(gtid);
205         team = __kmp_threads[gtid]->th.th_team;
206 #endif
207         KMP_DEBUG_ASSERT(team != NULL);
208         TCW_4(thr_bar->b_go, KMP_INIT_BARRIER_STATE);
209         KA_TRACE(20, ("__kmp_linear_barrier_release: T#%d(%d:%d) set go(%p) = %u\n",
210                       gtid, team->t.t_id, tid, &thr_bar->b_go, KMP_INIT_BARRIER_STATE));
211         KMP_MB();  // Flush all pending memory write invalidates.
212     }
213     KA_TRACE(20, ("__kmp_linear_barrier_release: T#%d(%d:%d) exit for barrier type %d\n",
214                   gtid, team->t.t_id, tid, bt));
215 }
216 
217 // Tree barrier
218 static void
219 __kmp_tree_barrier_gather(enum barrier_type bt, kmp_info_t *this_thr, int gtid, int tid,
220                           void (*reduce)(void *, void *)
221                           USE_ITT_BUILD_ARG(void *itt_sync_obj) )
222 {
223     KMP_TIME_DEVELOPER_BLOCK(KMP_tree_gather);
224     register kmp_team_t *team = this_thr->th.th_team;
225     register kmp_bstate_t *thr_bar = &this_thr->th.th_bar[bt].bb;
226     register kmp_info_t **other_threads = team->t.t_threads;
227     register kmp_uint32 nproc = this_thr->th.th_team_nproc;
228     register kmp_uint32 branch_bits = __kmp_barrier_gather_branch_bits[bt];
229     register kmp_uint32 branch_factor = 1 << branch_bits;
230     register kmp_uint32 child;
231     register kmp_uint32 child_tid;
232     register kmp_uint64 new_state;
233 
234     KA_TRACE(20, ("__kmp_tree_barrier_gather: T#%d(%d:%d) enter for barrier type %d\n",
235                   gtid, team->t.t_id, tid, bt));
236     KMP_DEBUG_ASSERT(this_thr == other_threads[this_thr->th.th_info.ds.ds_tid]);
237 
238 #if USE_ITT_BUILD && USE_ITT_NOTIFY
239     // Barrier imbalance - save arrive time to the thread
240     if(__kmp_forkjoin_frames_mode == 3 || __kmp_forkjoin_frames_mode == 2) {
241         this_thr->th.th_bar_arrive_time = this_thr->th.th_bar_min_time = __itt_get_timestamp();
242     }
243 #endif
244     // Perform tree gather to wait until all threads have arrived; reduce any required data as we go
245     child_tid = (tid << branch_bits) + 1;
246     if (child_tid < nproc) {
247         // Parent threads wait for all their children to arrive
248         new_state = team->t.t_bar[bt].b_arrived + KMP_BARRIER_STATE_BUMP;
249         child = 1;
250         do {
251             register kmp_info_t *child_thr = other_threads[child_tid];
252             register kmp_bstate_t *child_bar = &child_thr->th.th_bar[bt].bb;
253 #if KMP_CACHE_MANAGE
254             // Prefetch next thread's arrived count
255             if (child+1 <= branch_factor && child_tid+1 < nproc)
256                 KMP_CACHE_PREFETCH(&other_threads[child_tid+1]->th.th_bar[bt].bb.b_arrived);
257 #endif /* KMP_CACHE_MANAGE */
258             KA_TRACE(20, ("__kmp_tree_barrier_gather: T#%d(%d:%d) wait T#%d(%d:%u) "
259                           "arrived(%p) == %llu\n", gtid, team->t.t_id, tid,
260                             __kmp_gtid_from_tid(child_tid, team), team->t.t_id, child_tid,
261                             &child_bar->b_arrived, new_state));
262             // Wait for child to arrive
263             kmp_flag_64 flag(&child_bar->b_arrived, new_state);
264             flag.wait(this_thr, FALSE
265                       USE_ITT_BUILD_ARG(itt_sync_obj) );
266 #if USE_ITT_BUILD && USE_ITT_NOTIFY
267             // Barrier imbalance - write min of the thread time and a child time to the thread.
268             if (__kmp_forkjoin_frames_mode == 2) {
269                 this_thr->th.th_bar_min_time = KMP_MIN(this_thr->th.th_bar_min_time,
270                                                           child_thr->th.th_bar_min_time);
271             }
272 #endif
273             if (reduce) {
274                 KA_TRACE(100, ("__kmp_tree_barrier_gather: T#%d(%d:%d) += T#%d(%d:%u)\n",
275                                gtid, team->t.t_id, tid, __kmp_gtid_from_tid(child_tid, team),
276                                team->t.t_id, child_tid));
277                 (*reduce)(this_thr->th.th_local.reduce_data, child_thr->th.th_local.reduce_data);
278             }
279             child++;
280             child_tid++;
281         }
282         while (child <= branch_factor && child_tid < nproc);
283     }
284 
285     if (!KMP_MASTER_TID(tid)) { // Worker threads
286         register kmp_int32 parent_tid = (tid - 1) >> branch_bits;
287 
288         KA_TRACE(20, ("__kmp_tree_barrier_gather: T#%d(%d:%d) releasing T#%d(%d:%d) "
289                       "arrived(%p): %llu => %llu\n", gtid, team->t.t_id, tid,
290                       __kmp_gtid_from_tid(parent_tid, team), team->t.t_id, parent_tid,
291                       &thr_bar->b_arrived, thr_bar->b_arrived,
292                       thr_bar->b_arrived + KMP_BARRIER_STATE_BUMP));
293 
294         // Mark arrival to parent thread
295         /* After performing this write, a worker thread may not assume that the team is valid
296            any more - it could be deallocated by the master thread at any time.  */
297         kmp_flag_64 flag(&thr_bar->b_arrived, other_threads[parent_tid]);
298         flag.release();
299     } else {
300         // Need to update the team arrived pointer if we are the master thread
301         if (nproc > 1) // New value was already computed above
302             team->t.t_bar[bt].b_arrived = new_state;
303         else
304             team->t.t_bar[bt].b_arrived += KMP_BARRIER_STATE_BUMP;
305         KA_TRACE(20, ("__kmp_tree_barrier_gather: T#%d(%d:%d) set team %d arrived(%p) = %llu\n",
306                       gtid, team->t.t_id, tid, team->t.t_id,
307                       &team->t.t_bar[bt].b_arrived, team->t.t_bar[bt].b_arrived));
308     }
309     KA_TRACE(20, ("__kmp_tree_barrier_gather: T#%d(%d:%d) exit for barrier type %d\n",
310                   gtid, team->t.t_id, tid, bt));
311 }
312 
313 static void
314 __kmp_tree_barrier_release(enum barrier_type bt, kmp_info_t *this_thr, int gtid, int tid,
315                            int propagate_icvs
316                            USE_ITT_BUILD_ARG(void *itt_sync_obj) )
317 {
318     KMP_TIME_DEVELOPER_BLOCK(KMP_tree_release);
319     register kmp_team_t *team;
320     register kmp_bstate_t *thr_bar = &this_thr->th.th_bar[bt].bb;
321     register kmp_uint32 nproc;
322     register kmp_uint32 branch_bits = __kmp_barrier_release_branch_bits[bt];
323     register kmp_uint32 branch_factor = 1 << branch_bits;
324     register kmp_uint32 child;
325     register kmp_uint32 child_tid;
326 
327     // Perform a tree release for all of the threads that have been gathered
328     if (!KMP_MASTER_TID(tid)) { // Handle fork barrier workers who aren't part of a team yet
329         KA_TRACE(20, ("__kmp_tree_barrier_release: T#%d wait go(%p) == %u\n",
330                       gtid, &thr_bar->b_go, KMP_BARRIER_STATE_BUMP));
331         // Wait for parent thread to release us
332         kmp_flag_64 flag(&thr_bar->b_go, KMP_BARRIER_STATE_BUMP);
333         flag.wait(this_thr, TRUE
334                   USE_ITT_BUILD_ARG(itt_sync_obj) );
335 #if USE_ITT_BUILD && USE_ITT_NOTIFY
336         if ((__itt_sync_create_ptr && itt_sync_obj == NULL) || KMP_ITT_DEBUG) {
337             // In fork barrier where we could not get the object reliably (or ITTNOTIFY is disabled)
338             itt_sync_obj = __kmp_itt_barrier_object(gtid, bs_forkjoin_barrier, 0, -1);
339             // Cancel wait on previous parallel region...
340             __kmp_itt_task_starting(itt_sync_obj);
341 
342             if (bt == bs_forkjoin_barrier && TCR_4(__kmp_global.g.g_done))
343                 return;
344 
345             itt_sync_obj = __kmp_itt_barrier_object(gtid, bs_forkjoin_barrier);
346             if (itt_sync_obj != NULL)
347                 // Call prepare as early as possible for "new" barrier
348                 __kmp_itt_task_finished(itt_sync_obj);
349         } else
350 #endif /* USE_ITT_BUILD && USE_ITT_NOTIFY */
351         // Early exit for reaping threads releasing forkjoin barrier
352         if (bt == bs_forkjoin_barrier && TCR_4(__kmp_global.g.g_done))
353             return;
354 
355         // The worker thread may now assume that the team is valid.
356         team = __kmp_threads[gtid]->th.th_team;
357         KMP_DEBUG_ASSERT(team != NULL);
358         tid = __kmp_tid_from_gtid(gtid);
359 
360         TCW_4(thr_bar->b_go, KMP_INIT_BARRIER_STATE);
361         KA_TRACE(20, ("__kmp_tree_barrier_release: T#%d(%d:%d) set go(%p) = %u\n",
362                       gtid, team->t.t_id, tid, &thr_bar->b_go, KMP_INIT_BARRIER_STATE));
363         KMP_MB();  // Flush all pending memory write invalidates.
364     } else {
365         team = __kmp_threads[gtid]->th.th_team;
366         KMP_DEBUG_ASSERT(team != NULL);
367         KA_TRACE(20, ("__kmp_tree_barrier_release: T#%d(%d:%d) master enter for barrier type %d\n",
368                       gtid, team->t.t_id, tid, bt));
369     }
370     nproc = this_thr->th.th_team_nproc;
371     child_tid = (tid << branch_bits) + 1;
372 
373     if (child_tid < nproc) {
374         register kmp_info_t **other_threads = team->t.t_threads;
375         child = 1;
376         // Parent threads release all their children
377         do {
378             register kmp_info_t *child_thr = other_threads[child_tid];
379             register kmp_bstate_t *child_bar = &child_thr->th.th_bar[bt].bb;
380 #if KMP_CACHE_MANAGE
381             // Prefetch next thread's go count
382             if (child+1 <= branch_factor && child_tid+1 < nproc)
383                 KMP_CACHE_PREFETCH(&other_threads[child_tid+1]->th.th_bar[bt].bb.b_go);
384 #endif /* KMP_CACHE_MANAGE */
385 
386 #if KMP_BARRIER_ICV_PUSH
387             {
388                 KMP_TIME_DEVELOPER_BLOCK(USER_icv_copy);
389                 if (propagate_icvs) {
390                     __kmp_init_implicit_task(team->t.t_ident, team->t.t_threads[child_tid],
391                                              team, child_tid, FALSE);
392                     copy_icvs(&team->t.t_implicit_task_taskdata[child_tid].td_icvs,
393                               &team->t.t_implicit_task_taskdata[0].td_icvs);
394                 }
395             }
396 #endif // KMP_BARRIER_ICV_PUSH
397             KA_TRACE(20, ("__kmp_tree_barrier_release: T#%d(%d:%d) releasing T#%d(%d:%u)"
398                           "go(%p): %u => %u\n", gtid, team->t.t_id, tid,
399                           __kmp_gtid_from_tid(child_tid, team), team->t.t_id,
400                           child_tid, &child_bar->b_go, child_bar->b_go,
401                           child_bar->b_go + KMP_BARRIER_STATE_BUMP));
402             // Release child from barrier
403             kmp_flag_64 flag(&child_bar->b_go, child_thr);
404             flag.release();
405             child++;
406             child_tid++;
407         }
408         while (child <= branch_factor && child_tid < nproc);
409     }
410     KA_TRACE(20, ("__kmp_tree_barrier_release: T#%d(%d:%d) exit for barrier type %d\n",
411                   gtid, team->t.t_id, tid, bt));
412 }
413 
414 
415 // Hyper Barrier
416 static void
417 __kmp_hyper_barrier_gather(enum barrier_type bt, kmp_info_t *this_thr, int gtid, int tid,
418                            void (*reduce)(void *, void *)
419                            USE_ITT_BUILD_ARG(void *itt_sync_obj) )
420 {
421     KMP_TIME_DEVELOPER_BLOCK(KMP_hyper_gather);
422     register kmp_team_t *team = this_thr->th.th_team;
423     register kmp_bstate_t *thr_bar = &this_thr->th.th_bar[bt].bb;
424     register kmp_info_t **other_threads = team->t.t_threads;
425     register kmp_uint64 new_state = KMP_BARRIER_UNUSED_STATE;
426     register kmp_uint32 num_threads = this_thr->th.th_team_nproc;
427     register kmp_uint32 branch_bits = __kmp_barrier_gather_branch_bits[bt];
428     register kmp_uint32 branch_factor = 1 << branch_bits;
429     register kmp_uint32 offset;
430     register kmp_uint32 level;
431 
432     KA_TRACE(20, ("__kmp_hyper_barrier_gather: T#%d(%d:%d) enter for barrier type %d\n",
433                   gtid, team->t.t_id, tid, bt));
434 
435     KMP_DEBUG_ASSERT(this_thr == other_threads[this_thr->th.th_info.ds.ds_tid]);
436 
437 #if USE_ITT_BUILD && USE_ITT_NOTIFY
438     // Barrier imbalance - save arrive time to the thread
439     if(__kmp_forkjoin_frames_mode == 3 || __kmp_forkjoin_frames_mode == 2) {
440         this_thr->th.th_bar_arrive_time = this_thr->th.th_bar_min_time = __itt_get_timestamp();
441     }
442 #endif
443     /* Perform a hypercube-embedded tree gather to wait until all of the threads have
444        arrived, and reduce any required data as we go.  */
445     kmp_flag_64 p_flag(&thr_bar->b_arrived);
446     for (level=0, offset=1; offset<num_threads; level+=branch_bits, offset<<=branch_bits)
447     {
448         register kmp_uint32 child;
449         register kmp_uint32 child_tid;
450 
451         if (((tid >> level) & (branch_factor - 1)) != 0) {
452             register kmp_int32 parent_tid = tid & ~((1 << (level + branch_bits)) -1);
453 
454             KA_TRACE(20, ("__kmp_hyper_barrier_gather: T#%d(%d:%d) releasing T#%d(%d:%d) "
455                           "arrived(%p): %llu => %llu\n", gtid, team->t.t_id, tid,
456                           __kmp_gtid_from_tid(parent_tid, team), team->t.t_id, parent_tid,
457                           &thr_bar->b_arrived, thr_bar->b_arrived,
458                           thr_bar->b_arrived + KMP_BARRIER_STATE_BUMP));
459             // Mark arrival to parent thread
460             /* After performing this write (in the last iteration of the enclosing for loop),
461                a worker thread may not assume that the team is valid any more - it could be
462                deallocated by the master thread at any time.  */
463             p_flag.set_waiter(other_threads[parent_tid]);
464             p_flag.release();
465             break;
466         }
467 
468         // Parent threads wait for children to arrive
469         if (new_state == KMP_BARRIER_UNUSED_STATE)
470             new_state = team->t.t_bar[bt].b_arrived + KMP_BARRIER_STATE_BUMP;
471         for (child=1, child_tid=tid+(1 << level); child<branch_factor && child_tid<num_threads;
472              child++, child_tid+=(1 << level))
473         {
474             register kmp_info_t *child_thr = other_threads[child_tid];
475             register kmp_bstate_t *child_bar = &child_thr->th.th_bar[bt].bb;
476 #if KMP_CACHE_MANAGE
477             register kmp_uint32 next_child_tid = child_tid + (1 << level);
478             // Prefetch next thread's arrived count
479             if (child+1 < branch_factor && next_child_tid < num_threads)
480                 KMP_CACHE_PREFETCH(&other_threads[next_child_tid]->th.th_bar[bt].bb.b_arrived);
481 #endif /* KMP_CACHE_MANAGE */
482             KA_TRACE(20, ("__kmp_hyper_barrier_gather: T#%d(%d:%d) wait T#%d(%d:%u) "
483                           "arrived(%p) == %llu\n", gtid, team->t.t_id, tid,
484                           __kmp_gtid_from_tid(child_tid, team), team->t.t_id, child_tid,
485                           &child_bar->b_arrived, new_state));
486             // Wait for child to arrive
487             kmp_flag_64 c_flag(&child_bar->b_arrived, new_state);
488             c_flag.wait(this_thr, FALSE
489                         USE_ITT_BUILD_ARG(itt_sync_obj) );
490 #if USE_ITT_BUILD && USE_ITT_NOTIFY
491             // Barrier imbalance - write min of the thread time and a child time to the thread.
492             if (__kmp_forkjoin_frames_mode == 2) {
493                 this_thr->th.th_bar_min_time = KMP_MIN(this_thr->th.th_bar_min_time,
494                                                           child_thr->th.th_bar_min_time);
495             }
496 #endif
497             if (reduce) {
498                 KA_TRACE(100, ("__kmp_hyper_barrier_gather: T#%d(%d:%d) += T#%d(%d:%u)\n",
499                                gtid, team->t.t_id, tid, __kmp_gtid_from_tid(child_tid, team),
500                                team->t.t_id, child_tid));
501                 (*reduce)(this_thr->th.th_local.reduce_data, child_thr->th.th_local.reduce_data);
502             }
503         }
504     }
505 
506     if (KMP_MASTER_TID(tid)) {
507         // Need to update the team arrived pointer if we are the master thread
508         if (new_state == KMP_BARRIER_UNUSED_STATE)
509             team->t.t_bar[bt].b_arrived += KMP_BARRIER_STATE_BUMP;
510         else
511             team->t.t_bar[bt].b_arrived = new_state;
512         KA_TRACE(20, ("__kmp_hyper_barrier_gather: T#%d(%d:%d) set team %d arrived(%p) = %llu\n",
513                       gtid, team->t.t_id, tid, team->t.t_id,
514                       &team->t.t_bar[bt].b_arrived, team->t.t_bar[bt].b_arrived));
515     }
516     KA_TRACE(20, ("__kmp_hyper_barrier_gather: T#%d(%d:%d) exit for barrier type %d\n",
517                   gtid, team->t.t_id, tid, bt));
518 }
519 
520 // The reverse versions seem to beat the forward versions overall
521 #define KMP_REVERSE_HYPER_BAR
522 static void
523 __kmp_hyper_barrier_release(enum barrier_type bt, kmp_info_t *this_thr, int gtid, int tid,
524                             int propagate_icvs
525                             USE_ITT_BUILD_ARG(void *itt_sync_obj) )
526 {
527     KMP_TIME_DEVELOPER_BLOCK(KMP_hyper_release);
528     register kmp_team_t    *team;
529     register kmp_bstate_t  *thr_bar       = & this_thr -> th.th_bar[ bt ].bb;
530     register kmp_info_t   **other_threads;
531     register kmp_uint32     num_threads;
532     register kmp_uint32     branch_bits   = __kmp_barrier_release_branch_bits[ bt ];
533     register kmp_uint32     branch_factor = 1 << branch_bits;
534     register kmp_uint32     child;
535     register kmp_uint32     child_tid;
536     register kmp_uint32     offset;
537     register kmp_uint32     level;
538 
539     /* Perform a hypercube-embedded tree release for all of the threads that have been gathered.
540        If KMP_REVERSE_HYPER_BAR is defined (default) the threads are released in the reverse
541        order of the corresponding gather, otherwise threads are released in the same order. */
542     if (KMP_MASTER_TID(tid)) { // master
543         team = __kmp_threads[gtid]->th.th_team;
544         KMP_DEBUG_ASSERT(team != NULL);
545         KA_TRACE(20, ("__kmp_hyper_barrier_release: T#%d(%d:%d) master enter for barrier type %d\n",
546                       gtid, team->t.t_id, tid, bt));
547 #if KMP_BARRIER_ICV_PUSH
548         if (propagate_icvs) { // master already has ICVs in final destination; copy
549             copy_icvs(&thr_bar->th_fixed_icvs, &team->t.t_implicit_task_taskdata[tid].td_icvs);
550         }
551 #endif
552     }
553     else  { // Handle fork barrier workers who aren't part of a team yet
554         KA_TRACE(20, ("__kmp_hyper_barrier_release: T#%d wait go(%p) == %u\n",
555                       gtid, &thr_bar->b_go, KMP_BARRIER_STATE_BUMP));
556         // Wait for parent thread to release us
557         kmp_flag_64 flag(&thr_bar->b_go, KMP_BARRIER_STATE_BUMP);
558         flag.wait(this_thr, TRUE
559                   USE_ITT_BUILD_ARG(itt_sync_obj) );
560 #if USE_ITT_BUILD && USE_ITT_NOTIFY
561         if ((__itt_sync_create_ptr && itt_sync_obj == NULL) || KMP_ITT_DEBUG) {
562             // In fork barrier where we could not get the object reliably
563             itt_sync_obj = __kmp_itt_barrier_object(gtid, bs_forkjoin_barrier, 0, -1);
564             // Cancel wait on previous parallel region...
565             __kmp_itt_task_starting(itt_sync_obj);
566 
567             if (bt == bs_forkjoin_barrier && TCR_4(__kmp_global.g.g_done))
568                 return;
569 
570             itt_sync_obj = __kmp_itt_barrier_object(gtid, bs_forkjoin_barrier);
571             if (itt_sync_obj != NULL)
572                 // Call prepare as early as possible for "new" barrier
573                 __kmp_itt_task_finished(itt_sync_obj);
574         } else
575 #endif /* USE_ITT_BUILD && USE_ITT_NOTIFY */
576         // Early exit for reaping threads releasing forkjoin barrier
577         if (bt == bs_forkjoin_barrier && TCR_4(__kmp_global.g.g_done))
578             return;
579 
580         // The worker thread may now assume that the team is valid.
581         team = __kmp_threads[gtid]->th.th_team;
582         KMP_DEBUG_ASSERT(team != NULL);
583         tid = __kmp_tid_from_gtid(gtid);
584 
585         TCW_4(thr_bar->b_go, KMP_INIT_BARRIER_STATE);
586         KA_TRACE(20, ("__kmp_hyper_barrier_release: T#%d(%d:%d) set go(%p) = %u\n",
587                       gtid, team->t.t_id, tid, &thr_bar->b_go, KMP_INIT_BARRIER_STATE));
588         KMP_MB();  // Flush all pending memory write invalidates.
589     }
590     num_threads = this_thr->th.th_team_nproc;
591     other_threads = team->t.t_threads;
592 
593 #ifdef KMP_REVERSE_HYPER_BAR
594     // Count up to correct level for parent
595     for (level=0, offset=1; offset<num_threads && (((tid>>level) & (branch_factor-1)) == 0);
596          level+=branch_bits, offset<<=branch_bits);
597 
598     // Now go down from there
599     for (level-=branch_bits, offset>>=branch_bits; offset != 0;
600          level-=branch_bits, offset>>=branch_bits)
601 #else
602     // Go down the tree, level by level
603     for (level=0, offset=1; offset<num_threads; level+=branch_bits, offset<<=branch_bits)
604 #endif // KMP_REVERSE_HYPER_BAR
605     {
606 #ifdef KMP_REVERSE_HYPER_BAR
607         /* Now go in reverse order through the children, highest to lowest.
608            Initial setting of child is conservative here. */
609         child = num_threads >> ((level==0)?level:level-1);
610         for (child=(child<branch_factor-1) ? child : branch_factor-1, child_tid=tid+(child<<level);
611              child>=1; child--, child_tid-=(1<<level))
612 #else
613         if (((tid >> level) & (branch_factor - 1)) != 0)
614             // No need to go lower than this, since this is the level parent would be notified
615             break;
616         // Iterate through children on this level of the tree
617         for (child=1, child_tid=tid+(1<<level); child<branch_factor && child_tid<num_threads;
618              child++, child_tid+=(1<<level))
619 #endif // KMP_REVERSE_HYPER_BAR
620         {
621             if (child_tid >= num_threads) continue;  // Child doesn't exist so keep going
622             else {
623                 register kmp_info_t *child_thr = other_threads[child_tid];
624                 register kmp_bstate_t *child_bar = &child_thr->th.th_bar[bt].bb;
625 #if KMP_CACHE_MANAGE
626                 register kmp_uint32 next_child_tid = child_tid - (1 << level);
627                 // Prefetch next thread's go count
628 # ifdef KMP_REVERSE_HYPER_BAR
629                 if (child-1 >= 1 && next_child_tid < num_threads)
630 # else
631                 if (child+1 < branch_factor && next_child_tid < num_threads)
632 # endif // KMP_REVERSE_HYPER_BAR
633                     KMP_CACHE_PREFETCH(&other_threads[next_child_tid]->th.th_bar[bt].bb.b_go);
634 #endif /* KMP_CACHE_MANAGE */
635 
636 #if KMP_BARRIER_ICV_PUSH
637                 if (propagate_icvs) // push my fixed ICVs to my child
638                     copy_icvs(&child_bar->th_fixed_icvs, &thr_bar->th_fixed_icvs);
639 #endif // KMP_BARRIER_ICV_PUSH
640 
641                 KA_TRACE(20, ("__kmp_hyper_barrier_release: T#%d(%d:%d) releasing T#%d(%d:%u)"
642                               "go(%p): %u => %u\n", gtid, team->t.t_id, tid,
643                               __kmp_gtid_from_tid(child_tid, team), team->t.t_id,
644                               child_tid, &child_bar->b_go, child_bar->b_go,
645                               child_bar->b_go + KMP_BARRIER_STATE_BUMP));
646                 // Release child from barrier
647                 kmp_flag_64 flag(&child_bar->b_go, child_thr);
648                 flag.release();
649             }
650         }
651     }
652 #if KMP_BARRIER_ICV_PUSH
653     if (propagate_icvs && !KMP_MASTER_TID(tid)) { // copy ICVs locally to final dest
654         __kmp_init_implicit_task(team->t.t_ident, team->t.t_threads[tid], team, tid, FALSE);
655         copy_icvs(&team->t.t_implicit_task_taskdata[tid].td_icvs, &thr_bar->th_fixed_icvs);
656     }
657 #endif
658     KA_TRACE(20, ("__kmp_hyper_barrier_release: T#%d(%d:%d) exit for barrier type %d\n",
659                   gtid, team->t.t_id, tid, bt));
660 }
661 
662 // Hierarchical Barrier
663 
664 // Initialize thread barrier data
665 /* Initializes/re-initializes the hierarchical barrier data stored on a thread.  Performs the
666    minimum amount of initialization required based on how the team has changed.  Returns true if
667    leaf children will require both on-core and traditional wake-up mechanisms.  For example, if the
668    team size increases, threads already in the team will respond to on-core wakeup on their parent
669    thread, but threads newly added to the team will only be listening on the their local b_go. */
670 static bool
671 __kmp_init_hierarchical_barrier_thread(enum barrier_type bt, kmp_bstate_t *thr_bar, kmp_uint32 nproc,
672                                        int gtid, int tid, kmp_team_t *team)
673 {
674     // Checks to determine if (re-)initialization is needed
675     bool uninitialized = thr_bar->team == NULL;
676     bool team_changed = team != thr_bar->team;
677     bool team_sz_changed = nproc != thr_bar->nproc;
678     bool tid_changed = tid != thr_bar->old_tid;
679     bool retval = false;
680 
681     if (uninitialized || team_sz_changed) {
682         __kmp_get_hierarchy(nproc, thr_bar);
683     }
684 
685     if (uninitialized || team_sz_changed || tid_changed) {
686         thr_bar->my_level = thr_bar->depth-1; // default for master
687         thr_bar->parent_tid = -1; // default for master
688         if (!KMP_MASTER_TID(tid)) { // if not master, find parent thread in hierarchy
689             kmp_uint32 d=0;
690             while (d<thr_bar->depth) { // find parent based on level of thread in hierarchy, and note level
691                 kmp_uint32 rem;
692                 if (d == thr_bar->depth-2) { // reached level right below the master
693                     thr_bar->parent_tid = 0;
694                     thr_bar->my_level = d;
695                     break;
696                 }
697                 else if ((rem = tid%thr_bar->skip_per_level[d+1]) != 0) { // TODO: can we make this op faster?
698                     // thread is not a subtree root at next level, so this is max
699                     thr_bar->parent_tid = tid - rem;
700                     thr_bar->my_level = d;
701                     break;
702                 }
703                 ++d;
704             }
705         }
706         thr_bar->offset = 7-(tid-thr_bar->parent_tid-1);
707         thr_bar->old_tid = tid;
708         thr_bar->wait_flag = KMP_BARRIER_NOT_WAITING;
709         thr_bar->team = team;
710         thr_bar->parent_bar = &team->t.t_threads[thr_bar->parent_tid]->th.th_bar[bt].bb;
711     }
712     if (uninitialized || team_changed || tid_changed) {
713         thr_bar->team = team;
714         thr_bar->parent_bar = &team->t.t_threads[thr_bar->parent_tid]->th.th_bar[bt].bb;
715         retval = true;
716     }
717     if (uninitialized || team_sz_changed || tid_changed) {
718         thr_bar->nproc = nproc;
719         thr_bar->leaf_kids = thr_bar->base_leaf_kids;
720         if (thr_bar->my_level == 0) thr_bar->leaf_kids=0;
721         if (thr_bar->leaf_kids && (kmp_uint32)tid+thr_bar->leaf_kids+1 > nproc)
722             thr_bar->leaf_kids = nproc - tid - 1;
723         thr_bar->leaf_state = 0;
724         for (int i=0; i<thr_bar->leaf_kids; ++i) ((char *)&(thr_bar->leaf_state))[7-i] = 1;
725     }
726     return retval;
727 }
728 
729 static void
730 __kmp_hierarchical_barrier_gather(enum barrier_type bt, kmp_info_t *this_thr,
731                                   int gtid, int tid, void (*reduce) (void *, void *)
732                                   USE_ITT_BUILD_ARG(void * itt_sync_obj) )
733 {
734     KMP_TIME_DEVELOPER_BLOCK(KMP_hier_gather);
735     register kmp_team_t *team = this_thr->th.th_team;
736     register kmp_bstate_t *thr_bar = & this_thr->th.th_bar[bt].bb;
737     register kmp_uint32 nproc = this_thr->th.th_team_nproc;
738     register kmp_info_t **other_threads = team->t.t_threads;
739     register kmp_uint64 new_state;
740 
741     int level = team->t.t_level;
742 #if OMP_40_ENABLED
743     if (other_threads[0]->th.th_teams_microtask)    // are we inside the teams construct?
744         if (this_thr->th.th_teams_size.nteams > 1)
745             ++level; // level was not increased in teams construct for team_of_masters
746 #endif
747     if (level == 1) thr_bar->use_oncore_barrier = 1;
748     else thr_bar->use_oncore_barrier = 0; // Do not use oncore barrier when nested
749 
750     KA_TRACE(20, ("__kmp_hierarchical_barrier_gather: T#%d(%d:%d) enter for barrier type %d\n",
751                   gtid, team->t.t_id, tid, bt));
752     KMP_DEBUG_ASSERT(this_thr == other_threads[this_thr->th.th_info.ds.ds_tid]);
753 
754 #if USE_ITT_BUILD && USE_ITT_NOTIFY
755     // Barrier imbalance - save arrive time to the thread
756     if(__kmp_forkjoin_frames_mode == 3 || __kmp_forkjoin_frames_mode == 2) {
757         this_thr->th.th_bar_arrive_time = __itt_get_timestamp();
758     }
759 #endif
760 
761     (void)__kmp_init_hierarchical_barrier_thread(bt, thr_bar, nproc, gtid, tid, team);
762 
763     if (thr_bar->my_level) { // not a leaf (my_level==0 means leaf)
764         register kmp_int32 child_tid;
765         new_state = (kmp_uint64)team->t.t_bar[bt].b_arrived + KMP_BARRIER_STATE_BUMP;
766         if (__kmp_dflt_blocktime == KMP_MAX_BLOCKTIME && thr_bar->use_oncore_barrier) {
767             if (thr_bar->leaf_kids) { // First, wait for leaf children to check-in on my b_arrived flag
768                 kmp_uint64 leaf_state = KMP_MASTER_TID(tid) ? thr_bar->b_arrived | thr_bar->leaf_state : team->t.t_bar[bt].b_arrived | thr_bar->leaf_state;
769                 KA_TRACE(20, ("__kmp_hierarchical_barrier_gather: T#%d(%d:%d) waiting for leaf kids\n",
770                               gtid, team->t.t_id, tid));
771                 kmp_flag_64 flag(&thr_bar->b_arrived, leaf_state);
772                 flag.wait(this_thr, FALSE
773                           USE_ITT_BUILD_ARG(itt_sync_obj) );
774                 if (reduce) {
775                     for (child_tid=tid+1; child_tid<=tid+thr_bar->leaf_kids; ++child_tid) {
776                         KA_TRACE(100, ("__kmp_hierarchical_barrier_gather: T#%d(%d:%d) += T#%d(%d:%d)\n",
777                                        gtid, team->t.t_id, tid, __kmp_gtid_from_tid(child_tid, team),
778                                        team->t.t_id, child_tid));
779                         (*reduce)(this_thr->th.th_local.reduce_data, other_threads[child_tid]->th.th_local.reduce_data);
780                     }
781                 }
782                 (void) KMP_TEST_THEN_AND64((volatile kmp_int64 *)&thr_bar->b_arrived, ~(thr_bar->leaf_state)); // clear leaf_state bits
783             }
784             // Next, wait for higher level children on each child's b_arrived flag
785             for (kmp_uint32 d=1; d<thr_bar->my_level; ++d) { // gather lowest level threads first, but skip 0
786                 kmp_uint32 last = tid+thr_bar->skip_per_level[d+1], skip = thr_bar->skip_per_level[d];
787                 if (last > nproc) last = nproc;
788                 for (child_tid=tid+skip; child_tid<(int)last; child_tid+=skip) {
789                     register kmp_info_t *child_thr = other_threads[child_tid];
790                     register kmp_bstate_t *child_bar = &child_thr->th.th_bar[bt].bb;
791                     KA_TRACE(20, ("__kmp_hierarchical_barrier_gather: T#%d(%d:%d) wait T#%d(%d:%d) "
792                                   "arrived(%p) == %llu\n",
793                                   gtid, team->t.t_id, tid, __kmp_gtid_from_tid(child_tid, team),
794                                   team->t.t_id, child_tid, &child_bar->b_arrived, new_state));
795                     kmp_flag_64 flag(&child_bar->b_arrived, new_state);
796                     flag.wait(this_thr, FALSE
797                               USE_ITT_BUILD_ARG(itt_sync_obj) );
798                     if (reduce) {
799                         KA_TRACE(100, ("__kmp_hierarchical_barrier_gather: T#%d(%d:%d) += T#%d(%d:%d)\n",
800                                        gtid, team->t.t_id, tid, __kmp_gtid_from_tid(child_tid, team),
801                                        team->t.t_id, child_tid));
802                         (*reduce)(this_thr->th.th_local.reduce_data, child_thr->th.th_local.reduce_data);
803                     }
804                 }
805             }
806         }
807         else { // Blocktime is not infinite
808             for (kmp_uint32 d=0; d<thr_bar->my_level; ++d) { // Gather lowest level threads first
809                 kmp_uint32 last = tid+thr_bar->skip_per_level[d+1], skip = thr_bar->skip_per_level[d];
810                 if (last > nproc) last = nproc;
811                 for (child_tid=tid+skip; child_tid<(int)last; child_tid+=skip) {
812                     register kmp_info_t *child_thr = other_threads[child_tid];
813                     register kmp_bstate_t *child_bar = &child_thr->th.th_bar[bt].bb;
814                     KA_TRACE(20, ("__kmp_hierarchical_barrier_gather: T#%d(%d:%d) wait T#%d(%d:%d) "
815                                   "arrived(%p) == %llu\n",
816                                   gtid, team->t.t_id, tid, __kmp_gtid_from_tid(child_tid, team),
817                                   team->t.t_id, child_tid, &child_bar->b_arrived, new_state));
818                     kmp_flag_64 flag(&child_bar->b_arrived, new_state);
819                     flag.wait(this_thr, FALSE
820                               USE_ITT_BUILD_ARG(itt_sync_obj) );
821                     if (reduce) {
822                         KA_TRACE(100, ("__kmp_hierarchical_barrier_gather: T#%d(%d:%d) += T#%d(%d:%d)\n",
823                                        gtid, team->t.t_id, tid, __kmp_gtid_from_tid(child_tid, team),
824                                        team->t.t_id, child_tid));
825                         (*reduce)(this_thr->th.th_local.reduce_data, child_thr->th.th_local.reduce_data);
826                     }
827                 }
828             }
829         }
830     }
831     // All subordinates are gathered; now release parent if not master thread
832 
833     if (!KMP_MASTER_TID(tid)) { // worker threads release parent in hierarchy
834         KA_TRACE(20, ("__kmp_hierarchical_barrier_gather: T#%d(%d:%d) releasing T#%d(%d:%d) "
835                       "arrived(%p): %llu => %llu\n", gtid, team->t.t_id, tid,
836                       __kmp_gtid_from_tid(thr_bar->parent_tid, team), team->t.t_id, thr_bar->parent_tid,
837                       &thr_bar->b_arrived, thr_bar->b_arrived, thr_bar->b_arrived+KMP_BARRIER_STATE_BUMP));
838         /* Mark arrival to parent: After performing this write, a worker thread may not assume that
839            the team is valid any more - it could be deallocated by the master thread at any time. */
840         if (thr_bar->my_level || __kmp_dflt_blocktime != KMP_MAX_BLOCKTIME
841             || !thr_bar->use_oncore_barrier) { // Parent is waiting on my b_arrived flag; release it
842             kmp_flag_64 flag(&thr_bar->b_arrived, other_threads[thr_bar->parent_tid]);
843             flag.release();
844         }
845         else { // Leaf does special release on the "offset" bits of parent's b_arrived flag
846             thr_bar->b_arrived = team->t.t_bar[bt].b_arrived + KMP_BARRIER_STATE_BUMP;
847             kmp_flag_oncore flag(&thr_bar->parent_bar->b_arrived, thr_bar->offset);
848             flag.set_waiter(other_threads[thr_bar->parent_tid]);
849             flag.release();
850         }
851     } else { // Master thread needs to update the team's b_arrived value
852         team->t.t_bar[bt].b_arrived = new_state;
853         KA_TRACE(20, ("__kmp_hierarchical_barrier_gather: T#%d(%d:%d) set team %d arrived(%p) = %llu\n",
854                       gtid, team->t.t_id, tid, team->t.t_id, &team->t.t_bar[bt].b_arrived, team->t.t_bar[bt].b_arrived));
855     }
856     // Is the team access below unsafe or just technically invalid?
857     KA_TRACE(20, ("__kmp_hierarchical_barrier_gather: T#%d(%d:%d) exit for barrier type %d\n",
858                   gtid, team->t.t_id, tid, bt));
859 }
860 
861 static void
862 __kmp_hierarchical_barrier_release(enum barrier_type bt, kmp_info_t *this_thr, int gtid, int tid,
863                                    int propagate_icvs
864                                    USE_ITT_BUILD_ARG(void * itt_sync_obj) )
865 {
866     KMP_TIME_DEVELOPER_BLOCK(KMP_hier_release);
867     register kmp_team_t *team;
868     register kmp_bstate_t *thr_bar = &this_thr->th.th_bar[bt].bb;
869     register kmp_uint32 nproc;
870     bool team_change = false; // indicates on-core barrier shouldn't be used
871 
872     if (KMP_MASTER_TID(tid)) {
873         team = __kmp_threads[gtid]->th.th_team;
874         KMP_DEBUG_ASSERT(team != NULL);
875         KA_TRACE(20, ("__kmp_hierarchical_barrier_release: T#%d(%d:%d) master entered barrier type %d\n",
876                       gtid, team->t.t_id, tid, bt));
877     }
878     else { // Worker threads
879         // Wait for parent thread to release me
880         if (!thr_bar->use_oncore_barrier || __kmp_dflt_blocktime != KMP_MAX_BLOCKTIME
881             || thr_bar->my_level != 0 || thr_bar->team == NULL) {
882             // Use traditional method of waiting on my own b_go flag
883             thr_bar->wait_flag = KMP_BARRIER_OWN_FLAG;
884             kmp_flag_64 flag(&thr_bar->b_go, KMP_BARRIER_STATE_BUMP);
885             flag.wait(this_thr, TRUE
886                       USE_ITT_BUILD_ARG(itt_sync_obj) );
887             TCW_8(thr_bar->b_go, KMP_INIT_BARRIER_STATE); // Reset my b_go flag for next time
888         }
889         else { // Thread barrier data is initialized, this is a leaf, blocktime is infinite, not nested
890             // Wait on my "offset" bits on parent's b_go flag
891             thr_bar->wait_flag = KMP_BARRIER_PARENT_FLAG;
892             kmp_flag_oncore flag(&thr_bar->parent_bar->b_go, KMP_BARRIER_STATE_BUMP, thr_bar->offset,
893                                  bt, this_thr
894                                  USE_ITT_BUILD_ARG(itt_sync_obj) );
895             flag.wait(this_thr, TRUE);
896             if (thr_bar->wait_flag == KMP_BARRIER_SWITCHING) { // Thread was switched to own b_go
897                 TCW_8(thr_bar->b_go, KMP_INIT_BARRIER_STATE); // Reset my b_go flag for next time
898             }
899             else { // Reset my bits on parent's b_go flag
900                 ((char*)&(thr_bar->parent_bar->b_go))[thr_bar->offset] = 0;
901             }
902         }
903         thr_bar->wait_flag = KMP_BARRIER_NOT_WAITING;
904         // Early exit for reaping threads releasing forkjoin barrier
905         if (bt == bs_forkjoin_barrier && TCR_4(__kmp_global.g.g_done))
906             return;
907         // The worker thread may now assume that the team is valid.
908         team = __kmp_threads[gtid]->th.th_team;
909         KMP_DEBUG_ASSERT(team != NULL);
910         tid = __kmp_tid_from_gtid(gtid);
911 
912         KA_TRACE(20, ("__kmp_hierarchical_barrier_release: T#%d(%d:%d) set go(%p) = %u\n",
913                       gtid, team->t.t_id, tid, &thr_bar->b_go, KMP_INIT_BARRIER_STATE));
914         KMP_MB();  // Flush all pending memory write invalidates.
915     }
916 
917     nproc = this_thr->th.th_team_nproc;
918     int level = team->t.t_level;
919 #if OMP_40_ENABLED
920     if (team->t.t_threads[0]->th.th_teams_microtask ) {    // are we inside the teams construct?
921         if (team->t.t_pkfn != (microtask_t)__kmp_teams_master && this_thr->th.th_teams_level == level)
922             ++level; // level was not increased in teams construct for team_of_workers
923         if( this_thr->th.th_teams_size.nteams > 1 )
924             ++level; // level was not increased in teams construct for team_of_masters
925     }
926 #endif
927     if (level == 1) thr_bar->use_oncore_barrier = 1;
928     else thr_bar->use_oncore_barrier = 0; // Do not use oncore barrier when nested
929 
930     // If the team size has increased, we still communicate with old leaves via oncore barrier.
931     unsigned short int old_leaf_kids = thr_bar->leaf_kids;
932     kmp_uint64 old_leaf_state = thr_bar->leaf_state;
933     team_change = __kmp_init_hierarchical_barrier_thread(bt, thr_bar, nproc, gtid, tid, team);
934     // But if the entire team changes, we won't use oncore barrier at all
935     if (team_change) old_leaf_kids = 0;
936 
937 #if KMP_BARRIER_ICV_PUSH
938     if (propagate_icvs) {
939         __kmp_init_implicit_task(team->t.t_ident, team->t.t_threads[tid], team, tid, FALSE);
940         if (KMP_MASTER_TID(tid)) { // master already has copy in final destination; copy
941             copy_icvs(&thr_bar->th_fixed_icvs, &team->t.t_implicit_task_taskdata[tid].td_icvs);
942         }
943         else if (__kmp_dflt_blocktime == KMP_MAX_BLOCKTIME && thr_bar->use_oncore_barrier) { // optimization for inf blocktime
944             if (!thr_bar->my_level) // I'm a leaf in the hierarchy (my_level==0)
945                 // leaves (on-core children) pull parent's fixed ICVs directly to local ICV store
946                 copy_icvs(&team->t.t_implicit_task_taskdata[tid].td_icvs,
947                           &thr_bar->parent_bar->th_fixed_icvs);
948             // non-leaves will get ICVs piggybacked with b_go via NGO store
949         }
950         else { // blocktime is not infinite; pull ICVs from parent's fixed ICVs
951             if (thr_bar->my_level) // not a leaf; copy ICVs to my fixed ICVs child can access
952                 copy_icvs(&thr_bar->th_fixed_icvs, &thr_bar->parent_bar->th_fixed_icvs);
953             else // leaves copy parent's fixed ICVs directly to local ICV store
954                 copy_icvs(&team->t.t_implicit_task_taskdata[tid].td_icvs,
955                           &thr_bar->parent_bar->th_fixed_icvs);
956         }
957     }
958 #endif // KMP_BARRIER_ICV_PUSH
959 
960     // Now, release my children
961     if (thr_bar->my_level) { // not a leaf
962         register kmp_int32 child_tid;
963         kmp_uint32 last;
964         if (__kmp_dflt_blocktime == KMP_MAX_BLOCKTIME && thr_bar->use_oncore_barrier) {
965             if (KMP_MASTER_TID(tid)) { // do a flat release
966                 // Set local b_go to bump children via NGO store of the cache line containing IVCs and b_go.
967                 thr_bar->b_go = KMP_BARRIER_STATE_BUMP;
968                 // Use ngo stores if available; b_go piggybacks in the last 8 bytes of the cache line
969                 ngo_load(&thr_bar->th_fixed_icvs);
970                 // This loops over all the threads skipping only the leaf nodes in the hierarchy
971                 for (child_tid=thr_bar->skip_per_level[1]; child_tid<(int)nproc; child_tid+=thr_bar->skip_per_level[1]) {
972                     register kmp_bstate_t *child_bar = &team->t.t_threads[child_tid]->th.th_bar[bt].bb;
973                     KA_TRACE(20, ("__kmp_hierarchical_barrier_release: T#%d(%d:%d) releasing T#%d(%d:%d)"
974                                   " go(%p): %u => %u\n",
975                                   gtid, team->t.t_id, tid, __kmp_gtid_from_tid(child_tid, team),
976                                   team->t.t_id, child_tid, &child_bar->b_go, child_bar->b_go,
977                                   child_bar->b_go + KMP_BARRIER_STATE_BUMP));
978                     // Use ngo store (if available) to both store ICVs and release child via child's b_go
979                     ngo_store_go(&child_bar->th_fixed_icvs, &thr_bar->th_fixed_icvs);
980                 }
981                 ngo_sync();
982             }
983             TCW_8(thr_bar->b_go, KMP_INIT_BARRIER_STATE); // Reset my b_go flag for next time
984             // Now, release leaf children
985             if (thr_bar->leaf_kids) { // if there are any
986                 // We test team_change on the off-chance that the level 1 team changed.
987                 if (team_change || old_leaf_kids < thr_bar->leaf_kids) { // some old leaf_kids, some new
988                     if (old_leaf_kids) { // release old leaf kids
989                         thr_bar->b_go |= old_leaf_state;
990                     }
991                     // Release new leaf kids
992                     last = tid+thr_bar->skip_per_level[1];
993                     if (last > nproc) last = nproc;
994                     for (child_tid=tid+1+old_leaf_kids; child_tid<(int)last; ++child_tid) { // skip_per_level[0]=1
995                         register kmp_info_t   *child_thr = team->t.t_threads[child_tid];
996                         register kmp_bstate_t *child_bar = &child_thr->th.th_bar[bt].bb;
997                         KA_TRACE(20, ("__kmp_hierarchical_barrier_release: T#%d(%d:%d) releasing"
998                                       " T#%d(%d:%d) go(%p): %u => %u\n",
999                                       gtid, team->t.t_id, tid, __kmp_gtid_from_tid(child_tid, team),
1000                                       team->t.t_id, child_tid, &child_bar->b_go, child_bar->b_go,
1001                                       child_bar->b_go + KMP_BARRIER_STATE_BUMP));
1002                         // Release child using child's b_go flag
1003                         kmp_flag_64 flag(&child_bar->b_go, child_thr);
1004                         flag.release();
1005                     }
1006                 }
1007                 else { // Release all children at once with leaf_state bits on my own b_go flag
1008                     thr_bar->b_go |= thr_bar->leaf_state;
1009                 }
1010             }
1011         }
1012         else { // Blocktime is not infinite; do a simple hierarchical release
1013             for (int d=thr_bar->my_level-1; d>=0; --d) { // Release highest level threads first
1014                 last = tid+thr_bar->skip_per_level[d+1];
1015                 kmp_uint32 skip = thr_bar->skip_per_level[d];
1016                 if (last > nproc) last = nproc;
1017                 for (child_tid=tid+skip; child_tid<(int)last; child_tid+=skip) {
1018                     register kmp_info_t   *child_thr = team->t.t_threads[child_tid];
1019                     register kmp_bstate_t *child_bar = &child_thr->th.th_bar[bt].bb;
1020                     KA_TRACE(20, ("__kmp_hierarchical_barrier_release: T#%d(%d:%d) releasing T#%d(%d:%d)"
1021                                   " go(%p): %u => %u\n",
1022                                   gtid, team->t.t_id, tid, __kmp_gtid_from_tid(child_tid, team),
1023                                   team->t.t_id, child_tid, &child_bar->b_go, child_bar->b_go,
1024                                   child_bar->b_go + KMP_BARRIER_STATE_BUMP));
1025                     // Release child using child's b_go flag
1026                     kmp_flag_64 flag(&child_bar->b_go, child_thr);
1027                     flag.release();
1028                 }
1029             }
1030         }
1031 #if KMP_BARRIER_ICV_PUSH
1032         if (propagate_icvs && !KMP_MASTER_TID(tid)) // non-leaves copy ICVs from fixed ICVs to local dest
1033             copy_icvs(&team->t.t_implicit_task_taskdata[tid].td_icvs, &thr_bar->th_fixed_icvs);
1034 #endif // KMP_BARRIER_ICV_PUSH
1035     }
1036     KA_TRACE(20, ("__kmp_hierarchical_barrier_release: T#%d(%d:%d) exit for barrier type %d\n",
1037                   gtid, team->t.t_id, tid, bt));
1038 }
1039 
1040 // ---------------------------- End of Barrier Algorithms ----------------------------
1041 
1042 // Internal function to do a barrier.
1043 /* If is_split is true, do a split barrier, otherwise, do a plain barrier
1044    If reduce is non-NULL, do a split reduction barrier, otherwise, do a split barrier
1045    Returns 0 if master thread, 1 if worker thread.  */
1046 int
1047 __kmp_barrier(enum barrier_type bt, int gtid, int is_split, size_t reduce_size,
1048               void *reduce_data, void (*reduce)(void *, void *))
1049 {
1050     KMP_TIME_DEVELOPER_BLOCK(KMP_barrier);
1051     register int tid = __kmp_tid_from_gtid(gtid);
1052     register kmp_info_t *this_thr = __kmp_threads[gtid];
1053     register kmp_team_t *team = this_thr->th.th_team;
1054     register int status = 0;
1055     ident_t *loc = __kmp_threads[gtid]->th.th_ident;
1056 #if OMPT_SUPPORT
1057     ompt_task_id_t my_task_id;
1058     ompt_parallel_id_t my_parallel_id;
1059 #endif
1060 
1061     KA_TRACE(15, ("__kmp_barrier: T#%d(%d:%d) has arrived\n",
1062                   gtid, __kmp_team_from_gtid(gtid)->t.t_id, __kmp_tid_from_gtid(gtid)));
1063 
1064 #if OMPT_SUPPORT
1065     if (ompt_enabled) {
1066 #if OMPT_BLAME
1067         my_task_id = team->t.t_implicit_task_taskdata[tid].ompt_task_info.task_id;
1068         my_parallel_id = team->t.ompt_team_info.parallel_id;
1069 
1070 #if OMPT_TRACE
1071         if (this_thr->th.ompt_thread_info.state == ompt_state_wait_single) {
1072             if (ompt_callbacks.ompt_callback(ompt_event_single_others_end)) {
1073                 ompt_callbacks.ompt_callback(ompt_event_single_others_end)(
1074                     my_parallel_id, my_task_id);
1075             }
1076         }
1077 #endif
1078         if (ompt_callbacks.ompt_callback(ompt_event_barrier_begin)) {
1079             ompt_callbacks.ompt_callback(ompt_event_barrier_begin)(
1080                 my_parallel_id, my_task_id);
1081         }
1082 #endif
1083         // It is OK to report the barrier state after the barrier begin callback.
1084         // According to the OMPT specification, a compliant implementation may
1085         // even delay reporting this state until the barrier begins to wait.
1086         this_thr->th.ompt_thread_info.state = ompt_state_wait_barrier;
1087     }
1088 #endif
1089 
1090     if (! team->t.t_serialized) {
1091 #if USE_ITT_BUILD
1092         // This value will be used in itt notify events below.
1093         void *itt_sync_obj = NULL;
1094 # if USE_ITT_NOTIFY
1095         if (__itt_sync_create_ptr || KMP_ITT_DEBUG)
1096             itt_sync_obj = __kmp_itt_barrier_object(gtid, bt, 1);
1097 # endif
1098 #endif /* USE_ITT_BUILD */
1099         if (__kmp_tasking_mode == tskm_extra_barrier) {
1100             __kmp_tasking_barrier(team, this_thr, gtid);
1101             KA_TRACE(15, ("__kmp_barrier: T#%d(%d:%d) past tasking barrier\n",
1102                           gtid, __kmp_team_from_gtid(gtid)->t.t_id, __kmp_tid_from_gtid(gtid)));
1103         }
1104 
1105         /* Copy the blocktime info to the thread, where __kmp_wait_template() can access it when
1106            the team struct is not guaranteed to exist. */
1107         // See note about the corresponding code in __kmp_join_barrier() being performance-critical.
1108         if (__kmp_dflt_blocktime != KMP_MAX_BLOCKTIME) {
1109             this_thr->th.th_team_bt_intervals = team->t.t_implicit_task_taskdata[tid].td_icvs.bt_intervals;
1110             this_thr->th.th_team_bt_set = team->t.t_implicit_task_taskdata[tid].td_icvs.bt_set;
1111         }
1112 
1113 #if USE_ITT_BUILD
1114         if (__itt_sync_create_ptr || KMP_ITT_DEBUG)
1115             __kmp_itt_barrier_starting(gtid, itt_sync_obj);
1116 #endif /* USE_ITT_BUILD */
1117 #if USE_DEBUGGER
1118         // Let the debugger know: the thread arrived to the barrier and waiting.
1119         if (KMP_MASTER_TID(tid)) { // Master counter is stored in team structure.
1120             team->t.t_bar[bt].b_master_arrived += 1;
1121         } else {
1122             this_thr->th.th_bar[bt].bb.b_worker_arrived += 1;
1123         } // if
1124 #endif /* USE_DEBUGGER */
1125         if (reduce != NULL) {
1126             //KMP_DEBUG_ASSERT( is_split == TRUE );  // #C69956
1127             this_thr->th.th_local.reduce_data = reduce_data;
1128         }
1129 
1130         if (KMP_MASTER_TID(tid) && __kmp_tasking_mode != tskm_immediate_exec)
1131             __kmp_task_team_setup(this_thr, team, 0); // use 0 to only setup the current team if nthreads > 1
1132 
1133         switch (__kmp_barrier_gather_pattern[bt]) {
1134         case bp_hyper_bar: {
1135             KMP_ASSERT(__kmp_barrier_gather_branch_bits[bt]); // don't set branch bits to 0; use linear
1136             __kmp_hyper_barrier_gather(bt, this_thr, gtid, tid, reduce
1137                                        USE_ITT_BUILD_ARG(itt_sync_obj) );
1138             break;
1139         }
1140         case bp_hierarchical_bar: {
1141             __kmp_hierarchical_barrier_gather(bt, this_thr, gtid, tid, reduce
1142                                               USE_ITT_BUILD_ARG(itt_sync_obj));
1143             break;
1144         }
1145         case bp_tree_bar: {
1146             KMP_ASSERT(__kmp_barrier_gather_branch_bits[bt]); // don't set branch bits to 0; use linear
1147             __kmp_tree_barrier_gather(bt, this_thr, gtid, tid, reduce
1148                                       USE_ITT_BUILD_ARG(itt_sync_obj) );
1149             break;
1150         }
1151         default: {
1152             __kmp_linear_barrier_gather(bt, this_thr, gtid, tid, reduce
1153                                         USE_ITT_BUILD_ARG(itt_sync_obj) );
1154         }
1155         }
1156 
1157         KMP_MB();
1158 
1159         if (KMP_MASTER_TID(tid)) {
1160             status = 0;
1161             if (__kmp_tasking_mode != tskm_immediate_exec) {
1162                 __kmp_task_team_wait(this_thr, team
1163                                      USE_ITT_BUILD_ARG(itt_sync_obj) );
1164             }
1165 #if USE_DEBUGGER
1166             // Let the debugger know: All threads are arrived and starting leaving the barrier.
1167             team->t.t_bar[bt].b_team_arrived += 1;
1168 #endif
1169 
1170 #if USE_ITT_BUILD
1171             /* TODO: In case of split reduction barrier, master thread may send acquired event early,
1172                before the final summation into the shared variable is done (final summation can be a
1173                long operation for array reductions).  */
1174             if (__itt_sync_create_ptr || KMP_ITT_DEBUG)
1175                 __kmp_itt_barrier_middle(gtid, itt_sync_obj);
1176 #endif /* USE_ITT_BUILD */
1177 #if USE_ITT_BUILD && USE_ITT_NOTIFY
1178             // Barrier - report frame end (only if active_level == 1)
1179             if ((__itt_frame_submit_v3_ptr || KMP_ITT_DEBUG) && __kmp_forkjoin_frames_mode &&
1180 #if OMP_40_ENABLED
1181                 this_thr->th.th_teams_microtask == NULL &&
1182 #endif
1183                 team->t.t_active_level == 1)
1184             {
1185                 kmp_uint64 cur_time = __itt_get_timestamp();
1186                 kmp_info_t **other_threads = team->t.t_threads;
1187                 int nproc = this_thr->th.th_team_nproc;
1188                 int i;
1189                 switch(__kmp_forkjoin_frames_mode) {
1190                 case 1:
1191                     __kmp_itt_frame_submit(gtid, this_thr->th.th_frame_time, cur_time, 0, loc, nproc);
1192                     this_thr->th.th_frame_time = cur_time;
1193                     break;
1194                 case 2: // AC 2015-01-19: currently does not work for hierarchical (to be fixed)
1195                     __kmp_itt_frame_submit(gtid, this_thr->th.th_bar_min_time, cur_time, 1, loc, nproc);
1196                     break;
1197                 case 3:
1198                     if( __itt_metadata_add_ptr ) {
1199                         // Initialize with master's wait time
1200                         kmp_uint64 delta = cur_time - this_thr->th.th_bar_arrive_time;
1201                         // Set arrive time to zero to be able to check it in __kmp_invoke_task(); the same is done inside the loop below
1202                         this_thr->th.th_bar_arrive_time = 0;
1203                         for (i=1; i<nproc; ++i) {
1204                             delta += ( cur_time - other_threads[i]->th.th_bar_arrive_time );
1205                             other_threads[i]->th.th_bar_arrive_time = 0;
1206                         }
1207                         __kmp_itt_metadata_imbalance(gtid, this_thr->th.th_frame_time, cur_time, delta, (kmp_uint64)( reduce != NULL));
1208                     }
1209                     __kmp_itt_frame_submit(gtid, this_thr->th.th_frame_time, cur_time, 0, loc, nproc);
1210                     this_thr->th.th_frame_time = cur_time;
1211                     break;
1212                 }
1213             }
1214 #endif /* USE_ITT_BUILD */
1215         } else {
1216             status = 1;
1217 #if USE_ITT_BUILD
1218             if (__itt_sync_create_ptr || KMP_ITT_DEBUG)
1219                 __kmp_itt_barrier_middle(gtid, itt_sync_obj);
1220 #endif /* USE_ITT_BUILD */
1221         }
1222         if (status == 1 || ! is_split) {
1223             switch (__kmp_barrier_release_pattern[bt]) {
1224             case bp_hyper_bar: {
1225                 KMP_ASSERT(__kmp_barrier_release_branch_bits[bt]);
1226                 __kmp_hyper_barrier_release(bt, this_thr, gtid, tid, FALSE
1227                                             USE_ITT_BUILD_ARG(itt_sync_obj) );
1228                 break;
1229             }
1230             case bp_hierarchical_bar: {
1231                 __kmp_hierarchical_barrier_release(bt, this_thr, gtid, tid, FALSE
1232                                                    USE_ITT_BUILD_ARG(itt_sync_obj) );
1233                 break;
1234             }
1235             case bp_tree_bar: {
1236                 KMP_ASSERT(__kmp_barrier_release_branch_bits[bt]);
1237                 __kmp_tree_barrier_release(bt, this_thr, gtid, tid, FALSE
1238                                            USE_ITT_BUILD_ARG(itt_sync_obj) );
1239                 break;
1240             }
1241             default: {
1242                 __kmp_linear_barrier_release(bt, this_thr, gtid, tid, FALSE
1243                                              USE_ITT_BUILD_ARG(itt_sync_obj) );
1244             }
1245             }
1246             if (__kmp_tasking_mode != tskm_immediate_exec) {
1247                 __kmp_task_team_sync(this_thr, team);
1248             }
1249         }
1250 
1251 #if USE_ITT_BUILD
1252         /* GEH: TODO: Move this under if-condition above and also include in
1253            __kmp_end_split_barrier(). This will more accurately represent the actual release time
1254            of the threads for split barriers.  */
1255         if (__itt_sync_create_ptr || KMP_ITT_DEBUG)
1256             __kmp_itt_barrier_finished(gtid, itt_sync_obj);
1257 #endif /* USE_ITT_BUILD */
1258     } else { // Team is serialized.
1259         status = 0;
1260         if (__kmp_tasking_mode != tskm_immediate_exec) {
1261 #if OMP_41_ENABLED
1262             if ( this_thr->th.th_task_team != NULL ) {
1263                 void *itt_sync_obj = NULL;
1264 #if USE_ITT_NOTIFY
1265                 if (__itt_sync_create_ptr || KMP_ITT_DEBUG) {
1266                     itt_sync_obj = __kmp_itt_barrier_object(gtid, bt, 1);
1267                     __kmp_itt_barrier_starting(gtid, itt_sync_obj);
1268                 }
1269 #endif
1270 
1271                 KMP_DEBUG_ASSERT(this_thr->th.th_task_team->tt.tt_found_proxy_tasks == TRUE);
1272                 __kmp_task_team_wait(this_thr, team
1273                                                USE_ITT_BUILD_ARG(itt_sync_obj));
1274                 __kmp_task_team_setup(this_thr, team, 0);
1275 
1276 #if USE_ITT_BUILD
1277                 if (__itt_sync_create_ptr || KMP_ITT_DEBUG)
1278                     __kmp_itt_barrier_finished(gtid, itt_sync_obj);
1279 #endif /* USE_ITT_BUILD */
1280             }
1281 #else
1282             // The task team should be NULL for serialized code (tasks will be executed immediately)
1283             KMP_DEBUG_ASSERT(team->t.t_task_team[this_thr->th.th_task_state] == NULL);
1284             KMP_DEBUG_ASSERT(this_thr->th.th_task_team == NULL);
1285 #endif
1286         }
1287     }
1288     KA_TRACE(15, ("__kmp_barrier: T#%d(%d:%d) is leaving with return value %d\n",
1289                   gtid, __kmp_team_from_gtid(gtid)->t.t_id, __kmp_tid_from_gtid(gtid), status));
1290 
1291 #if OMPT_SUPPORT
1292     if (ompt_enabled) {
1293 #if OMPT_BLAME
1294         if (ompt_callbacks.ompt_callback(ompt_event_barrier_end)) {
1295             ompt_callbacks.ompt_callback(ompt_event_barrier_end)(
1296                 my_parallel_id, my_task_id);
1297         }
1298 #endif
1299         this_thr->th.ompt_thread_info.state = ompt_state_work_parallel;
1300     }
1301 #endif
1302 
1303     return status;
1304 }
1305 
1306 
1307 void
1308 __kmp_end_split_barrier(enum barrier_type bt, int gtid)
1309 {
1310     KMP_TIME_DEVELOPER_BLOCK(KMP_end_split_barrier);
1311     int tid = __kmp_tid_from_gtid(gtid);
1312     kmp_info_t *this_thr = __kmp_threads[gtid];
1313     kmp_team_t *team = this_thr->th.th_team;
1314 
1315     if (!team->t.t_serialized) {
1316         if (KMP_MASTER_GTID(gtid)) {
1317             switch (__kmp_barrier_release_pattern[bt]) {
1318             case bp_hyper_bar: {
1319                 KMP_ASSERT(__kmp_barrier_release_branch_bits[bt]);
1320                 __kmp_hyper_barrier_release(bt, this_thr, gtid, tid, FALSE
1321                                             USE_ITT_BUILD_ARG(NULL) );
1322                 break;
1323             }
1324             case bp_hierarchical_bar: {
1325                 __kmp_hierarchical_barrier_release(bt, this_thr, gtid, tid, FALSE
1326                                                    USE_ITT_BUILD_ARG(NULL));
1327                 break;
1328             }
1329             case bp_tree_bar: {
1330                 KMP_ASSERT(__kmp_barrier_release_branch_bits[bt]);
1331                 __kmp_tree_barrier_release(bt, this_thr, gtid, tid, FALSE
1332                                            USE_ITT_BUILD_ARG(NULL) );
1333                 break;
1334             }
1335             default: {
1336                 __kmp_linear_barrier_release(bt, this_thr, gtid, tid, FALSE
1337                                              USE_ITT_BUILD_ARG(NULL) );
1338             }
1339             }
1340             if (__kmp_tasking_mode != tskm_immediate_exec) {
1341                 __kmp_task_team_sync(this_thr, team);
1342             } // if
1343         }
1344     }
1345 }
1346 
1347 
1348 void
1349 __kmp_join_barrier(int gtid)
1350 {
1351     KMP_TIME_DEVELOPER_BLOCK(KMP_join_barrier);
1352     register kmp_info_t *this_thr = __kmp_threads[gtid];
1353     register kmp_team_t *team;
1354     register kmp_uint nproc;
1355     kmp_info_t *master_thread;
1356     int tid;
1357 #ifdef KMP_DEBUG
1358     int team_id;
1359 #endif /* KMP_DEBUG */
1360 #if USE_ITT_BUILD
1361     void *itt_sync_obj = NULL;
1362 # if USE_ITT_NOTIFY
1363     if (__itt_sync_create_ptr || KMP_ITT_DEBUG) // Don't call routine without need
1364         // Get object created at fork_barrier
1365         itt_sync_obj = __kmp_itt_barrier_object(gtid, bs_forkjoin_barrier);
1366 # endif
1367 #endif /* USE_ITT_BUILD */
1368     KMP_MB();
1369 
1370     // Get current info
1371     team = this_thr->th.th_team;
1372     nproc = this_thr->th.th_team_nproc;
1373     KMP_DEBUG_ASSERT((int)nproc == team->t.t_nproc);
1374     tid = __kmp_tid_from_gtid(gtid);
1375 #ifdef KMP_DEBUG
1376     team_id = team->t.t_id;
1377 #endif /* KMP_DEBUG */
1378     master_thread = this_thr->th.th_team_master;
1379 #ifdef KMP_DEBUG
1380     if (master_thread != team->t.t_threads[0]) {
1381         __kmp_print_structure();
1382     }
1383 #endif /* KMP_DEBUG */
1384     KMP_DEBUG_ASSERT(master_thread == team->t.t_threads[0]);
1385     KMP_MB();
1386 
1387     // Verify state
1388     KMP_DEBUG_ASSERT(__kmp_threads && __kmp_threads[gtid]);
1389     KMP_DEBUG_ASSERT(TCR_PTR(this_thr->th.th_team));
1390     KMP_DEBUG_ASSERT(TCR_PTR(this_thr->th.th_root));
1391     KMP_DEBUG_ASSERT(this_thr == team->t.t_threads[tid]);
1392     KA_TRACE(10, ("__kmp_join_barrier: T#%d(%d:%d) arrived at join barrier\n", gtid, team_id, tid));
1393 
1394 #if OMPT_SUPPORT
1395 #if OMPT_TRACE
1396     if (ompt_enabled &&
1397         ompt_callbacks.ompt_callback(ompt_event_barrier_begin)) {
1398         ompt_callbacks.ompt_callback(ompt_event_barrier_begin)(
1399             team->t.ompt_team_info.parallel_id,
1400             team->t.t_implicit_task_taskdata[tid].ompt_task_info.task_id);
1401     }
1402 #endif
1403     this_thr->th.ompt_thread_info.state = ompt_state_wait_barrier;
1404 #endif
1405 
1406     if (__kmp_tasking_mode == tskm_extra_barrier) {
1407         __kmp_tasking_barrier(team, this_thr, gtid);
1408         KA_TRACE(10, ("__kmp_join_barrier: T#%d(%d:%d) past taking barrier\n", gtid, team_id, tid));
1409     }
1410 # ifdef KMP_DEBUG
1411     if (__kmp_tasking_mode != tskm_immediate_exec) {
1412         KA_TRACE(20, ( "__kmp_join_barrier: T#%d, old team = %d, old task_team = %p, th_task_team = %p\n",
1413                        __kmp_gtid_from_thread(this_thr), team_id, team->t.t_task_team[this_thr->th.th_task_state],
1414                        this_thr->th.th_task_team));
1415         KMP_DEBUG_ASSERT(this_thr->th.th_task_team == team->t.t_task_team[this_thr->th.th_task_state]);
1416     }
1417 # endif /* KMP_DEBUG */
1418 
1419     /* Copy the blocktime info to the thread, where __kmp_wait_template() can access it when the
1420        team struct is not guaranteed to exist. Doing these loads causes a cache miss slows
1421        down EPCC parallel by 2x. As a workaround, we do not perform the copy if blocktime=infinite,
1422        since the values are not used by __kmp_wait_template() in that case. */
1423     if (__kmp_dflt_blocktime != KMP_MAX_BLOCKTIME) {
1424         this_thr->th.th_team_bt_intervals = team->t.t_implicit_task_taskdata[tid].td_icvs.bt_intervals;
1425         this_thr->th.th_team_bt_set = team->t.t_implicit_task_taskdata[tid].td_icvs.bt_set;
1426     }
1427 
1428 #if USE_ITT_BUILD
1429     if (__itt_sync_create_ptr || KMP_ITT_DEBUG)
1430         __kmp_itt_barrier_starting(gtid, itt_sync_obj);
1431 #endif /* USE_ITT_BUILD */
1432 
1433     switch (__kmp_barrier_gather_pattern[bs_forkjoin_barrier]) {
1434     case bp_hyper_bar: {
1435         KMP_ASSERT(__kmp_barrier_gather_branch_bits[bs_forkjoin_barrier]);
1436         __kmp_hyper_barrier_gather(bs_forkjoin_barrier, this_thr, gtid, tid, NULL
1437                                    USE_ITT_BUILD_ARG(itt_sync_obj) );
1438         break;
1439     }
1440     case bp_hierarchical_bar: {
1441         __kmp_hierarchical_barrier_gather(bs_forkjoin_barrier, this_thr, gtid, tid, NULL
1442                                           USE_ITT_BUILD_ARG(itt_sync_obj) );
1443         break;
1444     }
1445     case bp_tree_bar: {
1446         KMP_ASSERT(__kmp_barrier_gather_branch_bits[bs_forkjoin_barrier]);
1447         __kmp_tree_barrier_gather(bs_forkjoin_barrier, this_thr, gtid, tid, NULL
1448                                   USE_ITT_BUILD_ARG(itt_sync_obj) );
1449         break;
1450     }
1451     default: {
1452         __kmp_linear_barrier_gather(bs_forkjoin_barrier, this_thr, gtid, tid, NULL
1453                                     USE_ITT_BUILD_ARG(itt_sync_obj) );
1454     }
1455     }
1456 
1457     /* From this point on, the team data structure may be deallocated at any time by the
1458        master thread - it is unsafe to reference it in any of the worker threads. Any per-team
1459        data items that need to be referenced before the end of the barrier should be moved to
1460        the kmp_task_team_t structs.  */
1461     if (KMP_MASTER_TID(tid)) {
1462         if (__kmp_tasking_mode != tskm_immediate_exec) {
1463             __kmp_task_team_wait(this_thr, team
1464                                  USE_ITT_BUILD_ARG(itt_sync_obj) );
1465         }
1466 #if USE_ITT_BUILD
1467         if (__itt_sync_create_ptr || KMP_ITT_DEBUG)
1468             __kmp_itt_barrier_middle(gtid, itt_sync_obj);
1469 #endif /* USE_ITT_BUILD */
1470 
1471 # if USE_ITT_BUILD && USE_ITT_NOTIFY
1472         // Join barrier - report frame end
1473         if ((__itt_frame_submit_v3_ptr || KMP_ITT_DEBUG) && __kmp_forkjoin_frames_mode &&
1474 #if OMP_40_ENABLED
1475             this_thr->th.th_teams_microtask == NULL &&
1476 #endif
1477             team->t.t_active_level == 1)
1478         {
1479             kmp_uint64 cur_time = __itt_get_timestamp();
1480             ident_t * loc = team->t.t_ident;
1481             kmp_info_t **other_threads = team->t.t_threads;
1482             int nproc = this_thr->th.th_team_nproc;
1483             int i;
1484             switch(__kmp_forkjoin_frames_mode) {
1485             case 1:
1486                 __kmp_itt_frame_submit(gtid, this_thr->th.th_frame_time, cur_time, 0, loc, nproc);
1487                 break;
1488             case 2:
1489                 __kmp_itt_frame_submit(gtid, this_thr->th.th_bar_min_time, cur_time, 1, loc, nproc);
1490                 break;
1491             case 3:
1492                 if( __itt_metadata_add_ptr ) {
1493                     // Initialize with master's wait time
1494                     kmp_uint64 delta = cur_time - this_thr->th.th_bar_arrive_time;
1495                     // Set arrive time to zero to be able to check it in __kmp_invoke_task(); the same is done inside the loop below
1496                     this_thr->th.th_bar_arrive_time = 0;
1497                     for (i=1; i<nproc; ++i) {
1498                         delta += ( cur_time - other_threads[i]->th.th_bar_arrive_time );
1499                         other_threads[i]->th.th_bar_arrive_time = 0;
1500                     }
1501                     __kmp_itt_metadata_imbalance(gtid, this_thr->th.th_frame_time, cur_time, delta, 0);
1502                 }
1503                 __kmp_itt_frame_submit(gtid, this_thr->th.th_frame_time, cur_time, 0, loc, nproc);
1504                 this_thr->th.th_frame_time = cur_time;
1505                 break;
1506             }
1507         }
1508 # endif /* USE_ITT_BUILD */
1509     }
1510 #if USE_ITT_BUILD
1511     else {
1512         if (__itt_sync_create_ptr || KMP_ITT_DEBUG)
1513             __kmp_itt_barrier_middle(gtid, itt_sync_obj);
1514     }
1515 #endif /* USE_ITT_BUILD */
1516 
1517 #if KMP_DEBUG
1518     if (KMP_MASTER_TID(tid)) {
1519         KA_TRACE(15, ("__kmp_join_barrier: T#%d(%d:%d) says all %d team threads arrived\n",
1520                       gtid, team_id, tid, nproc));
1521     }
1522 #endif /* KMP_DEBUG */
1523 
1524     // TODO now, mark worker threads as done so they may be disbanded
1525     KMP_MB(); // Flush all pending memory write invalidates.
1526     KA_TRACE(10, ("__kmp_join_barrier: T#%d(%d:%d) leaving\n", gtid, team_id, tid));
1527 
1528 #if OMPT_SUPPORT
1529     if (ompt_enabled) {
1530 #if OMPT_BLAME
1531         if (ompt_callbacks.ompt_callback(ompt_event_barrier_end)) {
1532             ompt_callbacks.ompt_callback(ompt_event_barrier_end)(
1533                 team->t.ompt_team_info.parallel_id,
1534                 team->t.t_implicit_task_taskdata[tid].ompt_task_info.task_id);
1535         }
1536 #endif
1537 
1538         // return to default state
1539         this_thr->th.ompt_thread_info.state = ompt_state_overhead;
1540     }
1541 #endif
1542 }
1543 
1544 
1545 // TODO release worker threads' fork barriers as we are ready instead of all at once
1546 void
1547 __kmp_fork_barrier(int gtid, int tid)
1548 {
1549     KMP_TIME_DEVELOPER_BLOCK(KMP_fork_barrier);
1550     kmp_info_t *this_thr = __kmp_threads[gtid];
1551     kmp_team_t *team = (tid == 0) ? this_thr->th.th_team : NULL;
1552 #if USE_ITT_BUILD
1553     void * itt_sync_obj = NULL;
1554 #endif /* USE_ITT_BUILD */
1555 
1556     KA_TRACE(10, ("__kmp_fork_barrier: T#%d(%d:%d) has arrived\n",
1557                   gtid, (team != NULL) ? team->t.t_id : -1, tid));
1558 
1559     // th_team pointer only valid for master thread here
1560     if (KMP_MASTER_TID(tid)) {
1561 #if USE_ITT_BUILD && USE_ITT_NOTIFY
1562         if (__itt_sync_create_ptr || KMP_ITT_DEBUG) {
1563             // Create itt barrier object
1564             itt_sync_obj  = __kmp_itt_barrier_object(gtid, bs_forkjoin_barrier, 1);
1565             __kmp_itt_barrier_middle(gtid, itt_sync_obj);  // Call acquired/releasing
1566         }
1567 #endif /* USE_ITT_BUILD && USE_ITT_NOTIFY */
1568 
1569 #ifdef KMP_DEBUG
1570         register kmp_info_t **other_threads = team->t.t_threads;
1571         register int i;
1572 
1573         // Verify state
1574         KMP_MB();
1575 
1576         for(i=1; i<team->t.t_nproc; ++i) {
1577             KA_TRACE(500, ("__kmp_fork_barrier: T#%d(%d:0) checking T#%d(%d:%d) fork go == %u.\n",
1578                            gtid, team->t.t_id, other_threads[i]->th.th_info.ds.ds_gtid,
1579                            team->t.t_id, other_threads[i]->th.th_info.ds.ds_tid,
1580                            other_threads[i]->th.th_bar[bs_forkjoin_barrier].bb.b_go));
1581             KMP_DEBUG_ASSERT((TCR_4(other_threads[i]->th.th_bar[bs_forkjoin_barrier].bb.b_go)
1582                               & ~(KMP_BARRIER_SLEEP_STATE))
1583                              == KMP_INIT_BARRIER_STATE);
1584             KMP_DEBUG_ASSERT(other_threads[i]->th.th_team == team);
1585         }
1586 #endif
1587 
1588         if (__kmp_tasking_mode != tskm_immediate_exec) {
1589             __kmp_task_team_setup(this_thr, team, 0);  // 0 indicates setup current task team if nthreads > 1
1590         }
1591 
1592         /* The master thread may have changed its blocktime between the join barrier and the
1593            fork barrier. Copy the blocktime info to the thread, where __kmp_wait_template() can
1594            access it when the team struct is not guaranteed to exist. */
1595         // See note about the corresponding code in __kmp_join_barrier() being performance-critical
1596         if (__kmp_dflt_blocktime != KMP_MAX_BLOCKTIME) {
1597             this_thr->th.th_team_bt_intervals = team->t.t_implicit_task_taskdata[tid].td_icvs.bt_intervals;
1598             this_thr->th.th_team_bt_set = team->t.t_implicit_task_taskdata[tid].td_icvs.bt_set;
1599         }
1600     } // master
1601 
1602     switch (__kmp_barrier_release_pattern[bs_forkjoin_barrier]) {
1603     case bp_hyper_bar: {
1604         KMP_ASSERT(__kmp_barrier_release_branch_bits[bs_forkjoin_barrier]);
1605         __kmp_hyper_barrier_release(bs_forkjoin_barrier, this_thr, gtid, tid, TRUE
1606                                     USE_ITT_BUILD_ARG(itt_sync_obj) );
1607         break;
1608     }
1609     case bp_hierarchical_bar: {
1610         __kmp_hierarchical_barrier_release(bs_forkjoin_barrier, this_thr, gtid, tid, TRUE
1611                                            USE_ITT_BUILD_ARG(itt_sync_obj) );
1612         break;
1613     }
1614     case bp_tree_bar: {
1615         KMP_ASSERT(__kmp_barrier_release_branch_bits[bs_forkjoin_barrier]);
1616         __kmp_tree_barrier_release(bs_forkjoin_barrier, this_thr, gtid, tid, TRUE
1617                                    USE_ITT_BUILD_ARG(itt_sync_obj) );
1618         break;
1619     }
1620     default: {
1621         __kmp_linear_barrier_release(bs_forkjoin_barrier, this_thr, gtid, tid, TRUE
1622                                      USE_ITT_BUILD_ARG(itt_sync_obj) );
1623     }
1624     }
1625 
1626     // Early exit for reaping threads releasing forkjoin barrier
1627     if (TCR_4(__kmp_global.g.g_done)) {
1628         this_thr->th.th_task_team = NULL;
1629 
1630 #if USE_ITT_BUILD && USE_ITT_NOTIFY
1631         if (__itt_sync_create_ptr || KMP_ITT_DEBUG) {
1632             if (!KMP_MASTER_TID(tid)) {
1633                 itt_sync_obj = __kmp_itt_barrier_object(gtid, bs_forkjoin_barrier);
1634                 if (itt_sync_obj)
1635                     __kmp_itt_barrier_finished(gtid, itt_sync_obj);
1636             }
1637         }
1638 #endif /* USE_ITT_BUILD && USE_ITT_NOTIFY */
1639         KA_TRACE(10, ("__kmp_fork_barrier: T#%d is leaving early\n", gtid));
1640         return;
1641     }
1642 
1643     /* We can now assume that a valid team structure has been allocated by the master and
1644        propagated to all worker threads. The current thread, however, may not be part of the
1645        team, so we can't blindly assume that the team pointer is non-null.  */
1646     team = (kmp_team_t *)TCR_PTR(this_thr->th.th_team);
1647     KMP_DEBUG_ASSERT(team != NULL);
1648     tid = __kmp_tid_from_gtid(gtid);
1649 
1650 
1651 #if KMP_BARRIER_ICV_PULL
1652     /* Master thread's copy of the ICVs was set up on the implicit taskdata in
1653        __kmp_reinitialize_team. __kmp_fork_call() assumes the master thread's implicit task has
1654        this data before this function is called. We cannot modify __kmp_fork_call() to look at
1655        the fixed ICVs in the master's thread struct, because it is not always the case that the
1656        threads arrays have been allocated when __kmp_fork_call() is executed. */
1657     {
1658         KMP_TIME_DEVELOPER_BLOCK(USER_icv_copy);
1659         if (!KMP_MASTER_TID(tid)) {  // master thread already has ICVs
1660             // Copy the initial ICVs from the master's thread struct to the implicit task for this tid.
1661             KA_TRACE(10, ("__kmp_fork_barrier: T#%d(%d) is PULLing ICVs\n", gtid, tid));
1662             __kmp_init_implicit_task(team->t.t_ident, team->t.t_threads[tid], team, tid, FALSE);
1663             copy_icvs(&team->t.t_implicit_task_taskdata[tid].td_icvs,
1664                       &team->t.t_threads[0]->th.th_bar[bs_forkjoin_barrier].bb.th_fixed_icvs);
1665         }
1666     }
1667 #endif // KMP_BARRIER_ICV_PULL
1668 
1669     if (__kmp_tasking_mode != tskm_immediate_exec) {
1670         __kmp_task_team_sync(this_thr, team);
1671     }
1672 
1673 #if OMP_40_ENABLED && KMP_AFFINITY_SUPPORTED
1674     kmp_proc_bind_t proc_bind = team->t.t_proc_bind;
1675     if (proc_bind == proc_bind_intel) {
1676 #endif
1677 #if KMP_AFFINITY_SUPPORTED
1678         // Call dynamic affinity settings
1679         if(__kmp_affinity_type == affinity_balanced && team->t.t_size_changed) {
1680             __kmp_balanced_affinity(tid, team->t.t_nproc);
1681         }
1682 #endif // KMP_AFFINITY_SUPPORTED
1683 #if OMP_40_ENABLED && KMP_AFFINITY_SUPPORTED
1684     }
1685     else if (proc_bind != proc_bind_false) {
1686         if (this_thr->th.th_new_place == this_thr->th.th_current_place) {
1687             KA_TRACE(100, ("__kmp_fork_barrier: T#%d already in correct place %d\n",
1688                            __kmp_gtid_from_thread(this_thr), this_thr->th.th_current_place));
1689         }
1690         else {
1691             __kmp_affinity_set_place(gtid);
1692         }
1693     }
1694 #endif
1695 
1696 #if USE_ITT_BUILD && USE_ITT_NOTIFY
1697     if (__itt_sync_create_ptr || KMP_ITT_DEBUG) {
1698         if (!KMP_MASTER_TID(tid)) {
1699             // Get correct barrier object
1700             itt_sync_obj = __kmp_itt_barrier_object(gtid, bs_forkjoin_barrier);
1701             __kmp_itt_barrier_finished(gtid, itt_sync_obj);  // Workers call acquired
1702         } // (prepare called inside barrier_release)
1703     }
1704 #endif /* USE_ITT_BUILD && USE_ITT_NOTIFY */
1705     KA_TRACE(10, ("__kmp_fork_barrier: T#%d(%d:%d) is leaving\n", gtid, team->t.t_id, tid));
1706 }
1707 
1708 
1709 void
1710 __kmp_setup_icv_copy(kmp_team_t *team, int new_nproc, kmp_internal_control_t *new_icvs, ident_t *loc )
1711 {
1712     KMP_TIME_DEVELOPER_BLOCK(KMP_setup_icv_copy);
1713 
1714     KMP_DEBUG_ASSERT(team && new_nproc && new_icvs);
1715     KMP_DEBUG_ASSERT((!TCR_4(__kmp_init_parallel)) || new_icvs->nproc);
1716 
1717     /* Master thread's copy of the ICVs was set up on the implicit taskdata in
1718        __kmp_reinitialize_team. __kmp_fork_call() assumes the master thread's implicit task has
1719        this data before this function is called. */
1720 #if KMP_BARRIER_ICV_PULL
1721     /* Copy ICVs to master's thread structure into th_fixed_icvs (which remains untouched), where
1722        all of the worker threads can access them and make their own copies after the barrier. */
1723     KMP_DEBUG_ASSERT(team->t.t_threads[0]);  // The threads arrays should be allocated at this point
1724     copy_icvs(&team->t.t_threads[0]->th.th_bar[bs_forkjoin_barrier].bb.th_fixed_icvs, new_icvs);
1725     KF_TRACE(10, ("__kmp_setup_icv_copy: PULL: T#%d this_thread=%p team=%p\n",
1726                   0, team->t.t_threads[0], team));
1727 #elif KMP_BARRIER_ICV_PUSH
1728     // The ICVs will be propagated in the fork barrier, so nothing needs to be done here.
1729     KF_TRACE(10, ("__kmp_setup_icv_copy: PUSH: T#%d this_thread=%p team=%p\n",
1730                   0, team->t.t_threads[0], team));
1731 #else
1732     // Copy the ICVs to each of the non-master threads.  This takes O(nthreads) time.
1733     ngo_load(new_icvs);
1734     KMP_DEBUG_ASSERT(team->t.t_threads[0]);  // The threads arrays should be allocated at this point
1735     for (int f=1; f<new_nproc; ++f) { // Skip the master thread
1736         // TODO: GEH - pass in better source location info since usually NULL here
1737         KF_TRACE(10, ("__kmp_setup_icv_copy: LINEAR: T#%d this_thread=%p team=%p\n",
1738                       f, team->t.t_threads[f], team));
1739         __kmp_init_implicit_task(loc, team->t.t_threads[f], team, f, FALSE);
1740         ngo_store_icvs(&team->t.t_implicit_task_taskdata[f].td_icvs, new_icvs);
1741         KF_TRACE(10, ("__kmp_setup_icv_copy: LINEAR: T#%d this_thread=%p team=%p\n",
1742                       f, team->t.t_threads[f], team));
1743     }
1744     ngo_sync();
1745 #endif // KMP_BARRIER_ICV_PULL
1746 }
1747