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