1/* Copyright (C) 2005-2015 Free Software Foundation, Inc. 2 Contributed by Richard Henderson <rth@redhat.com>. 3 4 This file is part of the GNU Offloading and Multi Processing Library 5 (libgomp). 6 7 Libgomp is free software; you can redistribute it and/or modify it 8 under the terms of the GNU General Public License as published by 9 the Free Software Foundation; either version 3, or (at your option) 10 any later version. 11 12 Libgomp is distributed in the hope that it will be useful, but WITHOUT ANY 13 WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS 14 FOR A PARTICULAR PURPOSE. See the GNU General Public License for 15 more details. 16 17 Under Section 7 of GPL version 3, you are granted additional 18 permissions described in the GCC Runtime Library Exception, version 19 3.1, as published by the Free Software Foundation. 20 21 You should have received a copy of the GNU General Public License and 22 a copy of the GCC Runtime Library Exception along with this program; 23 see the files COPYING3 and COPYING.RUNTIME respectively. If not, see 24 <http://www.gnu.org/licenses/>. */ 25 26/* This file handles the ORDERED construct. */ 27 28#include "libgomp.h" 29 30 31/* This function is called when first allocating an iteration block. That 32 is, the thread is not currently on the queue. The work-share lock must 33 be held on entry. */ 34 35void 36gomp_ordered_first (void) 37{ 38 struct gomp_thread *thr = gomp_thread (); 39 struct gomp_team *team = thr->ts.team; 40 struct gomp_work_share *ws = thr->ts.work_share; 41 unsigned index; 42 43 /* Work share constructs can be orphaned. */ 44 if (team == NULL || team->nthreads == 1) 45 return; 46 47 index = ws->ordered_cur + ws->ordered_num_used; 48 if (index >= team->nthreads) 49 index -= team->nthreads; 50 ws->ordered_team_ids[index] = thr->ts.team_id; 51 52 /* If this is the first and only thread in the queue, then there is 53 no one to release us when we get to our ordered section. Post to 54 our own release queue now so that we won't block later. */ 55 if (ws->ordered_num_used++ == 0) 56 gomp_sem_post (team->ordered_release[thr->ts.team_id]); 57} 58 59/* This function is called when completing the last iteration block. That 60 is, there are no more iterations to perform and so the thread should be 61 removed from the queue entirely. Because of the way ORDERED blocks are 62 managed, it follows that we currently own access to the ORDERED block, 63 and should now pass it on to the next thread. The work-share lock must 64 be held on entry. */ 65 66void 67gomp_ordered_last (void) 68{ 69 struct gomp_thread *thr = gomp_thread (); 70 struct gomp_team *team = thr->ts.team; 71 struct gomp_work_share *ws = thr->ts.work_share; 72 unsigned next_id; 73 74 /* Work share constructs can be orphaned. */ 75 if (team == NULL || team->nthreads == 1) 76 return; 77 78 /* We're no longer the owner. */ 79 ws->ordered_owner = -1; 80 81 /* If we're not the last thread in the queue, then wake the next. */ 82 if (--ws->ordered_num_used > 0) 83 { 84 unsigned next = ws->ordered_cur + 1; 85 if (next == team->nthreads) 86 next = 0; 87 ws->ordered_cur = next; 88 89 next_id = ws->ordered_team_ids[next]; 90 gomp_sem_post (team->ordered_release[next_id]); 91 } 92} 93 94 95/* This function is called when allocating a subsequent allocation block. 96 That is, we're done with the current iteration block and we're allocating 97 another. This is the logical combination of a call to gomp_ordered_last 98 followed by a call to gomp_ordered_first. The work-share lock must be 99 held on entry. */ 100 101void 102gomp_ordered_next (void) 103{ 104 struct gomp_thread *thr = gomp_thread (); 105 struct gomp_team *team = thr->ts.team; 106 struct gomp_work_share *ws = thr->ts.work_share; 107 unsigned index, next_id; 108 109 /* Work share constructs can be orphaned. */ 110 if (team == NULL || team->nthreads == 1) 111 return; 112 113 /* We're no longer the owner. */ 114 ws->ordered_owner = -1; 115 116 /* If there's only one thread in the queue, that must be us. */ 117 if (ws->ordered_num_used == 1) 118 { 119 /* We have a similar situation as in gomp_ordered_first 120 where we need to post to our own release semaphore. */ 121 gomp_sem_post (team->ordered_release[thr->ts.team_id]); 122 return; 123 } 124 125 /* If the queue is entirely full, then we move ourself to the end of 126 the queue merely by incrementing ordered_cur. Only if it's not 127 full do we have to write our id. */ 128 if (ws->ordered_num_used < team->nthreads) 129 { 130 index = ws->ordered_cur + ws->ordered_num_used; 131 if (index >= team->nthreads) 132 index -= team->nthreads; 133 ws->ordered_team_ids[index] = thr->ts.team_id; 134 } 135 136 index = ws->ordered_cur + 1; 137 if (index == team->nthreads) 138 index = 0; 139 ws->ordered_cur = index; 140 141 next_id = ws->ordered_team_ids[index]; 142 gomp_sem_post (team->ordered_release[next_id]); 143} 144 145 146/* This function is called when a statically scheduled loop is first 147 being created. */ 148 149void 150gomp_ordered_static_init (void) 151{ 152 struct gomp_thread *thr = gomp_thread (); 153 struct gomp_team *team = thr->ts.team; 154 155 if (team == NULL || team->nthreads == 1) 156 return; 157 158 gomp_sem_post (team->ordered_release[0]); 159} 160 161/* This function is called when a statically scheduled loop is moving to 162 the next allocation block. Static schedules are not first come first 163 served like the others, so we're to move to the numerically next thread, 164 not the next thread on a list. The work-share lock should *not* be held 165 on entry. */ 166 167void 168gomp_ordered_static_next (void) 169{ 170 struct gomp_thread *thr = gomp_thread (); 171 struct gomp_team *team = thr->ts.team; 172 struct gomp_work_share *ws = thr->ts.work_share; 173 unsigned id = thr->ts.team_id; 174 175 if (team == NULL || team->nthreads == 1) 176 return; 177 178 ws->ordered_owner = -1; 179 180 /* This thread currently owns the lock. Increment the owner. */ 181 if (++id == team->nthreads) 182 id = 0; 183 ws->ordered_team_ids[0] = id; 184 gomp_sem_post (team->ordered_release[id]); 185} 186 187/* This function is called when we need to assert that the thread owns the 188 ordered section. Due to the problem of posted-but-not-waited semaphores, 189 this needs to happen before completing a loop iteration. */ 190 191void 192gomp_ordered_sync (void) 193{ 194 struct gomp_thread *thr = gomp_thread (); 195 struct gomp_team *team = thr->ts.team; 196 struct gomp_work_share *ws = thr->ts.work_share; 197 198 /* Work share constructs can be orphaned. But this clearly means that 199 we are the only thread, and so we automatically own the section. */ 200 if (team == NULL || team->nthreads == 1) 201 return; 202 203 /* ??? I believe it to be safe to access this data without taking the 204 ws->lock. The only presumed race condition is with the previous 205 thread on the queue incrementing ordered_cur such that it points 206 to us, concurrently with our check below. But our team_id is 207 already present in the queue, and the other thread will always 208 post to our release semaphore. So the two cases are that we will 209 either win the race an momentarily block on the semaphore, or lose 210 the race and find the semaphore already unlocked and so not block. 211 Either way we get correct results. 212 However, there is an implicit flush on entry to an ordered region, 213 so we do need to have a barrier here. If we were taking a lock 214 this could be MEMMODEL_RELEASE since the acquire would be coverd 215 by the lock. */ 216 217 __atomic_thread_fence (MEMMODEL_ACQ_REL); 218 if (ws->ordered_owner != thr->ts.team_id) 219 { 220 gomp_sem_wait (team->ordered_release[thr->ts.team_id]); 221 ws->ordered_owner = thr->ts.team_id; 222 } 223} 224 225/* This function is called by user code when encountering the start of an 226 ORDERED block. We must check to see if the current thread is at the 227 head of the queue, and if not, block. */ 228 229#ifdef HAVE_ATTRIBUTE_ALIAS 230extern void GOMP_ordered_start (void) 231 __attribute__((alias ("gomp_ordered_sync"))); 232#else 233void 234GOMP_ordered_start (void) 235{ 236 gomp_ordered_sync (); 237} 238#endif 239 240/* This function is called by user code when encountering the end of an 241 ORDERED block. With the current ORDERED implementation there's nothing 242 for us to do. 243 244 However, the current implementation has a flaw in that it does not allow 245 the next thread into the ORDERED section immediately after the current 246 thread exits the ORDERED section in its last iteration. The existance 247 of this function allows the implementation to change. */ 248 249void 250GOMP_ordered_end (void) 251{ 252} 253