| 1 |
/* Copyright (C) 2005-2025 Free Software Foundation, Inc. |
| 2 |
Contributed by Richard Henderson <[email protected]>. |
| 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 is a Linux specific implementation of a barrier synchronization |
| 27 |
mechanism for libgomp. This type is private to the library. This |
| 28 |
implementation uses atomic instructions and the futex syscall. */ |
| 29 |
|
| 30 |
#include <limits.h> |
| 31 |
#include "wait.h" |
| 32 |
|
| 33 |
|
| 34 |
void |
| 35 |
gomp_barrier_wait_end (gomp_barrier_t *bar, gomp_barrier_state_t state) |
| 36 |
{ |
| 37 |
if (__builtin_expect (state & BAR_WAS_LAST, 0)) |
| 38 |
{ |
| 39 |
/* Next time we'll be awaiting TOTAL threads again. */ |
| 40 |
bar->awaited = bar->total; |
| 41 |
__atomic_store_n (&bar->generation, bar->generation + BAR_INCR, |
| 42 |
MEMMODEL_RELEASE); |
| 43 |
futex_wake ((int *) &bar->generation, INT_MAX); |
| 44 |
} |
| 45 |
else |
| 46 |
{ |
| 47 |
do |
| 48 |
do_wait ((int *) &bar->generation, state); |
| 49 |
while (__atomic_load_n (&bar->generation, MEMMODEL_ACQUIRE) == state); |
| 50 |
} |
| 51 |
} |
| 52 |
|
| 53 |
void |
| 54 |
gomp_barrier_wait (gomp_barrier_t *bar) |
| 55 |
{ |
| 56 |
gomp_barrier_wait_end (bar, gomp_barrier_wait_start (bar)); |
| 57 |
} |
| 58 |
|
| 59 |
/* Like gomp_barrier_wait, except that if the encountering thread |
| 60 |
is not the last one to hit the barrier, it returns immediately. |
| 61 |
The intended usage is that a thread which intends to gomp_barrier_destroy |
| 62 |
this barrier calls gomp_barrier_wait, while all other threads |
| 63 |
call gomp_barrier_wait_last. When gomp_barrier_wait returns, |
| 64 |
the barrier can be safely destroyed. */ |
| 65 |
|
| 66 |
void |
| 67 |
gomp_barrier_wait_last (gomp_barrier_t *bar) |
| 68 |
{ |
| 69 |
gomp_barrier_state_t state = gomp_barrier_wait_start (bar); |
| 70 |
if (state & BAR_WAS_LAST) |
| 71 |
gomp_barrier_wait_end (bar, state); |
| 72 |
} |
| 73 |
|
| 74 |
void |
| 75 |
gomp_team_barrier_wake (gomp_barrier_t *bar, int count) |
| 76 |
{ |
| 77 |
futex_wake ((int *) &bar->generation, count == 0 ? INT_MAX : count); |
| 78 |
} |
| 79 |
|
| 80 |
void |
| 81 |
gomp_team_barrier_wait_end (gomp_barrier_t *bar, gomp_barrier_state_t state) |
| 82 |
{ |
| 83 |
unsigned int generation, gen; |
| 84 |
|
| 85 |
if (__builtin_expect (state & BAR_WAS_LAST, 0)) |
| 86 |
{ |
| 87 |
/* Next time we'll be awaiting TOTAL threads again. */ |
| 88 |
struct gomp_thread *thr = gomp_thread (); |
| 89 |
struct gomp_team *team = thr->ts.team; |
| 90 |
|
| 91 |
bar->awaited = bar->total; |
| 92 |
team->work_share_cancelled = 0; |
| 93 |
if (__builtin_expect (team->task_count, 0)) |
| 94 |
{ |
| 95 |
gomp_barrier_handle_tasks (state); |
| 96 |
state &= ~BAR_WAS_LAST; |
| 97 |
} |
| 98 |
else |
| 99 |
{ |
| 100 |
state &= ~BAR_CANCELLED; |
| 101 |
state += BAR_INCR - BAR_WAS_LAST; |
| 102 |
__atomic_store_n (&bar->generation, state, MEMMODEL_RELEASE); |
| 103 |
futex_wake ((int *) &bar->generation, INT_MAX); |
| 104 |
return; |
| 105 |
} |
| 106 |
} |
| 107 |
|
| 108 |
generation = state; |
| 109 |
state &= ~BAR_CANCELLED; |
| 110 |
do |
| 111 |
{ |
| 112 |
do_wait ((int *) &bar->generation, generation); |
| 113 |
gen = __atomic_load_n (&bar->generation, MEMMODEL_ACQUIRE); |
| 114 |
if (__builtin_expect (gen & BAR_TASK_PENDING, 0)) |
| 115 |
{ |
| 116 |
gomp_barrier_handle_tasks (state); |
| 117 |
gen = __atomic_load_n (&bar->generation, MEMMODEL_ACQUIRE); |
| 118 |
} |
| 119 |
generation |= gen & BAR_WAITING_FOR_TASK; |
| 120 |
} |
| 121 |
while (gen != state + BAR_INCR); |
| 122 |
} |
| 123 |
|
| 124 |
void |
| 125 |
gomp_team_barrier_wait (gomp_barrier_t *bar) |
| 126 |
{ |
| 127 |
gomp_team_barrier_wait_end (bar, gomp_barrier_wait_start (bar)); |
| 128 |
} |
| 129 |
|
| 130 |
void |
| 131 |
gomp_team_barrier_wait_final (gomp_barrier_t *bar) |
| 132 |
{ |
| 133 |
gomp_barrier_state_t state = gomp_barrier_wait_final_start (bar); |
| 134 |
if (__builtin_expect (state & BAR_WAS_LAST, 0)) |
| 135 |
bar->awaited_final = bar->total; |
| 136 |
gomp_team_barrier_wait_end (bar, state); |
| 137 |
} |
| 138 |
|
| 139 |
bool |
| 140 |
gomp_team_barrier_wait_cancel_end (gomp_barrier_t *bar, |
| 141 |
gomp_barrier_state_t state) |
| 142 |
{ |
| 143 |
unsigned int generation, gen; |
| 144 |
|
| 145 |
if (__builtin_expect (state & BAR_WAS_LAST, 0)) |
| 146 |
{ |
| 147 |
/* Next time we'll be awaiting TOTAL threads again. */ |
| 148 |
/* BAR_CANCELLED should never be set in state here, because |
| 149 |
cancellation means that at least one of the threads has been |
| 150 |
cancelled, thus on a cancellable barrier we should never see |
| 151 |
all threads to arrive. */ |
| 152 |
struct gomp_thread *thr = gomp_thread (); |
| 153 |
struct gomp_team *team = thr->ts.team; |
| 154 |
|
| 155 |
bar->awaited = bar->total; |
| 156 |
team->work_share_cancelled = 0; |
| 157 |
if (__builtin_expect (team->task_count, 0)) |
| 158 |
{ |
| 159 |
gomp_barrier_handle_tasks (state); |
| 160 |
state &= ~BAR_WAS_LAST; |
| 161 |
} |
| 162 |
else |
| 163 |
{ |
| 164 |
state += BAR_INCR - BAR_WAS_LAST; |
| 165 |
__atomic_store_n (&bar->generation, state, MEMMODEL_RELEASE); |
| 166 |
futex_wake ((int *) &bar->generation, INT_MAX); |
| 167 |
return false; |
| 168 |
} |
| 169 |
} |
| 170 |
|
| 171 |
if (__builtin_expect (state & BAR_CANCELLED, 0)) |
| 172 |
return true; |
| 173 |
|
| 174 |
generation = state; |
| 175 |
do |
| 176 |
{ |
| 177 |
do_wait ((int *) &bar->generation, generation); |
| 178 |
gen = __atomic_load_n (&bar->generation, MEMMODEL_ACQUIRE); |
| 179 |
if (__builtin_expect (gen & BAR_CANCELLED, 0)) |
| 180 |
return true; |
| 181 |
if (__builtin_expect (gen & BAR_TASK_PENDING, 0)) |
| 182 |
{ |
| 183 |
gomp_barrier_handle_tasks (state); |
| 184 |
gen = __atomic_load_n (&bar->generation, MEMMODEL_ACQUIRE); |
| 185 |
} |
| 186 |
generation |= gen & BAR_WAITING_FOR_TASK; |
| 187 |
} |
| 188 |
while (gen != state + BAR_INCR); |
| 189 |
|
| 190 |
return false; |
| 191 |
} |
| 192 |
|
| 193 |
bool |
| 194 |
gomp_team_barrier_wait_cancel (gomp_barrier_t *bar) |
| 195 |
{ |
| 196 |
return gomp_team_barrier_wait_cancel_end (bar, gomp_barrier_wait_start (bar)); |
| 197 |
} |
| 198 |
|
| 199 |
void |
| 200 |
gomp_team_barrier_cancel (struct gomp_team *team) |
| 201 |
{ |
| 202 |
gomp_mutex_lock (&team->task_lock); |
| 203 |
if (team->barrier.generation & BAR_CANCELLED) |
| 204 |
{ |
| 205 |
gomp_mutex_unlock (&team->task_lock); |
| 206 |
return; |
| 207 |
} |
| 208 |
team->barrier.generation |= BAR_CANCELLED; |
| 209 |
gomp_mutex_unlock (&team->task_lock); |
| 210 |
futex_wake ((int *) &team->barrier.generation, INT_MAX); |
| 211 |
} |