gcc/libgomp/work.c
Jakub Jelinek 57a957cb71 libgomp: Fix up two non-GOMP_USE_ALIGNED_WORK_SHARES related issues [PR105358]
Last fall I've changed struct gomp_work_share, so that it doesn't have
__attribute__((aligned (64))) lock member in the middle unless the target has
non-emulated aligned allocator, otherwise it just makes sure the first and
second halves are 64 bytes appart for cache line reasons, but doesn't make
the struct 64-byte aligned itself and so we can use normal allocators for it.

When the struct isn't 64-byte aligned, the amount of tail padding significantly
decreases, to 0 or 4 bytes or so.  The library uses that tail padding when
the ordered_teams_ids array (array of uints) and/or the memory for lastprivate
conditional temporaries (the latter wants to guarantee long long alignment).
The problem with it on ia32 darwin9 is that while the struct contains
long long members, long long is just 4 byte aligned while __alignof__(long long)
is 8.  That causes problems in gomp_init_work_share, where we currently rely on
if offsetof (struct gomp_work_share, inline_ordered_team_ids) is long long
aligned, then that tail array will be aligned at runtime and so no extra
memory for dynamic realignment will be needed (that is false when the whole
struct doesn't have long long alignment).  And also in the remaining hunks
causes another problem, where we compute INLINE_ORDERED_TEAM_IDS_OFF
as the above offsetof aligned up to long long boundary and subtract
sizeof (struct gomp_work_share) and INLINE_ORDERED_TEAM_IDS_OFF.
When unlucky, the former isn't multiple of 8 and the latter is 4 bigger
than that and as the subtraction is done in size_t, we end up with (size_t) -4,
so the comparison doesn't really work.

The fixes add additional conditions to make it work properly, but all of them
should be evaluated at compile time when optimizing and so shouldn't slow
anything.

2022-04-26  Jakub Jelinek  <jakub@redhat.com>

	PR libgomp/105358
	* work.c (gomp_init_work_share): Don't mask of adjustment for
	dynamic long long realignment if struct gomp_work_share has smaller
	alignof than long long.
	* loop.c (GOMP_loop_start): Don't use inline_ordered_team_ids if
	struct gomp_work_share has smaller alignof than long long or if
	sizeof (struct gomp_work_share) is smaller than
	INLINE_ORDERED_TEAM_IDS_OFF.
	* loop_ull.c (GOMP_loop_ull_start): Likewise.
	* sections.c (GOMP_sections2_start): Likewise.
2022-04-26 08:57:17 +02:00

321 lines
9.2 KiB
C

/* Copyright (C) 2005-2022 Free Software Foundation, Inc.
Contributed by Richard Henderson <rth@redhat.com>.
This file is part of the GNU Offloading and Multi Processing Library
(libgomp).
Libgomp is free software; you can redistribute it and/or modify it
under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 3, or (at your option)
any later version.
Libgomp is distributed in the hope that it will be useful, but WITHOUT ANY
WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
FOR A PARTICULAR PURPOSE. See the GNU General Public License for
more details.
Under Section 7 of GPL version 3, you are granted additional
permissions described in the GCC Runtime Library Exception, version
3.1, as published by the Free Software Foundation.
You should have received a copy of the GNU General Public License and
a copy of the GCC Runtime Library Exception along with this program;
see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
<http://www.gnu.org/licenses/>. */
/* This file contains routines to manage the work-share queue for a team
of threads. */
#include "libgomp.h"
#include <stddef.h>
#include <stdlib.h>
#include <string.h>
/* Allocate a new work share structure, preferably from current team's
free gomp_work_share cache. */
static struct gomp_work_share *
alloc_work_share (struct gomp_team *team)
{
struct gomp_work_share *ws;
unsigned int i;
/* This is called in a critical section. */
if (team->work_share_list_alloc != NULL)
{
ws = team->work_share_list_alloc;
team->work_share_list_alloc = ws->next_free;
return ws;
}
#ifdef HAVE_SYNC_BUILTINS
ws = team->work_share_list_free;
/* We need atomic read from work_share_list_free,
as free_work_share can be called concurrently. */
__asm ("" : "+r" (ws));
if (ws && ws->next_free)
{
struct gomp_work_share *next = ws->next_free;
ws->next_free = NULL;
team->work_share_list_alloc = next->next_free;
return next;
}
#else
gomp_mutex_lock (&team->work_share_list_free_lock);
ws = team->work_share_list_free;
if (ws)
{
team->work_share_list_alloc = ws->next_free;
team->work_share_list_free = NULL;
gomp_mutex_unlock (&team->work_share_list_free_lock);
return ws;
}
gomp_mutex_unlock (&team->work_share_list_free_lock);
#endif
team->work_share_chunk *= 2;
/* Allocating gomp_work_share structures aligned is just an
optimization, don't do it when using the fallback method. */
#ifdef GOMP_USE_ALIGNED_WORK_SHARES
ws = gomp_aligned_alloc (__alignof (struct gomp_work_share),
team->work_share_chunk
* sizeof (struct gomp_work_share));
#else
ws = gomp_malloc (team->work_share_chunk * sizeof (struct gomp_work_share));
#endif
ws->next_alloc = team->work_shares[0].next_alloc;
team->work_shares[0].next_alloc = ws;
team->work_share_list_alloc = &ws[1];
for (i = 1; i < team->work_share_chunk - 1; i++)
ws[i].next_free = &ws[i + 1];
ws[i].next_free = NULL;
return ws;
}
/* Initialize an already allocated struct gomp_work_share.
This shouldn't touch the next_alloc field. */
void
gomp_init_work_share (struct gomp_work_share *ws, size_t ordered,
unsigned nthreads)
{
gomp_mutex_init (&ws->lock);
if (__builtin_expect (ordered, 0))
{
#define INLINE_ORDERED_TEAM_IDS_SIZE \
(sizeof (struct gomp_work_share) \
- offsetof (struct gomp_work_share, inline_ordered_team_ids))
if (__builtin_expect (ordered != 1, 0))
{
size_t o = nthreads * sizeof (*ws->ordered_team_ids);
o += __alignof__ (long long) - 1;
if ((offsetof (struct gomp_work_share, inline_ordered_team_ids)
& (__alignof__ (long long) - 1)) == 0
&& __alignof__ (struct gomp_work_share)
>= __alignof__ (long long))
o &= ~(__alignof__ (long long) - 1);
ordered += o - 1;
}
else
ordered = nthreads * sizeof (*ws->ordered_team_ids);
if (ordered > INLINE_ORDERED_TEAM_IDS_SIZE)
ws->ordered_team_ids = team_malloc (ordered);
else
ws->ordered_team_ids = ws->inline_ordered_team_ids;
memset (ws->ordered_team_ids, '\0', ordered);
ws->ordered_num_used = 0;
ws->ordered_owner = -1;
ws->ordered_cur = 0;
}
else
ws->ordered_team_ids = ws->inline_ordered_team_ids;
gomp_ptrlock_init (&ws->next_ws, NULL);
ws->threads_completed = 0;
}
/* Do any needed destruction of gomp_work_share fields before it
is put back into free gomp_work_share cache or freed. */
void
gomp_fini_work_share (struct gomp_work_share *ws)
{
gomp_mutex_destroy (&ws->lock);
if (ws->ordered_team_ids != ws->inline_ordered_team_ids)
team_free (ws->ordered_team_ids);
gomp_ptrlock_destroy (&ws->next_ws);
}
/* Free a work share struct, if not orphaned, put it into current
team's free gomp_work_share cache. */
static inline void
free_work_share (struct gomp_team *team, struct gomp_work_share *ws)
{
gomp_fini_work_share (ws);
if (__builtin_expect (team == NULL, 0))
free (ws);
else
{
struct gomp_work_share *next_ws;
#ifdef HAVE_SYNC_BUILTINS
do
{
next_ws = team->work_share_list_free;
ws->next_free = next_ws;
}
while (!__sync_bool_compare_and_swap (&team->work_share_list_free,
next_ws, ws));
#else
gomp_mutex_lock (&team->work_share_list_free_lock);
next_ws = team->work_share_list_free;
ws->next_free = next_ws;
team->work_share_list_free = ws;
gomp_mutex_unlock (&team->work_share_list_free_lock);
#endif
}
}
/* The current thread is ready to begin the next work sharing construct.
In all cases, thr->ts.work_share is updated to point to the new
structure. In all cases the work_share lock is locked. Return true
if this was the first thread to reach this point. */
bool
gomp_work_share_start (size_t ordered)
{
struct gomp_thread *thr = gomp_thread ();
struct gomp_team *team = thr->ts.team;
struct gomp_work_share *ws;
/* Work sharing constructs can be orphaned. */
if (team == NULL)
{
#ifdef GOMP_USE_ALIGNED_WORK_SHARES
ws = gomp_aligned_alloc (__alignof (struct gomp_work_share),
sizeof (*ws));
#else
ws = gomp_malloc (sizeof (*ws));
#endif
gomp_init_work_share (ws, ordered, 1);
thr->ts.work_share = ws;
return true;
}
ws = thr->ts.work_share;
thr->ts.last_work_share = ws;
ws = gomp_ptrlock_get (&ws->next_ws);
if (ws == NULL)
{
/* This thread encountered a new ws first. */
struct gomp_work_share *ws = alloc_work_share (team);
gomp_init_work_share (ws, ordered, team->nthreads);
thr->ts.work_share = ws;
return true;
}
else
{
thr->ts.work_share = ws;
return false;
}
}
/* The current thread is done with its current work sharing construct.
This version does imply a barrier at the end of the work-share. */
void
gomp_work_share_end (void)
{
struct gomp_thread *thr = gomp_thread ();
struct gomp_team *team = thr->ts.team;
gomp_barrier_state_t bstate;
/* Work sharing constructs can be orphaned. */
if (team == NULL)
{
free_work_share (NULL, thr->ts.work_share);
thr->ts.work_share = NULL;
return;
}
bstate = gomp_barrier_wait_start (&team->barrier);
if (gomp_barrier_last_thread (bstate))
{
if (__builtin_expect (thr->ts.last_work_share != NULL, 1))
{
team->work_shares_to_free = thr->ts.work_share;
free_work_share (team, thr->ts.last_work_share);
}
}
gomp_team_barrier_wait_end (&team->barrier, bstate);
thr->ts.last_work_share = NULL;
}
/* The current thread is done with its current work sharing construct.
This version implies a cancellable barrier at the end of the work-share. */
bool
gomp_work_share_end_cancel (void)
{
struct gomp_thread *thr = gomp_thread ();
struct gomp_team *team = thr->ts.team;
gomp_barrier_state_t bstate;
/* Cancellable work sharing constructs cannot be orphaned. */
bstate = gomp_barrier_wait_cancel_start (&team->barrier);
if (gomp_barrier_last_thread (bstate))
{
if (__builtin_expect (thr->ts.last_work_share != NULL, 1))
{
team->work_shares_to_free = thr->ts.work_share;
free_work_share (team, thr->ts.last_work_share);
}
}
thr->ts.last_work_share = NULL;
return gomp_team_barrier_wait_cancel_end (&team->barrier, bstate);
}
/* The current thread is done with its current work sharing construct.
This version does NOT imply a barrier at the end of the work-share. */
void
gomp_work_share_end_nowait (void)
{
struct gomp_thread *thr = gomp_thread ();
struct gomp_team *team = thr->ts.team;
struct gomp_work_share *ws = thr->ts.work_share;
unsigned completed;
/* Work sharing constructs can be orphaned. */
if (team == NULL)
{
free_work_share (NULL, ws);
thr->ts.work_share = NULL;
return;
}
if (__builtin_expect (thr->ts.last_work_share == NULL, 0))
return;
#ifdef HAVE_SYNC_BUILTINS
completed = __sync_add_and_fetch (&ws->threads_completed, 1);
#else
gomp_mutex_lock (&ws->lock);
completed = ++ws->threads_completed;
gomp_mutex_unlock (&ws->lock);
#endif
if (completed == team->nthreads)
{
team->work_shares_to_free = thr->ts.work_share;
free_work_share (team, thr->ts.last_work_share);
}
thr->ts.last_work_share = NULL;
}