Hi!
struct gomp_team has struct gomp_work_share array inside of it.
If that latter structure has 64-byte aligned member in the middle,
the whole struct gomp_team needs to be 64-byte aligned, but we weren't
allocating it using gomp_aligned_alloc.
This patch fixes that, except that on gcn team_malloc is special, so
I've instead decided at least for now to avoid using aligned member
and use the padding instead on gcn.
Bootstrapped/regtested on x86_64-linux and i686-linux and in the PR Rainer
mentioned testing on Solaris, committed to trunk.
2021-11-18 Jakub Jelinek <ja...@redhat.com>
PR libgomp/102838
* libgomp.h (GOMP_USE_ALIGNED_WORK_SHARES): Define if
GOMP_HAVE_EFFICIENT_ALIGNED_ALLOC is defined and __AMDGCN__ is not.
(struct gomp_work_share): Use GOMP_USE_ALIGNED_WORK_SHARES instead of
GOMP_HAVE_EFFICIENT_ALIGNED_ALLOC.
* work.c (alloc_work_share, gomp_work_share_start): Likewise.
* team.c (gomp_new_team): If GOMP_USE_ALIGNED_WORK_SHARES, use
gomp_aligned_alloc instead of team_malloc.
--- libgomp/libgomp.h.jj 2021-11-11 14:35:37.699347142 +0100
+++ libgomp/libgomp.h 2021-11-16 11:57:26.657271188 +0100
@@ -95,6 +95,10 @@ enum memmodel
#define GOMP_HAVE_EFFICIENT_ALIGNED_ALLOC 1
#endif
+#if defined(GOMP_HAVE_EFFICIENT_ALIGNED_ALLOC) && !defined(__AMDGCN__)
+#define GOMP_USE_ALIGNED_WORK_SHARES 1
+#endif
+
extern void *gomp_malloc (size_t) __attribute__((malloc));
extern void *gomp_malloc_cleared (size_t) __attribute__((malloc));
extern void *gomp_realloc (void *, size_t);
@@ -348,7 +352,7 @@ struct gomp_work_share
are in a different cache line. */
/* This lock protects the update of the following members. */
-#ifdef GOMP_HAVE_EFFICIENT_ALIGNED_ALLOC
+#ifdef GOMP_USE_ALIGNED_WORK_SHARES
gomp_mutex_t lock __attribute__((aligned (64)));
#else
char pad[64 - offsetof (struct gomp_work_share_1st_cacheline, pad)];
--- libgomp/work.c.jj 2021-10-20 09:34:47.027331304 +0200
+++ libgomp/work.c 2021-11-16 11:58:10.136662003 +0100
@@ -78,7 +78,7 @@ alloc_work_share (struct gomp_team *team
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_HAVE_EFFICIENT_ALIGNED_ALLOC
+#ifdef GOMP_USE_ALIGNED_WORK_SHARES
ws = gomp_aligned_alloc (__alignof (struct gomp_work_share),
team->work_share_chunk
* sizeof (struct gomp_work_share));
@@ -191,7 +191,7 @@ gomp_work_share_start (size_t ordered)
/* Work sharing constructs can be orphaned. */
if (team == NULL)
{
-#ifdef GOMP_HAVE_EFFICIENT_ALIGNED_ALLOC
+#ifdef GOMP_USE_ALIGNED_WORK_SHARES
ws = gomp_aligned_alloc (__alignof (struct gomp_work_share),
sizeof (*ws));
#else
--- libgomp/team.c.jj 2021-11-11 14:35:37.699347142 +0100
+++ libgomp/team.c 2021-11-16 11:59:46.401311440 +0100
@@ -177,7 +177,12 @@ gomp_new_team (unsigned nthreads)
{
size_t extra = sizeof (team->ordered_release[0])
+ sizeof (team->implicit_task[0]);
+#ifdef GOMP_USE_ALIGNED_WORK_SHARES
+ team = gomp_aligned_alloc (__alignof (struct gomp_team),
+ sizeof (*team) + nthreads * extra);
+#else
team = team_malloc (sizeof (*team) + nthreads * extra);
+#endif
#ifndef HAVE_SYNC_BUILTINS
gomp_mutex_init (&team->work_share_list_free_lock);
Jakub
