openmp: Fix up struct gomp_work_share handling [PR102838]

If GOMP_HAVE_EFFICIENT_ALIGNED_ALLOC is not defined, the intent was to
treat the split of the structure between first cacheline (64 bytes)
as mostly write-once, use afterwards and second cacheline as rw just
as an optimization.  But as has been reported, with vectorization enabled
at -O2 it can now result in aligned vector 16-byte or larger stores.
When not having posix_memalign/aligned_alloc/memalign or other similar API,
alloc.c emulates it but it needs to allocate extra memory for the dynamic
realignment.
So, for the GOMP_HAVE_EFFICIENT_ALIGNED_ALLOC not defined case, this patch
stops using aligned (64) attribute in the middle of the structure and instead
inserts padding that puts the second half of the structure at offset 64 bytes.

And when GOMP_HAVE_EFFICIENT_ALIGNED_ALLOC is defined, usually it was allocated
as aligned, but for the orphaned case it could still be allocated just with
gomp_malloc without guaranteed proper alignment.

2021-10-20  Jakub Jelinek  <jakub@redhat.com>

	PR libgomp/102838
	* libgomp.h (struct gomp_work_share_1st_cacheline): New type.
	(struct gomp_work_share): Only use aligned(64) attribute if
	GOMP_HAVE_EFFICIENT_ALIGNED_ALLOC is defined, otherwise just
	add padding before lock to ensure lock is at offset 64 bytes
	into the structure.
	(gomp_workshare_struct_check1, gomp_workshare_struct_check2):
	New poor man's static assertions.
	* work.c (gomp_work_share_start): Use gomp_aligned_alloc instead of
	gomp_malloc if GOMP_HAVE_EFFICIENT_ALIGNED_ALLOC.

2 files changed, 40 insertions, 0 deletions

diff --git a/libgomp/libgomp.h b/libgomp/libgomp.h
index 33a9591..023d8a8 100644
--- a/libgomp/libgomp.h
+++ b/libgomp/libgomp.h
@@ -257,6 +257,30 @@ struct gomp_doacross_work_share
   unsigned int shift_counts[];
 };
 
+/* Like struct gomp_work_share, but only the 1st cacheline of it plus
+   flexible array at the end.
+   Keep in sync with struct gomp_work_share.  */
+struct gomp_work_share_1st_cacheline
+{
+  enum gomp_schedule_type sched;
+  int mode;
+  union {
+    struct {
+      long chunk_size, end, incr;
+    };
+    struct {
+      unsigned long long chunk_size_ull, end_ull, incr_ull;
+    };
+  };
+  union {
+    unsigned *ordered_team_ids;
+    struct gomp_doacross_work_share *doacross;
+  };
+  unsigned ordered_num_used, ordered_owner, ordered_cur;
+  struct gomp_work_share *next_alloc;
+  char pad[];
+};
+
 struct gomp_work_share
 {
   /* This member records the SCHEDULE clause to be used for this construct.
@@ -324,7 +348,12 @@ 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
   gomp_mutex_t lock __attribute__((aligned (64)));
+#else
+  char pad[64 - offsetof (struct gomp_work_share_1st_cacheline, pad)];
+  gomp_mutex_t lock;
+#endif
 
   /* This is the count of the number of threads that have exited the work
      share construct.  If the construct was marked nowait, they have moved on
@@ -362,6 +391,12 @@ struct gomp_work_share
   unsigned inline_ordered_team_ids[0];
 };
 
+extern char gomp_workshare_struct_check1
+  [offsetof (struct gomp_work_share_1st_cacheline, next_alloc)
+   == offsetof (struct gomp_work_share, next_alloc) ? 1 : -1];
+extern char gomp_workshare_struct_check2
+  [offsetof (struct gomp_work_share, lock) == 64 ? 1 : -1];
+
 /* This structure contains all of the thread-local data associated with 
    a thread team.  This is the data that must be saved when a thread
    encounters a nested PARALLEL construct.  */
diff --git a/libgomp/work.c b/libgomp/work.c
index 9d63eec..bf25591 100644
--- a/libgomp/work.c
+++ b/libgomp/work.c
@@ -191,7 +191,12 @@ gomp_work_share_start (size_t ordered)
   /* Work sharing constructs can be orphaned.  */
   if (team == NULL)
     {
+#ifdef GOMP_HAVE_EFFICIENT_ALIGNED_ALLOC
+      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;