Hi,
the patch below makes target_follows_kernelizable_pattern stricter by
adding a few checks for clauses that have to preclude kernelization.
Committed to the branch.
Thanks,
Martin
2015-09-02 Martin Jambor <mjam...@suse.cz>
* omp-low.c (target_follows_kernelizable_pattern): Parallel
num_thread clause and non-automatic loop schedule preclude
kernelization.
---
gcc/ChangeLog.hsa | 6 ++++++
gcc/omp-low.c | 32 ++++++++++++++++++++++++++++++--
2 files changed, 36 insertions(+), 2 deletions(-)
diff --git a/gcc/omp-low.c b/gcc/omp-low.c
index 6c2bbe7..d6c521f 100644
--- a/gcc/omp-low.c
+++ b/gcc/omp-low.c
@@ -2832,9 +2832,23 @@ target_follows_kernelizable_pattern (gomp_target
*target, tree *group_size_p,
gomp_parallel *par;
if (!stmt || !(par = dyn_cast <gomp_parallel *> (stmt)))
return NULL;
+
+ tree clauses = gimple_omp_parallel_clauses (par);
+ tree num_threads_clause = find_omp_clause (clauses, OMP_CLAUSE_NUM_THREADS);
+ if (num_threads_clause)
+ {
+ if (dump_enabled_p ())
+ dump_printf_loc (MSG_NOTE, tloc,
+ "Will not turn target construct into a "
+ "simple GPGPU kernel because there is a num_threads "
+ "clause of the parallel construct that "
+ "is likely to require looping \n");
+ return NULL;
+ }
+
stmt = single_stmt_in_seq_skip_bind (gimple_omp_body (par), tloc,
"parallel");
- /* FIXME: We are currently ignoring parallel clauses and potentially also
- sharing clauses of teams and distribute, if there are any. We need to
+ /* FIXME: We are currently ignoring parallel sharing clauses and potentially
+ also sharing clauses of teams and distribute, if there are any. We need to
check they can be skipped. */
gomp_for *gfor;
if (!stmt || !(gfor = dyn_cast <gomp_for *> (stmt)))
@@ -2859,6 +2873,20 @@ target_follows_kernelizable_pattern (gomp_target
*target, tree *group_size_p,
return NULL;
}
+ clauses = gimple_omp_for_clauses (gfor);
+ tree for_sched_clause = find_omp_clause (clauses, OMP_CLAUSE_SCHEDULE);
+
+ if (for_sched_clause
+ && OMP_CLAUSE_SCHEDULE_KIND (for_sched_clause) !=
OMP_CLAUSE_SCHEDULE_AUTO)
+ {
+ if (dump_enabled_p ())
+ dump_printf_loc (MSG_NOTE, tloc,
+ "Will not turn target construct into a simple GPGPU "
+ "kernel because the inner loop has non-automatic "
+ "scheduling clause\n");
+ return NULL;
+ }
+
if (teams)
gather_inner_locals (gimple_omp_body (teams), kri);
if (dist)
--
2.4.6
