On Sat, Dec 12, 2020 at 08:55:01AM +0100, Jakub Jelinek via Gcc-patches wrote:
> Actually, seems we diagnose goto out of it, so perhaps for trunk
> the best thing forward is to add the noexcept block around target data
> body, but I think we still don't really need the omp return, omp-expand
> doesn't really care, all interesting happens during gimplification and
> omplower, and adding the noexcept block is probably inappropriate for
> release branches.
That would be
2020-12-12 Jakub Jelinek <ja...@redhat.com>
* omp-low.c (lower_omp_target): Wrap even data region body
into maybe_catch_exception.
--- gcc/omp-low.c.jj 2020-12-12 08:35:51.182131932 +0100
+++ gcc/omp-low.c 2020-12-12 17:31:31.392537247 +0100
@@ -12978,11 +12978,10 @@ lower_omp_target (gimple_stmt_iterator *
gimple_seq_add_seq (&new_body, tgt_body);
gimple_seq_add_seq (&new_body, join_seq);
+ if (offloaded || data_region)
+ new_body = maybe_catch_exception (new_body);
if (offloaded)
- {
- new_body = maybe_catch_exception (new_body);
- gimple_seq_add_stmt (&new_body, gimple_build_omp_return (false));
- }
+ gimple_seq_add_stmt (&new_body, gimple_build_omp_return (false));
gimple_omp_set_body (stmt, new_body);
}
but thinking about it more, nothing in the OpenMP standard requires that
terminate is called, it is undefined behavior instead, and handling it the
way users expect is perhaps a reasonable choice when we easily can.
Testcase would be:
extern "C" void abort ();
int
main ()
{
int x = 1;
try
{
#pragma omp target data map (to: x)
{
throw 1;
}
}
catch (int y)
{
if (y == 1)
return 0;
}
abort ();
}
Tobias/Thomas, your thoughts on this?
Jakub
