On Tue, Sep 22, 2020 at 09:11:48AM +0200, Tobias Burnus wrote:
> > Okay – or do we find more issues?
I'm afraid so.
> + if (omp_declare_target_fn_p (decl)
> + || lookup_attribute ("omp declare target host",
> + DECL_ATTRIBUTES (decl)))
> + return NULL_TREE;
I'm worried that omp_declare_target_fn_p could be true and so this would
punt, but the intermediate aliases would be marked.
Or the aliases would be marked and the ultimate alias would not.
Consider:
int v;
#pragma omp declare target to (v)
void foo (void) { v++; }
void bar (void) __attribute__((alias ("foo")));
#pragma omp declare target to (bar)
void baz (void) __attribute__((alias ("foo")));
void qux (void) {
#pragma omp target
{
bar (); // Here the ultimate alias is not marked, so the code marks it,
// and adds another "omp declare target" attribute to bar,
// which it shouldn't.
baz (); // At this point, foo is marked, so the code wouldn't mark
// baz alias as "omp declare target".
}
}
So, I think it is fine to find the ultimate alias, but the loop to mark
the intermediate aliases should be invoked regardless of how decl is or is
not marked, and should test in each step whether it should or should not be
marked.
Jakub
