https://gcc.gnu.org/bugzilla/show_bug.cgi?id=121383
--- Comment #3 from Tamar Christina <tnfchris at gcc dot gnu.org> ---
(In reply to Richard Biener from comment #2)
> if-conversion does nothing to loops with multiple exits. One could think to
> sub-divide its work, but in this case there's a lack of PHIs and the
> expected if-conversion is on control dependences rather than data
> dependences.
>
> I would rather see this as an opportunity to look how to handle such cases
> in the vectorizer itself?
I had thought about that as well and that was my initial approach, however I
figured this particular case fits better in ifcvt.
I wanted ifcvt to split
int foo ()
{
for (int i = 0; i < N; i++)
{
if (a[i] > b[i] && a[i] > c[i])
return 1;
}
return 0;
}
into essentially
int foo ()
{
for (int i = 0; i < N; i++)
{
bool v = a[i] > b[i] && a[i] > c[i];
if (v != 0)
return 1;
}
return 0;
}
which because a, b and c all need to be speculated should be safe to do,
and I had a small prototype where it mostly worked.
We *could* do it in the vectorizer, but that means that every statement in
the second block needs to be predicated. So it'll only work for masked targets
whereas I think the original loop should work for non-masked ones too.
But will give the in-vectorizer one a try. I'll limit it only to if blocks for
now
rather than if-then-else as well.
