[Bug tree-optimization/111770] predicated loads inactive lane values not modelled

["rguenth at gcc dot gnu.org" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=111770

--- Comment #5 from Richard Biener <rguenth at gcc dot gnu.org> ---
(In reply to Alex Coplan from comment #4)
> (In reply to Richard Biener from comment #3)
> > Now, if-conversion could indeed elide the .COND_ADD for integers.  It's
> > problematic there only because of signed overflow undefinedness, so
> > you shouldn't see it for 'unsigned' already, and adding zero is safe.
> 
> Can you elaborate on this a bit? Do you mean to say that the .COND_ADD is
> only there to avoid if-conversion introducing UB due to signed overflow?

No, you are right.

> ISTM it's needed for correctness even without that, as the addend needn't be
> guaranteed to be zero in the general case.
> 
> > if-conversion would need to have an idea of all the ranges involved here
> > so it might be a bit sophisticated to get it right.
> 
> Does what I suggested above make any sense, or do you have in mind a
> different way of handling this in if-conversion? I'm wondering how ifcvt
> should determine that the addend is zero in the case where the predicate is
> false.

ifcvt would need to compute said fact, say, keep a lattice of the value
(or value-range) that's there when the block isn't executed (simulating
a disabled vector lane).  A load that's going to be replaced by a
.MASK_LOAD can be then known zero and this needs to be propagated through
regular stmts (like the multiply).  There's also .COND_* which if-conversion
could actually provide the else value for - like if we have a following
division to avoid dividing by zero.  But that would be propagating backwards
(I'd still have this usecase in mind).