[Bug tree-optimization/115629] Inefficient if-convert of masked conditionals

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

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115629

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2024-06-25
     Ever confirmed|0                           |1

--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
I think the transform is profitable for scalar for all cases as well.

Now, if-conversion records BB predicates and uses those to predicate
stmts/PHIs, I'm not sure how to go optimize it on that level.

It's more a CFG transform of a if (A) x; else { if (B) x; else y; } to
if (A || B) x; else y;  with the downside that in this case evaluating
'B' has side-effects (memory dereference which can trap).

Due to the trapping this needs to be somehow done in if-conversion
where you also need to determine that x == x' (both copies of x
evaluate to the same code).  Some scalar GIMPLE opt could have
de-duplicated the BB with 'x' (tail-merging to the rescue?  it considers
the cands but somehow doesn't merge - I think the PHI node isn't
supported).