https://gcc.gnu.org/bugzilla/show_bug.cgi?id=119577
--- Comment #4 from rguenther at suse dot de <rguenther at suse dot de> --- On Mon, 28 Apr 2025, rdapp at gcc dot gnu.org wrote: > https://gcc.gnu.org/bugzilla/show_bug.cgi?id=119577 > > --- Comment #3 from Robin Dapp <rdapp at gcc dot gnu.org> --- > I manage to have a quick look at the code now. It looks like we force live > every induction and build slp instances for the IV increments. > > I don't think adjusting the actual IV creation in vectorizable_induction would > be too involved, at least for the length-controlled case, but I wonder what to > do with the increment. > > Right now we just vectorize it with the fitting vector type (this limits VF > for > some of the examples I'm looking at BTW where we could vectorize at a higher > VF > without the vectorized int induction). > > Is there precedent of how to handle such a situation? I.e. can we keep track > of the affected stmts in the slp isntance and then "vectorize" with a scalar > type? Or would we need to circumvent that altogether and adjust the existing > scalar IV? Or rather create a new scalar IV and re-wire the uses? There is precedent of re-using a scalar IV with the in-order (fold-left) reduction vectorization which keeps the scalar reduction IV.
