https://gcc.gnu.org/bugzilla/show_bug.cgi?id=126550
--- Comment #3 from Robin Dapp <rdapp at gcc dot gnu.org> --- Possibly two issues. We expand the const permute index through a VLA route. I don't think that's useful for 99% of the cases, in particular if we load from the literal pool anyway... (and could have loaded the constant directly). But that's just a missed optimization. The second issue is that the VLA route for this constant is bogus.
