https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112325
--- Comment #5 from Richard Biener <rguenth at gcc dot gnu.org> --- Yes, as I said in comment#2. Note I specifically ended up not open-coding the reduction because of concerns of efficiency. So a target should only provide reduc_*_scal patterns when they are more efficient than open-coding with the obvious sequence. That possibly means fixing the issue I mentioned in comment#2 will not be enough to get this BB vectorized on x86.