https://gcc.gnu.org/bugzilla/show_bug.cgi?id=120996
--- Comment #14 from Richard Biener <rguenth at gcc dot gnu.org> --- The different handling of * 1. vs * 0. w/o -ffast-math makes SLP discovery difficult (that cannot currently compensate by inserting neutral element operations, thus re-insert the * 1.). Not sure how the rev in question changed this though.
