https://gcc.gnu.org/bugzilla/show_bug.cgi?id=123343
--- Comment #10 from Richard Biener <rguenth at gcc dot gnu.org> --- It's in-line with the previous comments - SLP for VLA ISAs is difficult, but we do not cost single-lane vs. multi-lane SLP but only fall back to single-lane SLP if analysis of multi-lane SLP failed. There's "heuristics" with respect to load/store-lane vs. SLP but indeed currently no heuristics that would apply this to reductions (but there's SLP loads there?).
