https://gcc.gnu.org/bugzilla/show_bug.cgi?id=126306
--- Comment #3 from Richard Biener <rguenth at gcc dot gnu.org> --- So this is only the first step. There is more inefficiency here in that we consider splitting only when there are excess elements for the vector type chosen, but for example for a group size of 6, if we chose v2<mode> we will not split but with v4<mode> we would and possibly get the better v4<mode> + v2<mode> parts vectorization.
