https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98235
--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> ---
Eh, so this is fully LSMed into a "PHI state machine" we try to vectorize.
Since the SLP discovery cache is not "permute aware" (thus it does not find
an existing SLP node for say {a, b} when looking for {b, a}) the discovery
follows backedges and re-discovers the flow with a different permutation.
That's the opportunity to change how we deal with permute limits as well I
guess.
