https://gcc.gnu.org/bugzilla/show_bug.cgi?id=69653
--- Comment #1 from amker at gcc dot gnu.org --- Since tree optimizer cannot handle cse opportunities as described here, backend needs to force register scaling expr (instead of loop invariant part address expression) out of memory reference, hoping RTL CSE can catch these. It would be better handled in tree optimizer because backend legitimize_address doesn't have context information and this is done at a cost of missing loop invariant opportunities.
