https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81611
--- Comment #16 from Alexandre Oliva <aoliva at gcc dot gnu.org> --- Even if create_mem_ref_raw created a MEM_REF, we'd still allocate a new pseudo for the reg - 1 at cfgexpand, and that ends up preventing the post_inc addressing mode from being selected. The more I think about it, the more I conclude we have to bite the bullet and support post_inc addressing modes in auto_inc_dec, even with increments saved on another pseudo before the memory uses. Not just for iv uses, but also for linear code, mainly because the IR we generate for post-inc, all the way from the beginning, makes it more likely that the increment will be logically before the use of the old value as an address. I was trying to avoid that, but at this point anything else feels like papering over the problem. A smarter addressing-mode-aware middle end could tentatively generate off-by-1 base addresses so as to enable otherwise invalid addresses and subsequent post_inc transformations, but... if the transformation doesn't take place, we may end up worse off.
