https://gcc.gnu.org/bugzilla/show_bug.cgi?id=55212
--- Comment #369 from Oleg Endo <olegendo at gcc dot gnu.org> ---
(In reply to Oleg Endo from comment #346)
>
> ... I've noticed that this is the same as the existing
> MAYBE_BASE_REGISTER_RTX_P.
>
> I've inserted a patch into the stash to tighten all the existing memory
> predicates and constraints to use MAYBE_BASE_REGISTER_RTX_P and
> MAYBE_INDEX_REGISTER_RTX_P. As a result your added check in
> 'general_mov{dst,src}_operand' becomes unnecessary, but I've left it in
> anyway.
>
> This is the patch
> https://github.com/olegendo/gcc/commit/
> 9636a55a071fd9b9acc3a38fb25799fc70177d28
>
> I think this kind of thing could be already upstreamed after reg-testing
> without LRA, it's a good catch.
I've tested this patch on mainline with LRA disabled and it doesn't work as-is.
I'm now trying to modify the patch.
If it doesn't work, then maybe we have to consider to drop non-LRA support
entirely as part of the migration to LRA. I was hoping to avoid that, and keep
the '-mno-lra' option around for one or two major versions ...
