https://gcc.gnu.org/bugzilla/show_bug.cgi?id=122869
--- Comment #4 from Robin Dapp <rdapp at gcc dot gnu.org> ---
My hunch is that an extra
(set (match_operand:SI 8 "register_operand" "=r, r, r, r")
(reg:SI VL_REGNUM))]
or similar would work, where operand[8] would be &d. That way the assignment
should be atomic enough?
When doing that we couldn't reuse use_contiguous_load_insn but would need a
preparation function but that shouldn't be a big deal.
Jeff, if you want me to take it, feel free to assign it as I have some state
now.