https://gcc.gnu.org/bugzilla/show_bug.cgi?id=122577
--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> --- So RTL expansion only ever reduces targets to precision (via REDUCE_BIT_FIELD), so arguably a conversion from non-mode precision does not need to be restricted and we'll catch that on other operations, for example in vectorizable_operation.
