CodeGen_ARM already does this (see here).

x86's options are far more limited (only i32 -> u8 and i32 -> i8 are possible), but a pair of saturating narrow instructions will be better than what we currently generate. Other backends might also benefit from this, but I haven't looked into it.