We never ran into this before, because replace_osuf was almost always
False (i.e. we were almost always using the vanilla way when using
bytecode, so adjust_ul wasn't needed). Now that we can have the dynamic
way on by default many tests were failing because of it.