Depending on specific values of restart_lsn or pg_current_wal_lsn()
is obviously unsafe.  The previous coding tried to dodge this issue
by rounding off, but that's not good enough, as shown by multiple
buildfarm members.  Nuke all the uses of these values except for
null-ness checks, pending some credible argument why we should think
something else could be usefully stable.

Kyotaro Horiguchi, further modified by me

Discussion: https://postgr.es/m/E1jM1Sa-0003mS-99@gemulon.postgresql.org