Re: [sqlsmith] ERROR: badly formatted node string "RESTRICTINFO...

I wrote:
> Apparently, postgres_fdw is trying to store RestrictInfos in the
> fdw_private field of a ForeignScan node. That won't do; those aren't
> supposed to be present in a finished plan tree, so there's no readfuncs.c
> support for them (and we're not adding it).

> Don't know if this is a new bug, or ancient but not previously reachable.
> It seems to be nearly the inverse of the problem you found yesterday,
> in which postgres_fdw was stripping RestrictInfos sooner than it really
> ought to. Apparently that whole business needs a fresh look.

Attached is a proposed patch that cleans up the mess here --- and it was
a mess. The comments for PgFdwRelationInfo claimed that the remote_conds
and local_conds fields contained bare clauses, but actually they could
contain either bare clauses or RestrictInfos or both, depending on where
the clauses had come from. And there was some seriously obscure and
undocumented logic around how the fdw_recheck_quals got set up for a
foreign join node. (BTW, said logic is assuming that no EPQ recheck is
needed for a foreign join. I commented it to that effect, but I'm not
very sure that I believe it. If there's a bug there, though, it's
independent of the immediate problem.)

Anybody want to review this, or shall I just push it?

(BTW, I've not yet looked to see if this needs to be back-ported.)

regards, tom lane