Re: [sqlsmith] Failed to generate plan on lateral subqueries

I wrote:
> As threatened, here's a patch on top of that that gets rid of
> LateralJoinInfo. I'm pretty happy with this, although I have an
> itchy feeling that we could dispense with the lateral_vars lists too.

I experimented with that and figured out what was bothering me: there is
no need for add_placeholders_to_base_rels to fool around with adding items
to lateral_vars anymore, because the code in create_lateral_join_info
that adds placeholders' ph_lateral bits to the evaluation locations'
lateral_relids sets now does everything that we needed to have happen.
We don't care exactly what's inside the PHV expression, only which rels
it comes from, and the ph_lateral bitmap is sufficient for that.

We still need the lateral_vars field in RelOptInfo, but it now exists
purely to carry the results of extract_lateral_references forward to
create_lateral_join_info. (We wouldn't need even that so far as Vars are
concerned, because we could just add their source rel to the referencing
rel's lateral_relids on-the-fly in extract_lateral_references. But we
can't handle PHVs that way, because at that stage we don't know precisely
where they'll be evaluated, so we don't know what to add to the
referencer's lateral_relids.)

Updated patch attached; compared to the previous one, this just removes a
big chunk of code in add_placeholders_to_base_rels and updates a couple of
related comments.

regards, tom lane