Re: Reduce LEFT/FULL JOIN to ANTI JOIN in more cases

I can't help pondering how 0001 collects the quals used to prove a
forced-null Var non-nullable. Pass 1 gathers them into every
subtree's state so that pass 2 doesn't have to re-walk the jointree at
decision time.

Two things bother me about that. First, we already have this
traversal: find_subquery_safe_quals() does the same job for
query_outputs_are_not_nullable(). Two copies with the same rule seems
like something we should avoid. Second, pass 1 builds a list at every
FromExpr and every inner or semi join, for every query containing an
outer join, and each level copies what its children collected, so in a
deep chain of inner joins the same quals get copied over and over on
the way up. What is worse, if there are no forced-null Vars in upper
quals (which is quite common), all of the lists we build in pass 1 are
just wasted.

So I think a better approach is to keep just the jointree node in the
pass 1 state, and have pass 2 call find_subquery_safe_quals() where it
needs the proof. The walk is cheap enough to do on demand.

Attached is a patch showing what I mean.

- Richard