I wrote:
> On a longer-term basis, I'm looking into what we could do with
> extracting stats from subqueries, but that doesn't seem like material
> for a backpatch. I have a draft patch that I've been playing with
> (attached).
I've committed a heavily rewritten version of that patch. Git HEAD
seems to do reasonably well on the test case you gave at the start of
this thread. I'm not sure yet how well the logic will hold up in
real-world queries as opposed to simplified test cases, but give it
a try.
regards, tom lane