I wrote:
> If you reverse out just that change you'll see why I added it: without it,
> the plan for the earlier "test a corner case in which we shouldn't apply
> the star-schema optimization" isn't optimized as much as I thought it
> should be.
Hmm ... looking at this closer, it seems like this patch probably breaks
what that regression test case was actually meant to test, ie once we've
deleted the VALUES subselects from the jointree, it's likely that the
planner join-ordering mistake that was testing for can no longer happen,
because the join just plain doesn't exist.
I'll plan to deal with that by running the test case with actual small
tables instead of VALUES subselects. It might be useful to run the test
case in its current form as an exercise for the RTE_RESULT optimizations,
but that would be separate from its current intention.
regards, tom lane
