I wrote:
> Richard Guo <guofenglinux@gmail.com> writes:
>> In b448f1c8 remove_useless_result_rtes will remove useless FromExprs and
>> merge its quals up to parent. This makes flag 'delay_upper_joins' not
>> necessary any more if the clauses between the two outer joins come from
>> FromExprs. However, if the clauses between the two outer joins come
>> from JoinExpr of an inner join, it seems we have holes in preserving
>> ordering.
> Hmm ... we'll preserve the ordering all right, but if we set commute_below
> or commute_above_x bits that don't match reality then we'll have trouble
> later with mis-marked varnullingrels, the same as we saw in b2d0e13a0.
> I don't think you need a JoinExpr, an intermediate multi-member FromExpr
> should have the same effect.
BTW, the presented test case doesn't fail anymore after the fix
for bug #17781. That's because build_joinrel_tlist() doesn't use
commute_above_l anymore at all, and is a bit more wary in its use of
commute_above_r. I'm not sure that that completely eliminates this
problem, but it at least makes it a lot harder to reach.
We might want to see if we can devise a new example (or wait for
Robins to break it ;-)) before expending a lot of effort on making
the commute_xxx bits more precise.
regards, tom lane