Re: Allowing join removals for more join types

David Rowley <dgrowley(at)gmail(dot)com> writes:
> On 6 July 2014 03:20, Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> wrote:
>> Just to note that I've started looking at this, and I've detected a rather
>> significant omission: there's no check that the join operator has anything
>> to do with the subquery's grouping operator.

> hmm, good point. If I understand this correctly we can just ensure that the
> same operator is used for both the grouping and the join condition.

Well, that's sort of the zero-order solution, but it doesn't work if the
join operators are cross-type.

I poked around to see if we didn't have some code already for that, and
soon found that not only do we have such code (equality_ops_are_compatible)
but actually almost this entire patch duplicates logic that already exists
in optimizer/util/pathnode.c, to wit create_unique_path's subroutines
query_is_distinct_for et al. So I'm thinking what this needs to turn into
is an exercise in refactoring to allow that logic to be used for both
purposes.

I notice that create_unique_path is not paying attention to the question
of whether the subselect's tlist contains SRFs or volatile functions.
It's possible that that's a pre-existing bug.

regards, tom lane