Re: Issue with point_ops and NaN

At Thu, 01 Apr 2021 09:34:40 +0900 (JST), Kyotaro Horiguchi <horikyota(dot)ntt(at)gmail(dot)com> wrote in
> I have to change almost all boolean-returning functions to
> tri-state-boolean ones. I'll give it try a bit futther.

The attached is a rush work of that, on top of the (rebased version of
the) base patch. Disregarding its uneffectiveness, it gives a rough
estimate of how large that would be and how that affects other parts.

Maybe one of the largest issue with that would be that GiST doesn't
seem to like NULL to be returned from comparison functions.

regression=# set enable_seqscan to off;
regression=# set enable_indexscan to on;
regression=# SELECT * FROM circle_tbl WHERE f1 && circle(point(1,-2), 1) ORDER BY area(f1);
ERROR: function 0x9d7bf6 returned NULL
(function 0x9d7bf6 is box_overlap())

That seems like the reason not to make the functions tri-state.

regards.

--
Kyotaro Horiguchi
NTT Open Source Software Center