Обсуждение: BUG #18835: spgist index fails to accept point with NaN

The following bug has been logged on the website:

Bug reference:      18835
Logged by:          Alexander Lakhin
Email address:      exclusion@gmail.com
PostgreSQL version: 17.4
Operating system:   Ubuntu 24.04
Description:

The following script:
CREATE TABLE t(p point);
INSERT INTO t SELECT point(g, g) FROM generate_series(1, 1000) g;
INSERT INTO t VALUES ('(0,NaN)'::point);

CREATE INDEX spgist_idx ON t USING spgist(p);
triggers:
ERROR:  XX000: getQuadrant: impossible case
LOCATION:  getQuadrant, spgquadtreeproc.c:77

Reproduced starting from 8daeb5ddd.

This anomaly is discovered with SQLsmith.

PG Bug reporting form <noreply@postgresql.org> writes:
> The following script:
> CREATE TABLE t(p point);
> INSERT INTO t SELECT point(g, g) FROM generate_series(1, 1000) g;
> INSERT INTO t VALUES ('(0,NaN)'::point);

> CREATE INDEX spgist_idx ON t USING spgist(p);
> triggers:
> ERROR:  XX000: getQuadrant: impossible case

Yeah.  There's been some past attempts to clean up our behavior
for NaNs in geometric values, but it's such a mess that not
much progress has been made.  So I can't get excited about
worrying about this case in particular (especially since I don't
see any easy fix).  I wonder if we shouldn't just ban NaNs in
the geometric types.

            regards, tom lane