Re: BUG #14676: neqsel is NULL dumb

I wrote:
> marko(at)joh(dot)to writes:
>> This seems to be because neqsel() doesn't take at all into account that both
>> operators will exclude NULL rows, and does a simple 1.0 - eqsel().

> Yeah, that's clearly broken. A localized fix would be to re-fetch the
> nullfrac statistic and subtract it off, but that seems pretty inefficient.
> I'd be inclined to refactor things so that eqsel() and neqsel() call a
> common routine that takes a "bool negate" argument, about like the way the
> patternsel() functions work, and then the common routine could handle
> the nullfrac correctly for both cases.

Here's a proposed patch for that. I also fixed patternsel, but elected
not to mess with neqjoinsel; partly because I think joining on inequality
is so rare as to not be worth sweating over, and partly because I wasn't
too sure what's the appropriate correction, especially for semijoins.

Although this is clearly a bug fix, I'm leaning towards committing it
only in HEAD; given the lack of previous field complaints, I fear
back-patching might yield more complaints about destabilizing plans
than kudos for fixing things.

regards, tom lane