Re: Which qsort is used

Tom,

On 12/12/05 2:47 PM, "Tom Lane" <tgl(at)sss(dot)pgh(dot)pa(dot)us> wrote:

> As those results suggest, there can be huge differences in sort
> performance depending on whether the input is random, nearly sorted,
> nearly reverse sorted, possesses many equal keys, etc. It would be very
> dangerous to say "implementation A is better than implementation B"
> without having tested all those scenarios.

Yes. The Linux glibc qsort is proven terrible in the general case by these
examples though.

Bruce's point on that thread was that we shouldn't be replacing the OS
routine in the general case. On the other hand, there is the precedent of
replacing Solaris' routine with the NetBSD version.

Based on the current testing, I think it would be a good idea to expose a
"--with-qsort" option in configure to allow for it's selection as suggested
by other posters.

> IIRC, the reason we reject
> Solaris' qsort is not that it is so bad in the typical case, but that it
> has some horrible corner-case behaviors.

Do you have a test suite you can recommend with those edge cases? I built
the one in the bizgres-general thread based on edge cases for Solaris that I
found on a sun mailing list. The edge case referred to there was the all
zero one, which does seem to have a significant advantage in the NetBSD.

- Luke