Re: Binary search in fmgr_isbuiltin() is a bottleneck.

On 2017-09-27 13:46:50 -0400, Tom Lane wrote:
> Andres Freund <andres(at)anarazel(dot)de> writes:
> > I'd been wondering about that too, but I'm not sure I buy that it's
> > worth the effort. The only real argument I see is that there's probably
> > multiple cases where it'd be potentially beneficial, not just here.
>
> The other question that ought to be answered is whether a gperf hash
> table would be faster. In principle it could be because of
> guaranteed-no-collisions, but I have no experience with how fast the
> constructed hash functions might be compared to our regular one.

The patch uses murmurhash32, i.e. a short and fast hash, already for
performance, and it shows up in profiles.

Ugh, hacking together a quick input file for gperf, I'm *far* from
convinced. The generated code does multiple lookups in significantly
sized arrays, and assumes string input. The latter seems like a complete
dealbreaker, and there doesn't seem to be an option to turn it off.

Greetings,

Andres Freund