On 2018-09-26 18:31:07 -0400, Tom Lane wrote:
> Andres Freund <andres@anarazel.de> writes:
> > On 2018-09-26 17:41:36 -0400, Tom Lane wrote:
> >> Well, if you're unhappy about snprintf.c's performance, you could review
> >> https://commitfest.postgresql.org/19/1763/
> >> so I can push that. In my tests, that got us down to circa 10% penalty
> >> for float conversions.
>
> > Uh, I can do that, but the fact remains that your commit slowed down
> > COPY and other conversion intensive workloads by a *significant* amount.
>
> [ shrug... ] There are other cases that got faster (particularly after
> the above-mentioned patch). I do not wish to consider floating-point
> conversion speed as the sole figure of merit for this change. If we
> are to consider only the worst-case, we should be reverting JIT.
Oh, come on. One can be disabled with a GUC, has (although not good
enough) intelligence when it switches on, the other has ... none of
that. Obviously performance is always a balancing act, but you'd be
pretty pissed at anybody else regressing performance in a non-fringe
case, and then refused responsibility. And as I said, I'm willing to
help.
Greetings,
Andres Freund