> Your test suggests that the performance differential is *at most*
> 2X --- probably much less in real-world situations where the disk
> pages aren't already cached. I can't get excited about introducing
> platform-dependent behavior and overflow risk for that. If it were
> 10X then I would, but right now I think we are OK as is. I think
> any speedup efforts here would be better put into making NUMERIC
> ops go faster ...
Another followup: on 7.0.2, with different optimizations etc,
sum(float8) takes 1.95 seconds, rather than the 5.2 on the current tree.
I'd better look at the compilation optimizations; is there another
explanation for the factor of 2.6 difference (!!)?
So I'd expect int4 to be closer to float8 in performance than my
previous mail suggested.
- Thomas