From: "Oliver Xymoron" <oxymoron@waste.org>
> On Wed, Aug 21, 2002 at 12:17:26AM +0200, johan.adolfsson@axis.com wrote:
>
> > I just compared the generated asm:
> > Accurate timestamp scaled to ns: 45 instructions (resolution actually 40
ns)
> > Approximate 40 ns resolution: 21 instructions
> > Approximate 40 us resolution: 9 instructions
> > For comparison one syscall path (gettimeofday()) is approx 400
instructions
> > and the add_timer_randomness() function that only uses jiffies is 76
> > instructions, so mayby I'm microoptimising here?
> > Is it worth the cycles to get 40 ns resolution instead of 40us ?
>
> Seems like it's probably worth the effort. In practice, such
> difference often are lost in the noise compared to cache flushes, etc.
> Does the 'correct' code suffer branch penalties or the like that might
> make it significantly worse than the quick code? If not, then I'd say
> definitely use it.
The correct code has two potential branches instead of one and also
need to stack one register when I have them in functions, but that might
change if the function is inlined in add_timer_randomness().
I can shave off a few instruction if I don't scale it to ns but instead
transform it to a plain 25MHz counter which is what we want anyway I guess.
BTW: I think the trust_pct approach looks nice adn I hope it gets included.
/Johan
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
This archive was generated by hypermail 2b29 : Fri Aug 23 2002 - 22:00:22 EST