Re: [RFC][PATCH] Introduce CLOCK_REALTIME_COARSE

[Arjan van de Ven <arjan@infradead.org>]

On Sat, 18 Jul 2009 23:00:55 -0400
Chris Snook <chris.snook@gmail.com> wrote:

> On Fri, Jul 17, 2009 at 7:39 PM, john stultz<johnstul@us.ibm.com>
> wrote:
> > -       if (likely(gtod->sysctl_enabled && gtod->clock.vread))
> > +       if (likely(gtod->sysctl_enabled))
> 
> This irks me.  If the sysctl is enabled and the codepath is getting
> used often enough that we care about performance, branch prediction
> should do the right thing without compiler hints.  On the other hand,
> if the sysctl is disabled, and the compiler is telling the cpu to
> ignore its branch predictor, it'll hurt.  I don't think we should be
> wrapping (un)likely annotations around configuration options, unless
> we're biasing against debug conditions where we definitely don't care
> about performance.  The patch is certainly no worse than the existing
> code, but while we have the hood up, it might be nice to remove the
> annotation, unless we're sure that it does no harm, and does some
> good.

it's on x86.. likely/unlikely don't impact the CPU (since there are no
"ignore the branch predictor" hints), only the code placement.....

(and that's probably a good thing; CPU branch predictors are pretty
good, I'd not be surprised if they're at least as good as the
programmers who think how they code is used)


--
Arjan van de Ven 	Intel Open Source Technology Centre
For development, discussion and tips for power savings, 
visit http://www.lesswatts.org