Re: 2.6.13/14 x86 Makefile - Pentiums penalized ?

[Margit Schubert-While <margitsw@t-online.de>]

At 09:59 14.09.2005 +0300, Denis Vlasenko wrote:
> On Wednesday 14 September 2005 09:48, Zwane Mwaikambo wrote:
> > On Wed, 14 Sep 2005, Chris White wrote:
> >
> > > That's correct, gcc 3.4 started the -mtune flag.  Chances are if you 
> really
> > > want the -mtune optimizations you're going to have to upgrade to gcc 
> 3.4 or
> > > greater.
> > >
> > > > This, of course, heavily penalizes P4's (the notorious inc/dec).
> > >
> > > Are you referring to cpu cycle counts?  Is there certain code that 
> causes the
> > > kernel to perform that unfavorably by a large scale?
> >
> > It's documented as being suboptimal to use inc/dec due to it modifying all
> > of eflags resulting in dependency related stalls. add/sub only modifies
> > one bit of eflags so is more optimal. However there is a problem of
>
> ?! add/sub doesn't modify "only one bit in eflags", it modifies all.
> In fact, it's dec/inc which does not modify all bits.
> It doesn't touch 'carry' bit (IIRC).
>
> If inc/dec is slower on P4, it must be just another P4 quirk.
>
> > increased code size with add/sub.
> >
> > But i've never benchmarked all of this ;)
>
> I don't even have one to test this.


In the Intel Architecture Optimization document it specifically states 
(Chapter 2.6) :
"Avoid instructions that unnecessarily introduce dependence-related
stalls: inc and dec instructions, .....".
And again on page 2-11 :
"The inc and dec instructions should always be avoided. Using add and
sub instructions instead avoids data dependence and improves performance".
And on page 2-71 :
"The inc and dec instructions modify only a subset of flags in the flag 
register.
This creates a dependence on all previous writes of the flag register.
This is especially problematic when these instructions are on the critical
path because they are used to change an address for a load on which
many other instructions depend. "

However, the kernel include and arch have a liberal sprinkling of inc/dec,
and AFAICT some of these in hot-path.

Margit







-
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/