march question on pentium4

march question on pentium4

[Martin Kahlert]

Hi!
I use gcc-3.2 release on a FreeBSD system running on a Pentium 4 machine.

My (integer based, no floating point at all) code shows a strange behaviour:
It is faster when compiled with -march=pentium3 than
with -march=pentium4. Is this a known issue or a problem that should be 
investigated further?

Thanks in advance.

Regards
Martin.

-- 
The early bird catches the worm. If you want something else for       
breakfast, get up later.

Re: march question on pentium4

[Tim Prince]

On Thursday 29 August 2002 23:04, Martin Kahlert wrote:
> Hi!
> I use gcc-3.2 release on a FreeBSD system running on a Pentium 4 machine.
>
> My (integer based, no floating point at all) code shows a strange
> behaviour: It is faster when compiled with -march=pentium3 than
> with -march=pentium4. Is this a known issue or a problem that should be
> investigated further?
>
If I am to speculate without an example, the pentium4 costs for shift and 
multiply are set so high that the compiler will always use the alternative of 
add sequences.  Intel's P4 Optimization Guide suggests they be set only high 
enough that a shift costs more than 3 adds, for example.  You could be 
getting excessive code expansion for multiplication by constant.  
While the expectation of gcc would be that the costs reflect accurately the 
performance of the instructions, high costs which produce long code may not 
be best in practice.
Evidently not speaking for anyone.
-- 
Tim Prince

Re: march question on pentium4

[Joe Buck]

On Thursday 29 August 2002 23:04, Martin Kahlert wrote:
> > Hi!
> > I use gcc-3.2 release on a FreeBSD system running on a Pentium 4 machine.
> >
> > My (integer based, no floating point at all) code shows a strange
> > behaviour: It is faster when compiled with -march=pentium3 than
> > with -march=pentium4. Is this a known issue or a problem that should be
> > investigated further?

Tim Prince writes:

> If I am to speculate without an example, the pentium4 costs for shift
> and multiply are set so high that the compiler will always use the
> alternative of add sequences...

It would be better, I think, to look at the actual example and determine
if there are any outright errors in costs for the P4 or P3.

Re: march question on pentium4

[Tim Prince]

On Friday 30 August 2002 12:08, Joe Buck wrote:
> On Thursday 29 August 2002 23:04, Martin Kahlert wrote:
> > > Hi!
> > > I use gcc-3.2 release on a FreeBSD system running on a Pentium 4
> > > machine.
> > >
> > > My (integer based, no floating point at all) code shows a strange
> > > behaviour: It is faster when compiled with -march=pentium3 than
> > > with -march=pentium4. Is this a known issue or a problem that should be
> > > investigated further?
>
> Tim Prince writes:
> > If I am to speculate without an example, the pentium4 costs for shift
> > and multiply are set so high that the compiler will always use the
> > alternative of add sequences...
>
> It would be better, I think, to look at the actual example and determine
> if there are any outright errors in costs for the P4 or P3.
This is not a question of outright error.  The point at which the switch 
between shift (or multiply) and add sequences should be made depends on more 
factors than the basic costs of the operations, and on the specific model of 
the CPU.  Extremely large code expansion in add sequences, even though it is 
in agreement with the basic costs, will not produce the expected performance 
benefit in practical context.  Even though a P4 shift costs much more than 4 
adds, in isolation, in practice it is seldom useful to use more than 3 adds 
to replace a shift.  The attempt to use the actual costs of the operations 
may be the right way on a processor which has no out-of-order facility.
-- 
Tim Prince