Cygnus Supported GCC vs. EGCS GCC

Cygnus Supported GCC vs. EGCS GCC

[Fred Richardson]

Hello-

I have the following question about the Cygnus Support version of GCC
and the EGCS version on the Sun "Ultra 1 Model 170".  I compiled and
ran an optimized version of a program with the following (approximate)
relative speed performances: 

COMPILER                                    SPEED
===================================================================
/opt/SUNWspro/bin/cc -fast:                 1.0         (baseline)
FSF-GCC 2.7.2  (-02):                       1.4
EGCS-GCC 971008 (-mcpu=ultrasparc -O2)      1.4


So it appears that FSF-GCC 2.7.2 and EGCS-GCC 971008 are both about
40% slower than the Sun compiler when using optimzation.  

Does anyone know if I should I expect about the same 40% performance
hit from the Cygnus Support version of GCC?  Put differently, does
anyone know if Cygnus has implemented more specific optimization for
the Ultra Sparc processor?

Another issue which I'm confused about is this: is there a common
branch point for Cygnus-GCC and EGCS-GCC?  If there is, how long ago
were the two versions the same?

        Thanks very much (in advance) for any answers to these
        questions, 

                -Fred Richardson

Re: Cygnus Supported GCC vs. EGCS GCC

[Jeffrey A Law]

  In message < 199710140306.XAA29034@bluebell.bbn.com >you write:
  > Does anyone know if I should I expect about the same 40% performance
  > hit from the Cygnus Support version of GCC?
Someone who deals with sparcs would have to answer this.  Doug?

  > Another issue which I'm confused about is this: is there a common
  > branch point for Cygnus-GCC and EGCS-GCC?  If there is, how long ago
  > were the two versions the same?
No, Cygnus gcc is still based on gcc2, not egcs.

jeff

Re: Cygnus Supported GCC vs. EGCS GCC

[Doug Evans]

   Date: Tue, 14 Oct 1997 11:27:30 -0600
   From: Jeffrey A Law <law@cygnus.com>

     In message < 199710140306.XAA29034@bluebell.bbn.com >you write:
     > Does anyone know if I should I expect about the same 40% performance
     > hit from the Cygnus Support version of GCC?
   Someone who deals with sparcs would have to answer this.  Doug?

Clearly one would want to be able to try the benchmark for oneself
if one was expected to give definitive answers.
Secondly, I don't have access to a recent version of Sunsoft's cc.
But in general there is a lot of tuning work that could be done
to the sparc port [scheduling for example, and if someone wanted
to spend time with the Haifa scheduler then by all means do it].

Re: Cygnus Supported GCC vs. EGCS GCC

[Joe Buck]

> I have the following question about the Cygnus Support version of GCC
> and the EGCS version on the Sun "Ultra 1 Model 170".  I compiled and
> ran an optimized version of a program with the following (approximate)
> relative speed performances: 
> 
> COMPILER                                    SPEED
> ===================================================================
> /opt/SUNWspro/bin/cc -fast:                 1.0         (baseline)
> FSF-GCC 2.7.2  (-02):                       1.4
> EGCS-GCC 971008 (-mcpu=ultrasparc -O2)      1.4
> 
> 
> So it appears that FSF-GCC 2.7.2 and EGCS-GCC 971008 are both about
> 40% slower than the Sun compiler when using optimzation.  

No, it appears that gcc is about 40% slower than the Sun compiler on
this particular program.  If the time is dominated by a single inner
loop, it's possible that a single missed optimization is the cause.
If scheduling is a problem, you might try seeing if the Haifa scheduler
does a better job.

> Does anyone know if I should I expect about the same 40% performance
> hit from the Cygnus Support version of GCC?

Unless you help solve the problem (do some profiling, help find the
cause), probably yes.  (Of course with such a vague report the actual
answer is more like "Who knows?").

Re: Cygnus Supported GCC vs. EGCS GCC

[Jim Wilson]

	Does anyone know if I should I expect about the same 40% performance
	hit from the Cygnus Support version of GCC?

Yes.  You should expect roughly the same performance from Cygnus gcc releases
as you get from other gcc releases in this case.

	Put differently, does
	anyone know if Cygnus has implemented more specific optimization for
	the Ultra Sparc processor?

Nothing that isn't already in EGCS and/or the eventual gcc 2.8 release.

	Another issue which I'm confused about is this: is there a common
	branch point for Cygnus-GCC and EGCS-GCC?  If there is, how long ago
	were the two versions the same?

The are both based on the FSF development gcc sources, so they are
substantially similar.  They have never been exactly the same, because they
contain different patches on top of the FSF development sources.  The plan
is to eventually eliminate these differences, but this will take some time.

	/opt/SUNWspro/bin/cc -fast:                 1.0         (baseline)
	EGCS-GCC 971008 (-mcpu=ultrasparc -O2)      1.4

I assume this is an FP program?  Gcc has historically been weak on FP
performance.

The new haifa-scheduler should help here.  Did you try using --enable-haifa?
I don't think the ultrasparc port has been tuned for the haifa scheduler
yet though, so this may not give best possible performance as yet.

The -fast option uses a different startup file that tweaks some bits in
the FPU status register so as to give less precise but faster arithmetic.
It should be possible to do the same thing with gcc, it is just that no one
has done the work yet.  You can get some similar results with -ffast-math
but it won't have as much affect as Sun's -fast.

You should also try the -fomit-frame-pointer, -funroll-loops, and
-finline-function options.

Jim

Re: Cygnus Supported GCC vs. EGCS GCC

[Neal Becker]

>>>>> "Jim" == Jim Wilson <wilson@cygnus.com> writes:


    Jim> The new haifa-scheduler should help here.  Did you try using --enable-haifa?

Really?  Isn't haifa enabled by default?

Re: Cygnus Supported GCC vs. EGCS GCC

[Jim Wilson]

	Really?  Isn't haifa enabled by default?

No, because not all ports have been tuned for it yet, and it may give
slower code for ports that have been specifically tuned for the old scheduler.

The plan is to make it the default for all targets, but there is no schedule
for when this will happen.

There are some ports that enable haifa by default as long as --disable-haifa
wasn't used.  This appears to includes all of the HPPA ports, but none of
the others.

Jim

Re: Cygnus Supported GCC vs. EGCS GCC

[Jeffrey A Law]

  In message < u9zpob6al2.fsf@neal.ctd.comsat.com >you write:
  > >>>>> "Jim" == Jim Wilson <wilson@cygnus.com> writes:
  > 
  > 
  >     Jim> The new haifa-scheduler should help here.  Did you try using --ena
  > ble-haifa?
  > 
  > Really?  Isn't haifa enabled by default?
Only on the PA right now.

First, haifa problably won't help too much without retuning the scheduling
parameters.  I did this for the PA, and I think someone submitted them
for the sparc, but did so after we went into a "slush" (as opposed to 
a "freeze" :-)

Second, we really haven't beat on haifa as much on other targets yet --
though that's improved over the last several weeks with folks reporting
with & without haifa test results.

jeff

Re: Cygnus Supported GCC vs. EGCS GCC

[Craig Burley]

>I have the following question about the Cygnus Support version of GCC
>and the EGCS version on the Sun "Ultra 1 Model 170".  I compiled and
>ran an optimized version of a program with the following (approximate)
>relative speed performances: 
>
>COMPILER                                    SPEED
>===================================================================
>/opt/SUNWspro/bin/cc -fast:                 1.0         (baseline)
>FSF-GCC 2.7.2  (-02):                       1.4
>EGCS-GCC 971008 (-mcpu=ultrasparc -O2)      1.4
>
>So it appears that FSF-GCC 2.7.2 and EGCS-GCC 971008 are both about
>40% slower than the Sun compiler when using optimzation.  

If I believed the "SPEED" heading, I'd say it was 40% *faster*.
(That is, a *speed* of 14 mph would be faster, to me, than one
of 10 mph.)

Perhaps you should change your header to say "TIME"?  ;-)

        tq vm, (burley)