[gentoo-dev] make.conf optimization flags for k6, athlon

[gentoo-dev] make.conf optimization flags for k6, athlon

[Dan Armak]

Hi all,

In make.conf we specify the optimization CXXFLAGS as follows:

# For Pentium and k6/2 machines 
# -----------------------------

#CHOST="i586-pc-linux-gnu"
#CFLAGS="-mpentium -march=pentium -O3 -pipe"
#CXXFLAGS="-mpentium -march=pentium -O3 -pipe"

Why don't we use -march=k6 -mcpu=k6? Wouldn't that allow use of mmx and 
3dnow, which a generic pentium doesn't have?


# For PII PIII and Athlon machines
# --------------------------------

#CHOST="i686-pc-linux-gnu"
#CFLAGS="-mpentiumpro -march=pentiumpro -O3 -pipe"
#CXXFLAGS="-mpentiumpro -march=pentiumpro -O3 -pipe"

The same here: I believe (though I'm not sure as I don't have an Athlon) that 
the flags -mcpu=athlon -march=athlon might be of use.

Dan Armak

Re: [gentoo-dev] make.conf optimization flags for k6, athlon

[Daniel Robbins]

On Fri, Jul 13, 2001 at 11:35:05PM +0300, Dan Armak wrote:
> Hi all,
> 
> In make.conf we specify the optimization CXXFLAGS as follows:
> 
> # For Pentium and k6/2 machines 
> # -----------------------------
> 
> #CHOST="i586-pc-linux-gnu"
> #CFLAGS="-mpentium -march=pentium -O3 -pipe"
> #CXXFLAGS="-mpentium -march=pentium -O3 -pipe"
> 
> Why don't we use -march=k6 -mcpu=k6? Wouldn't that allow use of mmx and 
> 3dnow, which a generic pentium doesn't have?

-march=k6 will cause the i586 build to *not* run on a Pentium Classic or
MMX, which is why we don't use it.  However, just the -mcpu=k6 should work
on anything.  Yes, we might want to think about using different optimization
settings for our default builds :)  I don't think that an -mathlon option
is available yet; I think that requires the athlon-gcc patches which really
don't seem to be actively developed.

Best Regards,

-- 
Daniel Robbins					<drobbins@gentoo.org>
President/CEO					http://www.gentoo.org 
Gentoo Technologies, Inc.			

Re: [gentoo-dev] make.conf optimization flags for k6, athlon

[Dan Armak]

On Friday 13 July 2001 23:54, you wrote:
> On Fri, Jul 13, 2001 at 11:35:05PM +0300, Dan Armak wrote:
> > Hi all,
> >
> > In make.conf we specify the optimization CXXFLAGS as follows:
> >
> > # For Pentium and k6/2 machines
> > # -----------------------------
> >
> > #CHOST="i586-pc-linux-gnu"
> > #CFLAGS="-mpentium -march=pentium -O3 -pipe"
> > #CXXFLAGS="-mpentium -march=pentium -O3 -pipe"
> >
> > Why don't we use -march=k6 -mcpu=k6? Wouldn't that allow use of mmx and
> > 3dnow, which a generic pentium doesn't have?
>
> -march=k6 will cause the i586 build to *not* run on a Pentium Classic or
> MMX, which is why we don't use it.  However, just the -mcpu=k6 should work
> on anything.  

I understand that, it reinforces what I said. People who compile from build 
snapshots should set more 'exact' optimization settings, and make.conf's 
comments should tell them to do so.

For example: I have a k6-2/500. When I first built Gentoo on top of a build 
snapshot, I used the make.conf instructions and now hae a system optimized 
for a generic pentium (i.e. it doesn't even use MMX). I'm not going to 
generate binary tarballs for use on other machines, so I'll now change my 
make.conf settings to have -march-k6 -mcpu=k6 and recompile all my packages 
for better performance. Doesn't that make sense?

make.conf should tell people this: i.e. that they can compile with better 
optimization but lose the ability to run the resulting executable on inferior 
CPUs.


Dan Armak

Re: [gentoo-dev] make.conf optimization flags for k6, athlon

[Daniel Robbins]

On Fri, Jul 13, 2001 at 11:59:53PM +0300, Dan Armak wrote:

> > -march=k6 will cause the i586 build to *not* run on a Pentium Classic or
> > MMX, which is why we don't use it.  However, just the -mcpu=k6 should work
> > on anything.  
> 
> I understand that, it reinforces what I said. People who compile from build 
> snapshots should set more 'exact' optimization settings, and make.conf's 
> comments should tell them to do so.

Ah, I misunderstood.  I agree.  Since you now have CVS access, feel free to add
commented-out K6 optimizations to
/usr/portage/sys-apps/portage/files/1.5/cnf/make.conf :)

Best Regards,

-- 
Daniel Robbins					<drobbins@gentoo.org>
President/CEO					http://www.gentoo.org 
Gentoo Technologies, Inc.			

Re: [gentoo-dev] make.conf optimization flags for k6, athlon

[Dan Armak]

On Saturday 14 July 2001 00:37, you wrote:
> On Fri, Jul 13, 2001 at 11:59:53PM +0300, Dan Armak wrote:
> > > -march=k6 will cause the i586 build to *not* run on a Pentium Classic
> > > or MMX, which is why we don't use it.  However, just the -mcpu=k6
> > > should work on anything.
> >
> > I understand that, it reinforces what I said. People who compile from
> > build snapshots should set more 'exact' optimization settings, and
> > make.conf's comments should tell them to do so.
>
> Ah, I misunderstood.  I agree.  Since you now have CVS access, feel free to
> add commented-out K6 optimizations to
> /usr/portage/sys-apps/portage/files/1.5/cnf/make.conf :)
>
 I added it. I also added an i386 option. Does anyone know why the options 
started from 486 initially? Is there a problem in using 386 (with a math 
coprocessor)?

Dan Armak

Re: [gentoo-dev] make.conf optimization flags for k6, athlon

[Daniel Robbins]

On Sat, Jul 14, 2001 at 10:10:09AM +0300, Dan Armak wrote:

>  I added it. I also added an i386 option. Does anyone know why the options 
> started from 486 initially? Is there a problem in using 386 (with a math 
> coprocessor)?

Nope, a 386 should work.  However, bzip2 on a 386 is, um, super-slow?
It's horribly slow on a 486; imagine a 386.

-- 
Daniel Robbins					<drobbins@gentoo.org>
President/CEO					http://www.gentoo.org 
Gentoo Technologies, Inc.			

Re: [gentoo-dev] make.conf optimization flags for k6, athlon

[Hendrik Visage]

On Sat, Jul 14, 2001 at 10:10:09AM +0300, Dan Armak wrote:
>  I added it. I also added an i386 option. Does anyone know why the options 
> started from 486 initially? Is there a problem in using 386 (with a math 
> coprocessor)?

There are some issues with the i386-i387 connections (Bad Intel decision,
causing us to have one less IRQ available :() and there are also an instruction
or two on the i486 that's not on the i386, especially cache related.

-- 
------------------------
Hendrik Visage
hvisage@envisage.co.za

Re: [gentoo-dev] make.conf optimization flags for k6, athlon

[Dan Armak]

On Saturday 14 July 2001 12:25, you wrote:
> On Sat, Jul 14, 2001 at 10:10:09AM +0300, Dan Armak wrote:
> >  I added it. I also added an i386 option. Does anyone know why the
> > options started from 486 initially? Is there a problem in using 386 (with
> > a math coprocessor)?
>
> There are some issues with the i386-i387 connections (Bad Intel decision,
> causing us to have one less IRQ available :() and there are also an
> instruction or two on the i486 that's not on the i386, especially cache
> related.

But as far as I understand all things properly compiled for the i386 should 
work, right?
I don't have a 386. Someone who does might want to test Gentoo on it. Though 
cross-compiling to 386 from a faster machine might be a good idea :-)


Dan Armak

Re: [gentoo-dev] make.conf optimization flags for k6, athlon

[Thomas Beaudry]

A few comments I've been meaning to make.

1.  According to the GCC documentation you shouldn't be using -m<processor> 
any more. It's been depreciated in favor of -mcpu<processor>.

2.  You don't need to use -mcpu usually with -march as -march sets -mcpu 
for you.  I say usually 'cause there's an AMD processor (can't remember 
which) that uses a different instruction set than the default set by the 
-march it uses.

3.  The usage of processor names has been depreciated as well.  Should 
use i586, etc. instead.

>>>>>>>>>>>>>>>>>> Original Message <<<<<<<<<<<<<<<<<<


> # For Pentium and k6/2 machines
> # -----------------------------

> #CHOST="i586-pc-linux-gnu"
> #CFLAGS="-mpentium -march=pentium -O3 -pipe"
> #CXXFLAGS="-mpentium -march=pentium -O3 -pipe"

Re: [gentoo-dev] make.conf optimization flags for k6, athlon

[Dan Armak]

On Sunday 15 July 2001 03:54, you wrote:
> A few comments I've been meaning to make.
>
> 1.  According to the GCC documentation you shouldn't be using -m<processor>
> any more. It's been depreciated in favor of -mcpu<processor>.
I know that. The mpentium etc. flags wre in the originl make.conf, the things 
I added were -march=x -mcpu=x. I'll change mpentium for march=pentium if I 
haven't already.

>
> 2.  You don't need to use -mcpu usually with -march as -march sets -mcpu
> for you.  I say usually 'cause there's an AMD processor (can't remember
> which) that uses a different instruction set than the default set by the
> -march it uses.
I think you're wrong. If I say march=i486 mcpu=i586 this means the binary 
will run on any 486 or better, but is optimized for the 586.


>
> 3.  The usage of processor names has been depreciated as well.  Should
> use i586, etc. instead.
For intel processors maybe. However AMD's k6, though a 586, is better than a 
Pentium, and the Athlon, though a 686, is better than a Pentium3.

Although it really would be better if gcc allowed for flags of instruction 
sets etc. so that you could define your 'custom processor' by saying, use 
MMX+3DNow!2+sse. Than we'd know exactly what optimization it was doing.


Dan Armak

Re: [gentoo-dev] make.conf optimization flags for k6, athlon

[Thomas Beaudry]

> > A few comments I've been meaning to make.
> >
> > 1.  According to the GCC documentation you shouldn't be using 
-m<processor>
> > any more. It's been depreciated in favor of -mcpu<processor>.
> I know that. The mpentium etc. flags wre in the originl make.conf, the 
things
> I added were -march=x -mcpu=x. I'll change mpentium for march=pentium if 
I
> haven't already.

Just making an observation of what I've seen to date.  Haven't seen your 
additions yet.  In fact, I just downloaded them.

> >
> > 2.  You don't need to use -mcpu usually with -march as -march sets -mcpu
> > for you.  I say usually 'cause there's an AMD processor (can't remember
> > which) that uses a different instruction set than the default set by the
> > -march it uses.
> I think you're wrong. If I say march=i486 mcpu=i586 this means the binary
> will run on any 486 or better, but is optimized for the 586.

You're correct.  I had the nagging feeling there was a second exception 
but decided I was crazy after thinking about it for a few seconds.  
Didn't have the time then to look it up.


> >
> > 3.  The usage of processor names has been depreciated as well.  Should
> > use i586, etc. instead.
> For intel processors maybe. However AMD's k6, though a 586, is better 
than a
> Pentium, and the Athlon, though a 686, is better than a Pentium3.

> Although it really would be better if gcc allowed for flags of 
instruction
> sets etc. so that you could define your 'custom processor' by saying, use
> MMX+3DNow!2+sse. Than we'd know exactly what optimization it was doing.

Now arguments there.  It really peaves me that it won't use the MMX 
extensions on my Pentium 167 +MMX machine.