[gentoo-dev] Automatic cflag downgrade from pentium4 to pentium3?

[gentoo-dev] Automatic cflag downgrade from pentium4 to pentium3?

[Dhruba Bandopadhyay]

Hello

Today I saw in a gentoo cvs server email the following.

2003-03-05 Wednesday 11:01  method

        * sys-devel/gcc/: ChangeLog, gcc-3.2.2-r4.ebuild,
        files/digest-gcc-3.2.2-r4: replace-flags "-march=pentium4"
        "-march=pentium3" to fix bug #16867

Now, I realise that the pentium4 is prone to errors but I hope this does
not begin a spree of changes where -march=pentium4 flags will
automatically downgraded to -march=pentium3 by the ebuild or otherwise.

I have used pentium4 for some time now with no problems whatsoever and
would like the raise the issue of whether downgrading of flags should be
manual and left to the user rather than being automatically modified.  A
choice would be nice too.

Also, what exactly is responsible for pentium4 bugs?  Is it gcc and is
there hope of this being resolved with future versions?  And will using
pentium3 result in a noticeable decline in performance?

Many thanks
Dhruba Bandopadhyay


--
gentoo-dev@gentoo.org mailing list

Re: [gentoo-dev] Automatic cflag downgrade from pentium4 to pentium3?

[Paul de Vrieze]

[-- Attachment #1: signed data --]
[-- Type: text/plain, Size: 2313 bytes --]

On Thursday 06 March 2003 13:19, Dhruba Bandopadhyay wrote:
> Hello
>
> Today I saw in a gentoo cvs server email the following.
>
> 2003-03-05 Wednesday 11:01  method
>
>         * sys-devel/gcc/: ChangeLog, gcc-3.2.2-r4.ebuild,
>         files/digest-gcc-3.2.2-r4: replace-flags "-march=pentium4"
>         "-march=pentium3" to fix bug #16867
>
> Now, I realise that the pentium4 is prone to errors but I hope this does
> not begin a spree of changes where -march=pentium4 flags will
> automatically downgraded to -march=pentium3 by the ebuild or otherwise.
>
> I have used pentium4 for some time now with no problems whatsoever and
> would like the raise the issue of whether downgrading of flags should be
> manual and left to the user rather than being automatically modified.  A
> choice would be nice too.
>
> Also, what exactly is responsible for pentium4 bugs?  Is it gcc and is
> there hope of this being resolved with future versions?  And will using
> pentium3 result in a noticeable decline in performance?

The problem with the pentium4 bugs is entirely in gcc. The packages which 
"downgrade" the pentium4 options do that because either:
- the program does not compile with pentium4
- the program does not run with pentium4
- the program is instable with pentium4

In summary the pentium4 flag makes the specific program instable. The loss in 
speed should be minimal provided that -mcpu=pentium4 is still provided. 
Downgrading flags per ebuild I believe is not an option for most users, as it 
makes automatic merging impossible. For this reason the ebuilds that are 
instable with pentium4 edit the cflags so that the generated code is still 
optimized up to the best stable level.

The only other option would be to have all those ebuilds warn the user about 
the problem with march=pentium4, and exit requesting the user to set his/her 
C(XX)FLAGS better. The thing is that most programs actually do work with 
-march=pentium4 so that people might want to use that.

Paul

ps. This flag downgrading concerns only gcc compilation. Not the code 
generated by this gcc. So other programs are still using -march=pentium4 
where it is not filtered too.

-- 
Paul de Vrieze
Researcher
Mail: pauldv@cs.kun.nl
Homepage: http://www.cs.kun.nl/~pauldv

[-- Attachment #2: signature --]
[-- Type: application/pgp-signature, Size: 189 bytes --]

Re: [gentoo-dev] Automatic cflag downgrade from pentium4 to pentium3?

[Martin Schlemmer]

On 06 Mar 2003 12:19:22 +0000
Dhruba Bandopadhyay <dhruba@codewordt.co.uk> wrote:

> Hello
> 
> Today I saw in a gentoo cvs server email the following.
> 
> 2003-03-05 Wednesday 11:01  method
> 
>         * sys-devel/gcc/: ChangeLog, gcc-3.2.2-r4.ebuild,
>         files/digest-gcc-3.2.2-r4: replace-flags "-march=pentium4"
>         "-march=pentium3" to fix bug #16867
> 
> Now, I realise that the pentium4 is prone to errors but I hope this
> does not begin a spree of changes where -march=pentium4 flags will
> automatically downgraded to -march=pentium3 by the ebuild or
> otherwise.
> 
> I have used pentium4 for some time now with no problems whatsoever and
> would like the raise the issue of whether downgrading of flags should
> be manual and left to the user rather than being automatically
> modified.  A choice would be nice too.
> 
> Also, what exactly is responsible for pentium4 bugs?  Is it gcc and is
> there hope of this being resolved with future versions?  And will
> using pentium3 result in a noticeable decline in performance?
> 

I use -march=pentium3, and my box is speedy (although haven't tried
with pentium4 for some time).  But then .. rather a bit slower than
unstable? 

I am with you on this.  I did only do the pentium4->pentium3 for some
critical ebuilds like xfree and mozilla, as else we might miss if
gcc actually gets it fixed.

Lastly I do still consider fixing glibc, as it really (math tests, etc)
break on -march=pentium4 :(


-- 

Martin Schlemmer
Gentoo Linux Developer, Desktop/System Team Developer
Cape Town, South Africa


--
gentoo-dev@gentoo.org mailing list