egcs: Problems compiling recen Linux kernels on intel

egcs: Problems compiling recen Linux kernels on intel

[Horst von Brand]

The FAQ mentions a problem with linux/arch/i386/kernel/process.c. This can
be worked around (without bad effects I can see) by specifying
-fno-strength-reduce for that file. Works up to linux-2.1.70 and
egcs-971201, at least.

Somebody on linux-kernel@vger.rutgers.edu mentioned that, I've no way to
know who it was now.
-- 
Dr. Horst H. von Brand                       mailto:vonbrand@inf.utfsm.cl
Departamento de Informatica                     Fono: +56 32 654431
Universidad Tecnica Federico Santa Maria              +56 32 654239
Casilla 110-V, Valparaiso, Chile                Fax:  +56 32 797513

Re: egcs: Problems compiling recen Linux kernels on intel

[Joe Buck]

> The FAQ mentions a problem with linux/arch/i386/kernel/process.c. This can
> be worked around (without bad effects I can see) by specifying
> -fno-strength-reduce for that file. Works up to linux-2.1.70 and
> egcs-971201, at least.

Please don't get the Linux people starting using -fno-strength-reduce
again.  You're just saying "I want my kernel to run more slowly".
That workaround is inferior to using the patch.

> Somebody on linux-kernel@vger.rutgers.edu mentioned that, I've no way to
> know who it was now.

The bug is in Linux and should be patched there.  Historically, Linux has
often used unsafe assembly language code that was experimentally verified
to work with particular gcc versions (as in: this code is OK if we trick
gcc into not optimizing it).  New gcc versions have broken this code.
This happened when 2.7.x came out, and it's happening again with egcs and
2.8.0.

Re: egcs: Problems compiling recen Linux kernels on intel

[Horst von Brand]

Joe Buck <jbuck@synopsys.com> said:
> "Horst von Brand" at Dec 4, 97 10:31 am said:

> > The FAQ mentions a problem with linux/arch/i386/kernel/process.c. This
> > can be worked around (without bad effects I can see) by specifying
> > -fno-strength-reduce for that file. Works up to linux-2.1.70 and
> > egcs-971201, at least.

> Please don't get the Linux people starting using -fno-strength-reduce
> again.  You're just saying "I want my kernel to run more slowly".
> That workaround is inferior to using the patch.

What patch? The FAQ asks for a workaround, this is one.

> > Somebody on linux-kernel@vger.rutgers.edu mentioned that, I've no way to
> > know who it was now.

> The bug is in Linux and should be patched there.  Historically, Linux has
> often used unsafe assembly language code that was experimentally verified
> to work with particular gcc versions (as in: this code is OK if we trick
> gcc into not optimizing it).  New gcc versions have broken this code.
> This happened when 2.7.x came out, and it's happening again with egcs and
> 2.8.0.

You are of course absolutely right; I'm going to take a closer look at this
(and hopefully propose a patch to linux). Just haven't got the time just
now...  But what do you propose to do in the meantime?
-- 
Dr. Horst H. von Brand                       mailto:vonbrand@inf.utfsm.cl
Departamento de Informatica                     Fono: +56 32 654431
Universidad Tecnica Federico Santa Maria              +56 32 654239
Casilla 110-V, Valparaiso, Chile                Fax:  +56 32 797513

Re: egcs: Problems compiling recen Linux kernels on intel

[Joe Buck]

> > Please don't get the Linux people starting using -fno-strength-reduce
> > again.  You're just saying "I want my kernel to run more slowly".
> > That workaround is inferior to using the patch.
> 
> What patch? The FAQ asks for a workaround, this is one.

I understand that there is a patch to fix this problem ... sorry I don't
have a pointer to it.  It has been applied to the 2.1.x kernels.

Re: egcs: Problems compiling recen Linux kernels on intel

[Jeffrey A Law]

  In message < 199712041803.KAA13273@atrus.synopsys.com >you write:
  > 
  > > > Please don't get the Linux people starting using -fno-strength-reduce
  > > > again.  You're just saying "I want my kernel to run more slowly".
  > > > That workaround is inferior to using the patch.
  > > 
  > > What patch? The FAQ asks for a workaround, this is one.
  > 
  > I understand that there is a patch to fix this problem ... sorry I don't
  > have a pointer to it.  It has been applied to the 2.1.x kernels.
Let's make sure we're talking about the same problem.

There's at least two problems with process.c -- one causes egcs
to crash building process.c, the other causes gcc-2.8 to mis-compile
process.c.

[ They may actually be the same bogus asm statement -- I don't know. ]

I know that newer 2.1.x kernels have the fix for the gcc-2.8 mis-
compilation.  I don't know if those kernels have a fix for the
egcs abort when compiling process.c.

[ Note the next egcs snapshot will have the ADDRESSOF optimizations
  so the mis-comilation problem will be an issue for egcs real soon.]


Jeff

Re: egcs: Problems compiling recen Linux kernels on intel

[Jeffrey A Law]

  In message < 199712041848.PAA05824@pincoya.inf.utfsm.cl >you write:
  > I'm talking about the crash here.
OK.

  > egcs-971201 still aborts on linux-2.1.70 for i586. I'm going to take a
  > closer look at this (thanks for the feedback you gave me earlier!).
OK.  If you find anything useful, let us know -- I'd like to be able say
something useful on the web page like (get this fix for process.c or something
similar).


  > > [ Note the next egcs snapshot will have the ADDRESSOF optimizations
  > >   so the mis-comilation problem will be an issue for egcs real soon.]
  > 
  > Care to elaborate?
Just a new optimization that Jason Merrill (and others) developed; it went into
gcc2 just after we made the egcs branch from gcc2.

We'll be picking it up (along with all the other gcc2 changes over the last 3
months) as we do merges from gcc2 into the mainline egcs sources.

jeff

Re: egcs: Problems compiling recen Linux kernels on intel

[Horst von Brand]

Jeffrey A Law <law@hurl.cygnus.com> said:
>   In message < 199712041803.KAA13273@atrus.synopsys.com >you write:

>   > > > Please don't get the Linux people starting using -fno-strength-reduce
>   > > > again.  You're just saying "I want my kernel to run more slowly".
>   > > > That workaround is inferior to using the patch.

>   > > What patch? The FAQ asks for a workaround, this is one.

>   > I understand that there is a patch to fix this problem ... sorry I don't
>   > have a pointer to it.  It has been applied to the 2.1.x kernels.

> Let's make sure we're talking about the same problem.

> There's at least two problems with process.c -- one causes egcs
> to crash building process.c, the other causes gcc-2.8 to mis-compile
> process.c.

I'm talking about the crash here.

> [ They may actually be the same bogus asm statement -- I don't know. ]

> I know that newer 2.1.x kernels have the fix for the gcc-2.8 mis-
> compilation.  I don't know if those kernels have a fix for the
> egcs abort when compiling process.c.

egcs-971201 still aborts on linux-2.1.70 for i586. I'm going to take a
closer look at this (thanks for the feedback you gave me earlier!).

> [ Note the next egcs snapshot will have the ADDRESSOF optimizations
>   so the mis-comilation problem will be an issue for egcs real soon.]

Care to elaborate?
-- 
Dr. Horst H. von Brand                       mailto:vonbrand@inf.utfsm.cl
Departamento de Informatica                     Fono: +56 32 654431
Universidad Tecnica Federico Santa Maria              +56 32 654239
Casilla 110-V, Valparaiso, Chile                Fax:  +56 32 797513