Re: config.guess and Linux/GNU

Re: config.guess and Linux/GNU

[Mike Stump]

> To: Mike Stump <mrs@windriver.com>
> Cc: config-patches@gnu.org, gcc-patches@gcc.gnu.org, gcc@gcc.gnu.org
> From: Per Bothner <per@bothner.com>
> Date: 12 Jun 2000 16:37:41 -0700

> Mike Stump <mrs@windriver.com> writes:

> > Hog wash.  As the author of config.guess let me sed [sic] some light.

> Is my mind getting feeble?  I distinctly remember being the original
> author and for some years the maintainer of config.guess.  ((Hurridly
> checks config.guess.))  Yep, that sure looks like my name.

This reminds me, could someone put up the old gcc2 archives (and the
older list as well) up on the web site, if nothing else, as big flat
files for download.  They are great for mining bits and pieces of
information.  I know some people have them, love to see them not get
lost.

gcc2 archives (was re: config.guess and Linux/GNU)

[Joe Buck]

Mike Stump writes:

> This reminds me, could someone put up the old gcc2 archives (and the
> older list as well) up on the web site, if nothing else, as big flat
> files for download.  They are great for mining bits and pieces of
> information.  I know some people have them, love to see them not get
> lost.

I'm not sure we can ethically do that.  The reason is that gcc2 was
operated as a closed list, and many people who posted to it believed that
their words were semi-private.  Even the location of the snapshots was
a secret.

Now, many of us did not like that policy and that's why we started egcs,
where we operate more in public (although the SC does occasionally have a
private discussion).  But because the conditions were different in the
gcc2 days, to publish the gcc2 archives would be to publish peoples' words
without their permission.

Re: config.guess and Linux/GNU

[Steven King]

Way back in june this was discussed.  I dont recall any resolution.  The reason
why I'm bringing it again is that I just noticed that among other things, it
breaks -V usage which seems likely to annoy atleast a few users.

-- 
Steven King
sxking@uswest.net

Re: config.guess and Linux/GNU

[Franz Sirl]

At 00:59 14.06.00, Bruce Korb wrote:
>Mike Stump wrote:
> > Bruce Korb wrote:
> > > No, wrong, do *not* delete the "elf_i?86" entry.
> >
> > [...] As [a] config.guess [contributor], let me shed some light.
>
>As the author of the patch in question, let me shed some history.
>Several years ago, I was working on a Linux box that had a working
>gcc, but not a working cc.  I posted some questions as to why it
>was necessary to specify the machine to configure, but got no
>response.  Eventually, I was sufficiently irritated to grope my
>way through the config.guess script to the point where I saw it
>was trying to invoke "cc", failing and giving up.  Hey, no problem:
>
>   ln -s /usr/local/bin/gcc /usr/local/bin/cc
>
>No luck.  Configuring "cc" turned out to be much harder.

What do you mean by that? The only requirement here is to have a working cc 
in your path or set $CC accordingly. AFAI can see this requirement was 
always there with config.guess and on a platform like Linux this is merely 
an installation problem and not a config.guess problem.

>Over time I noticed that every now and then someone else would
>stub their toe the same way with a similar result.  So, I
>looked deeper into config.guess and found that the first thing
>it attempts to do is to parse the output from "ld --help"
>and then simply ignore it if the supported emulation was i?86_elf.

What do you mean with "ignore"? That's rubbish, it's a normal fall-through 
to this part detecting the exact platform type:

           # Determine whether the default compiler is a.out or elf
           cat >$dummy.c <<EOF
#include <features.h>
#ifdef __cplusplus
#include <stdio.h>  /* for printf() prototype */
         int main (int argc, char *argv[]) {
#else
         int main (argc, argv) int argc; char *argv[]; {
#endif
#ifdef __ELF__
# ifdef __GLIBC__
#  if __GLIBC__ >= 2
     printf ("%s-${VENDOR}-linux-gnu\n", argv[1]);
#  else
     printf ("%s-${VENDOR}-linux-gnulibc1\n", argv[1]);
#  endif
# else
    printf ("%s-${VENDOR}-linux-gnulibc1\n", argv[1]);
# endif
#else
   printf ("%s-${VENDOR}-linux-gnuaout\n", argv[1]);
#endif
   return 0;
}
EOF
           $CC_FOR_BUILD $dummy.c -o $dummy 2>/dev/null && ./$dummy 
"${UNAME_MACHINE}" && rm $dummy.c $dummy && exit 0

I see nothing going wrong here, if you want to fix stuff, fix it here. Or, 
alternatively, move this test into your elf_i?86 case, if you happier with 
that. This distinction between different Linux libc versions is needed by 
at least the gcc, binutils and glibc, which means you just broke the heart 
of all GNU software :-(. You are welcome to change this test to use ld or 
other tools only, but this distinction has to stay.

>It then goes on and fails with the "cc" testing.  My first patch,
>approximately *TWO YEARS* ago was to emit the "i386-pc-linux-gnu"
>string that you GCC folks prefer.  It was rejected because of
>the big-sizean vs. little-sizean config name wars.  I submitted
>a patch that emitted the "i386-pc-linux" string, but *it* was
>rejected because the whole thing was going to be reworked
>"Real Soon Now"(tm).  After that, every now and then, someone
>hurt their toe and, if I were in the mood, I would post another
>"Come on, guys, let's clean this up" message.  A month or two ago,
>the config.guess maintainer finally relented and installed my
>patch.  Now, you're complaining.

Well, actually I could say "Let this crazy x86 lovers shoot them in their 
feet" and live happy ever after with my PPC, but after all, GNU software is 
a community effort.
I don't know about earlier discussions, but I know the original method 
worked fine for a long time and I just see no reason for your patch.

Franz.

Re: config.guess and Linux/GNU

[Bruce Korb]

Mike Stump wrote:
> Bruce Korb wrote:
> > No, wrong, do *not* delete the "elf_i?86" entry.
> 
> [...] As [a] config.guess [contributor], let me shed some light.

As the author of the patch in question, let me shed some history.
Several years ago, I was working on a Linux box that had a working
gcc, but not a working cc.  I posted some questions as to why it
was necessary to specify the machine to configure, but got no
response.  Eventually, I was sufficiently irritated to grope my
way through the config.guess script to the point where I saw it
was trying to invoke "cc", failing and giving up.  Hey, no problem:

  ln -s /usr/local/bin/gcc /usr/local/bin/cc

No luck.  Configuring "cc" turned out to be much harder.
Over time I noticed that every now and then someone else would
stub their toe the same way with a similar result.  So, I
looked deeper into config.guess and found that the first thing
it attempts to do is to parse the output from "ld --help"
and then simply ignore it if the supported emulation was i?86_elf.
It then goes on and fails with the "cc" testing.  My first patch,
approximately *TWO YEARS* ago was to emit the "i386-pc-linux-gnu"
string that you GCC folks prefer.  It was rejected because of
the big-sizean vs. little-sizean config name wars.  I submitted
a patch that emitted the "i386-pc-linux" string, but *it* was
rejected because the whole thing was going to be reworked
"Real Soon Now"(tm).  After that, every now and then, someone
hurt their toe and, if I were in the mood, I would post another
"Come on, guys, let's clean this up" message.  A month or two ago,
the config.guess maintainer finally relented and installed my
patch.  Now, you're complaining.

Re: config.guess and Linux/GNU

[Per Bothner]

Mike Stump <mrs@windriver.com> writes:

> Hog wash.  As the author of config.guess let me sed [sic] some light.

Is my mind getting feeble?  I distinctly remember being the original
author and for some years the maintainer of config.guess.  ((Hurridly
checks config.guess.))  Yep, that sure looks like my name.
-- 
	--Per Bothner
per@bothner.com   http://www.bothner.com/~per/

Re: config.guess and Linux/GNU

[Mike Stump]

> Date: Sun, 11 Jun 2000 12:04:33 -0700
> From: Bruce Korb <bkorb@sco.COM>
> Organization: Santa Cruz Operations

> No, wrong, do *not* delete the "elf_i?86" entry.

Hog wash.  As the author of config.guess let me sed some light.  If
the code was broken, bug fix it, without introducing new bugs.  If you
introduce new bugs, then to prevent bug pushing, which doesn't help
anyone, we should be allowed to demand the removal of the new breakage
to restore how it used to work.  If you're unable to bug fix it
(monotonically better), then hopefully someone else can, or provide
enough of a clue to get you going in the right direction.

Also, one need not run cc, if there are other ways to find that
information.  if [ -f /usr/include/newfile.h ] will find it equally
well (or better), then one might be tempted to use it.

> It turns out that the rest of the config.guess script depends upon
> "cc" doing the "right thing", which aint necessarily so.

We can see if cc does the right thing, and when it doesn't, then do as
you suggest.


The short of this is simple, we believe the code is wrong, you must
fix it.  It is less important exactly how you fix it, as long as it is
fixed.  The details matter not quite as much.  Do you agree the code
is wrong?  If it is, explain how you'r going to fix it.  If you don't
want to, then reverting a bad bug fix, is reasonable.

Re: config.guess and Linux/GNU

[Franz Sirl]

On Sun, 11 Jun 2000, Bruce Korb wrote:
> No, wrong, do *not* delete the "elf_i?86" entry.
> It turns out that the rest of the config.guess script depends upon
> "cc" doing the "right thing", which aint necessarily so.
> If you want to change the output for that entry, then
> go have a battle over whether triplets or quadruplets is
> correct, but leave this much of the output alone.  Too many
> people have too many problems because config.guess chokes and dies
> when "cc" does not do what it (config.guess) wants.
> Do *not* delete the "elf_i?86" entry.

Why not? If you have a problem with the original behaviour, you have to 
change it in a way that doesn't break other things.
I particular you have to ensure that config.guess reports

i?86-pc-linux-gnu

on Linux platforms with glibc2 and later, and on Linux platforms with libc5 
(aka glibc1) it _has_ to report:

i?86-pc-linux-gnulibc1

Otherwise you break configure scripts relying on that. In the GCC case this 
will break at least binary compatibility on libc5 platforms.

Please elaborate a bit on the problem you have with the original correct 
behaviour. I hope it's not just the simple missing link from gcc to cc you're 
trying to fix :-).

Franz.

Re: config.guess and Linux/GNU

[Mark Mitchell]

>>>>> "Bruce" == Bruce Korb <bkorb@sco.COM> writes:

    Bruce> No, wrong, do *not* delete the "elf_i?86" entry.

Bruce --

  I think it's a bad idea to change the config.guess output for
systems where a) it isn't "wrong" and b) it has been that way for
quite some time.  For example, changing `i686-pc-linux-gnu' to
`i686-pc-linux' seems pointless to me; the old entry is correct and it
has been that way for some time, so changing it will likely break
something.

  If there's a problem with config.guess output for some particular
system, then the output for that system, and only that system, should
be changed.  If you understand the issues here, then it would be very
helpful if you could submit a patch that restores the old behavior for
systems other than the ones you particularly wanted to change.

  Thanks,

--
Mark Mitchell                   mark@codesourcery.com
CodeSourcery, LLC               http://www.codesourcery.com

Re: config.guess and Linux/GNU

[Bruce Korb]

No, wrong, do *not* delete the "elf_i?86" entry.
It turns out that the rest of the config.guess script depends upon
"cc" doing the "right thing", which aint necessarily so.
If you want to change the output for that entry, then
go have a battle over whether triplets or quadruplets is
correct, but leave this much of the output alone.  Too many
people have too many problems because config.guess chokes and dies
when "cc" does not do what it (config.guess) wants.
Do *not* delete the "elf_i?86" entry.

Franz Sirl wrote:
> Index: config.guess
> ===================================================================
> RCS file: /cvs/gcc/egcs/config.guess,v
> retrieving revision 1.37
> diff -u -p -r1.37 config.guess
> --- config.guess        2000/06/05 23:23:43     1.37
> +++ config.guess        2000/06/08 10:40:14
> @@ -701,15 +701,11 @@ EOF
>                                      p'`
>           case "$ld_supported_emulations" in
>            *ia64)
> -               echo "${UNAME_MACHINE}-unknown-linux"
> +               echo "${UNAME_MACHINE}-unknown-linux-gnu"
>                  exit 0
>                  ;;
>            i?86linux)
>                  echo "${UNAME_MACHINE}-pc-linux-gnuaout"
> -               exit 0
> -               ;;
> -         elf_i?86)
> -               echo "${UNAME_MACHINE}-pc-linux"
>                  exit 0
>                  ;;
>            i?86coff)

Re: config.guess and Linux/GNU

[Jeffrey A Law]

  In message < 200006081738.MAA91022@latour.rsch.comm.mot.com >you write:
  > >         * config.guess: Detect Linux on ia64 as -linux-gnu.
  > >         Revert this patch:
  > >         * config.guess: Handle an ld emulation of elf_i?86. From Bruce
  > >         Korb <bkorb@sco.com>.
  > 
  > I have no comment on this patch in particular, but if it is applied to
  > gcc's tree it should *first* be submitted and accepted to the official
  > subversions.gnu.org CVS tree for config.
Agreed.  We don't want to diverge config.guess from the official
version (which is not maintained by the GCC group).

jeff

Re: config.guess and Linux/GNU

[Loren James Rittle]

>         * config.guess: Detect Linux on ia64 as -linux-gnu.
>         Revert this patch:
>         * config.guess: Handle an ld emulation of elf_i?86. From Bruce
>         Korb <bkorb@sco.com>.

I have no comment on this patch in particular, but if it is applied to
gcc's tree it should *first* be submitted and accepted to the official
subversions.gnu.org CVS tree for config.

If we make a special version of config.guess then it will just be
trashed the next time someone pulls in the official version.

Regards,
Loren
-- 
Loren J. Rittle
Staff Software Engineer, Distributed Object Technology Lab
Networks and Infrastructure Research Lab (IL02/2240), Motorola Labs
rittle@rsch.comm.mot.com, KeyID: 2048/ADCE34A5, FDC0292446937F2A240BC07D42763672

Re: config.guess and Linux/GNU

[Franz Sirl]

[-- Attachment #1: Type: text/plain, Size: 1537 bytes --]

At 00:37 08.06.00, Franz Sirl wrote:
>Not only it misses the -gnu, I think it also breaks support for x86 libc5
>platforms IMHO. I think it should be fixed. I wonder why this patch was
>necessary anyway, it looks superflous to me. Hmm, and for uniformity,
>ia64-linux should be fixed too.


I just verified that, libc5 on Linux is now detected as

i586-pc-linux

instead of the correct

i586-pc-linux-gnulibc1

I think this might break libstdc++ binary compatibility for libc5...

I suggest the appended patch.

Franz.


         * config.guess: Detect Linux on ia64 as -linux-gnu.
         Revert this patch:
         * config.guess: Handle an ld emulation of elf_i?86. From Bruce
         Korb <bkorb@sco.com>.


Index: config.guess
===================================================================
RCS file: /cvs/gcc/egcs/config.guess,v
retrieving revision 1.37
diff -u -p -r1.37 config.guess
--- config.guess        2000/06/05 23:23:43     1.37
+++ config.guess        2000/06/08 10:40:14
@@ -701,15 +701,11 @@ EOF
                                     p'`
          case "$ld_supported_emulations" in
           *ia64)
-               echo "${UNAME_MACHINE}-unknown-linux"
+               echo "${UNAME_MACHINE}-unknown-linux-gnu"
                 exit 0
                 ;;
           i?86linux)
                 echo "${UNAME_MACHINE}-pc-linux-gnuaout"
-               exit 0
-               ;;
-         elf_i?86)
-               echo "${UNAME_MACHINE}-pc-linux"
                 exit 0
                 ;;
           i?86coff)

[-- Attachment #2: config-guess.patch --]
[-- Type: text/x-diff, Size: 613 bytes --]

Index: config.guess
===================================================================
RCS file: /cvs/gcc/egcs/config.guess,v
retrieving revision 1.37
diff -u -p -r1.37 config.guess
--- config.guess	2000/06/05 23:23:43	1.37
+++ config.guess	2000/06/08 10:40:14
@@ -701,15 +701,11 @@ EOF
 				    p'`
         case "$ld_supported_emulations" in
 	  *ia64)
-		echo "${UNAME_MACHINE}-unknown-linux"
+		echo "${UNAME_MACHINE}-unknown-linux-gnu"
 		exit 0
 		;;
 	  i?86linux)
 		echo "${UNAME_MACHINE}-pc-linux-gnuaout"
-		exit 0
-		;;
-	  elf_i?86)
-		echo "${UNAME_MACHINE}-pc-linux"
 		exit 0
 		;;
 	  i?86coff)

Re: config.guess and Linux/GNU

[Mark Mitchell]

>>>>> "Martin" == Martin v Loewis <martin@loewis.home.cs.tu-berlin.de> writes:

    >> I have no personal opinion, but my understanding was the
    >> `linux-gnu' was the FSF way of spelling the system name, just
    >> as GNU/Linux is the way the FSF names the system(s) that some
    >> folks call just plain Linux.
    >> 
    >> Was this an intentional change?

    Martin> I don't know whether the change was intentional, but I
    Martin> believe your understanding is incorrect.

Very well.  Thank you for correcting me!

--
Mark Mitchell                   mark@codesourcery.com
CodeSourcery, LLC               http://www.codesourcery.com

Re: config.guess and Linux/GNU

[Franz Sirl]

On Wed, 07 Jun 2000, Martin v. Loewis wrote:
> > I have no personal opinion, but my understanding was the `linux-gnu'
> > was the FSF way of spelling the system name, just as GNU/Linux is the
> > way the FSF names the system(s) that some folks call just plain Linux.
> >
> > Was this an intentional change?
>
> I don't know whether the change was intentional, but I believe your
> understanding is incorrect. At some point, the target triplet was
> changed to optionally be a target quadruple, with the fourth part
> being the "environment". This was only used on Linux, where "gnu"
> really indicated presence of the GNU system, in particular of the GNU
> C library.
>
> At the time the feature was introduced, the rationale was that there
> could be also a sparc-sun-solaris-gnu (etc), which is a Sun Solaris
> kernel with a GNU system on top of it.

Ah, nice info on the history.

But I think this change of behaviour is wrong and buggy. It was introduced 
with:

revision 1.142
date: 2000/04/25 20:58:07;  author: bje;  state: Exp;  lines: +4 -0
2000-04-26  Ben Elliston  <bje@redhat.com>
 
        * config.guess: Handle an ld emulation of elf_i?86. From Bruce
        Korb <bkorb@sco.com>.                                                 

Not only it misses the -gnu, I think it also breaks support for x86 libc5 
platforms IMHO. I think it should be fixed. I wonder why this patch was 
necessary anyway, it looks superflous to me. Hmm, and for uniformity, 
ia64-linux should be fixed too.

Franz.

Re: config.guess and Linux/GNU

[Martin v. Loewis]

> I have no personal opinion, but my understanding was the `linux-gnu'
> was the FSF way of spelling the system name, just as GNU/Linux is the
> way the FSF names the system(s) that some folks call just plain Linux.
> 
> Was this an intentional change?  

I don't know whether the change was intentional, but I believe your
understanding is incorrect. At some point, the target triplet was
changed to optionally be a target quadruple, with the fourth part
being the "environment". This was only used on Linux, where "gnu"
really indicated presence of the GNU system, in particular of the GNU
C library.

At the time the feature was introduced, the rationale was that there
could be also a sparc-sun-solaris-gnu (etc), which is a Sun Solaris
kernel with a GNU system on top of it.

Regards,
Martin

config.guess and Linux/GNU

[Mark Mitchell]

I've noticed that config.guess is now calling my system:

  i686-pc-linux

instead of:

  i686-pc-linux-gnu

I have no personal opinion, but my understanding was the `linux-gnu'
was the FSF way of spelling the system name, just as GNU/Linux is the
way the FSF names the system(s) that some folks call just plain Linux.

Was this an intentional change?  

--
Mark Mitchell                   mark@codesourcery.com
CodeSourcery, LLC               http://www.codesourcery.com