Re: GCC build failed for native with your patch on 2001-12-29T23:23:46Z.

[Geoff Keating <geoffk at geoffk dot org>]

> cc: gcc-patches@gcc.gnu.org
> Date: Sun, 30 Dec 2001 17:17:15 -0500
> From: David Edelsohn <dje@watson.ibm.com>
> 
> 	Geoff, why are you sure that Kenner's patch is the cause of the
> failure?  I just spent the day backing out patches which are causing
> failures on AIX.  Kenner's patches had no effect.  Rth's patch earlier in
> the day was the culprit on AIX.  This is also how I interpret the tester
> error message, that the failure existed prior to Kenner's patch.
> 
> 	Kenner's patch may have problems, but I believe that you are
> blaming the wrong person for the bootstrap failure. 
> 
> 	If you look at the tester logs, the first message which reports a
> build error only include Jakub and Rth's patches, not Kenner's.  I have
> tracked the error to Rth's patch.  Please get your facts correct before
> you start blaming people.

There were two patches that introduced errors.  One took the form of a
comparison failure:

Bootstrap comparison failure!
cfgbuild.o differs
make[1]: *** [compare] Error 1

That seems to be introduced by rth's loop patch.

The other problem, which concealed the first problem, was:

/maat/heart/tbox/cvs-gcc/gcc/gcc/libgcc2.c: In function `__floatdidf':
/maat/heart/tbox/cvs-gcc/gcc/gcc/libgcc2.c:1058: Internal compiler
error in copy_to_mode_reg, at explow.c:711
Please submit a full bug report,
with preprocessed source if appropriate.
See <URL:http://www.gnu.org/software/gcc/bugs.html> for instructions.
make[3]: *** [libgcc/./_floatdidf.o] Error 1

This was introduced by Kenner's patch to copy_rtx_and_substitute, and
corrected by rth's later patch to that function.  If this (second)
problem is present, it's impossible to debug or even see the first
problem, as the build doesn't get that far.

Interestingly, I don't get the comparison failure on powerpc-linux.
I'd not be surprised that you don't get the copy_to_mode_reg failure
on AIX, since it seems to be address-related.

This is why the tester says "the problem might not be caused by your
patch" rather than "the problem is not caused by your patch"; it can't
be certain and requires a human to inspect the logs to see what's
going on.  Ideally I'd like the tester to know more about the nature
of the failure, so that it can warn whether the failure has moved
earlier in the build, but I can't think of a good design.  I have made
some improvements with respect to comparison failures now that
Kenner's failure has been fixed and they should help in future.

-- 
- Geoffrey Keating <geoffk@geoffk.org> <geoffk@redhat.com>