Re: floating point operations in gcc 3.3.3, glibc 2.3.2

[Ollie Wild <owild at ixiacom dot com>]

Dan,
 We've found the problem.  Turns out our build script was to blame, not 
the patches.  We were executing a loop like:

   for cpu in sh4 i686 powerpc-405 powerpc-750; do
           cd $TOP_DIR
                                                                                  

           # Load definitions for current CPU into environment and build
           eval `cat $cpu.dat ../$TOOLCOMBO.dat` buildit
   done

Interestingly, this behaves differently if executed with /bin/sh instead 
of /bin/bash.  In the former case, any variables set by $cpu.dat are 
preserved between iterations.  In the latter case, they are only valid 
for a single execution of buildit. 

We were all executing the script with "sh ./doit.sh", which was causing 
the ppc750 build to give weird floating point results *only* if it was 
built after sh4 and powerpc-405.

Anyway, thanks for your time, and sorry for the bogus bug report.

Ollie Wild

Dan Kegel wrote:

> Kunjan Naik wrote:
>
> > Target: ppc750
> > Toolchain 1: crosstool-0.28-rc19, gcc3.3.3 glibc-2.3.2
> > Toolchain 2: crosstool-0.28-rc19, gcc3.3.3 glibc-2.3.2 without
> > gcc-3.3.2-arm-softfloat.patch
> > Sample Testcase: modified struct-ret-1.c with printf statements (see
> > attached)
> > Complier flag: --static.
> >
> > Both the toolchains output the same .s file for the testcase. When 
> > executed on target, toolchain 1 binary prints garbage and toolchain 2
> > binary prints expected output.
> > For toolchain 1, the testcase compares garbage values and succeeds or 
> > aborts
> > and for toolchain 2 the testcase
> > compares valid output and always succeeds.
>
>
> I just wrote a shell script to try to reproduce your result.
> Before running it, I copied 
> patches/gcc-3.3.3/gcc-3.3.2-arm-softfloat.patch
> to the contrib directory, and put a copy of your struct-ret-1.c in the
> current directory.
> Does this look right?
>
> #!/bin/sh
> set -ex
> TARBALLS_DIR=$HOME/downloads
> export TARBALLS_DIR
> GCC_LANGUAGES="c"
> export GCC_LANGUAGES
> RESULT_TOP=/opt/crosstool/xxvfp
> export RESULT_TOP
> mkdir -p $RESULT_TOP
> cp contrib/gcc-3.3.2-arm-softfloat.patch patches/gcc-3.3.3
> eval `cat powerpc-750.dat gcc-3.3.3-glibc-2.3.2.dat`  sh all.sh --notest
> RESULT_TOP=/opt/crosstool/novfp
> export RESULT_TOP
> mkdir -p $RESULT_TOP
> rm -f patches/gcc-3.3.3/gcc-3.3.2-arm-softfloat.patch
> eval `cat powerpc-750.dat gcc-3.3.3-glibc-2.3.2.dat`  sh all.sh --notest
> /opt/crosstool/xxvfp/powerpc-750-linux-gnu/gcc-3.3.3-glibc-2.3.2/bin/powerpc-750-linux-gnu-gcc 
> -static struct-ret-1.c -o xxvfp.out
> /opt/crosstool/novfp/powerpc-750-linux-gnu/gcc-3.3.3-glibc-2.3.2/bin/powerpc-750-linux-gnu-gcc 
> -static struct-ret-1.c -o novfp.out
> cmp xxvfp.out novfp.out
> echo Finished
>
> It does say that xxvfp.out and novfp.out differ, but if you run
> $ cmp -l xxvfp.out novfp.out
> you will see that the only differences are that 'no' has been
> replaced with 'xx'.  Thus I doubt that the two programs will
> behave differently.  I can send them to you if you like, but
> perhaps you should run this shell script yourself and report
> back the results.
>
> - Dan

------
Want more information?  See the CrossGCC FAQ, http://www.objsw.com/CrossGCC/
Want to unsubscribe? Send a note to crossgcc-unsubscribe@sources.redhat.com