[SciPy-dev] Python (Enthought Edition) for Windows test release

[Pearu Peterson]

On Wed, 19 Jan 2005, Nils Wagner wrote:

> AFAIK, there is still a mysterious segmentation fault in scipy.test() ? How 
> about that ?

These segfaults would be present also in 0.3.2 if the new features in CVS 
would be added to 0.3.2. As far as I understand, these segfaults are due 
to gcc bugs and to avoid segfaults with bugy gcc, one should either 
rebuild lapack libraries with -O2, or we can disable these new features in 
scipy. The former may be a pain for most users but the latter does not 
make much sense to me. Just for the records, I cannot reproduce the 
segfaults in my debian sid box, and probably because debian sid provides 
rather uptodate software, including gcc 3.4.4 that I am using for building 
scipy. And, obviously, if I cannot reproduce a failure, there's little I 
can to do to find a workaround for it.

> Is ATLAS recommended or shall we use
>
>  export BLAS_SRC=~/src/blas
>  export LAPACK_SRC=~/src/lapack
>  mkdir $BLAS_SRC
>  cd $BLAS_SRC
>  wget http://www.netlib.org/blas/blas.tgz
>  tar xzf blas.tgz
>
>  cd $LAPACK_SRC/..
>  wget http://www.netlib.org/lapack/lapack.tgz
>  tar xzf lapack.tgz
>
>  export BLAS=None
>  export LAPACK=None
>  export ATLAS=None

If that works for you then building lapack with -O2 and completing ATLAS 
lapack with it should fix the segfault.

Pearu