[SciPy-Dev] segfault in test_qhull.TestRidgeIter2D.test_complicated

Pauli Virtanen pav@iki...
Mon Nov 15 19:27:09 CST 2010


On Mon, 15 Nov 2010 12:07:16 +0000, Pauli Virtanen wrote:

> Mon, 15 Nov 2010 11:10:40 +0100, Jens Nielsen wrote:
>> I get a segmentation fault at the same point on ubuntu 10.10 both 64
>> bit and 32 bit versions but only when linking against atlas. Linking
>> against blas and lapack only is fine.
> 
> Ok, this bug is absent if Qhull is compiled without optimizations, and
> present if compiled with -O2.
> 
> Qhull apparently breaks gcc's strict-aliasing assumptions, and compiling
> with 'OPT="-fno-strict-aliasing -O2" python setup.py ...' makes it work
> again. This should probably be enabled by default, or one would need to
> check if the aliasing issues can be fixed on the source code level.

Ok, source code level "fix" here:

https://github.com/pv/scipy-work/commit/8efed6e403597783243f476c7bdb1c506d99f22b

I guess this is not completely robust, but it seems unlikely the
compiler will go swapping the two statements after this.

-- 
Pauli Virtanen



More information about the SciPy-Dev mailing list