[SciPy-dev] errors in linalg test_basic

[William Reade]

Hi

I'm developing a package called Ironclad; its purpose is to get
compiled CPython extensions to work transparently in IronPython. SciPy
looks pretty healthy at the moment (if rather slow in places): I can
run 2608 tests*, and I get 160 errors and 17 failures. Many of those
problems are actually inconsequential, but a number of them are all
too real; hence this post.

The problem I'm currently investigating is in linalg.basic.lstsq(), at
the call to the 'gelss' lapack function: when overwrite_a and
overwrite_b are False, I get a segfault somewhere (it works fine when
they're both True; haven't tried mixed). By way of investigation, I
hunted through the source as far as generic_flapack.pyf; then I got
scared and decided to ask for help. My thought process so far is as
follows:

Presumably, those arguments signify that the results should be written
into a new block of memory somewhere; when we're overwriting
pre-existing arrays, everything is fine, so I imagine there's a
problem with the new result array(s?). However, I have no idea by what
mechanism those arrays should be created, so I don't really know where
to start looking. Can anyone suggest a good direction?

Thanks in advance
William