[SciPy-dev] Core dump in PyArray_Return

[Travis Oliphant]

Sasha wrote:

>I cannot reproduce the problem with a small example, but the following
>change fixes it:
>
>Index: numpy/core/src/arrayobject.c
>===================================================================
>--- numpy/core/src/arrayobject.c        (revision 1855)
>+++ numpy/core/src/arrayobject.c        (working copy)
>@@ -915,8 +915,7 @@
>                 Py_XDECREF(mp);
>                 return NULL;
>         }
>-
>-       if (mp->nd == 0) {
>+       if (mp->nd == 0 && !PyArray_IsScalar(mp, Generic)) {
>                PyObject *ret;
>                ret = PyArray_ToScalar(mp->data, mp);
>                Py_DECREF(mp);
>
>Apparently in my program PyArray_Return somehow gets an object that is
>already a scalar  and calls PyArray_ToScalar on it.
>
I've seen this before during development.  It typically is due to a 
PyArray_Return on something that already returns a scalar.  It would be 
nice to know where the core dump occurred.  I think a check like this is 
probably in order though...  just added it.

I also undid the size-0 check.  It was added to fix the "problem" that 
several places in the code where an array can be copied into another 
array many times.  That is a size-10 array can be used to fill a 
size-100 array by repeating over it 10 times.  The code uses iterators 
to do it, and there are a few places where you can even put a size-20 
array into a size-13 array and a fraction of it will be iterated over.

I guess the universal check is to stringent and the size-0 check should 
be placed in those places where copying data from a size-0 array makes 
no sense...

-Travis