[Numpy-discussion] numarray CVS seems to have an __array_struct__ bug

Andrew Straw strawman at astraw.com
Thu Nov 17 03:42:06 EST 2005


I think there's a bug in the new __array_struct__ code in numarray CVS.

Consider an object a which implements the __array_struct__ interface. If
one constructs a numarray "view" of this array using aview =
numarray.asarray(a), the numarray.ndarray instance does not keep a
reference to a.__array_struct__. Thus, a segfault may occur because
a.data can be freed although aview may attempt to access it.

In particular code like this will crash:

buf = numarray.asarray( array_struct_factory() ) # buffer is freed
buf[1] = 2 # attempt is made to write to the non-existant buffer,
causing a segfault

I've "fixed" this issue with the following, which I'm sure is not the
right way, but it's all I can do at this point.

diff -u -r1.130 numarraycore.py
--- Lib/numarraycore.py 4 Nov 2005 20:27:11 -0000       1.130
+++ Lib/numarraycore.py 17 Nov 2005 11:38:29 -0000
@@ -357,6 +357,7 @@
         a = None
         if hasattr(sequence, "__array_struct__"):
             a =
+            a._hack_to_hold_reference_ = sequence.__array_struct__
         elif (hasattr(sequence, "__array_shape__") and
             hasattr(sequence, "__array_typestr__")):
             typestr = sequence.__array_typestr__

And while I'm at it, can the "from distutils.command.config import log"
be turned off in setup.py? I like to see those errors, warnings, and
status strings whiz past...

More information about the NumPy-Discussion mailing list