FIX: PyErr_SetObject does not steal reference (Ticket 2216)

[seberg]

It seems that PyErr_SetObj is assumed ot steal a references in most parts of numpy, but appearently it does not. I assume that in /doc/pyrex/numpyx.c it is correct, but am not quite sure about the occurances in /numpy/random/mtrand/mtrand.c. Seems to also fix Ticket 2125.

[seberg]

This would be a test, but since it only covers broadcasting error like this... also I am not sure if the refcount may not be a bit different between python versions, since it seems to be 4 and not 2 here because of function scope or such.

    def test_error_refcount(self):
        # ticket #2216
        a = np.arange(3)
        try:
            a *= a[:-1] # broadcasting error
        except ValueError, e:
            pass
        m = e.args[0]
         # not sure why its 4 in func scope
        assert_equal(sys.getrefcount(m), 4)

[teoliphant]

You are correct. We should DECREF the error message because PyErr_SetObject doesn't steal the reference. Good catch!

[njsmith]

Looks good to me. I don't know of any good way to test this either. And really you'd need a different test for every possible error message, since whether or not we the programmer forgot the DECREF is determined separately for every call-site. So I guess we should let this in without tests, unless someone has a better idea.

[njsmith]

(The test failure is unrelated.)

[njsmith]

Well, looks like no-one has a better idea, so let's do it :-)