Try to make backend_gdk compatible with numpy 1.6

[jkseppan]

Fixes #5078, at least according to a quick test with import matplotlib.backends.backend_gdk

[OceanWolf]

Hmm, trying to restart the 3.4 test but it doesn't want to... Travis says it restarted successfully, but it hasn't restarted...

[WeatherGod]

Do we need to cast the py_pixbuf pointer into a PyObject pointer? py_pixbuf is type PyGObject*.

[mdehoon]

Yes, because PyArray_SetBaseObject expects a PyObject pointer as its second argument.

[jkseppan]

On 16 Sep 2015, at 17:18, mdehoon notifications@github.com wrote:

In src/_backend_gdk.c #5080 (comment):

@@ -52,7 +52,15 @@ static PyObject *pixbuf_get_pixels_array(PyObject *self, PyObject *args)

 /* the array holds a ref to the pixbuf pixels through this wrapper*/
 Py_INCREF(py_pixbuf);

• PyArray_SetBaseObject(array, (PyObject *)py_pixbuf);
  +#if NPY_API_VERSION >= 0x00000007
• if (PyArray_SetBaseObject(array, (PyObject *)py_pixbuf) == -1) {

•    Py_DECREF(py_pixbuf);
  

•    Py_DECREF(array);
  

•    return NULL;
  

• }
  +#else
• PyArray_BASE(array) = py_pixbuf;
  Yes, because PyArray_SetBaseObject expects a PyObject pointer as its second argument.

I guess the question is whether we need to do the cast in the #else branch too. It’s not required in C, but assigning to a pointer to a different type can cause a compiler warning, so it would be slightly cleaner to cast in that branch too.

[WeatherGod]

Correct. That was the question I was asking.

[tacaswell]

All the tests pass and cleans up a corner-case failure mode.