There are a lot of places this function should be used, and more in the offing.

Why not do:

*cache = NULL;

if there is an error?

Use of this helper function would then typically look like:

static PyObject *cache = NULL;

npy_cache_pyfunc("the.module.name", "function", &cache);
if (cache == NULL) {
    goto fail;
}

That way you can handle errors more gracefully.

I had that in there originally, but in most (all?) cases numpy is unusable when there is a failure. The intent is to load python implementations of multiarray functions, and later have explicit loads of ufuncs (n_ops currently set on loading the ufunc module) as implementation of the ndarray operators becomes set. I'd actually prefer to call abort on failure to avoid debug dependency, but assert is probably adequate.

Note that cache will be NULL on failure as is.

I'm open to argument on this point, however.

I think on failure, if the assert statements are turned off, you will never make it outside of the function call, because you are going to either PyModule_GetDict on a NULL pointer, or going to Py_INCREF a NULL pointer, which is going to segfault big time. That's probably the least helpful of error messages we can provide...

You could also make it return an int, so that the standard use would be:

if (npy_cache_pyfunc("the.module.name", "function", &cache) < 0) {
    goto fail;
}

Yeah, I had that one too ;) I'm leaning towards your first solution.

OK, changed. Most of the current uses already check for NULL so that is an easy upgrade.

In it goes, thanks!

If I use it i.e. for the VisibleDeprecationWarning, do you mind if I rename it to _pyobject (honestly don't care much)? ;)

I'm thinking of renaming it npy_cache_attr since it uses the PyObject_GetAttr function ;)