Pythran fallback not used for numpy call

[serge-sans-paille]

Investigating the C++ code generated from

# cython: np_pythran=True

from __future__ import absolute_import

import numpy as np
cimport numpy as np
cimport cython

def roll(np.int8_t[::1] state, Py_ssize_t idx):
    return np.roll(state, -idx, axis=0)

shows that Pythran is not used to implement np.roll, while there's support for that in Pythran.

[serge-sans-paille]

cc @aguinet !

[serge-sans-paille]

Got it! It's because of the keyword argument.

@scoder would that make sense to turn keyword argument into normal argument based on the inspect.getfullargspec information for pythran candidates?

[scoder]

Why not use the signatures in pythran.tables.MODULES instead? They don't appear to know the names of their arguments, but – shouldn't they?

Also see GeneralCallNode.map_to_simple_call_node(), which does this transformation (for keyword arguments to C functions), so the main thing that you need is a Cython function type for the target signature.