Providing np.int64 as subdomain id to a measure leads to wrong ffc code with empty measures

[Johannes Neumann]

Providing a np.int64 value with an empty cache to a Dolfin measure leads to zero measures. Even providing a Python int afterwards keeps the wrong code and does neither trigger an error or a recompile.

Clearing the cache and first calling the measure with an int creates the correct ffc code which then can be called with a np.int64 as well and producing in the correct results.

Minimal Working Example for FEniCS 2019.1.0

from multiprocessing import Process

import numpy as np

import dijitso
import dolfin as df


def square_perimeter(data_type, clear_cache):
    if clear_cache:
        dijitso.main(("clean",))
    mesh = df.UnitSquareMesh(10, 10)
    mf = df.MeshFunction("size_t", mesh, 1, 0)
    measure = df.Measure("ds")(subdomain_data=mf, domain=mesh)
    perimeter = df.assemble(1.0 * measure(data_type(0)))
    print(f"{data_type}: {perimeter}")


def call(*args):
    p = Process(target=square_perimeter, args=args)
    p.start()
    p.join()


if __name__ == "__main__":
    call(int, True)  # result correct
    call(np.int64, False)  # result correct
    call(np.int64, True)  # result wrong
    call(int, False)  # result wrong

Output

<class 'int'>: 4.000000000000003
<class 'numpy.int64'>: 4.000000000000003
<class 'numpy.int64'>: 0.0
<class 'int'>: 0.0

‌