cpdef override of a cdef method can produce incorrect vtables

[saraedum]

There seems to be a problem in the way that cython checks for declarations of cpdef overrides of cdef methods. It seems that this is related to the feature introduced in #545.

Consider the following base.pxd, base.pyx, and bug.pyx:

cdef class A:
    cdef a(self)

cdef class B(A):
    pass
    #cpdef a(self)

cdef class C(B):
    cdef c(self)

cdef class A:
    cdef a(self):
        return "A.a()"

cdef class B(A):
    cpdef a(self):
        return "B.a()"

cdef class C(B):
    cdef c(self):
        return "C.c()"

from base cimport C

def call():
    return C().c()

Calling bug.call() prints B.a(). Enabling the #cpdef a(self) fixes the problem, and it prints C.c() as expected.

The problem here is that without the cpdef, the vtable used by bug.pyx is missing the entry for B.a.

To solve this problem, one would probably just have to add a check to Cython so it requires cpdef overrides of cdef methods to be declared in their .pxd files.

[robertwb]

Thanks for the report and concise repro.

[roed314]

Thanks for the quick fix!