Blogpost SageMath Sphinx Cython

One of the deliverables (D4.13) of the OpenDreamKit project is refactoring the documentation system of SageMath. The SageMath documentation is built using a heavily customized Sphinx. Many of the customizations are neccessary to support autodoc (automatically generated documentation from docstrings) for Cython files.

Thanks to some changes I made to Sphinx, autodoc for Cython now works provided that:

  1. You use Sphinx version 1.6 or later.

  2. The Cython code is compiled with the binding=True directive. See How to set directives in the Cython documentation.

  3. A small monkey-patch is applied to inspect.isfunction. You can put this in your Sphinx for example:

     def isfunction(obj):
         return hasattr(type(obj), "__code__")
     import inspect
     inspect.isfunction = isfunction

This was used successfully for the documentation of cysignals and fpylll. There is ongoing work to do the same for SageMath.

Implementation of functions in Python

To understand why items 2 and 3 on the above list are needed, we need to look at how Python implements functions. In Python, there are two kinds of functions (we really mean functions here, not methods or other callables):

  1. User-defined functions, defined with def or lambda:

     >>> def foo(): pass
     >>> type(foo)
     <class 'function'>
     >>> type(lambda x: x)
     <class 'function'>
  2. Built-in functions such as len, repr or isinstance:

     >>> type(len)
     <class 'builtin_function_or_method'>

In the CPython implementation, these are completely independent classes with different behaviours.

User-defined functions binding as methods

Just to give one example, built-in functions do not have a __get__ method, which means that they do not become methods when used in a class.

Let’s consider this class:

class X(object):
    def printme(self):
        return repr(self)

This is essentially equivalent to

class X(object):
    printme = (lambda self: repr(self))

>>> X().printme()
'<__main__.X object at 0x7fb342f960b8>'

However, directly putting the built-in function repr in the class does not work as expected:

class Y(object):
    printme = repr

>>> Y().printme()
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
TypeError: repr() takes exactly one argument (0 given)

This is simply something that built-in functions do not support.

User-defined vs. built-in functions

Here is a list of the main differences between user-defined and built-in functions:

Cython generates C code, so Cython functions must be built-in functions. This has unfortunate disadvantages, such as the lack of introspection support, which is particularly important for Sphinx.

The Cython function type: cyfunction

Luckily, the Cython developers came up with a solution: they invented a completely new function type (called cyfunction), which is implemented like built-in functions but which behaves as much as possible like user-defined functions.

By default, functions in Cython are built-in functions. With the directive binding=True, functions in Cython become cyfunctions. Since cyfunctions are not specifically optimized by CPython, this comes with a performance penalty. More precisely, calling cyfunctions from Python is slower than calling built-in functions from Python. The slowdown can be significant for simple functions. Within Cython, cyfunctions are as fast as built-in functions.

Since a cyfunction is not a built-in function nor a user-defined function (those two types are not subclassable), the inspect module (and hence Sphinx) does not recognize it as being a function. So, to have full inspect support for Cython functions, we need to change inspect.isfunction. After various attempts, I came up with hasattr(type(obj), "__code__") to test whether the object obj is a function (for introspection purposes). This will match user-defined functions and cyfunctions but not built-in functions, nor any other Python type that I know of.

The future: a PEP to change the function types?

I have some vague plans for a Python Enhancement Proposal (PEP) to change the implementation of the Python function types. The goal is that Cython functions can be implemented on top of some standard Python function type, with all features that cyfunctions currently have, the performance of built-in functions and introspection support of user-defined functions.

At this point, it is too early to say anything about the implementation of this hypothetical future Python function type. If anything happens, I will surely post an update.