Doctest fails to find doctests defined in extension modules. With tools
like cython (http://cython.org) it's trivially easy to add docstrings to
extension code, a task that is much less pleasant with hand-written
extensions. The following patch is a minimal fix for the problem:

--- doctest_ori.py      2008-06-20 19:22:56.000000000 -0700
+++ doctest.py  2008-06-20 19:23:53.000000000 -0700
@@ -887,7 +887,8 @@
             for valname, val in obj.__dict__.items():
                 valname = '%s.%s' % (name, valname)
                 # Recurse to functions & classes.
-                if ((inspect.isfunction(val) or inspect.isclass(val)) and
+                if ((inspect.isfunction(val) or inspect.isclass(val) or
+                     inspect.isbuiltin(val) ) and
                     self._from_module(module, val)):
                     self._find(tests, val, valname, module, source_lines,
                                globs, seen)

However, it is likely not sufficient as it doesn't take into account the
__test__ dict, for which probably the same change would work, just a few
lines later. Furthermore, the real issue is in my view in the
distinction made by inspect between isfunction() and isbuiltin() for the
sake of analyzing docstrings. isfunction() returns false for a function
that is defined in an extension module (though it *is* a function) while
isbuiltin returns True (though it is *not* a builtin).

For purposes of doctesting, doctest should simply care:

• That it is a function.
• That it has a docstring that can be parsed.

But in too many places in doctest there are currently assumptions about
being able to extract full source, line numbers, etc. Hopefully this
quick fix can be applied as it will immediately make doctest work with
large swaths of extension code, while a proper rethinking of doctest is
made.

BTW, in that process doctest will hopefully be made more modular and
flexible: its current structure forces massive copy/paste subclassing
for any kind of alternate use, since it has internally hardwired use of
its own classes. Doctest is tremendously useful, but it really could
use with some structural reorganization to make it more flexible (cleanly).

For me, a 'function' is written in Python, a 'builtin' is written in C.

The fact that it is defined in an extension module is irrelevant, and
depends on the implementation: zlib is an extension module on Unix, but
is linked inside python26.dll on Windows.

maybe inspect.isroutine() is the correct test here.

I think there are two issues that need to be separated:

1. The doctest bug. I'm happy with any resolution for it, and I'm not
   claiming that my patch is the best approach. isroutine() indeed works
   in my case, and if that approach works well in general for doctest, I'm
   perfectly happy with it.

2. Terminology. I really disagree with the idea that

• 'function' describes the implementation language of an object instead
  of whether it's a standalone callable (vs an object method).

• 'builtin' doesn't mean the object is "built into the shipped Python"
  but instead that it's "written in C".

The language exposes its builtins via the __builtin__ module precisely
to declare what is part of itself, and it even has in the documentation:

http://docs.python.org/lib/built-in-funcs.html

a section that starts:

"""2.1 Built-in Functions

The Python interpreter has a number of functions built into it that are
always available."""

Nowhere does it say that "builtins are written in C and functions in
Python".

In summary, I'm happy with any fix for the bug, but I very strongly
disagree with a use of terminology that is confusing and misleading (and
which unfortunately is enshrined in the inspect and types modules in how
they distinguish 'Function' from 'BuiltinFunctionType').

And by the way, by 'extension module' I mean to describe C-extensions,
since that is how most C code is shipped by third-party authors, those
affected by this bug (since the stdlib doesn't seem to use doctests
itself for its own testing of C code).

I've added Tim Peters to the nosy list. Is there anyone else who should
be considered as doctest maintainer? I've also checked further Python
versions - a quick a look at latest doctest source shows that the
problem is still there.

There are some details (and a workaround) in Cython FAQ:
http://wiki.cython.org/FAQ#HowcanIrundoctestsinCythoncode.28pyxfiles.29.3F

In looking at test_decimal for bpo-16748, I discovered that not all of the doctests for _decimal are being tested. So, I fixed it (or at least tried to :).

This patch does the following:

• Replace most inspect.isfunction checks in DocTestFinder with inspect.isroutine

• Add a check to DocTestFinder._from_module for method_descriptors

• Correct a doctest on float.fromhex (which is incorrect back to 2.7) due to the new float repr

• Add a couple doctests to the builtins module because previously there were no functions in builtins with docstrings for testing against. I chose to add to bin(), hex(), and oct(); I believe those three can benefit from having the docstring show the format of the output string. I'm not terribly attached to those, though, so if anyone has a better suggestion for testing, I'm all ears.

• Add tests using the builtins module (since it's a C module that's definitely going to be available).

• Add doctests to test_builtin (because the tests aren't actually run in test_doctest, just found). While at it, convert test_builtin to unittest.main(). I believe test_builtin should grow doctest running even if the doctests I added to bin, hex, and oct aren't kept; float and int have some doctests that should be kept up to date.

• Add notes to the docs.

Ping!

Anyone able to do a review of this patch? It still applies cleanly to default (or even 3.3, if this qualifies as a bug rather than a new feature).

Added some review comments.

Because it could cause possibly buggy doctest fragments to run that previously did not run, I don't think it should be backported as a bug fix.

Here's a new version of the patch that I think addresses the points in your review (which I've also replied to on Rietveld).

And I agree about not backporting.

Does this qualify as a new feature that needs to be in before beta 1?

Larry: thoughts?

In the absence of input, I'm going to go ahead and commit this just in case in needs to be in before feature freeze, but I'll leave the issue open at "commit review" stage for a few days in case anyone does have something to say about it.

New changeset 95dc3054959b by Zachary Ware in branch 'default':
Issue bpo-3158: doctest can now find doctests in functions and methods
http://hg.python.org/cpython/rev/95dc3054959b

One-year-olds don't like productivity. Committed, 3 hours after I said I would :). I'll leave this open for a couple days just in case.

New changeset 30b95368d253 by Zachary Ware in branch 'default':
Issue bpo-3158: Relax new doctests a bit.
http://hg.python.org/cpython/rev/30b95368d253

I think there's an issue with this change - ismethoddescriptor() doesn't guarantee that that the object has an __objclass__ attribute. Unbound PyQt4 signals appear to be a case where this goes wrong.

This came up testing IPython on Python 3.4 - we subclass DocTestFinder, which creates other problems, but it looks like it would run into trouble even with the base implementation.
ipython/ipython#4892

Does this patch fix things for you?

the patch seems to work for me in ipython.

New changeset c964b6b83720 by Zachary Ware in branch 'default':
Issue bpo-3158: Provide a couple of fallbacks for in case a method_descriptor
http://hg.python.org/cpython/rev/c964b6b83720

Done.

New changeset 8520e0ff8e36 by R David Murray in branch 'default':
whatsnew: doctest finds tests in extension modules (bpo-3158)
http://hg.python.org/cpython/rev/8520e0ff8e36