On Python 3.0:

>>> class C:
...     def __len__(self): return "foo"
...
>>> len(C())
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
SystemError: Objects/longobject.c:433: bad argument to internal function

On Python 2.6 the behaviour is different for old and new-style classes,
with old-style classes giving the more informative error message and
both accepting (and truncating) floats.

I attached a patch for Python 3.0, which refuses everything but ints and
gives an informative error message. Or does the float-truncating
behaviour of Python 2.x need to be preserved?

What about using PyNumber_AsSsize_t? it uses PyNumber_Index to accept
integral-like types, and refuses floats.

Sounds ok, but then you get a more generic "object cannot be interpreted
as an integer" at the point where len() is called, instead of the
clearer "__len__() should return an int". I'm not sure if this could be
confusing.

Of course we can do both: Accept integral-like types and reset the
exception text. The new patch does this and adds a test for the new
behaviour.

Review carefully - I'm a newbie! ;-)

Not bad! some remarks though:

• It's better to avoid the "expensive" call to PyErr_Occurred() when
  possible. Here, an exception is set if (and only if) len==-1.
  For example, it is enough to add these lines after the "__len__() should
  return >= 0" message:

+ else if (PyErr_ExceptionMatches(PyExc_TypeError))
+ PyErr_SetString(PyExc_TypeError,
+ "__len__() should return an int");

• Please clarify (that is: add tests for) the case where __len__ returns
  1<<50 or -1<<50. If I remember correctly, PyNumber_AsSsize_t(res, NULL)
  clips the values to MAXINT. Is this wanted?

That is *not* wanted. We had a discussion on the list about changing the
return value of the sq_length slot to allow larger lengths to be
reported, and while I don't recall why this wasn't done, I do recall
that the consensus was that if Py_ssize_t remains, len() should raise
rather than lie for larger values-

In the latest list message I could find Guido wanted len() to lie:

http://mail.python.org/pipermail/python-3000/2008-May/013387.html

Has this been resolved in bpo-2723?

True. However, it's no pronouncement either. I suggest bringing it up on
the list again; probably other people would want to voice their opinions
too.

There seems to be a pronouncement now
(http://mail.python.org/pipermail/python-3000/2008-September/014692.html),
so I'm attaching a patch which clips to sys.maxsize and includes
Amaury's suggestions.

Fixed in bpo-5137.