Code in example abuses the fact that _as_parameter_ is passed
recursively. Not sure if this is for fixing or not.

I've got a hunch that any crashes you cause with ctypes are your own
fault. :-)

Confirmed in trunk and py3k. Maybe we should treat ctypes segfaults that
don't involve external C code more strictly as bugs?

I'm not convinced that this needs doing, so I'll close in a couple of weeks unless anyone objects.

I concur with Daniel, this strikes me as a legitimate bug.  Crashing is obviously possibly by calling into arbitrary C code, but in this case control never leaves the runtime.

So has anyone got an idea on how to fix this without solving the halting problem?

No need to solve the halting problem, just to detect a cycle in data.  I've got a patch, however it requires objects to be hashable.  Another case for identity dict I suppose :)  Not sure if that's a reasonable requirement, in any event, proof of concept patch, against trunk.

This function is called quite often. Building a set each time will be expensive.
What about a simpler recursion check with Py_EnterRecursiveCall()?

Good idea Amaury, seems to work just fine.

New changeset 4945a45b8be4 by Benjamin Peterson in branch '3.1':
check possible recursive _as_parameter_ to prevent segfault (closes #1838)
http://hg.python.org/cpython/rev/4945a45b8be4

New changeset 4bd06503eaca by Benjamin Peterson in branch '2.7':
check possible recursive _as_parameter_ to prevent segfault (closes #1838)
http://hg.python.org/cpython/rev/4bd06503eaca