Quite by accident, I came across a case where Python would quit
generating traceback text and skip printing the actual exception
information. Here is a sample program:

    exec(compile("spam()", ".", "exec"))

and the output in Python 3.1 ("spam" is undefined):

    $ python3.1 test.py
    Traceback (most recent call last):
      File "test.py", line 1, in <module>
        exec(compile("spam()", ".", "exec"))
      File ".", line 1, in <module>
    $

This was bewildering until I realised that the traceback generator was
unable to read from the filename passed to compile() (my original
example was using a name other than "." that wasn't intended to be a
file name, but just happened to also be a directory name). I didn't
really mind the lack of source text, but the lack of the actual
exception message was most disturbing.

This appears to be a failure mode common to both traceback.c and
traceback.py, through slightly different mechanisms. In traceback.c, if
the source filename refers to a directory, the C open() succeeds but an
error occurs when trying to read from the directory handle. In 
traceback.py, the Python open() call fails with an IOError exception, 
and the exception wasn't handled gracefully.

I have attached a patch that creates the following output instead:

    $ ./python.exe test.py
    Traceback (most recent call last):
      File "test.py", line 1, in <module>
        exec(compile("spam()", ".", "exec"))
      File ".", line 1, in <module>
        [Errno 21] Is a directory: '.'
    NameError: name 'spam' is not defined
    $
    
I have tested the patch against Python 3.1, but it applies cleanly to 
the trunk (for some reason I couldn't make the trunk build, but that's 
unrelated). This patch may need some finesse for a Win32 build; I don't
have the ability to test that at the moment.

Patch still applies to py3k, not applying cleanly to trunk anymore. Tests pass with patch on py3k.

3.1 gets only security fixes. In 2.7 and 3.2+ the bug has already fixed.