ISTM that if frame->f_exc_type == NULL then exc_value
and exc_traceback will also be NULL.  I didn't see that
this is documented, perhaps I missed it or there is
some case when this can occur.  If it can occur, we
shoul develop a test for it.

Assuming this condition is invariant, some
simplifications can be made in reset_exc_info which is
called once per eval_frame (on function exit).

Also, I think there is currently an extra Py_INCREF on
Py_None.  This occurs when tstate->exc_type == NULL.

This patch seems to have little to no effect on
performance.  I did measure a 0.3% speed improvement.

Logged In: YES 
user_id=31435

Note that the patch is out of date.

I agree the invariant you deduced should hold, but in fact
it doesn't now, at least due to insane initialization
problems in exceptions.c:

http://mail.python.org/pipermail/python-dev/2006-May/065248.html

I'd like to ensure & exploit a stronger invariant:

http://mail.python.org/pipermail/python-dev/2006-May/065231.html

but that's stuck for now.  I put my work in progress on a
new branch:

svn+ssh://svn.python.org/python/branches/tim-exc_sanity

BTW, I don't agree that the incref on Py_None wasn't needed.
 Py_None is getting assigned to two new pointers
(tstate->exc_type and frame->f_exc_type), so should be
incremented twice.

Logged In: YES 
user_id=31435

I got a very tiny bit more out of this and added it to the
trunk.  Thanks!  Note that I left in the disputed Py_INCREF.

The more ambitious tim-exc_sanity branch is looking like
more  trouble than it's worth.