1) New policy for what can and cannot be converted during the Derby
First, let me apologize for only figuring this out now. The Derby
has been a learning process, discovering things that Argument Clinic
didn't handle. And there were a lot of funny edge cases that we
weren't going to discover until we tried doing the conversion. And
to be honest I was pushing harder than I should have.
*ahem*
The new policy for conversion work for Python 3.4:
We may only convert functions that have signatures that
we can represent 100% accurately in an inspect.Signature object.
IF the function has a default value that can't be represented in
Python (e.g. NULL), but we can find a value in Python that behaves
identically to not passing in that parameter (e.g. _sha1.sha1(b'')
== _sha1.sha1()), then we may convert that function using that
clever default value (e.g. "string: object(c_default='NULL') = b''
").
IF the function has parameters with default values that are
dynamic or cannot be represented accurately in Python, it cannot
be converted without changing its semantics. So it cannot be
converted for 3.4.
IF the function *requires* "optional groups" (as in, the original
function used switch(PyTuple_GET_SIZE(args)) and multiple calls to
PyArg_ParseTuple()), then it's permissible to convert it for 3.4,
but they are low priority. Such functions have semantics that are
so weird, we will have to modify inspect.Parameter to support
them, and that will only happen in 3.5. However, I will ensure
that they otherwise convert correctly for 3.4. (They won't
actually generate signatures. They will however generate the
first line of the docstring.)
For 3.5, I expect we'll have more leeway in doing things like "this
should accept an int or None". But we should talk about that then.
If you have patches outstanding that convert functions that
shouldn't be converted for 3.4, please put them aside. We can
almost certainly use them for 3.5.
If any conversions have been committed that change the semantics of
the function, someone will have to back them out. I'd appreciate it
if the person who checked it in could do it. I expect to make a
pass before rc1 to check all the conversions myself. (Which I hope
will be quite time-consuming, as hopefully there will be lots of
converted functions by then!)
_______________________________________________________________________________
2) New recommendation for marking to-do functions
If you examine a function and determine that it can't be converted
to Argument Clinic right now, please add a comment to that effect.
The comment should be one line, and contain a special marker so we
can find it easily with searches.
I nominate two different markers:
"AC 3.4" means "It's okay to convert this function to Argument
Clinic in 3.4, but it can't be converted right now because of a bug
or missing feature in Argument Clinic."
"AC 3.5" means "This function can't be converted to Argument Clinic
in 3.4, it must wait for 3.5."
I encourage you to add a little text saying why, like:
/* AC 3.4: waiting for *args support */
/* AC 3.5: value parameter has default value of NULL */
_______________________________________________________________________________
Finally, I've realized that right now there's no good way to stay
abreast of what's new and changing in clinic.py. I check in a patch
just about every day for clinic.py, and sometimes I don't remember
to update the "howto". The best way unfortunately is to read the
output of "hg log".
If you have questions, you can email me directly at "larry at
hastings dot org", or you can find me in the #python-dev IRC
channel.
Thank you, everybody who's participating in the Derby, and again I'm
sorry I didn't realize sooner this *had* to be the policy for 3.4.
Hope to see your issues in the tracker,
/arry