Take for example select.epoll.__new__():

static PyObject *
pyepoll_new(PyTypeObject *type, PyObject *args, PyObject *kwds)
{
    int flags = 0, sizehint = FD_SETSIZE - 1;
    static char *kwlist[] = {"sizehint", "flags", NULL};

    if (!PyArg_ParseTupleAndKeywords(args, kwds, "|ii:epoll", kwlist,
                                     &sizehint, &flags))
        return NULL;

How should that be handled? Currently I can't specify FD_SETSIZE - 1 as the default value, but I don't want a magic -909 either :)

See also a discussion in bpo-20193. zlib needs this in many functions.

Yes, this is supported, just undocumented (for now). The magic you want is, for example in Modules/_sre.c:

endpos: Py_ssize_t(c_default="PY_SSIZE_T_MAX") = sys.maxsize

Nice, thanks.

The attached patch extends Argument Clinic and the inspect module: now you may use simple (!) expressions as default values. So for example "sys.maxsize - 1" should now work fine.

• Names may be in the current module; as a backup they will also be looked up in sys.modules.
• You may look up attributes of names.
• You may subscript into names.
• You may use math operators, both unary and binary.
• int/float/string literals are (still) fine.
• You may not use: function calls, list/set/dict/tuple/bytes literals, generator expressions or comprehensions, if expressions, anything fun.

Other changes in this diff:

• "doc_default" is gone, it no longer makes sense.
• "py_default" and "c_default" handling is simpler and more predictable.
• Minor doc fix suggested on tracker.

The docs still need work. More tests would be good too but I'm already overwhelmed by work--could one of you guys contribute some tests?

Larry, nice. I am verified that your latest patch fixes my third sqlite3
nit from bpo-20178.

Just confirmed that this patch fixes the traceback I was getting from winreg in bpo-20172, but this has a new oddity:

>>> help(winreg)
...
    ConnectRegistry(computer_name=<object object at 0x0190E148>, key=<object object at 0x0190E148>)
...

Every param of every function has this kind of thing, positional or keyword, default or no. I'll try to look into it, but my hopes aren't high (much of clinic's inner workings are still mysterious to me).

That's okay, email me a sample at

larry at hastings dot org

and I'll take a look at it. Be sure to tell me which patch(es) were applied at the time.

(Note: all you need to send is the Clinic block.)

Actually, I can reproduce on tip with just this patch applied; os.chmod shows it. And I was wrong, params that have a default are correct, it's just ones without.

I'd prefer it if you emailed me as I asked. That will help me get to it a lot quicker.

Okay, I have a fix for the help(os.chmod) problem, that'll be in the next refreshed patch.

Why not allow arbitrary string as py_default?

Isn't that what it does?

I may actually remove py_default. Nobody uses it, and I don't think you need it--you just specify what you want as the real default value.

Second rollup patch. More small fixes, many suggested by other folks.

It would be better if you commit a patch which changes PyTuple_Size to PyTuple_GET_SIZE in separate commit.

Excellent! I'm rewriting the zlib module and the code becomes much cleaner with this patch.

There is one problem left -- Py_buffer doesn't support default value at all.

Such code

/[clinic input]
zlib.compressobj
zdict: Py_buffer = unspecified
[clinic start generated code]/

produces error:

When you specify a named constant ('unspecified') as your default value,
you MUST specify a valid c_default.

But this code

/[clinic input]
zlib.compressobj
zdict: Py_buffer(c_default="{NULL, NULL}") = unspecified
[clinic start generated code]/

produces error too:

The only legal default value for Py_buffer is None.

And specifying c_default=None returns first error.

Py_buffer is currently inflexible. The only default value it accepts is None. If you want it to be required, give it no default value. If you want it to be optional, give it a default value of None.

Do you have a use case for any other default value?

Third patch, changes based on Georg's review. Thanks, Georg!

Rietveld doesn't like your new patch?

Sorry, forgot to update trunk, here you go.

Thx! (without soundcheck)

New changeset abf87e1fbc62 by Larry Hastings in branch 'default':
Issue bpo-20226: Major improvements to Argument Clinic.
http://hg.python.org/cpython/rev/abf87e1fbc62

Done!

Georg: Thanks for the suggestion, and for the reviews! Argument Clinic is improved this day.

New changeset cd3fdf21a6e4 by Larry Hastings in branch 'default':
Issue bpo-20226: Added tests for new features and regressions.
http://hg.python.org/cpython/rev/cd3fdf21a6e4