homepage

This issue is part of the Great Argument Clinic Conversion Derby,
where we're trying to convert as much of Python 3.4 to use
Argument Clinic as we can before Release Candidate 1 on January 19.
This issue asks you to change the following bundle of files:
 Modules/_datetimemodule.c: 20 sites
 Modules/timemodule.c: 8 sites
 Modules/_decimal/_decimal.c: 24 sites
Talk to me (larry) if you only want to attack part of a bundle.
For instructions on how to convert a function to work with Argument
Clinic, read the "howto":
 http://docs.python.org/dev/howto/clinic.html 

Please take out _decimal.c. It has a huge test suite that would break entirely.

Absolutely. decimal is a very special case.
Whoever takes over this issue: please ignore the part about _decimal.c.

I have started working on this (partial patch attached), but I am unsure how to handle cases where the docstring definition utilizes a macro. For example:
PyDoc_STRVAR(strftime_doc,
"strftime(format[, tuple]) -> string\n\
\n\
Convert a time tuple to a string according to a format specification.\n\
See the library reference manual for formatting codes. When the time tuple\n\
is not present, current time as returned by localtime() is used.\n\
\n" STRFTIME_FORMAT_CODES);
I don't think STRFTIME_FORMAT_CODES will be expanded if I put it into the [clinic] comment section. 
Should I include the contents verbatim? Or do not convert these functions to argument clinic?

That macro, STRFTIME_FORMAT_CODES, is only used in two functions. If it were me I'd just copy and paste the text into both functions.

Will do. Another question: what should I do with return annotations, e.g. the "-> dict" part in
"get_clock_info(name: str) -> dict\n
Should I drop that part? In the clinic howto, I couldn't find anything about return annotations (only that parameter annotations are unsupported).

One more question: what's a good way to find sites that need changes? Searching for "PyArg_" finds both converted and unconverted functions...

Yes, you should drop things that look like return annotations in the original docstring.
I don't have a magic formula for finding unchanged functions, sorry.

Here is a patch for timemodule.c. I have converted everything except: 
* ctime() uses the gettmarg() function. Converting gettmarg() itself doesn't seem possible to me (because it's not Python callable), and duplicating the functionality in a single complicated clinic section just for ctime() did not seem reasonable either.
* gmtime(), localtime(), ctime() use a the parse_time_t_args() utility function. I have not yet figured out how to replace this with argument clinic (sent a mail to python-dev about that).

Hmm. After reading some of the threads on python-dev, I'm now confused if I did the conversion of optional arguments correctly.
Should something like "strftime(format[, tuple])" be converted using optional groups, or should tuple become a parameter with a default value? (so far, I have done the latter).

Optional groups really are only for cases when that's the behavior
of the original code. If the original function used a switch statement
examining the size of the tuple, then had different parsing functions based on that size, use optional groups.
But if the function can be expressed with normal positional parameters and default values, that's best.

As discussed on devel, here's an updated patch for timemodule.c that
uses a custom C converter to handle the parse_time_t_args() utility
function. The only function I did not convert is mktime, because I did
not find a way to do so without duplicating the gettmarg() utility
function (which cannot be converted because it is also called by other
C functions).
You probably want to take a close look at what I did to the strptime
function. I didn't see a way to include the actual default value
(rather than None) in the signature without reconstructing the
argument tuple for calling the Python implementation.
As far as I can see, this completes the conversion for timemodule.c. I'll
work on _datetimemodule.c next, but I'm not sure I'll have time before
next weekend.
This is probably already planned, but since I haven't seen it in
Misc/NEWS yet: I think it would be a good idea to entry informing the
user about the conversion from strictly optional parameters to
parameters with default values. Eg.:
 * Issues xxx, yyy, zzz: many functions in the standard library now
 accept `None` for optional arguments. Previously, specifying `None`
 mostly resulted in a `TypeError`. Starting with this version,
 passing `None` is equivalent to not specfying the parameter.

(I already sent this to python-dev, but maybe it makes more sense to have these thoughts together with the patch)
After looking at the conversion of parse_time_t_args again, I think I lost track of what we're actually gaining with this procedure. If all we need is a type that can distinguish between a valid time_t value and a default value, why don't we simply use PyObject?
In other words, what's the advantage of the extra custom strict, plus the custom C converter function, plus the custom converter python class over:
static int
PyObject_to_time_t(PyObject *obj, time_t *when)
{
 if (obj == NULL || obj == Py_None) 
 *when = time(NULL);
 else if (_PyTime_ObjectToTime_t(obj, when) == -1)
 return 0;
 return 1;
}
/*[clinic input]
time.gmtime
 seconds: object=NULL
 /
[...]
static PyObject *
time_gmtime_impl(PyModuleDef *module, PyObject *seconds)
{
 PyObject *return_value = NULL;
 time_t when;
 if(!PyObj_to_time_t(seconds, &when))
 return NULL;
[...]
To me this version looks shorter and clearer.

I've attached updated patches to reflect the recent changes in the derby policy.
part1 is the part of the patch that is suitable for 3.4.
part2 are the changes that have to wait for 3.5 (mostly because of default values).

I'll wait with working on _datetimemodule.c until someone had a chance to look over my _timemodule.c patch. (I want to make sure I'm following the right procedure).

Regarding the handling of time_t arguments which can be None, I agree that the second version (without custom convertors) is simpler and clearer while having no disadvantage that I can see.
I'd like to review the rest of the patches, but you mention changes to the AC policy regarding 3.4 and 3.5 and I can't find a reference to these changes or to the new policy.

Tal, I was referring to this mail: https://mail.python.org/pipermail/python-dev/2014-January/132066.html 

All the Derby patches should only go into trunk at this point.

The modern workflow requires creating a PR on GitHub. Nikolaus, do you mind to convert your patch to a PR?

Sorry, no. I have long lost context and interest in this.

I have created a PR for Modules/timemodule.c.