When a function has only positional arguments and at least one argument is optional, the expected signature is:

  func(mandatory_arg1, mandatory_arg2[, optional_arg3[, optinal_arg4]])

For example, the signature of format() is inconsistent with its documentation.

Signature:
---
$ python3 -c 'help(format)'|cat
Help on built-in function format in module builtins:

format(value, format_spec='', /)
    Return value.__format__(format_spec)
    
    format_spec defaults to the empty string
---

Documentation:
---
.. function:: format(value[, format_spec])
---

Attached patch is a first attempt to fix the issue. The problem is that my heuristic to check if an argument is "optional" doesn't seem to work as expected in all cases. I chose to check if the C default is NULL.

The problem is that some functions defines a C default to NULL whereas the Python default is set to a different value and is correct.

Example with _io.FileIO.truncate:

    /*[clinic input]
    _io.FileIO.truncate
        size as posobj: object = NULL
        /

whereas the documentation says that the default is None:

   .. method:: truncate(size=None)

It's easy to fix the default, but in this case my change doesn't fix the signature anymore since the C default is still NULL:

    /*[clinic input]
    _io.FileIO.truncate
        size as posobj: object(c_default="NULL") = None
        /

We need a different heuristic than C default is NULL, or we should fix functions where the heuristic fails.

List of functions modified when Argument Clinic is run to regenerate .c.h files:

builtin_format
builtin_getattr
builtin_input
builtin_sum
cmath_log
_codecs_ascii_decode
_codecs_ascii_encode
_codecs_charmap_decode
_codecs_charmap_encode
_codecs_code_page_encode
_codecs_escape_decode
_codecs_escape_encode
_codecs_latin_1_decode
_codecs_latin_1_encode
_codecs_mbcs_encode
_codecs_oem_encode
_codecs_raw_unicode_escape_decode
_codecs_raw_unicode_escape_encode
_codecs_readbuffer_encode
_codecs_unicode_escape_decode
_codecs_unicode_escape_encode
_codecs_unicode_internal_decode
_codecs_unicode_internal_encode
_codecs_utf_16_be_encode
_codecs_utf_16_le_encode
_codecs_utf_32_be_encode
_codecs_utf_32_le_encode
_codecs_utf_7_encode
_codecs_utf_8_encode
_dbm_dbm_get
_dbm_dbm_setdefault
fcntl_fcntl
_imp_create_dynamic
_io_FileIO_truncate
os_putenv
os_unsetenv
pyexpat_xmlparser_ExternalEntityParserCreate
_sre_SRE_Match_end
_sre_SRE_Match_span
_sre_SRE_Match_start
_tkinter_create
unicodedata_UCD_decimal
unicodedata_UCD_digit
unicodedata_UCD_name
unicodedata_UCD_numeric
unicode_lstrip
unicode_maketrans
unicode_rstrip

See also issue #20291: "Argument Clinic should understand *args and **kwargs parameters".

This issue is blocking me to convert more functions to Argument Clinic. See for example attached getattr_ac.patch which converts getattr() to AC. Without ac_optional_positional.patch, AC generates the signature:

   "getattr($module, object, name, default=None, /)\n"

whereas the following signature is expected:

   "getattr($module, object, name[, default])\n"

The problem is that

  func(mandatory_arg1, mandatory_arg2[, optional_arg3[, optinal_arg4]])

is not compatible with the inspect module.

In many case a meaningful default value was added if this is possible. For example the Python default shown in the signature can be set to '', 'utf-8' or 'strict' while the C default value is NULL for performance. If the parameter is upper index in the sequence it can be set to sys.maxsize (Py_SSIZE_T_MAX in C).

This is not always possible. For example there is not default value for dict.pop().

Please don't change this part of Argument Clinic without Larry. There were several attempts to solve this problem, I don't like current status, but this is Larry's decision.

Argument Clinic can currently only generate signatures for functions whose signatures can be represented in Python.  This is because the signatures are parsed by the inspect module, which in turn uses the ast module to generate a parse tree; it then examines the parse tree and uses that to generate the Signature object.

(By the time you see a signature of a function using inspect, the signature has actually made a round-trip through the ast module, been turned into a Signature object, then str() has been run on it to re-generate the text representation from scratch.  The fact that it's usually identical to the original text signature buried in the function's docstring is a happy accident.)

Changing Argument Clinic to generate signatures with square brackets in them to signify optional parameters is insufficient.  You'd also have to upgrade inspect to support this new syntax--otherwise there'd be no point.  And *that* would be a lot of code.

Thanks for the explanation.

> And *that* would be a lot of code.

Would an effort to make inspect support this be appreciated or do you think this is not important?