Argument Clinic: Better internal implementation for @getter and @setter

@erlend-aasland

Even though we successed to provide @getter and @setter features from #112205
The current internal implementation is full of hacky workaround.

We need to improve the internal implementation for better generated C codes and maintenance.
@erlend-aasland wrote some ideas for this: #112205 (comment)

Linked PRs

See also: #113160 (comment)

Thanks for creating this issue, Donghee. See #112205 (comment) for more info.

Not only internal implementation needs improvement, but the interface too.

Getter should support return converter.
Setter should support converter for the single parameter (for example, ssl.SSLContext.options would benefit from this).
Support deleter.

My 2c:

aliases aren't supported properly
"ifdef trick" doesn't work with get/setters. Trying to create a getter/setter in #ifdef block, I get following error from the clinic.py:

Traceback (most recent call last):
  File "/home/sk/src/cpython/Tools/clinic/clinic.py", line 11, in <module>
    main()
    ~~~~^^
  File "/home/sk/src/cpython/Tools/clinic/libclinic/cli.py", line 226, in main
    run_clinic(parser, args)
    ~~~~~~~~~~^^^^^^^^^^^^^^
  File "/home/sk/src/cpython/Tools/clinic/libclinic/cli.py", line 218, in run_clinic
    parse_file(filename, output=ns.output,
    ~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^
               verify=not ns.force, limited_capi=ns.limited_capi)
               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/sk/src/cpython/Tools/clinic/libclinic/cli.py", line 84, in parse_file
    cooked = clinic.parse(raw)
  File "/home/sk/src/cpython/Tools/clinic/libclinic/app.py", line 193, in parse
    parser.parse(block)
    ~~~~~~~~~~~~^^^^^^^
  File "/home/sk/src/cpython/Tools/clinic/libclinic/dsl_parser.py", line 492, in parse
    block.output.extend(self.clinic.language.render(self.clinic, block.signatures))
                        ~~~~~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/sk/src/cpython/Tools/clinic/libclinic/clanguage.py", line 83, in render
    return self.render_function(clinic, function)
           ~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^
  File "/home/sk/src/cpython/Tools/clinic/libclinic/clanguage.py", line 521, in render_function
    s = template.format_map(template_dict)
KeyError: 'methoddef_name'

#155778 fixes several bugs. A larger feature PR will follow.

@Getter

gh-113318: Fix @Getter and @Setter in Argument Clinic (GH-155778)

@Getter

[3.13] gh-113318: Fix @Getter and @Setter in Argument Clinic (GH-155778…

@Getter

[3.14] gh-113318: Fix @Getter and @Setter in Argument Clinic (GH-155778…

Thank you, and sorry for the delay reviewing. The rework looks too big to backport, but, that's done now.

Are you planning to update the documentation in the devguide?

It was a small bugfix. I have a larger feature branch on my hands, in process of polishing.

Documentation: python/devguide#1886.

Merge remote-tracking branch 'upstream/main' into pythongh-113318-get…

#156066 is a feature PR. It changes how properties are implemented internally and how the code is generated. It also adds support of input and output converters.

I have also a follow-up which adds support of separate @deleter, but it adds quite complexity, and currently there is no need of such feature. So I will hold it until there was a real need.

python/devguide#1887 is the documentation part.

As I noted on the docs PR, I'm concerned about using a "default value" as placeholder for del. From:

    /*[clinic input]
    @critical_section
    @setter
    @deleter
    MyType.count
        value: int = -1
    [clinic start generated code]*/

it is not obvious that the -1 is what's used when deleting. I fear that this is replacing one footgun with another (though a much smaller one of course).

I'd be interested in the separate deleter functions. Could we use that when converters are needed? I guess the NULL for pointers is needed for backcompat/familiarity, but it seems really awkward for other types.

I don't want to encourage things like making obj.count = -1 equivalent to del obj.count.

See also #156106 which adds tests for all writable attributes.

Added restriction: only NULL is allowed as default value for setter-and-deleter. This restriction is artificial, it does not simplify the code. For now, there are no properties with converters which do not accept NULL as default value, so the issue is rather hypothetical.

I can publish a PR for the separate deleter functions after merging #156066. It adds complexity without any benefits for current code, so I do not recommend to accept it now.

Most of the current deleters raise TypeError. If we accept AttributeError instead, we can remove these deleters -- only 6 will remain. And we can then deprecate these 6 -- they do nothing that the setter cannot do.

Merge remote-tracking branch 'upstream/main' into pythongh-113318-get…

@Getter

gh-113318: Better implementation of @Getter and @Setter in Argument C…

@Getter

[3.15] gh-113318: Fix @Getter and @Setter in Argument Clinic (GH-155778…

Navigation Menu

Repository navigation

Argument Clinic: Better internal implementation for `@getter` and `@setter` #113318

Linked PRs

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

Argument Clinic: Better internal implementation for @getter and @setter #113318

Description

Linked PRs

Activity

corona10 commented on Dec 20, 2023

erlend-aasland commented on Dec 20, 2023

serhiy-storchaka commented on Apr 29, 2025

skirpichev commented on Aug 12, 2025

serhiy-storchaka commented on Aug 14, 2026

encukou commented on Aug 19, 2026

serhiy-storchaka commented on Aug 19, 2026

serhiy-storchaka commented on Aug 19, 2026

serhiy-storchaka commented on Aug 20, 2026

encukou commented on Aug 20, 2026

serhiy-storchaka commented on Aug 20, 2026

serhiy-storchaka commented on Aug 20, 2026

Metadata

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

Issue actions

Argument Clinic: Better internal implementation for `@getter` and `@setter` #113318