Replace PyEval_GetFuncName/PyEval_GetFuncDesc

[jdemeyer]

BPO	37645
Nosy	@vstinner, @encukou, @rlamy, @jdemeyer, @miss-islington
PRs	python/cpython#14890 python/cpython#15295

^{Note: these values reflect the state of the issue at the time it was migrated and might not reflect the current state.}

Show more details

GitHub fields: ```python assignee = None closed_at = None created_at = labels = ['interpreter-core', '3.9'] title = 'Replace PyEval_GetFuncName/PyEval_GetFuncDesc' updated_at = user = 'https://github.com/jdemeyer' ``` bugs.python.org fields: ```python activity = actor = 'petr.viktorin' assignee = 'none' closed = False closed_date = None closer = None components = ['Interpreter Core'] creation = creator = 'jdemeyer' dependencies = [] files = [] hgrepos = [] issue_num = 37645 keywords = ['patch'] message_count = 13.0 messages = ['348255', '348257', '349007', '349690', '349691', '349807', '349903', '349904', '351946', '352019', '352028', '356045', '356134'] nosy_count = 5.0 nosy_names = ['vstinner', 'petr.viktorin', 'Ronan.Lamy', 'jdemeyer', 'miss-islington'] pr_nums = ['14890', '15295'] priority = 'normal' resolution = None stage = 'patch review' status = 'open' superseder = None type = None url = 'https://bugs.python.org/issue37645' versions = ['Python 3.9'] ```

[jdemeyer]

PyEval_GetFuncName is bad API because

1. It hardcodes a limited number of function classes (which doesn't even include all function classes in the core interpreter) instead of supporting duck-typing.
2. In case of a "function" object, it relies on a borrowed reference to the function.
3. It is returning a C string instead of a Python string, so we cannot return a new reference even if we wanted to.

Since PyEval_GetFuncName and PyEval_GetFuncDesc are always used together, we might as well replace them by a single function with a proper API.

[jdemeyer]

4. It uses the __name instead of the __qualname

[jdemeyer]

Another solution would be to change the __str__ of various function objects to a prettier output. For example, we currently have

>>> def f(): pass
>>> print(f)
<function f at 0x7f9f4bbe5e18>

We could change this to

>>> def f(): pass
>>> print(f)
f()

and then use "%S" to display the functions in error messages. But I have a feeling that this is a more controversial change than PR 14890.

[vstinner]

Maybe repr(func) should be left unchanged, but str(func) can be enhanced?

[jdemeyer]

Maybe repr(func) should be left unchanged, but str(func) can be enhanced?

Yes, that is what I meant.

[encukou]

I am not convinced.

I'm wary of making error messages depend on the str representation of a function; that would prevent us from changing it later. I'm wary of "%S" used in error messages. Those are for the programmer, not the user, so they should prefer __repr__.

I train beginners to recognize "\<function f at 0x7f9f4bbe5e18>" as a sign of omitted parentheses. The ugliness is useful: it shows you're dealing with an internal object, not a data value.

So, I think "\<function f>" is much better than just "f()". I wouldn't mind "\<function f()>" (maybe even with the full signature), but that doesn't quite help this case. (I don't care much for the "at 0x7f9f4bbe5e18" part, but that's not the issue here.)

[jdemeyer]

I'm wary of making error messages depend on the str representation of a function; that would prevent us from changing it later.

Why wouldn't we be able to change anything? Typically, the exact string of an error message is NOT part of the API (the exception *type* is, but we're not talking about that).

I'm wary of "%S" used in error messages. Those are for the programmer, not the user

I'm not following here. Given that Python is a programming language, the user *is* the programmer.

Anyway, you don't have to be convinced. I'm trying to solve a problem here and I have two approaches (PR 14890 and PR 15295). I'm more inclined towards PR 15295, but if you like the idea of PR 14890 better, we can go with that instead.

[jdemeyer]

I'm wary of "%S" used in error messages.

Maybe you're misunderstanding something. The goal is not really to change error messages, only the way how they are produced. For example, we currently have

>>> def f(): pass
>>> f(**1)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
TypeError: f() argument after ** must be a mapping, not int

This is about how the "f()" in the error message is produced. Currently, this uses PyEval_GetFuncName() and PyEval_GetFuncDesc(). For the reasons explained in this issue, I want to replace that.

I see two ways of doing this:

1. (PR 14890) Write a new function _PyObjectFunctionStr(func) which returns func.\_qualname + "()" with a suitable fallback if there is no __qualname attribute. At some point, we could also introduce a %F format character for this.

2. (PR 15295) Use str(func) in the error message and change various __str__ methods (really tp_str functions) to give a more readable output.

[encukou]

I like PR 14890 better. I like the separation of representation for error messages (where it's clearer that this is a callable) and for __str__.

Also, changing the __str__ of functions would need much wider discussion than on issues/PRs.

I left some comments on the PR.

[jdemeyer]

I left some comments on the PR.

I don't see anything. Either I'm doing something wrong or GitHub is messed up.

[encukou]

My bad, I didn't publish the comments. They should be there now.

[miss-islington]

New changeset bf17d41826a8bb4bc1e34ba6345da98aac779e41 by Miss Islington (bot) (Jeroen Demeyer) in branch 'master': bpo-37645: add new function _PyObject_FunctionStr() (GH-14890) https://github.com/python/cpython/commit/bf17d41826a8bb4bc1e34ba6345da98aac779e41

[encukou]

With this change, CPython no longer uses PyEval_GetFuncName/PyEval_GetFuncDesc internally. Shall we deprecate/discourage them? Shall we expose PyObject_FunctionStr publicly?

[sobolevn]

I've recently covered these two functions in tests, so this issue caught my eye. Can we start the deprecation process? Or is there anything holding us back?

[encukou]

Before deprecating them, we should provide a good replacement for the use case they served. It could mean exposing PyObject_FunctionStr, which would mean documenting it, and deciding what kind of guarantees we have for it. Is the output gonna change? Or is this something like “the perferred way to reference callable objects” -- and in that case, why isn't it the same as str? Or instead of exposing PyObject_FunctionStr there could just be a “recipe” with hints on how to replace them.

[vstinner]

I found 3 projects using PyEval_GetFuncName() in PyPI top 8,000 projects (at 2024-03-16):

• Nuitka (2.1.2)
• pyuwsgi-2.0.23.post0
• uwsgi (2.0.24)

Code:

Nuitka-2.1.2.tar.gz: Nuitka-2.1.2/nuitka/build/static_src/HelpersCalling.c: return PyEval_GetFuncName(PyMethod_GET_FUNCTION(object));
pyuwsgi-2.0.23.post0.tar.gz: pyuwsgi-2.0.23.post0/plugins/python/profiler.c: PyEval_GetFuncName(arg), code->co_argcount, code->co_stacksize);
uwsgi-2.0.24.tar.gz: uwsgi-2.0.24/plugins/python/profiler.c: PyEval_GetFuncName(arg), code->co_argcount, code->co_stacksize);

I found no matches for PyEval_GetFuncDesc().