[Regression] TypeVar crashes when name is specified via keyword parameters

[mxschmitt]

Crash report

What happened?

from typing import TypeVar
TypeVar(name="Handler")

Good: 3.11.X
Bad: 3.12.0

Downstream issue: jfhbrook/pyee#134

CPython versions tested on:

3.12

Operating systems tested on:

Linux, macOS

Output from running 'python -VV' on the command line:

Python 3.12.0 (main, Oct 12 2023, 00:42:11) [GCC 12.2.0]

Linked PRs

• gh-110782: Fix crash when TypeVar is constructed with keyword args #110784
• [3.12] gh-110782: Fix crash when TypeVar is constructed with keyword args (GH-110784) #110787

[mxschmitt]

Investigation: Changing TypeVar(name="Handler") to TypeVar("Handler") fixes it for me.

[JelleZijlstra]

This is possibly a bug in Argument Clinic (see

cpython/Objects/typevarobject.c

Line 326 in e733136

typevar.__new__ as typevar_new

): if name is passed as a positional argument and there are no constraints, constraints is an empty tuple, but if it's a keyword argument, constraints is NULL.

I'll send a PR in a moment with a workaround in typevarobject.c.

[JelleZijlstra]

Fix is merged, and will be in 3.12 soon; 3.12.1 should include the fix.

I'll note that using a keyword argument for the name of a TypeVar is not an officially supported part of the type system, so to make your code portable across type checkers and Python versions, you should use a positional argument for the name. It appears that pyright supports the keyword argument form, but mypy does not. We'll continue to support the keyword argument form at runtime for backwards compatibility.

[sobolevn]

Thanks!

@JelleZijlstra would you accept

cpython/Objects/typevarobject.c

Lines 329 to 330 in e733136

	*constraints: object
	*

to be changed to just

*constraints

?

Because this seems like a fortunate event that this notation works. I was not able to find that this is a documented approach.

If so, I would send a PR :)

[AlexWaygood]

Because this seems like a fortunate event that this notation works. I was not able to find that this is a documented approach.

We should try to make sure AC raises an error if it encounters syntax like this which isn't supported! Cc. @erlend-aasland

[sobolevn]

Ok, will do after #110864

[JelleZijlstra]

Thanks @sobolevn for investigating further!

I'm fine with whatever syntax works here, but ideally AC should support annotations on *args (e.g., I should be able to write *args: bool to get any number of bool args), and as Alex says it should error on unsupported constructs.