+--------------+--------+----------------------+
| Benchmark    | master | vectorcall           |
+==============+========+======================+
| bench filter | 191 ns | 151 ns: 1.26x faster |
+--------------+--------+----------------------+

Like reversed(https://bugs.python.org/issue41922), it looks okay to update filter() to use PEP 590.

+0 I don't see any downside.

Note, the benchmark only times instantiation of the filter object.  It doesn't actually run the iterator which is where most of the runtime cost is spent.  So in actual code there is almost zero benefit.  For example, add "list" to the statement:   stmt="b = list(filter(lambda x: x % 2 == 0, a))" and the improvement disappears.

> +0 I don't see any downside.

Thank you Raymond :)
I agree with your view, At first, I thought that instantiation time reducing is also meaningful and we already applied PEP 590 for some of the types to reduce instantiation. (e.g range(), list(), dict(), bool(), reversed(), type()..)

And also I thought that filter is one of the well-used features in Python so it was one of the candidates to apply from my sight. ;)

But if this PR only consume maintenance cost, I am okay with not to apply it doesn't matter :)

And happy new year Raymond(Sorry I am late ;) But the lunar new year was only 2weeks ago so not too late lol)

Happy new year to you as well :-)

FWIW, I don't think there is much of a maintenance burden; hence, the +0.  You've already done the work.

See also bpo-43447: "Generate vectorcall code to parse arguments using Argument Clinic".

New changeset 9a9c11ad41d7887f3d6440e0f0e8966156d959b5 by Dong-hee Na in branch 'master':
bpo-43287: Use PEP 590 vectorcall to speed up filter() (GH-24611)
https://github.com/python/cpython/commit/9a9c11ad41d7887f3d6440e0f0e8966156d959b5