Regression in the AC generated code

[serhiy-storchaka]

Argument Clinic now generates less efficient code using _PyArg_ParseStackAndKeywords instead of _PyArg_UnpackKeywords. It is a regression introduced after establishing module state. For example see #23393. It may be a cause of #91404.

@erlend-aasland

[erlend-aasland]

FYI GH-23393 is also in 3.10, not just 3.11 (seeing that GH-91404 has the title Possible slowdown of regex searching in 3.11)

[erlend-aasland]

Did you benchmark the difference?

[serhiy-storchaka]

No, but _PyArg_ParseStackAndKeywords adds some overhead, and I do not see reason of not using more efficient code here. It seems that adding the implicit cls parameter broke some checks in Argument Clinic. These checks should ignore that parameter (as they ignore self), it does not affect the parsing code.

[erlend-aasland]

I do not see reason of not using more efficient code here

+1

It seems that adding the implicit cls parameter broke some checks in Argument Clinic. These checks should ignore that parameter (as they ignore self), it does not affect the parsing code.

In sqlite3, for the sqlite3.Connection class, I store a pointer to the module state in the class context. A module state lookup in any of the connection methods is simply a class context member lookup. We could do a similar optimisation here.

[serhiy-storchaka]

METH_METHOD was introduced in 3.9, and Argument Clinic support was added at the same time.