json: make "memo" dict local to scan_once call

[colesbury]

Feature or enhancement

The Modules/_json.c parser is mostly stateless (or the state is immutable). The one exception is the "memo" dictionary, which is used to avoid duplicate PyUnicodeObject instances for the same JSON C strings.

cpython/Modules/_json.c

Lines 696 to 700 in 289af86

	memokey = PyDict_SetDefault(s->memo, key, key);
	if (memokey == NULL) {
	goto bail;
	}
	Py_SETREF(key, Py_NewRef(memokey));

The memo dictionary is already cleared after each call scan_once:

cpython/Modules/_json.c

Line 1118 in 289af86

PyDict_Clear(self->memo);

We should move the creation and destruction of the memo dict to the invocation of scan_once instead of having it as part of the module state. This will avoid contention on the dictionary locks in --disable-gil builds if multiple threads are concurrently parsing JSON strings.

For an example modification, see colesbury/nogil-3.12@964bb33962.

Linked PRs

• gh-111928: make "memo" dict local to scan_once call #112005

[aisk]

I just took a look at the example modification, and run the test on it locally and it succeded. I think there is no more works to do, so just cherry-picked it and made a PR.

If anyone think there need some improves, I want to investigate on it.

[colesbury]

Thanks @aisk!