Although ga_call determines whether origin has a vectorcall, it needs to be unpacked the parameters that are already packed.
/-> origin.vectorcall(unpacked)
MakeTpCall(packed) -> ga_call -> PyObject_Call
-> origin.tp_call

We can advance the vectorcall judgment to the setup phase.

ga_vectorcall -> origin.vectorcall

or

ga_make_tp_call -> _PyObject_MakeTpCall(packed argument) -> origin.tp_call

This will have no effect on tp_call, which still only needs to be packed once, while vectorcall does not need packed/unpacked

We decided not to add it
see bpo-40369

The point of bpo-40369 issue seems to be to provide vectorcall to the GenericAliasType, means vectorcall for Queue[int]

However, my idea is to add vectorcall to gaobject, like this:

d = dict[int]
d(a=1, b=2) <-- vectorcall

The idea came from my desire to implement different subclasses of a type when implementing some kind of factory model：

class ChildClassBuilder:
    ...

    def set_type(self, t: Type):
        self.t = t

    def build():
        instance = self.t()
        ...
        return instance

In the case of high type-hint coverage, the self.t is often a gaobject.

Benchmark example：

def test():
    d = dict[int]
    for _ in range(1000000):
        d(a=1, b=2)

Speedup 20% in my wsl2

opt:
   ncalls  tottime  percall  cumtime  percall filename:lineno(function)
        1    0.000    0.000    1.728    1.728 <string>:1(<module>)
        1    1.728    1.728    1.728    1.728 test.py:4(test)

master:
   ncalls  tottime  percall  cumtime  percall filename:lineno(function)
        1    0.000    0.000    2.271    2.271 <string>:1(<module>)
        1    2.271    2.271    2.271    2.271 test.py:4(test)

There is a small problem with the example:

d = dict[int]

should

d = dict[str, int]

@dennis, if/when the PR looks good to you, you can merge it.

I profiled dict[str, int](a=1, b=2), and it looks like a decent chunk of time comes from PyUnicode_New as used by PyObject_SetAttrString.

You could also try replacing PyObject_SetAttrString with PyObject_SetAttr and adding "__orig_class__" to the global strings with Tools/scripts/generate_global_objects.py, probably for a later PR.

You could also try replacing PyObject_SetAttrString with PyObject_SetAttr and adding "__orig_class__" to the global strings with Tools/scripts/generate_global_objects.py, probably for a later PR.

Already done via make regen-global-objects

Also, I found that regen-global-objects depends on regen-deepfreeze, which should be a bug. This causes the regen-global-objects to be executed with the compile task first, and it does not compile successfully at this point because of the new symbols(_Py_ID(orig_class)) added.

New changeset 1ea055b by penguin_wwy in branch 'main':
bpo-47067: Optimize calling GenericAlias objects (GH-31996)
1ea055b

Closing as the associated PR is merged.