Windows build importlib performance regression

まず、提供された import_models.py と some_class.py を使い、-X importtime を指定して Windows 10 でのタイミング差を再現します。次に、python -m compileall . の前後、および -B を指定した場合の実行を比較します。モジュールの読み込みに関係する importlib と Windows 固有のインポートパスを確認します。キャッシュ済みのインポートが遅くなる原因を特定し、検証済みの比較によってリグレッションを文書化または修正できれば完了です。

Bug report

I've a file import_models.py:

import time
from ctypes import c_longdouble
from multiprocessing import Value


def import_models_func(_, import_value, exec_value):
    init = time.process_time()
    import some_class
    import_value.value = time.process_time() - init
    exec_value.value = 0


if __name__ == '__main__':
    import_models_func(0, Value(c_longdouble, lock=False), Value(c_longdouble, lock=False))

The some_class.py is:

import time


class SomeClass:
    __slots__ = ('a', 'b', 'c')

    def __init__(self):
        self.a = time.time()
        self.b = self.a / 100
        self.c = self.b ** 2

    B = 2
    for _ in range(10):
        B **= 2

    @property
    def a_val(self):
        return self.a

    @a_val.getter
    def a_val(self):
        return self.a

    @property
    def b_val(self):
        return self.b

    @b_val.getter
    def b_val(self):
        return self.b

    @property
    def c_val(self):
        return self.c

    @c_val.getter
    def c_val(self):
        return self.c

The directory containing both files has no __pycache__. I run python -X importtime import_models.py and receive:

import time:       521 |        521 | some_class

Then I run python -m compileall . and re-run the command above, the output is:

import time:       548 |        548 | some_class

I've tried the sampling with multiprocessing governor, which is run with -B. The governor runs this scenario for 100 times, both mean and median statistics values of the timing measured mostly provide with the same results: the timing without __pycache__ is less than when the source files are pre-compiled. Which is a non-sense keeping in mind what the pre-compiled byte-code cache is built for.

I've searched for Windows-related fixes through the versions 3.9 to 3.12 and found nothing about this regression.

Linux build in a native (non-WSL, non-emulated) environment does not have such an issue (the bytecode-compiled version gets imported faster, up to 10 times, than a raw non-compiled source)

Your environment

• CPython versions tested on: Python 3.9.13 (tags/v3.9.13:6de2ca5, May 17 2022, 16:36:42) [MSC v.1929 64 bit (AMD64)] on win32
• Operating system and architecture: Windows 10 Enterprise 21H2 19044.2251