Thank you for PR. Can you verify if this solution works if MBCS is not enabled (Based on this comment the issue is not always reproducible)?

Thank you so much for your review!

Do you mean to test the utf8 mode introduced since Windows 10?

On my computer, when this feature is enabled, python starts abnormally (because the username contains unicode characters).

I checked python's source and it should work fine in utf8 mode.

With one exception, utf8 mode was introduced in April 2018, and CP_ACP was made equal to CP_UTF8 in May 2019 (1903).

Very few people use utf8 mode though, especially before version 1903 (because that would break most software).

In #18305 (comment) I explained why the code in #18305 (comment) cannot be reproduced .

You can try the tests in this PR, it should be 100% reproducible.

I don't have an access to the computer with Windows right now, but overall idea is to check if this patch don't introduce the regression to Windows users with enabled UTF-8 mode.
I would like to suggest the following schema. For Windows users first try MBCS conversion and if it fails then fallback to UTF8. For all other users - nothing new.

static inline PyObject* PyString_FromString(const char* u)
{
#if defined(_WIN32)
    PyObject* str = PyUnicode_DecodeMBCS(u, strlen(u), NULL);
    if (str) {
        return str;
    }
    // clears error indicator from failed conversion
    PyErr_Clear();
#endif
    return PyUnicode_FromString(u);
}

// PyString_FromStringAndSize looks similar to the above implementation.

and

static inline PyObject* PyUnicode_AsBytes(PyObject* unicode) {
#if defined(_WIN32)
    PyObject* bytes = PyUnicode_AsMBCSString(obj);
    if (bytes) {
        return bytes;
    }
     // clears error indicator from failed conversion
    PyErr_Clear();
#endif
    return PyUnicode_AsUTF8String(obj);
}

You are right!
But I'm worried that sometimes it will convert incorrectly without throwing an exception, maybe we can use https://docs.microsoft.com/en-us/windows/win32/api/winnls/nf-winnls-getacp.

it will convert incorrectly without throwing an exception

Do you mean the case when MBCS conversion fails, but UTF8 produces wrong result? But in this case it means that user provided a string in unnatural encoding for the OS. In my opinion, the only way to resolve such situation is to require to provide an encoding.

Can you check whenever MBCS encoding fails if provided string is in UTF8 on you Windows machine?

Can you check whenever MBCS encoding fails if provided string is in UTF8 on you Windows machine?

That's what I meant! I'm worried about garbled characters instead of throwing exceptions when utf8 mode is enabled (when passing a utf8 string to PyUnicode_DecodeMBCS).

But it's solved now!
And I tested it with windows 1803 + utf8 mode using a virtual machine, it works fine!

I've run into the same issue, and it looks like this PR just got left hanging. @VadimLevin , do you have the bandwidth to take another look since you looked at it in the past?

The CI failure after rebasing doesn't seem to be related to this PR, it succeeds on windows, and this PR only changes things under the _WIN32 macro.🤔

@asmorkalov Hi, sorry to bother you, can you approve the CI to run? Thanks!

Thanks!
The cv.VideoWriter test for linux fails, this seems to be a different problem that exists in c.
As mentioned in https://github.com/opencv/opencv/pull/21961/files#r871041445, other tests ensure that the conversion between OpenCV -> Python -> OpenCV is correct.
Maybe just mark this test as skipped on linux so it can be fixed in the future?