bpo-46303: Don't define _Py_stat() and _Py_wstat() on Windows

[vstinner]

https://bugs.python.org/issue46303

[eryksun]

GetFileAttributesW() doesn't follow a reparse point, such as a symlink, in the final component of the path. If tcl_library_path is the path of a symlink, checking existence of the target requires CreateFileW(wpath, 0, 0, NULL, OPEN_EXISTING, FILE_FLAG_BACKUP_SEMANTICS, NULL). If the call succeeds, i.e. the call doesn't return INVALID_HANDLE_VALUE, the target exists and is accessible; close the handle.

[eryksun]

Here's the suggested implementation:

static int
file_exists(PyObject *path)
{
    wchar_t *wpath = PyUnicode_AsWideCharString(path, NULL);
    if (wpath == NULL) {
        return -2;
    }

    HANDLE hfile = CreateFileW(wpath, 0, 0, NULL, OPEN_EXISTING,
                        FILE_FLAG_BACKUP_SEMANTICS, NULL);
    if (hfile != INVALID_HANDLE_VALUE) {
        CloseHandle(hfile);
    }
    
    PyMem_Free(wpath);

    return (hfile == INVALID_HANDLE_VALUE) ? -1 : 0;
}

This returns -1 when the file doesn't exist, since file_exists() replaces _Py_stat().

[vstinner]

I didn't expect that checking if a file exists would be so "complicated" :-) I'm waiting for @pacampbell to see if further changes are needed. Maybe this PR is not needed to build Python with clang.

[eryksun]

The GetFileAttributesW() call is like checking for existence with lstat() instead of stat(). That's probably good enough for tcl_library_path. Python's r"tcl\tclX.Y" is installed as a normal directory. However, I don't think the CreateFileW() call is complicated, and it makes file_exists() generally correct.

For the current implementation, the use of res == -1 in _get_tcl_lib_path() is wrong. The current implementation returns 1 if the file exists and 0 if it doesn't exist. This has to be fixed even if you keep the GetFileAttributesW() call.

[vstinner]

https://bugs.python.org/issue46303#msg410377 is now fixed, this change is not needed.