The _io module imports the os module purely for the use of os.device_encoding(). That function, though, is defined by posix and thus is already available to C code. To avoid this import dependency it would be better to simply expose the function somehow so that _io can just directly call the function.

Antoine has suggested Python/fileutils.c as a possible place, although I'm personally happy simply exposing the function from posix and making it a METH_O function so that as little code needs to change for _io to use it.

If it can help to bootstrap importlib, you can add:

PyObject* _Py_device_encoding(int fd);

And reuse it in posixmodule.c.

Either that or PyObject *_Py_device_encoding(PyObject *fd) would help since both _io and posixmodule.c will have a PyObject already for the fd.

I think fileutils.c is a good place. Then we can kill the import of "os" in io, which would be nice.

OK, then I will make this happen. People care whether it take an int or a PyObject? That seems to be the only open question.

2012/2/28 Brett Cannon <report@bugs.python.org>:
>
> Brett Cannon <brett@python.org> added the comment:
>
> OK, then I will make this happen. People care whether it take an int or a PyObject? That seems to be the only open question.

Probably int if it goes in fileutils.

Attached is a patch which is trying to do the refactor, but I'm somehow causing a something to be gc'ed when it shouldn't. If anyone spots something obvious let me know, else I will try to debug some more on my way home.

I don't seen any refcounting problem here, but your patch is buggy.
Before patch:

>>> os.device_encoding(0)
'UTF-8'

After patch:

>>> os.device_encoding(0) is None
True

Interestingly, it seems there is no test for os.device_encoding. Perhaps add one to test_os?

> your patch is buggy

It should be a missing #include. Maybe this one:

#ifdef HAVE_LANGINFO_H
#include <langinfo.h>
#endif

Victor's point fixed my issue.

As for a test, I will happily write one, but what can I test? A string for fd 0 and None for fd -1 (or is there some os call I can make to find the largest fd so I can test fd+1)?

> As for a test, I will happily write one, but what can I test? A string
> for fd 0 and None for fd -1 (or is there some os call I can make to
> find the largest fd so I can test fd+1)?

For fd 0, it should be a string if the fd is a tty.
(it should also point to a valid codec)
You can also try fd 42 (same, call isatty() on it).

> For fd 0, it should be a string if the fd is a tty.

os.device_encoding() returns None on a non-Windows OS without langinfo support.

locale.nl_langinfo(locale.CODESET) may be used in such test. (to check
for nl_langinfo() support?)

OK, so if on windows or isatty + nl_langinfo + CODESET, check fd 0 for a returned string that codecs.lookup and find something for. fd 42 (if not a tty) should return None.

New changeset c80d9a0bd5a0 by Brett Cannon in branch 'default':
Issue #14153 Create _Py_device_encoding() to prevent _io from having to import
http://hg.python.org/cpython/rev/c80d9a0bd5a0