It would be good to have the library work unchanged on both implementations.

If subprocess only failed later, it would be less good, as the stdlib would then set an example that doesn't actually succeed.

Note that the attached patch (by flox) does NOT implement the discussed "or" tests on sysm.platform; it instead checks whether _subprocess it importable.  Is the assumption of an accelerator module itself too implementation-specific?  I'm also not sure that the test in the patch isn't too broad -- is windows the only platform with _subprocess?  Because if not, then the new test would mess up logic later in the file, such as that at line 635 in Popen.__init__