[Python-Dev] [Python-3000] Stabilizing the C API of 2.6 and 3.0

[Gregory P. Smith]

On Fri, May 30, 2008 at 1:37 AM, M.-A. Lemburg <mal at egenix.com> wrote:
> On 2008-05-30 00:57, Nick Coghlan wrote:
>>
>> M.-A. Lemburg wrote:
>>>
>>> * Why can't we have both PyString *and* PyBytes exposed in 2.x,
>>> with one redirecting to the other ?
>>
>> We do have that - the PyString_* names still work perfectly fine in 2.x.
>> They just won't be used in the Python core codebase anymore - everything in
>> the Python core will use either PyBytes_* or PyUnicode_* regardless of which
>> branch (2.x or 3.x) you're working on. I think that's a good thing for ease
>> of maintenance in the future, even if it takes people a while to get their
>> heads around it right now.
>
> Sorry, I probably wasn't clear enough:
>
> Why can't we have both PyString *and* PyBytes exposed as C
> APIs (ie. visible in code and in the linker) in 2.x, with one redirecting
> to the other ?
>
>>> * Why should the 2.x code base turn to hacks, just because 3.x wants
>>> to restructure itself ?
>>
>> With the better explanation from Greg of what the checked in approach
>> achieves (i.e. preserving exact ABI compatibility for PyString_*, while
>> allowing PyBytes_* to be used at the source code level), I don't see what
>> has been done as being any more of a hack than the possibly more common
>> "#define <oldname> <newname>" (which *would* break binary compatibility).
>>
>> The only things that I think would tidy it up further would be to:
>> - include an explanation of the approach and its effects on API and ABI
>> backward and forward compatibility within 2.x and between 2.x and 3.x in
>> stringobject.h
>> - expose the PyBytes_* functions to the linker in 2.6 as well as 3.0
>
> Which is what I was suggesting all along; sorry if I wasn't
> clear enough on that.
>
> The standard approach is that you provide #define redirects from the
> old APIs to the new ones (which are then picked up by the compiler)
> *and* add function wrappers to the same affect (to make linkers,
> dynamic load APIs such ctypes and debuggers happy).
>
>
> Example from pythonrun.h|c:
> ---------------------------
>
> /* Use macros for a bunch of old variants */
> #define PyRun_String(str, s, g, l) PyRun_StringFlags(str, s, g, l, NULL)
>
> /* Deprecated C API functions still provided for binary compatiblity */
>
> #undef PyRun_String
> PyAPI_FUNC(PyObject *)
> PyRun_String(const char *str, int s, PyObject *g, PyObject *l)
> {
>        return PyRun_StringFlags(str, s, g, l, NULL);
> }
>

Okay, how about this?  http://codereview.appspot.com/1521

Using that patch, both PyString_ and PyBytes_ APIs are available using
function stubs similar to the above.  I opted to define the stub
functions right next to the ones they were stubbing rather than
putting them all at the end of the file or in another file but they
could be moved if someone doesn't like them that way.

> I still believe that we should *not* make "easy of merging" the
> primary motivation for backporting changes in 3.x to 2.x. Software
> design should not be guided by restrictions in the tool chain,
> if not absolutely necessary.
>
> The main argument for a backport needs to be general usefulness
> to the 2.x users, IMHO... just like any other feature that
> makes it into 2.x.
>
> If merging is difficult then this needs to be addressed, but
> there are more options to that than always going back to the
> original 2.x trunk code. I've given a few suggestions on how
> this could be approached in other emails on this thread.

I am not the one doing the merging or working on merge tools so I'll
leave this up to those that are.

-gps