[Python-Dev] Tweak to PEP 523 for storing a tuple in co_extra

[Yury Selivanov]

On 2016-09-03 5:19 PM, Brett Cannon wrote:
>
>
> On Sat, 3 Sep 2016 at 16:43 Yury Selivanov <yselivanov.ml at gmail.com 
> <mailto:yselivanov.ml at gmail.com>> wrote:
>
>
>
>     On 2016-09-03 4:15 PM, Christian Heimes wrote:
>     > On 2016-09-04 00:03, Yury Selivanov wrote:
>     >>
>     >> On 2016-09-03 12:27 PM, Brett Cannon wrote:
>     >>> Below is the `co_extra` section of PEP 523 with the update
>     saying that
>     >>> users are expected to put a tuple in the field for easier
>     simultaneous
>     >>> use of the field.
>     >>>
>     >>> Since the `co_extra` discussions do not affect CPython itself I'm
>     >>> planning on landing the changes stemming from the PEP probably
>     on Monday.
>     >> Tuples are immutable.  If you have multiple co_extra users then
>     they
>     >> will have to either mutate tuple (which isn't always possible, for
>     >> instance, you can't increase size), or to replace it with
>     another tuple.
>     >>
>     >> Creating lists is a bit more expensive, but item access speed
>     should be
>     >> in the same ballpark.
>     >>
>     >> Another question -- sorry if this was discussed before -- why
>     do we want
>     >> a PyObject* there at all?  I.e. why don't we create a dedicated
>     struct
>     >> CoExtraContainer to manage the stuff in co_extra? My
>     understanding is
>     >> that the users of co_extra are C-level python optimizers and
>     profilers,
>     >> which don't need the overhead of CPython API.
>
>
> As Chris pointed out in another email, the overhead is only in the 
> allocation, not the iteration/access if you use the PyTuple macros to 
> get the size and index into the tuple the overhead is negligible.

Yes, my point was that it's as cheap to use a list as a tuple for 
co_extra.  If we decide to store PyObject in co_extra.

>     >>
>     >> This way my work to add an extra caching layer (which I'm very much
>     >> willing to continue to work on) wouldn't require another set of
>     extra
>     >> fields for code objects.
>     > Quick idea before I go to bed:
>     >
>     > You could adopt a similar API to OpenSSL's CRYPTO_get_ex_new_index()
>     > API,
>     >
>     https://www.openssl.org/docs/manmaster/crypto/CRYPTO_get_ex_new_index.html
>     >
>     >
>     > static int code_index = 0;
>     >
>     > int PyCodeObject_NewIndex() {
>     >      return code_index++;
>     > }
>     >
>     > A library like Pyjion has to acquire an index first. In further
>     calls it
>     > uses the index as offset into the new co_extra field. Libraries
>     don't
>     > have to hard-code their offset and two libraries will never
>     conflict.
>     > PyCode_New() can pre-populate co_extra with a PyTuple of size
>     > code_index. This avoids most resizes if you load Pyjion early. For
>     > code_index == 0 leaf the field NULL.
>
>     Sounds like a very good idea!
>
>
> The problem with this is the pre-population. If you don't get your 
> index assigned before the very first code object is allocated then you 
> still have to manage the size of the tuple in co_extra. So what this 
> would do is avoid the iteration but not the allocation overhead.
>
> If we open up the can of worms in terms of custom functions for this 
> (which I was trying to avoid), then you end up with Py_ssize_t 
> _PyCode_ExtraIndex(), PyObject *
>   _PyCode_GetExtra(PyCodeObject *code, Py_ssize_t index), and int 
> _PyCode_SetExtra(PyCodeObject *code, Py_ssize_t index, PyObject *data) 
> which does all the right things for creating or resizing the tuple as 
> necessary and which I think matches mostly what Nick had proposed 
> earlier. But the pseudo-code for _PyCode_GetExtra() would be::
>
>   if co_extra is None:
>     co_extra = (None,) * _next_extra_index;
>     return None
>   elif len(co_extra) < index - 1:
>     ... pad out tuple
>     return None
>    else:
>      return co_extra[index]
>
> Is that going to save us enough to want to have a custom API for this?

But without that new API (basically what Christian proposed) you'd need 
to iterate over the list in order to find the object that belongs to 
Pyjion.  If we manage to implement my opcode caching idea, we'll have at 
least two known users of co_extra.  Without a way to claim a particular 
index in co_extra you will have some overhead to locate your objects.

Yury