[Numpy-discussion] Fix to #789 maybe not right.

Charles R Harris charlesr.harris@gmail....
Thu May 22 00:09:10 CDT 2008


On Wed, May 21, 2008 at 10:55 PM, Travis E. Oliphant <oliphant@enthought.com>
wrote:

> Charles R Harris wrote:
> >
> >
> > On Wed, May 21, 2008 at 9:32 PM, Travis E. Oliphant
> > <oliphant@enthought.com <mailto:oliphant@enthought.com>> wrote:
> >
> >     Charles R Harris wrote:
> >     > Really, all the increments and decrements should be inside
> >     > PyArray_FromArray, but calls to this function are scattered all
> >     over.
> >     I don't understand what you mean by this statement.    All functions
> >     that return an object and take a PyArray_Descr object steal a
> >     reference
> >     to the descriptor (even if it fails).   That's the rule.
> >
> >
> > Why should it not increment the reference itself? Note that calls to
> > this function are normally preceded by incrementing the reference,
> > probably because one wants to keep it around.
> I wouldn't say normally.   I would say sometimes.
>
> Normally you create a reference to the data-type and want
> PyArray_FromAny and friends to steal it (i.e. PyArray_DescrFromType).
>
> I wouldn't call stealing a reference count a side effect, but even if
> you want to call it that, it can't really change without a *huge*
> re-working effort for almost zero gain.   You would also have to re-work
> all the macros that take type numbers and construct data-type objects
> for passing to these functions.   I don't see the benefit at all.


I  agree with all that, which is why I'm not advocating a change. But it
does raise the bar for working with the C code and I think the current case
is an example of that.


> > I think it would be clearer to have the rule: you increment it, you
> > decrement it.
> Maybe, but Python's own C-API doesn't always follow that rule, there are
> functions that steal references.   Remember, PyArray_Descr was
> retrofitted as a Python Object.  It didn't use to be one.   This steal
> rule was the cleanest I could come up with --- i.e. it wasn't an idle
> decision.
>

I realize that Python does this too, I also note that getting reference
counting right is one of the more difficult aspects of writing Python
extension code. The more programmers have to know and keep in mind, the more
likely they are to make mistakes.

Chuck
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://projects.scipy.org/pipermail/numpy-discussion/attachments/20080521/af87aaac/attachment-0001.html 


More information about the Numpy-discussion mailing list