[Python-Dev] [numpy wishlist] Interpreter support for temporary elision in third-party classes

[Nathaniel Smith]

On Sat, Jun 7, 2014 at 1:37 AM, Greg Ewing <greg.ewing at canterbury.ac.nz> wrote:
> Julian Taylor wrote:
>>
>> tp_can_elide receives two objects and returns one of three values:
>> * can work inplace, operation is associative
>> * can work inplace but not associative
>> * cannot work inplace
>
>
> Does it really need to be that complicated? Isn't it
> sufficient just to ask the object potentially being
> overwritten whether it's okay to overwrite it?
> I.e. a parameterless method returning a boolean.

For the numpy case, we really need to see all the operands, *and* know
what the operation in question is. Consider

tmp1 = np.ones((3, 1))
tmp2 = np.ones((1, 3))
tmp1 + tmp2

which returns an array with shape (3, 3).

Both input arrays are temporaries, but neither of them can be stolen
to use for the output array.

Or suppose 'a' is an array of integers and 'b' is an array of floats,
then 'a + b' and 'a += b' have very different results (the former
upcasts 'a' to float, the latter has to either downcast 'b' to int or
raise an error). But the casting rules depend on the particular input
types and the particular operation -- operations like & and << want to
cast to int, < and > return bools, etc. So one really needs to know
all the details of the operation before one can determine whether
temporary elision is possible.

-n

-- 
Nathaniel J. Smith
Postdoctoral researcher - Informatics - University of Edinburgh
http://vorpus.org