On Fri, Apr 6, 2012 at 12:19 AM, Travis Oliphant <tra...@continuum.io>wrote:
> > On Apr 6, 2012, at 1:01 AM, Charles R Harris wrote: > > > > On Thu, Apr 5, 2012 at 11:57 PM, Travis Oliphant <tra...@continuum.io>wrote: > >> As of 1.5.1 this worked: >> >> >>> numpy.__version__ >> 1.5.1 >> >>> numpy.uint64(5) & 3 >> 1L >> >> >> So, this is a regression and a bug. It should be fixed so that it >> doesn't raise an error. I believe the scalars were special cased so that a >> raw 3 would not be interpreted as a signed int when it is clearly unsigned. >> The casting rules were well established over a long period. They had >> issues, but they should not have been changed like this in a 1.X release. >> >> Fortunately there are work-arounds and these issues arise only in corner >> cases, but we should strive to do better. >> >> > I disagree, promoting to object kind of destroys the whole idea of bitwise > operations. I think we *fixed* a bug. > > > That is an interesting point of view. I could see that point of view. > But, was this discussed as a bug prior to this change occurring? > > I just heard from a very heavy user of NumPy that they are nervous about > upgrading because of little changes like this one. I don't know if this > particular issue would affect them or not, but I will re-iterate my view > that we should be very careful of these kinds of changes. > > In this particular case, what should the behavior be? It would be ideal > if the scalar math did not just re-use the array-math machinery. Let's say > that scalars had their own bit-wise operator. What should the output of > > numpy.uint64(5) & 3 actually be? I think it should interpret the 3 as > unsigned and perform the operation (but not promote to an object). > > > Yes, I agree with that. We should think about how the scalar types combine in the common operations. It looks like it could be made more consistent. Chuck
_______________________________________________ NumPy-Discussion mailing list NumPy-Discussion@scipy.org http://mail.scipy.org/mailman/listinfo/numpy-discussion
