ENH: Add bitwise_and identity

[charris]

Couple of simple fixes for the bitwise operators.

• Give bitwise_xor the identity zero rather than none.
• Give bitwise_and an identity of all bits set (-1).

Still needs tests, so do not merge.

[madphysicist]

That last sentence is not strictly necessary. There could potentially be other ufuncs that use this.

[charris]

There might be others, but I could not think of any. Hence the bit of explanation as to why this otherwise useless macro is there. I suppose it could be "Introduced" instead of "Intended".

[jaimefrio]

Nice! Last week I went to a "Hacker Garten" event, and spent an hour or so introducing NumPy to one of the "hacker gardeners", and this came up in the "easy fix" issue list. We spent about 10 minutes seeing if it would be a good fit for his first PR, which it obviously wasn't.

[charris]

For reference, the other option is #7061 which uses ReorderableNone as the identity. That should only make a difference for empty arrays. Another corner case to deal with is empty object arrays. Object arrays are already special cased:

            if (PyArray_ISOBJECT(arr) && PyArray_SIZE(arr) != 0) {
                assign_identity = NULL;
            }

We can choose what to use as the return value in that case, if any, or raise an error.

[seberg]

Hmmm forgot about that. I admit the object cases are always a little weird. But that is also the case for half the current identities. For non-object, to me this option is "obviously right".

[charris]

Updated tests.

[charris]

also might be good to add a test for bitwise_and.reduce for non-boolean inputs,

@madphysicist The ones isn't boolean, it is all bits set. We only need to check that everything works correctly bitwise for every bit, and that is done.

[madphysicist]

Good point. I rescind my comment.

[charris]

Testing with PyInt_FromLong to see what happens.

[charris]

OK, works, so squashing that commit into others. The only open questions is what to do with empty object arrays. Because non-empty object arrays do not use the identity, it might be most consistent to raise an error in that case, but currently the identity is returned.

[seberg]

Thanks. @jaimefrio since you already looked at it, do you have a preference whether we should -1 return for empty object arrays, or rather make the identity "crash" on object arrays if used? I can't make up my mind and either way has its downsides ;).

[njsmith]

I guess I prefer erroring out on empty object arrays because "in the face of ambiguity refuse to guess". But it doesn't seem hugely important either way.

[charris]

Note that erroring out will reguire a deprecation cycle, so is probably independent of this PR.

[seberg]

In that case, I guess we should put this in as a bug fix and then open an issue about the other stuff? Will merge in a bit if nothing crops up.

[seberg]

OK, thanks Chuck! Will open an issue about the object identity.

[xflr6]

If this is finally resolved (great!) then I guess #4696 and #5250 can also be closed.

[charris]

@xflr6 Thanks for the heads up.