BUG: Overflow reported when negating uint64

[bashtage]

Describe the issue:

A usually valid operation, negation of an unsigned integer, is raising an underflow warning.

Reproduce the code example:

import numpy as np

a = np.uint64(0xffffffff)
b = -a

Error message:

`RuntimeWarning: overflow encountered in scalar negative`

Runtime information:

1.25.2
3.11.5 | packaged by Anaconda, Inc. | (main, Sep 11 2023, 13:26:23) [MSC v.1916 64 bit (AMD64)]

Context for the issue:

Wrapping unsigned integer values is fairly common in C using negation. This used to be free from warnings, so was surprised to see this.

[ngoldbaum]

I can reproduce this. It looks like this changed sometime after NumPy 1.23, with the 1.23.3 wheel:

In [1]: import numpy as np

In [2]: a = np.uint64(0xffffffff)

In [3]: b = -a

and the NumPy 1.24.3 wheel:

In [1]: import numpy as np
   ...: 
   ...: a = np.uint64(0xffffffff)

In [2]: b = -a
<ipython-input-2-064521cdc161>:1: RuntimeWarning: overflow encountered in scalar negative

[charris]

Works for me: gcc 13.2.1 in the development branch.

Apparently C++ does have a preference, the hex literal has the type of the first type where it fits:

• int
• unsigned int
• long int
• unsigned long int
• long long int
• unsigned long long int

Maybe that is related? I assume the conversion is done by Python or a library.

[ngoldbaum]

I bisected this to #21648, ping @Micky774.

(I posted an incorrect comment earlier that said a different PR was the cause and I deleted that comment, sorry for the duplicate e-mail)

[seberg]

This needs a decision that we want to allow some scalar operators to wrap silently to users while we warn for others.
I.e. there was a reason for this change, that if you have an unsigned int and don't realize it, getting the wrong result deserves a warning. (something that NEP 50 may make slightly more common)

[bashtage]

Would it be possible to separate unary negation from multiplication by a negative number, where unary would not warn?

[seberg]

Yes, we can do all of that. We just need to decide to do it.

[Micky774]

IMO it's safer to offer a warning that can be suppressed than to remove the warning because the underlying operation is sometimes intended. The behavior is still noteworthy in some cases and I think it is safe to optimize for those cases.

[RubTalha]

https://stackoverflow.com/questions/77467422/runtimewarning-overflow-encountered-in-scalar-negative

[mattip]

In the latest triage meeting we decided to close this for now. In the future, we would prefer that scalars perform more like 0d arrays (and not warn), but that would be part of a larger effort and may even need a NEP.