Support div_wrapping/rem_wrapping for numeric arithmetic kernels

[wForget]

Which issue does this PR close?

Closes #7158.

Rationale for this change

The div/rem operations may also overflow, so I want to add div_wrapping/rem_wrapping operations to wrapping on overflow. Make it consistent with the logic of add/sub/mul so that we can implement fail_on_overflow feature for div/rem in downstream datafusion.

What changes are included in this PR?

add div_wrapping/rem_wrapping functions

Are there any user-facing changes?

yes, newly added div_wrapping/rem_wrapping functions

[tustvold]

The absence of these kernels is intentional - #2647

[wForget]

The absence of these kernels is intentional - #2647

Thank you for the information, if I understand correctly, it is to avoid inconsistent behavior caused by unchecked divide by zero returning NaN. This PR uses mod_wrapping method directly which keeps panics on divide by zero, only wrapping on overflow for signed_integer::MIN / -1.

[tustvold]

The motivation for the wrapping variants is solely performance, as they vectorise better and are therefore orders of magnitude faster. This does not apply to the division operators, I'm therefore unclear on why one would want an unchecked version, when it will be no faster

[wForget]

The motivation for the wrapping variants is solely performance, as they vectorise better and are therefore orders of magnitude faster. This does not apply to the division operators, I'm therefore unclear on why one would want an unchecked version, when it will be no faster

Thank you for your explanation. My motivation was just to provide div/rem implementation similar to add_wrapping/sub_wrapping that does not fail when overflow occurs. That way we can easily provide consistent behavior for binary operators in Datafusion. BinaryExpr in Datafusion defines a fail_on_overflow property which when false means it will not fail on overflow, and there is a similar behavior in Spark.

[tustvold]

If I'm not mistaken, these new kernels will panic on divide by zero?

[wForget]

If I'm not mistaken, these new kernels will panic on divide by zero?

Yes, should divide by zero be considered as overflow?

[tustvold]

Yes, should divide by zero be considered as overflow?

At the very least it shouldn't panic 😅

I personally feel it would be better to simply document that fail_on_overflow doesn't impact division operators. It seems a pretty tough pill to effectively double the codegen in order to handle a single set of arguments that overflows... Especially when in general overflow leads to an error (e.g. in the cast kernels).