Potential type instability for integer promotions?

[mbauman]

In changing the documentation for #6254, I realized that even the more complicated description of integer promotion is incorrect because promote(x) = (x,). While I haven't experienced this myself in real code, I think the current behavior here is surprising, particularly when calling promote with a splatted variable-length iterable. This PR adds specialized methods for promote with one SmallSigned/SmallUnsigned argument so the behavior matches the documentation.

With this PR, typeof(promote(uint8(3))) == (Uint,)

[StefanKarpinski]

@JeffBezanson, thoughts?

[mbauman]

I realize this is pretty core-machinery level here, so perhaps it would have been better to simply open an issue. But I hacked this together and the tests pass on my 64-bit machine. (That said, a quick and basic grep doesn't show any splatted promote calls in all of Base/Test.)

[StefanKarpinski]

No, this is not an unreasonable thing to do. The core promotion machinery is super delicate in terms of overall performance and @JeffBezanson is the expert there, so we should wait for his eyeballs.

[mbauman]

You know, it just dawned on me that this isn't a good solution, either. There's no promotion when the types are the same, so we'd have the opposite problem with homogenous collections of small integers:

julia> typeof(promote(int8(1),int8(1)))
(Int8,Int8)
julia> typeof(promote(int8(1))) # with this patch
(Int64,)

[StefanKarpinski]

I'm warming to the idea that our basic arithmetic promotions need some reconsideration.

[johnmyleswhite]

I'm warming to the idea that our basic arithmetic promotions need some reconsideration.

😈

[mbauman]

Also, I suppose that any time you're calling promote with a variable-length iterable, you're going to have some sort of type instability by definition (unless of course, it's a homogenous array, in which case the status quo is correct). So the title is a little ridiculous.

But I'd be in support of simplifying things here. It's telling that it's difficult to concisely describe the promotion behavior for integers (or keep in my head, for that matter!).

[mbauman]

Given that this is not going to be the direction to take here, I'll close this in favor of a more relevant issue. Perhaps re-open #1078?