change widen to strict 2x widening

[sbromberger]

Per #14407 (comment) et seq., change the behavior of widen() as follows:

for all types with bit size n, widen() will produce a similarly-signed type with bit size 2n regardless of underlying word size / architecture. This changes the following:

Type	Current widen()	This PR
UInt8	UInt	UInt16
UInt16	UInt	UInt32
Int8	Int	Int16
Int16	Int	Int32

In addition, added explicit type tests.

[simonbyrne]

Given the relatively small amount of code churn in the tests, I don't think this should be too breaking.

[StefanKarpinski]

LGTM. It's a good sign that the only things that need to change seem to be explicit tests for the return type. Also, you gotta love how easy this kind of change is in Julia 😁

[simonbyrne]

Maybe explain this in the docs?

[JeffBezanson]

There's an item about widen in NEWS.md that should be changed to reflect this.

[sbromberger]

I didn't know how to update NEWS without an issue/PR number. Should I include the NEWS.md changes as another commit here now that we know it's #28045 ?

[simonbyrne]

Yes, it's fine to add as another commit. We can squash it all on merge.

[sbromberger]

OK, what doc changes do people suggest? I think the current docstring is fairly clear. Do we really need to change it? (ref https://docs.julialang.org/en/latest/base/base/#Base.widen)

[StefanKarpinski]

It is fairly clear and this is still in line with it. Might be good to explicitly state that for fixed-size integer types less than 128-bits widen doubles the bit size since that's now the rule.

[simonbyrne]

Something like:

For fixed-width integer types, this will a type

[simonbyrne]

Sorry, I pressed the wrong button.

[sbromberger]

OK, I made changes to NEWS.md and to the docstring in base/operators.jl and fixed up the commits.

[sbromberger]

@StefanKarpinski

It's a good sign that the only things that need to change seem to be explicit tests for the return type.

I was actually puzzled as to why the return types weren't being tested in the first place (except for BigInt).

[JeffBezanson]

Trailing whitespace on this line and the next.

[sbromberger]

That's a weird test in numbers.jl. Why is that there?

[sbromberger]

It doesn’t look like the appveyor failure is related. Could someone please confirm?

[martinholters]

Yep, looks like a problem with httpbin.org which happens quite a bit lately.

[StefanKarpinski]

The Win32 failure is indeed the httpbin flakiness and is safe to ignore. So this has passed on all platforms and is ready to merge as far as I can tell. I'll merge later today unless there are objections.

[StefanKarpinski]

Eh, why wait. Just went for it.

[sbromberger]

Thank you all for the guidance! This was an easy PR.