Some more typeintersect tests

[martinholters]

Conversion of the list from #19998 (comment) to tests. For those which have been fixed in the meantime, I have use the @testintersect macro. There is one @test_skip at the end because that one loops infinitely (?) or sometimes throws a stack overflow.

A word about the fuzz testing that produced those:
I start with creating a random type that is an Int64, a Float64, a Val{T} with T being a random type again, or a Tuple{T1,T2, ...} of a random number of random types T1, T2, etc. For the random types inside a Val or Tuple, a constant integer is also possible. Note that these types may be large and deep, but don't contain Unions, Varargs, or UnionAlls, so I hoped these would still be easy ones... Anyway, the type thus obtained is x.
I then apply random widening by replacing random parts of the type x with TypeVars, potentially introducing the same TypeVar multiple times if the same type is to be replaced more than once. All introduced TypeVars are bound by UnionAlls at the outermost layer. This process is done twice to obtain a and b, such that x<:a and x<:b.
With these types, it is verified that

• typeintersect(a,b) == typeintersect(b,a)
• x <: typeintersect(a,b)
• typeintersect(a,b) <: a
• typeintersect(a,b) <: b

If any of these conditions is violated, the combination x, a, b is reduced while maintaining the same violation. The reduction tries steps such as removing tuple elements or unwrapping (i.e. Tuple{T} -> T or Val{T} -> T).

Once we get (most of) these right, I will add things like Unions, Varargs, inner UnionAlls, TypeVar bounds other than Bottom <: T <: Any, ...

[timholy]

I like the fuzzing, and I wonder if it's worth making it a repo. Cherry-picking the failures is great for coming up with a not-too-large test set, but of course there's always the risk that fixes to those specific problems will break other things that used to pass. Some kind of overall statistics that can be run episodically might be helpful.

[JeffBezanson]

Thanks!

[martinholters]

I like the fuzzing, and I wonder if it's worth making it a repo.

Up at https://github.com/martinholters/TypeFuzz.jl

[Keno]

Might be worth running the subtype predicate through AFL with appropriate mutators.