gf: support more dispatch on abstract types

[vtjnash]

This removes the restriction on defining dispatch over user-defined abstract types. This fixes #14919 for 99% of cases.

abstract type A14919; end
struct B14919 <: A14919; end
struct C14919 <: A14919; end
struct D14919 <: Function; end
(::A14919)() = "It's a brand new world"
(::Union{C14919,D14919})() = "Boo."
@test B14919()() == "It's a brand new world"
@test C14919()() == D14919()() == "Boo."

The "cannot add methods to an abstract type" error is now only applicable to a couple types (Any, Function, and functions), and instead now gives a "not implemented yet" message (that's perhaps not entirely right, since in several instances of this, I think they would be ambiguous with calling a Builtin function—and we don't need to ever enable that particular foot-gun).

This PR actually hilariously similar to my branch #14919 (comment) from 3 years ago, but adds a few more error checks so that performance is unaffected.

[JeffBezanson]

I think not being able to define methods for Any is definitely a feature :)

[JeffreySarnoff]

ya' know the Community will be dancing while this becomes second nature 🌞

[vtjnash]

@nanosoldier runbenchmarks(ALL, vs=":master")

[JeffBezanson]

This benchmark from #21760 goes from .034 sec to .042 sec on this branch:

struct f{t}
end
(::f{t})(x) where {t} = t
@time for i=1:1000; f{i}()(1); end

[nanosoldier]

Your benchmark job has completed - possible performance regressions were detected. A full report can be found here. cc @ararslan

[vtjnash]

@JeffBezanson I ran that code locally, and saw no performance difference. Probably just noise or a different random seed.

[JeffBezanson]

This will not catch the case of defining struct Foo <: Function, which is probably fine, just want to check that that's the intent.

[vtjnash]

Correct, that was my intention (it's even part of my examples at the top). Eventually the goal is probably just to merge all of the tables, but for now, I just merge the subset that I expect will be small.

[JeffBezanson]

Delete?

[vchuravy]

Running into a segmentation fault that bisect blames on this, JuliaGPU/CUDAnative.jl#407

[MasonProtter]

The "cannot add methods to an abstract type" error is now only applicable to a couple types (Any, Function, and functions), and instead now gives a "not implemented yet" message (that's perhaps not entirely right, since in several instances of this, I think they would be ambiguous with calling a Builtin function—and we don't need to ever enable that particular foot-gun).

What if we were to allow this, but only in cases where a method error would be thrown? Ie. make (::Function)(x::MyType) = ... the lowest possible priority such that it always loses priority to standard definitions?

[caseykneale]

Awesome I've been waiting for this: https://github.com/caseykneale/ChemometricsTools.jl/blob/master/src/RegressionModels.jl

How can developers maintain backward compatibility with previous julia versions with the implementation of this(once its streamlined)? Maybe the correct solution is to tell users to use the latest version of julia.

[StefanKarpinski]

This is a language feature that cannot be implemented in a package, so the only option of one needs this is to use a newer Julia version.