Standardize access to type parameters

[Tokazama]

foo(@nospecialize(T::Type{<:Foo})) = T.parameters[1] is helpful when Foo is complicated and could potentially generate excessive methods otherwise (e.g., foo(::Type{<:Foo{T1,T2,...}})). Keno recently pointed out this isn't safe. I propose we standardize access to a types parameters with something like this:

getparam(x::DataType, i::Int) = x.parameters[i]
getparam(x::UnionAll, i::Int) = getparam(x.body, i)
getparam(x::TypeVar, i::Int) = getparam(x.ub, i)
getparam(x::Union, i::Int) = _combine_params(getparam(x.a, i), getparam(x.b, i))
_combine_params(a::Type, b::Type) = typejoin(a, b)
_combine_params(@nospecialize(a), @nospecialize(b)) = a === b ? a : Any

[vtjnash]

All of those would be wrong, since the ith parameter is generally meaningless (as noted by Keno already in the issue you linked)

[Tokazama]

So we can't even ensure that Dict{Int,Float64}.parameters[1] returns Int?

[Tokazama]

Sorry to keep pestering, but I want to make sure I don't keep making the same mistake. I've seen T.parameters[i] throughout julia and would like to at least know when that is permissible (and potentially add documentation so others don't make similar mistakes).

I was under the impression that given if T refers to Foo{T1,T2} then T.parameters[1] <: T1 and T.parameters[2] <: T2. I see T.parameters[i] most often when T<:Tuple so perhaps it is something that works when all of the parameters refer to concrete field types? On the other hand I've seen parameters that don't refer to a field. For example, NamedTuple's first parameters is used to get its field names here. Is that one okay because it's manually designed in Base?

[vtjnash]

People sometimes use it when they absolutely should not (e.g. often in @generated functions), but it will do the expected thing in certain limited cases. If you look around base, you will notice most of the uses are involved with reflection in some way or, equivalently, in the compiler.

[Tokazama]

I used the code from typejoin as guidance for what I originally posted but it appears that's still not valid. I assumed that at least for DataType` this would be valid though. Is there any other way to programmatically extract that information?