Broaden array constructor to allow any fill value

[rayegun]

It seems odd to me that Array{Union{Missing, T}}(missing, n...) and Array{Union{Nothing, T}}(nothing, n...) exist but the more general case Array{T}(x::T, n...) does not.

There are some potential issues I'm sure. For one it makes Array{T}(undef, n...) seem even more of an odd one out, rather than just the only strange one out of three options (the other two being nothing and missing). But I'm inclined to think this reduces the inconsistency since there are only two options, x::UndefInitializer and x::T after the change.

[fredrikekre]

See #24595 and a bunch of linked issues/PRs in there.

[rayegun]

I believe this syntax was actually proposed by you (4 years ago) deep in that issue.

This seems like a relatively limited subset of the proposal(s) there, while still allowing for all the extensions by restricting the type of x.

[fredrikekre]

I believe this syntax was actually proposed by you (4 years ago) deep in that issue.

I am sure, but that's not really a convincing argument. In particular, this PR makes it impossible to ever support something like Vector{T}(it, 4) (iterate by default), Vector{T}(Rep(x), 4) (repeat x at every location), #33840, etc.

[JeffBezanson]

Another issue is that the 0-d case is ambiguous, e.g. Array{T}(other_array) already works to construct one array from another.

[rayegun]

I am sure, but that's not really a convincing argument.

Oh, no I was just noting that.

In particular, this PR makes it impossible to ever support something like Vector{T}(it, 4) (iterate by default), Vector{T}(Rep(x), 4) (repeat x at every location), #33840, etc.

Isn't the type restriction sufficient? Rep(x) and it shouldn't have the type T right?

Another issue is that the 0-d case is ambiguous, e.g. Array{T}(other_array) already works to construct one array from another.

I think this is the same right? The other array isn't type T.

[JeffBezanson]

I don't think it's a good idea for this to depend on the element type, but we could disambiguate it by requiring a dimension argument, e.g. Array(val, ()) for 0-d.

[rayegun]

I don't think it's a good idea for this to depend on the element type

So it should (hypothetically) be broadened further to Array{T}(x, n...) instead of Array{T}(x::T, n...)? I don't get a method ambiguity for the latter and Array{T}(other_array), at least in my limited testing.

[JeffBezanson]

Ideally everything should work the same if the element type is widened; e.g. Array{Any}(x, ...) should do the same thing for all x.