Type instabilities when applying a parametric type constructor to a collection using map

[ffevotte]

Hi,

I'm filing this issue following a discourse thread, in case it actually is a bug. Please feel free to close it if this behavior is actually intended.

Consider the following (minimal?) example, demonstrating a type-instable application of map to the constructor of a parametric type:

julia> struct A{N}
           A(::NTuple{N,Int}) where {N} = new{N}()
       end

julia> t = (1,2)
(1, 2)

julia> T = typeof(t)
Tuple{Int64, Int64}

julia> using Test

julia> @inferred A(t) # the constructor is type-stable and correctly inferred
A{2}()

julia> @inferred map(A, [t]) # ... but map is not type-stable
ERROR: return type Vector{A{2}} does not match inferred return type Union{Vector{Any}, Vector{A{2}}}

julia> map(A, T[]) # indeed, on an empty vector it returns Any[]
Any[]

This behavior can be observed with Julia v1.9 as well as v1.10(beta).

As uncovered by @Seelengrab in the aforementioned discourse thread, an easy workaround consists in applying map to a standard function instead of a constructor, e.g.:

julia> @inferred map(x->A(x), T[])
A{2}[]

I'm paraphrasing / summarizing below what they uncovered while trying to understand this; the discourse thread contains much more details.

It seems that this issue arises in Base.@default_eltype: because the applied function (A) is a type, the element type of the resulting array is determined to be A itself, even though it is a union-all type and it would be possible to infer the more precise element type (A{2} in this case). A possible fix would therefore consist in handling union-all types like regular functions, rather than special-casing them like the "regular" types. But doing this might have unintended consequences, so I'm not sure what the best way forward would be.

[Seelengrab]

The additional thing with this example in particular is that you can't construct an A through A{2}; it needs to go through A, so there's no way to map A directly and have it be type stable, in the current implementation.

[jishnub]

Duplicate of #46331 I think, although in that case the parameter was a type rather than a value