of_eltype does not work for arrays of eltype Any (?)

[lassepe]

julia> import MappedArrays

julia> a = Any[1,2,3]
3-element Vector{Any}:
 1
 2
 3

julia> b = MappedArrays.of_eltype(Int, a)
3-element mappedarray(x->MappedArrays.convert($(Expr(:static_parameter, 1)), x), y->MappedArrays.convert($(Expr(:static_parameter, 1)), y), ::Vector{Any}) with eltype Any:
 1
 2
 3

julia> eltype(b)
Any

[johnnychen94]

The reason behind this is the type of x->convert(Int, x) is not inferrable:

julia> Base.return_types(x->convert(Int, x), (Any,))
1-element Vector{Any}:
 Any

MappedArrays.jl/src/MappedArrays.jl

Line 116 in 9fbda81

of_eltype(::Type{T}, data::AbstractArray{S}) where {S,T} = mappedarray(x->convert(T,x), y->convert(S,y), data)

I don't know how to propagate the type information to Base.return_types; a "bad" workaround is:

of_eltype(::Type{T}, data::AbstractArray{S}) where {S, T<:Number} = mappedarray(T, y->convert(S, y), data)

It's bad because it's not very extensible to other cases.

[lassepe]

Wow, I was not aware that convert is not inferrable. I would have assumed that this should always be trivially inferrable because it has the return type as an input?

Also, @code_warntype reports:

julia> f = x -> convert(Int, x)
#1 (generic function with 1 method

julia> @code_warntype f(first(Any[1.0]))
Variables
  #self#::Core.Const(var"#1#2"())
  x::Float64

Body::Int64
1 ─ %1 = Main.convert(Main.Int, x)::Int64
└──      return %1

Edit: So there it seems inferrable. See answer below.

[cstjean]

@code_warntype evaluates the arguments of the function. It's assessing f(1.0)...