type unstability with container generation

[etiennedeg]

julia> model = Model(Gurobi.Optimizer)

julia> @variable(model, a[x in 1:10; x < 5], Bin)
JuMP.Containers.SparseAxisArray{VariableRef, 1, Tuple{Int64}} with 4 entries:

julia> @variable(model, b[x in 1:10; x > 15], Bin)
JuMP.Containers.SparseAxisArray{Any, 1, Tuple{Any}} with 0 entries

Here, I would expect the second example to return a JuMP.Containers.SparseAxisArray{VariableRef, 1, Tuple{Int64}}
I don't know if the current behaviour is expected, but it can cause type unstability if we don’t check the emptiness of the resulting container beforehand.

[odow]

Thanks for opening an issue. Empty dictionaries seem to be well-typed, so maybe there is something we can do:

julia> Dict(x => x for x in 1:10 if x > 15)
Dict{Int64, Int64}()

Here is the backstory on Discourse: https://discourse.julialang.org/t/type-unstability-with-container-generation-in-jump/58433

[odow]

Here's the problem:

JuMP.jl/src/Containers/container.jl

Lines 117 to 123 in 41e4bf0

	if isempty(dict)
	# If `dict` is empty, it was not able to determine the type
	# of the key hence the type of `dict` is `Dict{Any, Any}`.
	# This is an issue since `SparseAxisArray` needs the key
	# type to be a tuple.
	dict = Dict{_eltype_or_any(indices),Any}()
	end

[blegat]

Dict(x => x for x in 1:10 if x > 15) is using some dark magic to work. See the _iterator_upper_bound function:
https://github.com/JuliaLang/julia/blob/8218480f059b7d2ba3388646497b76759248dd86/base/abstractdict.jl#L541
https://github.com/JuliaLang/julia/blob/8218480f059b7d2ba3388646497b76759248dd86/base/array.jl#L628-L656
We could use Base.@default_eltype as well.