Type-stability issues

[odow]

I have some type stability issues that I think are the cause of some performance issues I'm having.

Is it reasonable to expect that calls such as value and dual will return Float64s?

model = JuMP.Model(with_optimizer(Gurobi.Optimizer, OutputFlag=0))
@variable(model, x)
JuMP.optimize!(model)

@code_warntype JuMP.value(x)
Body::Any
705 1 ─ %1 = MathOptInterface.get::Core.Compiler.Const(MathOptInterface.get, false)
    │   %2 = (Base.getfield)(v, :model)::Model                   │╻  owner_model
    │   %3 = %new(MathOptInterface.VariablePrimal, 1)::MathOptInterface.VariablePrimal
    │   %4 = invoke %1(%2::Model, %3::MathOptInterface.VariablePrimal, _2::VariableRef)::Any
    └──      return %4                                           │

[blegat]

Is it due to the fact that the moi_backend type is not concrete (e.g. what is addressed by #1348) ?
JuMP only support Float64 arithmetic at the moment so adding a type assert seems reasonable.

[odow]

Yeah it's probably due to the non-concreteness. Here is the function that fails to infer properly when is_set_by_optimize is true:

https://github.com/JuliaOpt/JuMP.jl/blob/1d7c5030ce50300f564b1cde1e1ecfabb82c5494/src/JuMP.jl#L573-L581