Construct Device is a Runtime call

[jd-lara]

^  %185 = < constprop > getproperty(::TimerOutput{…},::Symbol{…})::Bool
   get_component_type(device_model::DeviceModel)
   "$(get_component_type(device_model::DeviceModel)::Type{D} where D<:De…
   TimerOutputs.@timeit BUILD_PROBLEMS_TIMER "$(get_component_type(devic…
   %194 = gc_bytes()::Int64
   %195 = time_ns()::UInt64
   validate_available_devices(device_model::DeviceModel, sys)
   ArgumentConstructStage()
 • %207 = runtime < construct_device!(…,…,…,…,…)::Any >
v  %221 = < concrete eval > getproperty(::LogLevel{…},::Symbol{…})::…
[ Info: This is a runtime call. You cannot descend into it. 

Making an exploration with Ctuchulu.jl I got to this point when looking for performance issues.

[luke-kiernan]

I strongly suspect this is a consequence of how we're declaring construct_device! as a stub in IOM then implementing it in POM.

edit: might be useful to create a minimum working example, to see if that indeed is the case. The fact that there's a huge number of different definitions of construct_device! may play a role, too.

edit^2: nope, AI disagrees, and instead blames it on iterating over a heterogeneously-typed container.