[Regression] RuntimeGeneratedFunctions tagged under ModelingToolkit instead of precompiled user module

[dpad]

I am writing a package that builds some ODEFunctions at precompile time dependent on some ModelingToolkit models.
Supporting this is what ultimately drove my previous PR #477.

It seems that since ModelingToolkit's v4.0 (I upgraded from v3.20.1), the above functionality is once again broken.

Using the first test from https://github.com/SciML/ModelingToolkit.jl/blob/master/test/function_registration.jl as an example, the regression I'm seeing is as follows:

julia> # NOTE: this is an illustrative example, the keys shown may not match exactly.
julia> # I have a package MyModule (as per the test case above).
julia> using MyModule  
[precompiling MyModule...]

julia> # As per the test, I expect an answer of [30.0]
julia> MyModule.fun([0.5], [5.0], 0.)  
ERROR: KeyError: key (0xd6493547, 0xaf3999e1, 0x66674bbf, 0xb716e034, 0xa6bb5c0f) not found
Stacktrace:
 [1] getindex(::Dict{Any,Any}, ::NTuple{5,UInt32}) at .\dict.jl:467
 [2] (::RuntimeGeneratedFunctions.var"#5#6"{DataType,NTuple{5,UInt32}})() at C:\Users\DPad\.julia\packages\RuntimeGeneratedFunctions\uvbuv\src\RuntimeGeneratedFunctions.jl:133
 [3] lock(::RuntimeGeneratedFunctions.var"#5#6"{DataType,NTuple{5,UInt32}}, ::Base.Threads.SpinLock) at .\lock.jl:161
 [4] _lookup_body(::Type{T} where T, ::NTuple{5,UInt32}) at C:\Users\DPad\.julia\packages\RuntimeGeneratedFunctions\uvbuv\src\RuntimeGeneratedFunctions.jl:131
 [5] generated_callfunc_body(::NTuple{4,Symbol}, ::Type{T} where T, ::NTuple{5,UInt32}, ::NTuple{4,DataType}) at C:\Users\DPad\.julia\packages\RuntimeGeneratedFunctions\uvbuv\src\RuntimeGeneratedFunctions.jl:76
 [6] #s12#1 at C:\Users\DPad\.julia\packages\RuntimeGeneratedFunctions\uvbuv\src\RuntimeGeneratedFunctions.jl:163 [inlined]     
 [7] #s12#1(::Any, ::Any, ::Any, ::Any, ::Any) at .\none:0
 [8] (::Core.GeneratedFunctionStub)(::Any, ::Vararg{Any,N} where N) at .\boot.jl:527
 [9] RuntimeGeneratedFunction at C:\Users\DPad\.julia\packages\RuntimeGeneratedFunctions\uvbuv\src\RuntimeGeneratedFunctions.jl:68 [inlined]
 [10] f at C:\Users\DPad\.julia\packages\ModelingToolkit\HTjKG\src\systems\diffeqs\abstractodesystem.jl:133 [inlined]
 [11] ODEFunction at C:\Users\DPad\.julia\packages\DiffEqBase\b38LC\src\diffeqfunction.jl:248 [inlined]
...

julia> # It seems my function is tagged as being in ModelingToolkit...
julia> typeof(MyModule.fun.f.f_oop).parameters[2:3] 
svec(ModelingToolkit.var"#_RGF_ModTag", (0xd6493547, 0xaf3999e1, 0x66674bbf, 0xb716e034, 0xa6bb5c0f))

julia>   # The ModelingToolkit's function cache is empty.
julia> using ModelingToolkit
julia> using RuntimeGeneratedFunctions
julia> getfield(ModelingToolkit, RuntimeGeneratedFunctions._cachename)
Dict{Any,Any}()

julia> # If I define the module at runtime, it works fine.
julia> module MyModule   
               ...
               fun = ODEFunction(sys)
       end
julia> MyModule.fun([0.5], [5.0], 0.)
1-element Array{Float64,1}:
 30.0
julia> getfield(ModelingToolkit, RuntimeGeneratedFunctions._cachename)
Dict{Any,Any} with 2 entries:
  (0x7091df0e, 0xaddd82b1, 0xaa32db83, 0x… => WeakRef(quote…
  (0x61321fab, 0x744b53c8, 0xe99e5a49, 0x… => WeakRef(quote…

Note that the regression does not seem to apply when a module is defined at runtime, as I assume is the case for the actual test cases above.

I would expect my function MyModule.fun to actually get tagged under MyModule instead of ModelingToolkit (and that I also should initialise my own cache using RuntimeGeneratedFunctions.init(MyModule) inside my module definition, at least if I understood correctly).