RuntimeGeneratedFunctions.jl

RuntimeGeneratedFunctions are functions generated at runtime without world-age issues and with the full performance of a standard Julia anonymous function. This builds functions in a way that avoids eval.

Note that RuntimeGeneratedFunction does not handle closures. Please use the GeneralizedGenerated.jl package for more fixable staged programming. While GeneralizedGenerated.jl is more powerful, RuntimeGeneratedFunctions.jl handles large expressions better.

Tutorials and Documentation

For information on using the package, see the stable documentation. Use the in-development documentation for the version of the documentation, which contains the unreleased features.

Simple Example

Here's an example showing how to construct and immediately call a runtime generated function:

using RuntimeGeneratedFunctions
RuntimeGeneratedFunctions.init(@__MODULE__)

function no_worldage()
    ex = :(function f(_du, _u, _p, _t)
        @inbounds _du[1] = _u[1]
        @inbounds _du[2] = _u[2]
        nothing
    end)
    f1 = @RuntimeGeneratedFunction(ex)
    du = rand(2)
    u = rand(2)
    p = nothing
    t = nothing
    f1(du, u, p, t)
end
no_worldage()

Changing how global symbols are looked up

If you want to use helper functions or global variables from a different module within your function expression you'll need to pass a context_module to the @RuntimeGeneratedFunction constructor. For example

RuntimeGeneratedFunctions.init(@__MODULE__)

module A
using RuntimeGeneratedFunctions
RuntimeGeneratedFunctions.init(A)
helper_function(x) = x + 1
end

function g()
    expression = :(f(x) = helper_function(x))
    # context module is `A` so that `helper_function` can be found.
    f = @RuntimeGeneratedFunction(A, expression)
    @show f(1)
end

Precompilation and setting the function expression cache

For technical reasons RuntimeGeneratedFunctions needs to cache the function expression in a global variable within some module. This is normally transparent to the user, but if the RuntimeGeneratedFunction is evaluated during module precompilation, the cache module must be explicitly set to the module currently being precompiled. This is relevant for helper functions in some module which construct a RuntimeGeneratedFunction on behalf of the user. For example, in the following code, any third party user of HelperModule.construct_rgf() user needs to pass their own module as the cache_module if they want the returned function to work after precompilation:

RuntimeGeneratedFunctions.init(@__MODULE__)

# Imagine HelperModule is in a separate package and will be precompiled
# separately.
module HelperModule
using RuntimeGeneratedFunctions
RuntimeGeneratedFunctions.init(HelperModule)

function construct_rgf(cache_module, context_module, ex)
    ex = :((x) -> $ex^2 + x)
    RuntimeGeneratedFunction(cache_module, context_module, ex)
end
end

function g()
    ex = :(x + 1)
    # Here cache_module is set to the module currently being compiled so that
    # the returned RGF works with Julia's module precompilation system.
    HelperModule.construct_rgf(@__MODULE__, @__MODULE__, ex)
end

f = g()
@show f(1)

Retrieving Expressions

From a constructed RuntimeGeneratedFunction, you can retrieve the expressions using the RuntimeGeneratedFunctions.get_expression command. For example:

ex = :((x) -> x^2)
rgf = @RuntimeGeneratedFunction(ex)
julia> RuntimeGeneratedFunctions.get_expression(rgf)
#=
quote
    #= c:\Users\accou\OneDrive\Computer\Desktop\test.jl:39 =#
    x ^ 2
end
=#

This can be used to get the expression even if drop_expr has been performed.

Example: Retrieving Expressions from ModelingToolkit.jl

ModelingToolkit.jl uses RuntimeGeneratedFunctions.jl for the construction of its functions to avoid issues of world-age. Take for example its tutorial:

using ModelingToolkit, RuntimeGeneratedFunctions
using ModelingToolkit: t_nounits as t, D_nounits as D

@mtkmodel FOL begin
    @parameters begin
        τ # parameters
    end
    @variables begin
        x(t) # dependent variables
    end
    @equations begin
        D(x) ~ (1 - x) / τ
    end
end

using DifferentialEquations: solve
@mtkbuild fol = FOL()
prob = ODEProblem(fol, [fol.x => 0.0], (0.0, 10.0), [fol.τ => 3.0])

If we check the function:

julia> prob.f
(::ODEFunction{true, SciMLBase.AutoSpecialize, ModelingToolkit.var"#f#697"{RuntimeGeneratedFunction{(:ˍ₋arg1, :ˍ₋arg2, :t), ModelingToolkit.var"#_RGF_ModTag", ModelingToolkit.var"#_RGF_ModTag", (0x2cce5cf2, 0xd20b0d73, 0xd14ed8a6, 0xa4d56c4f, 0x72958ea1), Nothing}, RuntimeGeneratedFunction{(:ˍ₋out, :ˍ₋arg1, :ˍ₋arg2, :t), ModelingToolkit.var"#_RGF_ModTag", ModelingToolkit.var"#_RGF_ModTag", (0x7f3c227e, 0x8f116bb1, 0xb3528ad5, 0x9c57c605, 0x60f580c3), Nothing}}, UniformScaling{Bool}, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, ModelingToolkit.var"#852#generated_observed#706"{Bool, ODESystem, Dict{Any, Any}, Vector{Any}}, Nothing, ODESystem, Nothing, Nothing}) (generic function with 1 method)

It's a RuntimeGeneratedFunction. We can find the code for this system using the retrieval command on the function we want. For example, for the in-place function:

julia> RuntimeGeneratedFunctions.get_expression(prob.f.f.f_iip)

:((ˍ₋out, ˍ₋arg1, ˍ₋arg2, t)->begin
          #= C:\Users\accou\.julia\packages\SymbolicUtils\c0xQb\src\code.jl:373 =#
          #= C:\Users\accou\.julia\packages\SymbolicUtils\c0xQb\src\code.jl:374 =#
          #= C:\Users\accou\.julia\packages\SymbolicUtils\c0xQb\src\code.jl:375 =#
          begin
              begin
                  begin
                      #= C:\Users\accou\.julia\packages\Symbolics\HIg7O\src\build_function.jl:546 =#
                      #= C:\Users\accou\.julia\packages\SymbolicUtils\c0xQb\src\code.jl:422 =# @inbounds begin
                              #= C:\Users\accou\.julia\packages\SymbolicUtils\c0xQb\src\code.jl:418 =#
                              ˍ₋out[1] = (/)((+)(1, (*)(-1, ˍ₋arg1[1])), ˍ₋arg2[1])
                              #= C:\Users\accou\.julia\packages\SymbolicUtils\c0xQb\src\code.jl:420 =#
                              nothing
                          end
                  end
              end
          end
      end)

or the out-of-place function:

julia> RuntimeGeneratedFunctions.get_expression(prob.f.f.f_oop)
:((ˍ₋arg1, ˍ₋arg2, t)->begin
          #= C:\Users\accou\.julia\packages\SymbolicUtils\c0xQb\src\code.jl:373 =#
          #= C:\Users\accou\.julia\packages\SymbolicUtils\c0xQb\src\code.jl:374 =#
          #= C:\Users\accou\.julia\packages\SymbolicUtils\c0xQb\src\code.jl:375 =#
          begin
              begin
                  begin
                      #= C:\Users\accou\.julia\packages\SymbolicUtils\c0xQb\src\code.jl:468 =#
                      (SymbolicUtils.Code.create_array)(typeof(ˍ₋arg1), nothing, Val{1}(), Val{(1,)}(), (/)((+)(1, (*)(-1, ˍ₋arg1[1])), ˍ₋arg2[1]))
                  end
              end
          end
      end)