Generated functions no longer re-tracable with Symbolics

[baggepinnen]

A recent change (maybe PR #536 @YingboMa ?) causes functions produced by build_function to fail when called with symbolic variables:

julia> @variables x
1-element Vector{Num}:
 x

julia> f = build_function([sin(x)], [x], expression=Val{false});

julia> f[1]([1])
1-element Vector{Float64}:
 0.8414709848078965

julia> f[1]([x])
ERROR: StackOverflowError:
Stacktrace:
 [1] sin(x::Num) (repeats 79984 times)
   @ NaNMath ~/.julia/packages/NaNMath/ceWIc/src/NaNMath.jl:11

This is a problem when generated functions are used to build, e.g., cost functions for optimization problems, after which symbolic tracing is used to find sparsity patterns or generate jacobian functions etc.

[shashi]

        ($f)(x::Real) = ($f)(float(x))

They have this, and assume float will always return a float which it should, but @YingboMa added this:

https://github.com/JuliaSymbolics/Symbolics.jl/commit/c890de9bb#diff-b8c0c7fd71b608d8ede7feac325fbfa572e06bd50bc32f0b835f3dbe61bb7fa5R32

which will make it go into an infinite loop.

Makes me think #536 is not really working...

[shashi]

@baggepinnen just curious why is the sparsity detection implemented like this? Is it beneficial to do this over just calling jacobian_sparsity before you generate the code? (I need legit examples where we trace generated code symbolically again.)

[baggepinnen]

I my use case, I have dynamics functions generated by MTK, those are then used to compute a Lagrangian for an optimal control problem, the Hessian of which I need for the optimization. It's this Hessian, and other similar functions, I am "re tracing". I could keep everything symbolic without generating the intermediate dynamics function, but then I'd need separate logic to handle MTK models and manually written models, so I opted for always going through a function.

[shashi]

Awesome! Thanks that's a great example of code reuse.