AutoSpecialize for NonlinearProblem

[oscardssmith]

since NonlinearProblem doesn't have a t field, most of the time, we can just use Void for all of the functions of inplace ones without a performance penalty. This still needs to be benchmarked to make sure it doesn't have a performance impact.

[oscardssmith]

interestingly, even with this, NonlinearSolve is still 99% compilation time...

[ChrisRackauckas]

@avik-pal test it a bit?

[avik-pal]

interestingly, even with this, NonlinearSolve is still 99% compilation time...

Do you mean the current precompile workload? I think it manually sets the specialization so it has no effect.

Also we would need the corresponding DiffEqBase update for this to work I think

[oscardssmith]

I changed the workload to make p=nothing and AutoSpecialize (since it is going to be hard to precompile meaningfully for non-null parameters). What DiffEqBase changes would be needed?

[avik-pal]

https://github.com/SciML/DiffEqBase.jl/pull/1101/files

[oscardssmith]

That very specifically shouldn't be needed. The key observation of this PR is that you don't actually need to FunctionWrap in place functions since you can just do a dynamic dispatch instead.