Use precompilation

[jkosata]

Just-in-time compilation times are getting pretty crazy (minutes). Most of the time, similar functions are used - get_harmonic_equations and get_steady_states. Precompilation could make this much more convenient.

Starting resource: https://julialang.org/blog/2021/01/precompile_tutorial/

[jkosata]

Closer inspection shows things are quite complex.

@time_imports using HarmonicBalance shows that out of the ~60s, only 8s is HarmonicBalance, the rest are dependencies. A flamegraph shows that little time is spent on time inference, ruling out an easy fix.

More resources: https://discourse.julialang.org/t/taking-ttfx-seriously-can-we-make-common-packages-faster-to-load-and-use/74949/49

[jkosata]

using SnoopCompile.@snoopr reveals there is a huge amount of invalidation coming from Arblib triggering the recompilation of some parts of OrdinaryDiffEq .

[oameye]

That most time was spent on compilation on dependencies rather than our own code I already suspected. We use the most notorious packages for precompilation time. Luckily, on Juliacon 2022 I picked up that a lot of the Julia masterminds are working on the problem. https://www.youtube.com/watch?v=N4h46_TCmGc

[oameye]

Looked into SnoopCompile.jl and the PR. Very cool stuff. Handy how it all just can be done by one macro. Nice that you also only added the less overhead SnoopPrecompile.jl version. Did you do a full analyses of the type instability with JET.jl as the documentation of SnoopCompile.jl suggests?