ChainRulesCore dependency causes 4x load time regression

[KristofferC]

The quite recent addition of ChainRulesCore (#238) causes a 4x load time regression for this package. This is quite unfortunate since this package is a very common dependency. This has the effect that e.g. complaints are made upstream of slow loading packages (for example JuliaDiff/ForwardDiff.jl#518).

Current master:

julia> @time using SpecialFunctions
  0.256169 seconds (627.68 k allocations: 37.144 MiB, 4.97% gc time, 2.64% compilation time)

Without ChainRulesCore:

julia> @time using SpecialFunctions
  0.053946 seconds (51.40 k allocations: 4.460 MiB, 7.50% compilation time)

There is a bit of irony here in that ForwardDiff (or rather DiffRules) already has derivatives defined for SpecialFunctions (and its tests pass perfectly fine without the ChainRulesCore definitions). Is there any way you could make the people that don't need the (big) ChainRulesCore overhead able to avoid it?

In addition, ChainRulesCore does a bunch of quite exotic stuff (https://github.com/JuliaDiff/ChainRulesCore.jl/blob/63de708a85611f4a65604bec582c71b53d6d5d43/src/ruleset_loading.jl#L84-L126), for example adding hooks that run at every package load (which call into reflection of all things) and it seems unwanted to have such things in such a core dependency as SpecialFunctions.jl is.

[stevengj]

@ChrisRackauckas, this is a bit unfortunate. Why can't ChainRulesCore be a more lightweight dependency? Don't we just need a package defining a few functions that we can overload?

[ChrisRackauckas]

@oxinabox

[oxinabox]

We are working on it.
Will be fixed in v0.8, might be fixed before then.

See JuliaDiff/ChainRulesCore.jl#340

[oschulz]

This has the effect that e.g. complaints are made upstream of slow loading packages (for example JuliaDiff/ForwardDiff.jl#518).

It wasn't a complaint, really, I was just wondering if ForwardDiff might not profit from a lightweight abstract interface. :-) But I had indeed not realized that the load time of ForwardDiff had regressed recently.