Fix deprecation warning introduced by DiffRules v1.4

[odow]

DiffRules 1.4.0 introduced a new deprecation warning: JuliaDiff/DiffRules.jl#69

This caused the JuMP tests to fail since we run with --depwarn=error:
https://github.com/jump-dev/JuMP.jl/runs/4110519060?check_suite_focus=true#step:6:26

x-ref: jump-dev/JuMP.jl#2790

[devmotion]

I guess one could use the opportunity and already prepare this code for future releases of DiffRules. As stated in the deprecation warning it will change to filter_modules=nothing, ie. it will load all rules defined in DiffRules. We only had to introduce this keyword argument because many downstream packages, including ForwardDiff, did not actually check if a module M with a function f was defined when they tried to load the rules.

So an alternative approach here that would also add derivative definitions for functions in LogExpFunctions would be to:

• add LogExpFunctions >= 0.3 as explicit dependency (it's already an indirect dependency via DiffRules and SpecialFunctions)
• use DiffRules.diffrules(; filter_modules=nothing) and load all rules, including the ones for LogExpFunctions
• check if M and M.f are defined before trying to eval the rule definitions

[odow]

Is this a reasonable way to check if M and f are defined? Seems a bit clunky.

[odow]

Why are the tests non-deterministic? Failure on 1.0 isn't caused by this: https://github.com/JuliaDiff/ForwardDiff.jl/runs/4111622597?check_suite_focus=true#step:6:293

[odow]

Tests are passing now thanks to #554.

[devmotion]

LGTM, thanks! 🙂

[odow]

Any chance we could get this merged and a new release tagged?

[odow]

Thanks!