`NaNMath` (and `SpecialFunctions`) as extensions?

[longemen3000]

`NaNMath` (and `SpecialFunctions`) as extensions?

[mcabbott]

seems really weird

Can you explain what you think is weird here?

[longemen3000]

nevermind that line, i got trip up with the !, but if i understand correctly, NaNMath is loaded to overload NaNMath.pow right?

ForwardDiff.jl/src/dual.jl

Line 536 in 2263749

for f in (:(Base.:^), :(NaNMath.pow))

[mcabbott]

NaNMath needs to be loaded before overloading any of its functions. Mostly that's done by the loop above, which runs over these definitions:

https://github.com/JuliaDiff/DiffRules.jl/blob/2001650a4a009d2136f89496f2d22fe7fb04dfbd/src/rules.jl#L251-L280

[longemen3000]

is it possible to load those in a extension instead? i suppose that DiffRules does not load NaNMath and shouldn't load any extensions

[mcabbott]

I think this could be done. But what would be gained? The time saved would be small compared to StaticArrays:

julia> @time_imports using ForwardDiff
      5.5 ms  IrrationalConstants
      0.7 ms  DiffRules
      0.7 ms  StaticArraysCore
      0.7 ms  DiffResults
               ┌ 1.4 ms OpenLibm_jll.__init__() 
     18.3 ms  OpenLibm_jll 90.51% compilation time
      0.3 ms  NaNMath
               ┌ 0.0 ms DocStringExtensions.__init__() 
      0.8 ms  DocStringExtensions
      0.3 ms  LogExpFunctions
      0.2 ms  JLLWrappers
               ┌ 2.3 ms OpenSpecFun_jll.__init__() 86.08% compilation time
      2.7 ms  OpenSpecFun_jll 72.07% compilation time
      5.6 ms  SpecialFunctions
      5.5 ms  MacroTools
      0.3 ms  CommonSubexpressions
     19.6 ms  ForwardDiff

julia> @time_imports using StaticArrays
    171.0 ms  StaticArrays
      0.5 ms  ForwardDiff → ForwardDiffStaticArraysExt