NLsolve significantly slower on 0.5

[KristofferC]

Making this to track regression in master and to make people aware

using NLsolve

    function f!(x::Vector, fvec::Vector)
        fvec[1] = 1 - x[1]
        fvec[2] = 10(x[2]-x[1]^2)
    end
    function g!(x::Vector, fjac::Matrix)
        fjac[1,1] = -1
        fjac[1,2] = 0
        fjac[2,1] = -20x[1]
        fjac[2,2] = 10
    end


df! = DifferentiableMultivariateFunction(f!, g!)
x0 = [-1.2; 1.]

@time for i in 1:10^4 nlsolve(df!, x0, method = :newton) end

On 0.4.3: 0.292429 seconds (5.31 M allocations: 123.749 MB, 10.00% gc time)

On master: 1.336571 seconds (5.53 M allocations: 101.471 MB, 2.06% gc time)

I believe this is because we are mutating a variable inside a closure which is slow in 0.5.. see JuliaLang/julia#15276

[KristofferC]

This is sligtly better now on master.

julia> @time for i in 1:10^4 nlsolve(df!, x0, method = :newton) end
  0.856573 seconds (5.28 M allocations: 96.436 MB, 2.41% gc time)

Still bad though.

[KristofferC]

Commenting out these two lines which are the only problematic ones (inferred as Any) on master gives:

julia> @time for i in 1:10^4 nlsolve(df!, x0, method = :newton) end
  0.264198 seconds (4.62 M allocations: 97.046 MB, 4.19% gc time)

Without them commented out:

julia> @time for i in 1:10^4 nlsolve(df!, x0, method = :newton) end
  0.397250 seconds (4.67 M allocations: 98.114 MB, 2.91% gc time)

[KristofferC]

This is fixed now.