A very weird method dispatch bug

[lindahua]

This bug is due to a recent Base update, and it causes the test failure of Distributions.jl

Related code: (https://github.com/JuliaStats/Distributions.jl/blob/master/src/univariate/categorical.jl#L225)

function fit_mle{T<:Integer}(::Type{Categorical}, k::Integer, x::Array{T}) 
    Categorical(normalize!(add_categorical_counts!(zeros(k), x), 1), NoArgCheck())
end

fit_mle{T<:Integer}(::Type{Categorical}, data::(Int, Array{T})) = 
      fit_mle(Categorical, data...)

Then in Julia:

julia> using Distributions

julia> x = rand(1:3, 10);

julia> fit_mle(Categorical, 3, x)
Categorical( K=3 prob=[0.7000000000000001, 0.30000000000000004, 0.0] )

julia> fit_mle(Categorical, (3, x))
Binomial( size=3 prob=0.43333333333333335 )

However, it I change the last fit_mle method to

fit_mle{T<:Integer}(::Type{Categorical}, data::(Int, Array{T})) = 
      fit_mle(Categorical, data[1], data[2])

Then it works correctly.

It is very weird why fit_mle(Categorical, data...) calls a wrong method, while fit_mle(Categorical, data[1], data[2]) is correct.

Related to JuliaStats/Distributions.jl#198

cc: @JeffBezanson @StefanKarpinski

[lindahua]

Using git bisect, I figured the commit that causes this bug: 9718fcf

[JeffBezanson]

Thanks for doing the bisect.