Avoid type inference problem in chol which caused a 10x slowdown in stockcorr benchmark without breaking anything

[andreasnoack]

@JeffBezanson I could solve the problem by calling another constructor. For solve reason Triangular{eltype(C),typeof(C),uplo,false}(C) worked better than Triangular(C,uplo,false) although the latter method is defined by

function Triangular{T}(A::AbstractMatrix{T}, uplo::Symbol, isunit::Bool=false)
    chksquare(A)
    Triangular{T,typeof(A),uplo,isunit}(A)
end

Is there something wrong with the latter definition that breaks type inference?

[tkelman]

Should there be a test of chol on a non-BlasFloat type?

[andreasnoack]

Actually, there is a test of chol! for BigFloat, but the new chol bypassed chol!. I'm working towards a reorganisation of the linear algebra tests that has the goal of covering 100pct of the methods. To keep track of that, I want to organise the tests according to the name of the file in base.

[JeffBezanson]

Ah, the key seems to be manually inlining false into the Triangular{ } type application instead of passing it through a function.