std() type instability?

[alyst]

Julia 0.5.2

julia> @code_warntype std(randn(5))
Variables:
  #self#::Base.#std
  A::Array{Float64,1}

Body:
  begin 
      return (Base.sqrt)($(Expr(:invoke, LambdaInfo for (::Base.#kw##var)(::Array{Any,1}, ::Base.#var, ::Array{Float64,1}), Base.#var, :($(Expr(:invoke, LambdaInfo for vector_any(::Any, ::Vararg{Any,N}), :(Base.vector_any), :(:corrected), true, :(:mean), :(Base.nothing)))), :(Base.var), :(A))))::Any
  end::Any

Julia 0.6rc-1

julia> @code_warntype std(randn(5))
Variables:
  #self#::Base.#std
  A::Array{Float64,1}

Body:
  begin 
      return (Base.sqrt)($(Expr(:invoke, MethodInstance for (::Base.#kw##var)(::Array{Any,1}, ::Base.#var, ::Array{Float64,1}), :($(QuoteNode(Base.#var))), :($(Expr(:invoke, MethodInstance for vector_any(::Any, ::Vararg{Any,N} where N), :(Base.vector_any), :(:corrected), true, :(:mean), :(Base.nothing)))), :(Base.var), :(A))))::Any
  end::Any

The 1st problem is type instability.

The 2nd problem is that for functions taking keyword arguments I cannot get the meaningful output since it only shows the wrapper call (but note that var() result is correctly inferred as Float64):

julia> @code_warntype var(randn(5))
Variables:
  #self#::Base.#var
  A::Array{Float64,1}

Body:
  begin 
      return $(Expr(:invoke, MethodInstance for #var#533(::Bool, ::Void, ::Function, ::Array{Float64,1}), :(Base.#var#533), true, :(Base.nothing), :(#self#), :(A)))
  end::Float64

[alyst]

Ah, looks like a dup of #21794. Well, at least it's a reminder to explicitly specify the result type of std()

[nalimilan]

But why does it work for var then?

[andreasnoack]

It doesn't if you supply the keyword so the version without keywords is probably handled in a special way. However, to pass the keyword values from std to var we'll have to use keywords even in the default case. Maybe we should get rid of the mean keyword as we have already done for cov and cor.

[alyst]

The "Performance tips" section of the manual is somewhat cryptic and doesn't describe the problem that we see here. The more relevant details are in #10443, #15468.
I'm a little bit lost, because to my understanding the behaviour of std() doesn't fit what is described in the issues: it doesn't look like the specialization on the default keyword argument value is working well.
Would be nice to clarify the situation either by updating the manual or updating the plan for #9551.

[KristofferC]

julia> using Test

julia> @inferred std(randn(5))
1.369033101041574