Type instability of the std function?

[goszlanyi]

Please consider the following code:

function fun1(x)
    delta = std(x)
    
    for i in eachindex(x)
        if x[i]<delta
           x[i] = -x[i]
        end
    end

    return x
end

function fun2(x)
    delta::Float64 = std(x)
    
    for i in eachindex(x)
        if x[i]<delta
           x[i] = -x[i]
        end
    end

    return x
end


x = rand(100_000);

@time for i=1:1000 fun1(x); end
@time for i=1:1000 fun2(x); end

2.233152 seconds (100.01 M allocations: 1.490 GiB, 4.51% gc time)
0.404369 seconds (8.43 k allocations: 360.187 KiB)

This is a significant speed difference.
Does it come from from the type instability of the std function?

julia> versioninfo()
Julia Version 0.6.0
Commit 9036443* (2017-06-19 13:05 UTC)
Platform Info:
OS: Windows (x86_64-w64-mingw32)
CPU: Intel(R) Core(TM) i7-4790S CPU @ 3.20GHz
WORD_SIZE: 64
BLAS: libopenblas (USE64BITINT DYNAMIC_ARCH NO_AFFINITY Haswell)
LAPACK: libopenblas64_
LIBM: libopenlibm
LLVM: libLLVM-3.9.1 (ORCJIT, haswell)

[StefanKarpinski]

I think this is coming from the fact that std passes a function as a keyword argument (mean=mean) to var and I don't think that we can infer keyword function arguments like we can positional function arguments. Nothing fundamental here, it's just a matter of optimization elbow grease.

[goszlanyi]

Thank you for the clear explanation.
Is there already a more general issue for the inference of functions as keyword arguments?

[yuyichao]

Ref #9551 #21794 #23548 #21915 and all linked issues.

Dup of #21811

[goszlanyi]

Huh. This is a substantial history.
Thank you for linking it here.