Inlining kwarg functions decreases performance

[cstjean]

I know that keyword arguments are slow (#9551), but I was surprised that not only are they not properly inlinable, @inline is actually slower and allocates more. On Julia 0.4.5

plus(x) = x+x
plus2(args...; kwargs...) = plus(args...; kwargs...)
@inline plus3(args...; kwargs...) = plus(args...; kwargs...)

function pluser2()
    for _ in 1:1000 plus2(2) end
end

function pluser3()
    for _ in 1:1000 plus3(2) end
end

# After warming up...
@time pluser2()
> 0.000584 seconds (3.00 k allocations: 140.781 KB)

@time pluser3()
> 0.001023 seconds (5.00 k allocations: 172.031 KB)

[KristofferC]

Dup of #12219.

[yuyichao]

Not really. #12219 is about inlining heuristics and this is not.

[KristofferC]

julia> @time pluser2()
  0.000001 seconds (4 allocations: 160 bytes)

julia> @time pluser3()
  0.000002 seconds (4 allocations: 160 bytes)