Make doc collection 600x faster in genstdlib.jl

[yuyichao]

This line takes ~0.12s on 0.4 and ~70s on master. It spends all it's time doing type inference on string(::Function) and add_all_docs(_, ::Function) (where Function is a concrete gf type). This patch bring the time down to ~0.12s.

@JeffBezanson Seems that some specialization heuristic is messed up here?

[yuyichao]

A stand alone test case,

on 0.4

julia> function f(m)
           for i in names(m)
               v = m.(i)
               isa(v, Function) && string(v)
           end
       end
f (generic function with 1 method)

julia> @time f(Base)
  0.019418 seconds (10.62 k allocations: 595.977 KB)

on master

julia> function f(m)
           for i in names(m)
               v = m.(i)
               isa(v, Function) && string(v)
           end
       end
f (generic function with 1 method)

julia> @time f(Base)
 47.295612 seconds (33.63 M allocations: 1.330 GB, 0.83% gc time)

[StefanKarpinski]

Well, that's an optimization if I've ever seen one.

[StefanKarpinski]

It seems like the heuristic of not specializing on function values when they're not called should kick in here though, so maybe this is a case for heuristic improvement? cc: @JeffBezanson

[yuyichao]

Right, and that's the heuristic I suspect is somehow messed up. Looks like AppVeyor has problem finishing the test and wants to fail at totally unrelated places....

[yuyichao]

AppVeyor is finally green so I'm merging this now. I'll open an issue to track the specialization heuristic later (unless Jeff fixes it before I'm able to do that =) ).

I've also updated the standalone example above to show a more dramatic difference (2000x). A profile of it with only points with >= 1000 samples is available here.

@jrevels for adding to the benchmark suite. Note that this benchmark shouldn't be rerun without restarting julia.