OrderedDict: improve performance for non-concrete values and keys

[KristofferC]

Using the following benchmark

  using OrderedCollections
  using BenchmarkTools

  const N = 1000
  pairs = [i => (iseven(i) ? i : Float64(i)) for i in 1:N]

  d_base = Dict{Int,Number}(pairs)
  d_ord  = OrderedDict{Int,Number}(pairs)

  function sumvals(d)
      s = 0.0
      for (k, v) in d
          s += v
      end
      return s
  end

  println(Base.return_types(iterate, (Dict{Int,Number},))[1])
  println(Base.return_types(iterate, (OrderedDict{Int,Number},))[1])

  @btime sumvals($d_base)
  @btime sumvals($d_ord)

I get:

  # Base:
  #  16.441 μs (1000 allocations: 15.62 KiB)
  # Ordered Before PR
  # 179.508 μs (6214 allocations: 143.97 KiB)
  # Ordered After PR:
  # 9.180 μs (1000 allocations: 15.62 KiB)

This mirrors how Base does it https://github.com/JuliaLang/julia/blob/3e1bf51ccd6f5cb8a547ca99d0fd33de24ac85e8/base/dict.jl#L701

[timholy]

Sad to see more rather than fewer @inbounds annotations but the performance gain is compelling.

One optional tweak. Good to go once you decide how you want to proceed.

[timholy]

I know you're just copying what was already there, but I wonder about switching to first and similarly checkbounds(Bool, t.keys, i) for the i-generic one; those are more precisely aligned to what's actually being tested. It's the same thing for Vector, of course. Up to you, I'd merge this without that change.

[KristofferC]

I'd rather minimize the diff in this PR to focus on the specifics that changed the behavior.

[KristofferC]

Sad to see more rather than fewer @inbounds annotations but the performance gain is compelling.

Most (pretty much all) perf comes from the type of the Pair but there is already a lot of inbounds in here that asserts the same things and Base uses it as well.