Speed up BitArray packing

[carlobaldassi]

Addresses #13914.

[tkelman]

this is the kind of thing we should try and track perf for, ref #13893

[carlobaldassi]

I managed to speed up the whole process of BitArray packing.
Example:

a = rand(1:5, 300, 200);

julia> b = 2;

julia> function bcmp(n, a, b)
           for _ = 1:n
               a .!= b
           end
       end;

julia> function cmp(n, a, b)
           for _ = 1:n
               o = Array(Bool, size(a))
               for i in eachindex(a)
                   o[i] = a[i] != b
               end
            end
       end;

Timings:
on julia-0.4:

julia-0.4> @time bcmp(10_000, a, b)
  2.066846 seconds (41.74 k allocations: 113.155 MB, 0.56% gc time)

julia-0.4> @time cmp(10_000, a, b)
  0.407888 seconds (34.97 k allocations: 573.515 MB, 2.91% gc time)

with this PR:

julia-0.5> @time bcmp(10_000, a, b)
  0.675820 seconds (341.83 k allocations: 124.603 MB, 1.14% gc time)

julia-0.5> @time cmp(10_000, a, b)
  0.407919 seconds (34.30 k allocations: 573.476 MB, 3.04% gc time)

If everything works, and after some grace period, this could also be backported.

this is the kind of thing we should try and track perf for, ref #13893

Yes, definitely.

[carlobaldassi]

Also, a ~2x speedup in unsafe_setindex! by avoiding branching.

[timholy]

Very impressive!

[JeffBezanson]

Nice!