Large performance regression with skipchars on small/medium inputs

[fcard]

Julia 0.5:

julia> mktemp() do _, file
                println(file, "G")
                flush(file)
                seek(file, 0)
                skipchars(file, islower)
                @time for i in 1:1000000; skipchars(file, islower) end
              end
  0.016611 seconds

Julia 0.6 (alpha.182):

julia> mktemp() do _, file
                       println(file, "G")
                       flush(file)
                       seek(file, 0)
                       skipchars(file, islower)
                       @time for i in 1:1000000; skipchars(file, islower) end
                     end
  0.951664 seconds (1000.00 k allocations: 30.518 MiB, 0.33% gc time)

Booo! Whoever introduced this sucks!

Just so people won't forget about it. This problem happens until there are about 40/50 characters to be skipped with an inexpensive predicate. (islower, isspace, etc)

I've got a first draft of a fix >>here<< (code under MIT license, etc etc), but I am not willing to make a PR at the moment.

[nalimilan]

What do you mean by "not willing"? Do you need some time, or do you need somebody else to fix this?

Cf. #18752

[fcard]

Someone else, plz.

[fcard]

I was convinced to give a PR a go, because the solution turned out to be really simple. There it is: #21117

Still not perfect but much better.