Firstly, you are misinterpreting the output. In your example, wit (where the limit is being changed), runs 35% faster than non as it should.

And as you increase the number of iterations of the for loop, thereby reducing the significance of the benchmarking overhead, so the difference gets closer and closer to the 2x faster you are expecting:

c:\test>junk
        Rate  non  wit
non 381539/s   -- -32%
wit 559590/s  47%   --

c:\test>junk -N=100
       Rate  non  wit
non 46194/s   -- -47%
wit 87401/s  89%   --

c:\test>junk -N=1000
      Rate  non  wit
non 4763/s   -- -50%
wit 9557/s 101%   --

c:\test>junk -N=1e5
      Rate  non  wit
non 48.8/s   -- -50%
wit 96.8/s  98%   --

c:\test>junk -N=1e6
      Rate  non  wit
non 4.80/s   -- -50%
wit 9.55/s  99%   --
[download]