Improve the range comparison

Improve the range comparison

As mentioned in #29734, the range comparison closure can be improved.

The LLVM IR and the assembly from the new version are much simpler and
unfortunately we cannot rely on the compiler to optimise this much, as
it would need to know that `lo <= hi`.

Besides from simpler code, there might also be a performance
advantage, although it is unlikely to appear on benchmarks, as we are
doing a binary search, which should always involve few comparisons.

The code is available on the playpen for ease of comparison:
http://is.gd/4raMmH