I don't have an assembly output for this particular case, but how I understand it is that the re-write basically turns it into two separate conditions, which means the branch predictor is free to model their outcomes separately.

In this case, the data is bimodal, depending on the chosen input, two likely outcomes exists. Either no looping is needed, or much looping is needed. This seemingly confuses the branch predictor when it's the same branch dealing with both scenarios.