My guess is, that in the first case two different branches end up in the same branch-prediction slot on the CPU. If these two branches predict different each time the code will slow down.

In the second loop, the added code may just be enough to move one of the branches to a different branch prediction slot.

To be sure you can give the Intel VTune analyzer or the AMD CodeAnalyst tool a try. These tools will show you what's exactly going on in your code.

However, keep in mind that it's most probably not worth to optimize this code further. If you tune your code to be faster on your CPU it may at the same time become slower on a different brand.

---

EDIT:

If you want to read on the branch-prediction give Agner Fog's excellent web-site a try: http://www.agner.org/optimize/

This pdf explains the branch-prediction slot allocation in detail: http://www.agner.org/optimize/microarchitecture.pdf