My guess/intuition was that walking backwards might defeat any memory pre-caching at various layers (CPU, OS).

I wrote this quick-and-dirty code [http://codepad.org/d9ilykxA] to try and test the effect.

To me, it's a little interesting how it various with optimisation level, but I've not picked apart the assembly to see what's going on.

Under gcc 4.8.1, 64bit ubuntu):

  -O0 forwards 4.37s, backwards 4.88s (fowards faster)
  -O1 forwards 2.67s, backwards 2.66s (same/backwards faster)
  -O2 forwards 2.67s, backwards 2.66s (same/backwards faster)
  -Os forwards 3.33s, backwards 3.00s (backwards faster)

and the same under clang:

  -O0 forwards 3.35s, backwards 4.11s
  -O1 forwards 2.67s, backwards 2.67s
  -O2 forwards 0.69s, backwards 0.77s
  -Os forwards 2.67s, backwards 3.00s

So...it's not as simple as "forwards or backwards always faster", unless the test code is simple enough to be defeated by the optimiser in some cases which real code wouldn't.

Also - what is going on with clang -O2?