Add "#pragma UNROLL(1)" (or another small unroll factor, like 2 or 3) in front of the inner loops, to prevent complete unrolling. Using a local scalar temp variable for the outer-loop trip count may also help (and tends to help performance in many cases anyway). Adding "restrict" declarations to some of the pointers may also help. Finally, if the areas copied are contiguous, using memcpy() instead of an explicit loop is likely to be faster, less troublesome, and easier to maintain.