Microbenchmark measurements that are faster than the memory system are frequently a symptom that the compiler optimized most or all of the unit under test into a no-op.
Also, I perhaps could more fully motivate my time equation parenthetical. With such an equation in hand, one can easily compare hash functions with differing trade offs, such as Wang Yi's hash and Farm Hash:
mem-wy : tCC = (0.1573 +- 0.0039)*bytes + (7.10 +- 0.13)
mem-farm: tCC = (0.3017 +- 0.0014)*bytes + (2.347 +- 0.044)
t1 = t2 => m1*n + b1 = m2*n + b2 => n = (b1-b2)/(m2-m1)
n = (7.1 - 2.347) / (.3017 - .1573)
= (4.753 +- .137) / (0.1444 +- 0.0041)
= 32.9 +- 1.3 bytes
Of course full charts/graphs/yadda-yadda can be even better, but just two summary numbers can also be nice.