[slavik81]

I bet the performance would look different in a real program anyways. The random input data probably results in more branch predict misses than would occur in normal text, and the tight loop of the benchmark basically ensures the lookup table is always in cache.

[BeeOnRope]

Definitely, but I think it's unlikely that it will close the gap much between the fast and slow std::toupper() compilations (my own toupper is mostly just in there as an interesting reference point).

In addition to more instructions and a function call, the slow version has many more memory dereferences, so if everything is very cold it is likely to suffer more misses.