It does but the heuristics for loop unrolling without any runtime data as basis is very difficult and thus very much a hit or miss affair (and missing is expensive) which is why no compiler I know of (GCC included) enables -funroll-loops or equivalent by default in any of the standard optimization levels (-On).

In GCC, the only option which enables -funroll-loops (apart from explicitly enabling it) is -fprofile-generate which is GCC's profile guided optimization.

The reason it enables -funroll-loops is that since it gathers runtime statistics during the profiling run it has enough information to accurately perform loop unrolling without risking performance degradation.