c73846e418
This now either outperforms, or performs as well within the region of measurement noise as, FreeBSD's libc memcpy, which is hand-written assembly. This uses a jump table for small copies with a sequence of power-of-two loads and stores for each, a vector-register copy with an overlapping copy for the last chunk for medium copies and, on x86, rep movsb for large copies. The checked version still incurs a noticeable overhead.