https://gcc.gnu.org/bugzilla/show_bug.cgi?id=95079
--- Comment #5 from Mathias Stearn <redbeard0531 at gmail dot com> --- @François Dumont: Sorry I didn't see your question earlier. The reason that unordered_map perf hurts on 64-bit platforms is because it is designed to do a size_t modulus-by-prime on every lookup, and on most platforms that is *very* expensive (up to 100 cycles for 64 bits vs 20ish for 32 bits). Some very modern CPUs have made improvements here, but it is still much more expensive than just using power-of-2 buckets and masking, even if you need to hash the hash if you don't trust the low order bits to have enough entropy. Unfortunately, fixing this is a pretty big ABI break, so it isn't going to change any time soon.
