libcxx

64-bit machines: https://github.com/llvm-mirror/libcxx/blob/master/include/__... https://github.com/llvm-mirror/libcxx/blob/master/include/ut... But hashes all sizes of integers to themselves: https://github.com/llvm-mirror/libcxx/blob/master/include/ex... Changing that to a Fibonacci hash, or a simpler shift-and-xor, could be a quick win. Provided that libc++'s unordered_map uses power-of-two table sizes, that is. The code is labyrinthine, but i think, rather gloriously, sometimes it does, and sometimes it doesn't: https://github.com/llvm-mirror/libcxx/blob/master/include/__... https://github.com/llvm-mirror/libcxx/blob/master/include/__... __constrain_hash is a simple but entertaining bit of bit-dickery (reformatted slightly): size_t __constrain_hash(size_t __h, size_t __bc) { return !(__bc & (__bc - 1)) ? __h & (__bc - 1) : (__h < __bc ? __h : __h % __bc); } The x & (x - 1) tests whether a number is a power of two, because for any number that is not a power of two, subtracting one leaves the top bit set, so the bitwise and will contain at least one set bit. If the bucket count (number of slots) is a power of two, use a mask to extract the bottom bits of the hash. If it's not, do a modulus - but spend a branch to avoid that if the hash is already in the right range, which i'm surprised is a win.