Thanks. FNV looks fast because it's so simplistic. Just 1 mul and 1 xor. CPU Multiply isn't as slow compared to Add/Shift as it used to be, so that helps FNV, compared to One-At-a-Time trying to avoid MUL by using Add/Shift instead. I would expect Jenkins' Lookup3 to be slower, because it uses 12-byte blocks, so it wastes a lot of effort for item names whose lengths aren't at/just-below a multiple of 12.
Does the switch to FNV mean we'll have a new $fnv() identifier too? If so, would it be limited to the 32-bit variant due to avoiding big-number arithmetic? It looks like FNV-32 multiplies to a quotient as large as 56 bits, but FNV-64 would have a quotient up to 104 bits. I know some mIRC interal functions are able to have accuracy better than the 2^52 limit in $calc().