Lookup tables hit the load/store units, IIRC only 2 loads per clock tick on modern Intel or AMD-Zen chips (though if anyone remembers, correct me if I'm wrong).

Lookup tables hit the load/store units, IIRC only 2 loads per clock tick on modern Intel or AMD-Zen chips (though if anyone remembers, correct me if I'm wrong). Modern bitshifts/add instructions can instead perform 3x per clock tick on Intel, and 4x per clock tick on AMD Zen. Multiplication is 5-cycles of latency, but at least 1-per-clock tick, maybe 2-per-clock-tick depending on architecture. The /128 is just a "bitshift-right 7". ------- EDIT: So you need 8x lookups (at least 4 clock ticks worth of resources, assuming the CPU can perform 2-lookups per clock tick) for your lookup table approach. But this "*(unsigned long long*)out = 3472328296227680304ULL + (((c * 9241421688590303745ULL) / 128) & 72340172838076673ULL)" operation only has 3x bit operations + 1x multiply, using 2-clock ticks worth of resources. The multiply and lookups have multiple clock-ticks of latency, but throughput tends to be the more important metric rather than latency in high-speed CPU code in my experience. (CPUs are very smart about rearranging code out-of-order to become throughput bound rather than latency bound)