Billion-query-per-second lookup tables in C++

Many programs repeatedly map a small, fixed set of strings - such as HTTP methods, protocol names, or compiler keywords - to values. A conventional C++ unordered_map handles these lookups at runtime by hashing the full string and navigating buckets, even though its keys never change. The proposed alternative is a perfect hash: a function tailored to the known keys so each one maps to a unique table slot, eliminating collisions, probing, and chains. The slot is calculated from the string’s length plus small values drawn from selected character positions, then reduced with a bit mask. For GET, PUT, POST, and DELETE, for example, using the first character can assign each key a distinct slot. A lookup still compares the candidate string with the slot’s stored key, which also lets it reject unknown inputs.

A C++23 library performs the search for this hash function during compilation, using a bounded greedy algorithm that tries association values without revising earlier choices. This avoids a separate code-generation tool and leaves a constant map for runtime use; the hash calculation can require only a few instructions. The remaining cost is string equality. Rather than compare bytes one at a time with a branch per byte, the approach treats keys as vectors and aims to use processor vector operations on ARM and x64. The core argument is that fixed-key lookups can be much cheaper when both hashing and comparison are designed around the known set.