GitHub engineers have improved the speed of case folding used in its code search engine, Blackbird, to more than 45 GB/s on an Apple M4 processor. The optimization removes early‑exit branches and makes the ASCII case conversion arithmetic and branch‑free, allowing the loop to be fully vectorized.

Case folding normalizes text for matching by removing case distinctions, a requirement for search engines, regular expressions, and case‑insensitive usernames. At scale, with over 180 million repositories indexed, even a simple operation must execute at memory bandwidth.

The new implementation sweeps the entire input buffer, detecting any non‑ASCII byte with a single accumulator. Upper‑case ASCII letters are converted to lower case by setting bit 5 using a branchless arithmetic expression. This eliminates conditional stores that would otherwise write only when a change occurs.

For Unicode characters, the design uses a 64‑code‑point page bitmap and run‑length encoding to avoid decoding each character. A lookup determines whether a character folds, and if it does, the fold is applied by adding a per‑run delta to the byte representation, skipping UTF‑8 decode and re‑encode steps. The output buffer is allocated once with a 1.5× size bound, and unchanged byte runs are copied with a single operation.

The crate, named casefold, is open source and includes the generated table and full design notes. Reported performance shows the ASCII path running at memory bandwidth, while the Unicode path remains efficient because folds are rare. The approach demonstrates that removing branches and performing arithmetic in byte space can achieve superior throughput compared with traditional decode‑then‑fold methods.