tolower() small string performance

Tony Finch's recent blog post discusses the performance of the `tolower()` function for small strings, particularly focusing on strings up to 100 bytes long. This analysis builds on previous work that examined larger strings and aims to identify the crossover point between scalar code and AVX-512 with masked loads and stores. Finch's testing involved various implementations of `memcpy` and `tolower`, with performance measured using Linux's `perf_event_open(2)`. The results indicate that for very small strings, scalar code is faster for strings less than 5 bytes, while AVX-512 outperforms scalar code for strings longer than 5 bytes. The study also highlights that Clang's AVX-256 implementation is more efficient for strings between 32 and 256 bytes. Finch encountered challenges in benchmarking, particularly with implausibly low timing measurements for small strings, which he resolved by incorporating memory fences. He concludes that benchmarking is complex and invites suggestions for improving measurement accuracy. The findings emphasize the need for further optimization in handling small strings in performance-critical applications.

- The crossover point for performance between scalar code and AVX-512 is at 5 bytes.

- Scalar code is faster for strings less than 5 bytes, while AVX-512 is better for longer strings.

- Clang's AVX-256 implementation is more efficient for strings between 32 and 256 bytes.

- Benchmarking small string performance presents significant challenges.

- The results may not accurately reflect performance in real-world applications due to inlining and optimization.