Re: [RFC] Locale-independent case conversion

On 23/9/21 4:53 pm, Pierre Joye wrote: > I wonder if either JIT could be used for the intrinsics support, > adding neon, sse[2-4.2] or avc256/512 (the latter would basically > allow most common strings to be converted in one go. > > If not, maybe split implementation however runtime cpu support would > be better. Many distributions built with SSE2 flags but are actually > ran on much recent CPUs, same for ARM (ie. graviton/neon f.e.). > Thoughts? SSE2 is already scary-fast. I benchmarked the SSE2 tolower code on my laptop: it was chewing through strings at 18 GiB/s, i.e. 50ps per byte. This was a benchmark written in C -- you would have a lot of trouble making a PHP tight loop in which SSE2 case conversion is the slow part. -- Tim Starling