Since SIMD remains unstable, I'm playing around with the SIMD code to see how I can translate it into code that autovectorizes instead. It's not trivial, and some optimizations are bizarre, but it is definitely possible. Getting another 10–30% performance gain on parts of parsing!