Where do those speed ups come from? 10x seems extraordinary for a task that’s pretty IO-bound. Was Pyre ineffectively parallelised or something?

IO bound, how you define that? From my perspective, modern consumer grade SSDs/NVMes producing 7GB/s-8GB/s (yep, Giga Byte per second) for reads gives me 1/7-1/8 of second to load that data in RAM the rest of the 1 second is for CPU.

"pretty" was doing too much work, but there's the twin cost of reading and parsing the file, which usually needs to be done twice (once for static reflection, and once for analysis). The reading and parsing eats up about 16% of the analysis phase, and slightly more overall — maybe 20% of the total runtime.

Parsing is not IO and clearly an area you can get much better performance in-program.