V8 has a hand rolled JSON parser which feeds directly into our object model without any intermediate data structures. This allows us to take advantage of structural knowledge, e.g. we assume that objects in an array have the same shape (and just verify this before falling back to pointer chasing through transition trees).

I keep meaning to look into SIMD for both this and ordinary tokenisation, last time I tried it the bookkeeping overheads and alignment requirements weren't worth the speedup but it's always worth re-evaluating.

Here's the JSON parser for V8. I'd never looked at it before, but it's surprisingly straight-forward and readable.

https://github.com/v8/v8/blob/master/src/json/json-parser.cc

Most JSON parses are small. Does this have bigger spin-up overhead? Also the memory handling requirements of working with GC'd JS data are very particular, this sounds like it has its own way with memory.

This, perhaps: https://chromium.googlesource.com/chromium/src/+/refs/heads/...?