A bit of one, a bit of the other. It's fairly well-optimised. I think the big remaining optimisation will be to de-allocate memory less frequently — the program creates and destroys millions of structs representing types. 20% of the runtime is taken up iterating over and dropping collections of type-related structs.
Yes — that's after a ton of Rust-centric optimisation. About 25% of the runtime is currently consumed with deallocating memory (there are a lot of heavily-nested data structures getting cleaned up), so there's definitely some more work to be done to reduce cloning.