I've resolved the conflicts and am preparing this PR to be pulled in, since I'd like it to be in before deploying.

Is there a reason to choose Arc over Rc? Are we going to multithread the compiler?

I've rebased this off of master.

Is there a reason to choose Arc over Rc? Are we going to multithread the compiler?

I figured it would be good to leave us with the option to. I don't know how big the performance hit is from using Arc vs Rc, but I could measure it.

Is there a reason to choose Arc over Rc? Are we going to multithread the compiler?

I figured it would be good to leave us with the option to. I don't know how big the performance hit is from using Arc vs Rc, but I could measure it.

The only experience I have is with C++ std::shared_ptr in which case it was worth worrying about when used non-trivially...

This is rebased (again) and CI fixed.

The only experience I have is with C++ std::shared_ptr in which case it was worth worrying about when used non-trivially...

C++'s shared_ptr<T> is equivalent to Rust's Rc<T>. Rust's Arc<T> is equivalent to C++'s atomic<shared_ptr<T>>. There will be a performance hit switching to either. I don't know how big the difference will be between Arc and Rc.

C++'s shared_ptr<T> is equivalent to Rust's Rc<T>. Rust's Arc<T> is equivalent to C++'s atomic<shared_ptr<T>>. There will be a performance hit switching to either. I don't know how big the difference will be between Arc and Rc.

It's beside the point here, but shared_ptr is thead safe and maintains a lock. I experienced horrendous perf once in a project which used hundreds of them and it was fixed when refactoring to unique_ptr. So it depends on the execution profile, if we end up with 100s of our Arc<str> and copy them a lot then I imagine perf will be impacted. But only if this is the case. We might not copy these guys around that much.

Looking at the CI runs for this branch vs. master, I don't think there's a significant performance difference. Considering we are single threaded, there's essentially just an extra semaphore and a u64 counting references. I'm gonna go ahead and merge this in, if we want to bench against plain Rc later we can.

+1 on @sezna's comment to bench first, and if using Arc is a bottleneck then we can optimize when it's an issue.