This does look at bit expensive. It's probably a good idea to measure the performance impact of it. It might be a better idea to write the error to a mutex / atomic instead.

Originally posted by @Zoxc in #116849 (comment)

Once we have enabled parallel-compiler by default, we can run multithreaded benchmarks of the current impl against an impl that instead writes to an AtomicBool (as Result<(), ErrorGuaranteed> is effectively just a bool)