When running the doctests for the core crate on my 28-core system the system unfortunately never really gets above 40% cpu usage, yet the doctests being compiled are embarassingly parallel and should result in full usage of the cores available to execute.

For now this can be reproduced with ./x.py test src/libcore --doc and watching CPU usage via python src/ci/cpu-usage-over-time.py

One suspicion I have for this is that rustdoc has always use in-process compilation for all the doctests, but there may be an excessive amount of synchronization in LLVM or rustc which prevents all CPU cores from being used. I'm not sure if this is the cause though.