Another way incremental performance can be improved is to split compiler passes into queries operating on modules. Where we only run them if the module change instead of running them for the entire crate. This is useful for cheap passes where having a query per item would have too much overhead. I did try to introduce this for some passes, but ran into some issues. This probably needs some attention from @michaelwoerister. I also have a local branch which applies this to a nice number of passes, some of them in a very questionable way. It’s this approach very applicable to lints in particular, since the are usually cheap and local.
Doing this would also allow running the passes on the modules in parallel, without per-item overhead. I think this approach is probably one of the easiest ways to improve performance.