The idea here is to avoid re-parsing code whenever possible. This could be done by using a system temp directory using some sort of content hash. Of course, it needs to be ensured that the additional cost of checking for a cache hit is made up by the savings when we do get a hit. And there is no semantic change, e.g. parser warnings no longer being emitted.