So it kicks this constraint to the end of the queue. Maybe you have (Num a) => a but I do not know what a is yet. GHC can reach a type class constraint that it cannot solve right now. It lets you have variants like HM(=) for languages like Elm and HM(≤) for languages like Scala with it is true about using a greedy algorithm, but that refers to the order of solving constraints. It needs a certain data structure, so you end up writing code a certain way and small details are not sure! I believe the approach I use is called HM(X). Maybe that would have been okay though! Not that doesn’t sound plausible to me. But I’m not sure it would have been ideal to learn how to write compilers with a language that lets you get so deep into perf. If I was starting again knowing everything I know now, I might consider Rust as well. Very very frustrating! Thinking of all the tradeoffs though, I still think it is a good option for my needs! So you may make a cool concurrent system to compile files, but it all “gets evaluated lazily” when you actually write the assets to disk, thereby becoming sequential. compilation is not necessarily going to happen on the light-weight thread you want unless you are very careful and vigilant. That said, a huge amount of my time has been wasted by laziness in my experience. It has good portability to different OSes, it can be made to go quite fast, and it is pretty good for concurrently compiling files. I think Haskell is a good fit for this kind of problem.

And I know the language isn’t changing to accommodate such a particular scenario at such high cost to the rest of the ecosystem! I believe that is the only thing that I use mutation for when compiling individual files in the Elm compiler! I think it’d be cool to write a compiler in Elm, but I also wouldn’t want to worry about that aspect.

Without it, you can get really bad asymptotics that are obvious when you start getting to 200+ line files. Specifically the union-find data structure as outlined here. Only technical note that stands out to me is that the type inference algorithm absolutely needs mutation to run in roughly linear time for large files.

Writing a compiler