So there is a basic introduction to generating LLVM in Haskell there, but what's so great about it is that generating a lexer/parser/(interpreter|compiler) is almost magic in Haskell. The total code of that project is remarkably small. For that reason, it's a pretty standard Haskell workflow to do everything in Haskell and then have it emit LLVM/Cuda/whatever to do the heavy-lifting, and then you get to reason in Haskell and your back-end takes care of the rest. As much as that sounds complicated, it's really, really not.

Think of it as a template for this new essay. Of all of the hard optimization work occurring, most of it is embedded in LLVM. But the process of going from stringly syntax to intermediate language to back-end compilation is an important journey and doing it as Steven does using Haskell is an elegant way to scope out those steps.