Exactly, the best compilation technique involves dozens of small passes, ideally with a new AST after every pass. And rewriting the whole thing so many times is

Exactly, the best compilation technique involves dozens of small passes, ideally with a new AST after every pass. And rewriting the whole thing so many times is very error-prone and clumsy in general. If you add a new variant to the first AST you may end up amending dozens of the lower ones. My ideal approach is the one of Nanopass (or MBase, which was developed independently), where you can derive an AST from an existing one using simple rules. For example, my usual trick for a typing pass is to enumerate all the expression nodes, giving them unique ids (that will become free variables in the type equations). The new AST is usually generated with a couple of lines of code: ast typed: simple (simpleexpr -> osimpleexpr) recform { simpleexpr = t(ident:ttag, osimpleexpr:e);} And the rest of the AST can be very complex and elaborate, but we only wanted to rewrite `simpleexpr` nodes here and nothing else.