> that don't use an AST at all but generate machine code directly during parsing It is not any different. A "sufficiently smart" compiler must be capable of fu

> that don't use an AST at all but generate machine code directly during parsing It is not any different. A "sufficiently smart" compiler must be capable of fusing the lowering passes straight into a parser. Parsing is not that different from the other kinds of passes. > that becomes harder the more stages you introduce My usual trick is to have a "unit test" simple AST which, when building in a certain mode, is fed into the compilation pipeline and displayed with all the annotations. It helps a lot to understand what exactly each pass is doing. In the next version of my DSL for building compilers (currently a work in progress) it will even be a part of an IDE. > the more language dialects you need to manage to mentally separate when working on the stages Ideally, one pass = one language. Nanopass handles it nicely. I am using this approach for pretty much all kinds of languages, and never had any issues with too many IRs. In my book, the more distinct IRs, the better.