The diff thing sounds good, but there are many layers of optimizations in between the top and the bottom, and so a small change can cascade into a tremendous nu

The diff thing sounds good, but there are many layers of optimizations in between the top and the bottom, and so a small change can cascade into a tremendous number of changes. Also, holding the entire tree of intermediate data, at each step, uses a lot of memory. That said, check out Julia: for any function, you can see how it looks at each step of compilation, down to assembly: source, optimized source, type inferred source (for given input types; it is a dynamic language), LLVM bytecode, and assembly. All visible and introspectable at runtime from either the REPL or scripts.