One thing I tried to do in dfsch is to use the same IR format (in the sense of being interpretable by the interpreter) accross all the stages. What I found out

One thing I tried to do in dfsch is to use the same IR format (in the sense of being interpretable by the interpreter) accross all the stages. What I found out is that this is surprisingly hard to do in Lisp-1 language while preserving the semantics to the extent of somehow abandoning the language and thinking about making it CL-style fully lexically scoped (ie. with suport for labels and similar special forms) Lisp-2. (CPS-style IR does not make much sense for dfsch, as it explicitly and intentionally does not support call/cc, as the main use-case is easy and straightforward C interoperability) Edit: dfsch distinguishes mutable, immutable (which are somewhat redundant and in fact consume one combination of tag bits which can be used for something more useful, like strings) pairs and CDR-coded lists, which are always immutable and CDR-coded lists have additional slot that records where they came from in case of things that either was parsed from source or is result of macro-expansion, which in my opinion meets the intention of R5RS. On the other hand the very aggressive constant folding done by the macro-expansion stage probably does not.