A small nanopass exampleback

The nanopass framework is a scheme framework for writing compilers using tiny passes transforming the AST. The framework helps you immensely by generating transformers for you, in particular it handles recursively calling the appropriate transformers for the different kinds of terms in your AST.

Its implementations support Chez scheme and Racket. The Racket implementation also has some accompanying documentation.

However I found more answers to my hands-on questions by looking at the scheme-to-c compiler source. Although it's very well commented, it's quite a difficult read when not used to the framework.

So in order to showcase some of the framework's features and perhaps help new users, I decided to write down a small example of transforming a silly arithmetic language into Church encoded lambda calculus, which is trivial to then transform back into scheme.

For this example I've used Chez scheme for no other reason than that I saw Friedman and Byrd use it when showing of miniKanren.

Demo time!

The example code has a REPL, where we can evaluate our highly advanced scientifically important arithmetic calculations:

silly-church> (+ (- 3 1) (* 2 4))

For a moment one might have the desire to know the resulting Church encoding, that urge can be satisfied as well:

> (encode '(+ (- 3 1) (* 2 4)))
(((lambda (m)
    (lambda (n) (lambda (f) (lambda (x) ((m f) ((n f) x))))))
   ((((lambda (pred) (lambda (m) (lambda (n) ((n pred) m))))
       (lambda (n)
         (lambda (f)
           (lambda (x)
             (((n (lambda (g) (lambda (h) (h (g f))))) (lambda (u) x))
               (lambda (u) u))))))
      (lambda (f) (lambda (x) (f (f (f x))))))
     (lambda (f) (lambda (x) (f x)))))
  (((lambda (m) (lambda (n) (lambda (f) (m (n f)))))
     (lambda (f) (lambda (x) (f (f x)))))
    (lambda (f) (lambda (x) (f (f (f (f x))))))))

but without further stalling for time...

Show me the code!

The code is kept here and as an overview, the small "compiler" goes through these passes:

The passes are primarily designed to showcase different feature of the nanopass framework, or secondarily perhaps to keep your black box hot.

The language Lsrc and the pass ast-to-Lsrc

(define-language Lsrc
    (number (n))
    (operator (op)))
  (Expr (e)
    (op e0 e1)))

The terminals all need to have corresponding predicates, and since number? is already defined we only need to define:

(define (operator? x) (memq x '(+ - *)))

The first and last pass of a compiler written in the nanopass framework are quite different from the passes in between. Or to be precise, any pass to or from anything other than a language defined in the framework's syntax.

The first pass has to transform a regular scheme list, represented in the nanopass framework as the * language, into our Lsrc language. Here is how that is done:

(define-pass ast-to-Lsrc : * (ast) -> Lsrc ()
  (parse : * (e) -> Expr ()
      [(number? e) e]
      [(and (list? e) (= 3 (length e)))
       (let ([op (car e)] [e0 (parse (cadr e))] [e1 (parse (caddr e))])
       `(,op ,e0 ,e1))]))
  (parse ast))

Noteworthy things are:

The language L1 and the pass encode-numbers

Here we encode the numbers into lambdas and applications, so the next language reflects exactly that:

(define-language L1
  (extends Lsrc)
    (- (number (n)))
    (+ (variable (v))))
  (Expr (e)
    (- n)
    (+ v)
    (+ (lambda (v) e))
    (+ (apply e0 e1))))

To look at the full definition of a language do:

> (language->s-expression L1)
(define-language L1
  (entry Expr)
  (terminals (variable (v)) (operator (op)))
  (Expr (e) (apply e0 e1) (lambda (v) e) v (op e0 e1)))

The encode-numbers pass encodes numbers using a small recursive function:

(define-pass encode-numbers : Lsrc (ast) -> L1 ()
  (Expr : Expr (e) -> Expr ()
    [,n (letrec ([go (lambda (n) (if (= n 0) `x `(apply f ,[go (- n 1)])))])
        `(lambda (f) (lambda (x) ,[go n])))]))

Note that since the (op e0 e1) expression is not affected by this transformation it's not explicitly stated in the transformer. Here's the automagic, the nanopass framework generates cases for these with the appropriate recursive calls. That is, even deeply nested numbers get encoded by this pass without us having to bother to take care of the recursion over the AST.

The language L2 and the pass curry-operators

Since our target lambda calculus only support lambda abstractions with arity one we need to expand the dyadic (op e0 e1) into two applications.

The L2 language reflects this change:

(define-language L2
  (extends L1)
  (Expr (e)
    (+ op)
    (- (op e0 e1))))

Note that the op terminal became an Expr on its own, since it will take the place of one the e:s in (apply e0 e1).

The curry-operators pass does the actual work:

(define-pass curry-operators : L1 (ast) -> L2 ()
  (Expr : Expr (e) -> Expr ()
    [(,op ,e0 ,e1)
     `(apply (apply ,op ,[Expr e0]) ,[Expr e1])]))

Note that we need to explicitly invoke our Expr-transformer recursively to transform the two sub-expressions e0 and e1. That is, the framework does not automagically transform the input expression e0 of type L1 Expr into L2 Expr, but that's exactly the type of the defined Expr transformer.

The language L3 and the pass encode-operators

Now we're quite ready to encode the operators:

(define-language L3
  (extends L2)
    (- (operator (op))))
  (Expr (e)
    (- op)))

and what we're left with is just:

> (language->s-expression L3)
(define-language L3
  (entry Expr)
  (terminals (variable (v)))
  (Expr (e) (apply e0 e1) (lambda (v) e) v))

The actual pass transforms the operators following this table. Since this in itself is not a nanopass related exercise, I've taken the opportunity to showcase the with-output-language syntax, which interprets quasiquoted lists as the chosen output language's non-terminal (here the Expr):

(define-pass encode-operators : L2 (ast) -> L3 ()
    (with-output-language (L3 Expr)
      (define plus
        `(lambda (m) (lambda (n) (lambda (f) (lambda (x)
           (apply (apply m f) (apply (apply n f) x)))))))
      (define pred
        `(lambda (n) (lambda (f) (lambda (x)
               (apply n (lambda (g) (lambda (h) (apply h (apply g f)))))
               (lambda (u) x))
             (lambda (u) u))))))
      (define minus
           (lambda (pred)
             (lambda (m) (lambda (n) (apply (apply n pred) m))))
      (define multiply
        `(lambda (m) (lambda (n) (lambda (f)
           (apply m (apply n f))))))))
  (Expr : Expr (e) -> Expr ()
     (case op
       [+ plus]
       [- minus]
       [* multiply]
       [else (error 'encode-operators "unsupported operator" op)])]))

The pass output-scheme

The final pass simply transforms the L3 language into scheme, quite effortlessly:

(define-pass output-scheme : L3 (ast) -> * ()
  (Expr : Expr (e) -> * ()
    [,v v]
    [(apply ,e0 ,e1) `(,[Expr e0] ,[Expr e1])]
    [(lambda (,v) ,e) `(lambda (,v) ,[Expr e])]))

and this pass completes the journey from * to *! Happy passing!