Note: monadic interpreter available at the monad-transformers branch.

$ nix run github:siraben/r5rs-denot # with Nix
$ cabal run # with Cabal

The R5RS Scheme specification is a 50-page beauty, outlining the syntax and semantics of Scheme in easy to understand prose, and concludes with a denotational semantics. The semantics looks a lot like Haskell because in a sense it is Haskell! We can turn the above image into a Haskell code fragment:

-- Evaluate a lambda expression with an environment e0, continuation κ
-- and store σ.
eval (Lambda is gs e0) ρ κ =
  \σ ->
    send
      (Ef
         ( new σ
         , \εs κ' ->
             if length εs == length is
               then tievals
                      ((\ρ' -> evalc gs ρ' (eval e0 ρ' κ')) . extends ρ is)
                      εs
               else wrong "wrong number of arguments"))
      κ
      (update (new σ) (Em Unspecified) σ)

Currently, the standard environment contains the following primitive procedures:

+ * - / modulo < > = >= <= cons car cdr list eqv? boolean? symbol?
procedure? pair? number? set-car! set-cdr! null? apply
call-with-values values call-with-current-continuation call/cc
string? symbol->string string->symbol string-append number->string
recursive

And the following core special forms:

(if <expr> <expr> <expr>)
(if <expr> <expr>)
(set! <id> <expr>)
(lambda <id>* <expr>*)
(lambda <id> <expr>*)
(lambda (<id>* . <id>) <expr>*)
(<expr> <expr>*)

The following derived forms have been implemented. Currently they are parsed and expanded in Haskell, but a hygienic macro system will take their place.

(define <id> <expr>)
(define (<id> <id>*) <expr>*)
(define (<id>* . <id>) <expr>*)
'<expr>
(begin <expr>*)
(cond (<expr> <expr>)*)
(let ((<id> <expr>)*) <expr>*)
(letrec ((<id> <expr>)) <expr>*)
(and <expr>*)
(or <expr>*)

Ensure Cabal is installed and build this project by running cabal run. The REPL will boot up. Type an expression and hit ENTER to evaluate it.

r5rs-denot> (define (fact n) (if (= n 0) 1 (* n (fact (- n 1))))) (fact 6)
720
Memory used: 10 cells

Alternatively, read it from a file in the demo folder, using GHCi:

-- Factorial
λ> repf "demo/factorial.scm"
720
Memory used: 10 cells

-- Primes via streams
λ> repf "demo/primes.scm"
(2 3 5 7 11 13 17 19 23 29 31 37 41 43 47 53 59 61 67 71)
Memory used: 3196 cells

-- Mutable state
λ> repf "demo/counter.scm"
(1 2 3 4)
Memory used: 15 cells

The Scheme program:

((lambda (x) (+ x x)) 10)

Can be written in the Haskell Expr datatype as:

prog = App (Lambda ["x"] [] (App (Id "+") [Id "x", Id "x"])) [Const (Number 10)]

Which can be evaluated in a standard environment and infinite store by running.

evalStd prog

The result is a triple (String, [E], S), consisting of a string, a list of results, and the final store (up to the first empty cell).

The performance has been greatly improved due to the use of the Data.IntMap library to implement the store. Thus, the Scheme interpreter runs reasonably fast.

λ> repf "demo/primes.scm"
(2 3 5 7 11 13 17 19 23 29 31 37 41 43 47 53 59 61 67 71)
Memory used: 6410 cells
λ> repf "demo/eval/eval-define.scm"
a
Memory used: 409 cells
λ> repf "demo/counter.scm"
(1 2 3 4)
Memory used: 15 cells
λ> repf "demo/meta-circ.scm"
"\n720"
Memory used: 16142 cells

• Strings
• define program declarations
  • Desugar into let and set!
• Escaping characters in strings
• quasiquote/unquote
• Hygienic macro expansion, according to this paper
• Vectors (using IntMap as representation)
• Rewrite eval from a store-passing, continuation-passing interpreter to a monadic style
  • In progress: at monad-transformers branch, Scheme monad Scheme m u r s a representing a Scheme computation in environment u, store s, continuation r, over monad m returning a value of type a
  • Inject IO computations into the Scheme monad
    • Implement ports, and user I/O
• Automated test suite

This is also a good example of the problems with "unstable denotations", as the specification is quite rigid and would have to be re-written from scratch in the face of new effects.

Furthermore, the semantics does not cover all of the language, and several language features are missing semantics. Thus, we can only guess at the appropriate semantics for this, and test them against the examples in the earlier sections of the paper.