Happy Halloween everybody, here’s a little report on how I’ve added a macro parser to yasih (Yet Another Scheme in Haskell), my own R5RS Scheme implementation written in Haskell.

At first I’ve had to define parsers for quote, quasiquote, and unquote and add them to the expression parser. You can learn more about how scheme quoting works in the MIT-scheme documentation.

-- |Parse a Quoted Expression 'a
parseQuoted :: Parser LispVal
parseQuoted = do
    char '\''
    x <- parseExpr
    return $ List [Atom "quote", x]

-- |Parse a QuasiQuoted Expression
-- See https://schemers.org/Documents/Standards/R5RS/HTML/r5rs-Z-H-7.html#%_sec_4.2.6
parseQuasiQuoted :: Parser LispVal
parseQuasiQuoted = do
    char '`'
    x <- parseExpr
    return $ List [Atom "quasiquote", x]

parseUnQuote :: Parser LispVal
parseUnQuote = do
    char ','
    x <- parseExpr
    return $ List [Atom "unquote", x]

Expression parser:

-- |Parse an Expression
parseExpr :: Parser LispVal
parseExpr = try parseComplex
        <|> try parseRatio
        <|> try parseFloat
        <|> try parseNumber
        <|> try parseAtom
        <|> parseString
        <|> try parseBool
        <|> try parseCharacter
        <|> try parseQuoted
        <|> try parseQuasiQuoted
        <|> try parseUnQuote
        <|> try parseVector
        <|> try parseParens

The symbols 'expr, `expr, and ,expr are shorthands and are correspondingly transformed by the parser into the AST equivalent of the expressions: (quote expr), (quasiquote expr) and (unquote expr). In fact, the shorthand and list syntaxes are interchangeable. unquote makes sense only inside a quasiquoted block

After defining the parsing rules for quoting and unquoting, it is time to define the corresponding evaluation rules as eval pattern matching cases:

-- Evalaute a quoted expression, just return the value.
eval env (List [Atom "quote", val]) = return val
-- ...

-- Evaluate quasiquotation. AKA macro expander.
-- Recursively evaluate unquote forms, or just return the items if not unquoted
eval env (List [Atom "quasiquote", form]) =
    evalUnquotes form
    where evalUnquotes form =
            case form of
                List [Atom "unquote", form] -> eval env form
                List items -> do
                    results <- traverse evalUnquotes items
                    return $ List results
                _ -> return form

Here’s an example, showing what happens when entering quoted expressions in yasih’s REPL.

λ> ; here, the evaluator tries to apply
λ> (1 2 3) ; 2 and 3 as arguments to a function "1"
Not a function: : "1"
λ> ; the result is clearly not a function
λ> ; the list here is now quoted and will be treated as a literal value
λ> '(1 2 3)
(1 2 3)
λ> `(1 2 3) ; same here but using quasiquote
(1 2 3)
λ> `(1 2 ,(+ 4 5)) ; the unquoted expression is evaluated! (un-quoted)!
(1 2 9)

After we have working quotation and quasiquotation it is time to add actual macros to the interpreter. I had to change the expression data type by adding an isMacro field.

-- |Lisp Value data type
data LispVal =
    -- Other data types here...
    | Func {
        isMacro :: Bool,
        params :: [String], -- Parameters name
        vararg :: Maybe String, -- name of a variable-length argument list
        body :: [LispVal], -- list of expressions
        closure :: Env -- the environment where the function was created
    | IOFunc ([LispVal] -> IOThrowsError LispVal) -- A dirty function that performs IO
    | Port Handle -- Represents input and output devices

The important bit was changing eval’s function application clause by checking if a called function is a macro, and if so changing from eager evaluation order of the arguments to a normal-order evaluation. Read more about evaluation models here (huge thanks to Rei for the link).

-- Function application clause
-- Run eval recursively over args then apply func over the resulting list
eval env (List (function : args)) = do
    func <- eval env function
    case func of
        Func {isMacro = True} -> apply func args >>= eval env
        _ -> mapM (eval env) args >>= apply func

You can see that if the applied function is a macro, then eval env is not called over the arguments. Minor changes were needed also in apply, just adding an isMacro deconstructor to the case where Func is passed as an argument. It is not needed by now inside apply’s body.

apply (Func isMacro params varargs body closure) args =
    -- more code...

Now, it’s time to add some helper functions in Environment.hs to make a macro

-- |Helper functions to create function objects in IOThrowsError monad
makeFunc :: Bool -> Maybe String -> Env -> [LispVal] -> [LispVal] -> IOThrowsError LispVal
makeFunc isMacro varargs env params body = return $ Func isMacro (map showVal params) varargs body env
makeNormalFunc = makeFunc False Nothing
makeVarargs = makeFunc False . Just . showVal
makeMacro = makeFunc True Nothing

Finally, a define-syntax form can be defined to make macros. I have not yet implemented syntax-rules but i surely will in the next yasih updates.

-- Define a macro
eval env (List (Atom "define-syntax" : List (Atom var : params) : body)) =
    makeMacro env params body >>= defineVar env var

Here’s an example on how you can make a macro and use it as a practical construct for a programming language. Here, the let statement is defined as a macro that will expand to an anonymous function (lambda) application, in this way a closure will be made and lexical scoping will be respected, correctly reaching the let statement’s goals. The implementation is a little sketchy and will surely get better as syntax-rules gets implemented.

(define-syntax (bind-vars bindings)
    `(map car bindings))

(define-syntax (bind-vals bindings)
    `(map cadr bindings))

(define-syntax (let bindings body)
    `(apply (lambda ,(bind-vars bindings) ,body) ',(bind-vals bindings)))
(let ((x 1) (y 4)) (- (* y y) (+ 1 x)))
; will expand to
((lambda (x y) (- (* y y) (+ 1 x))) 1 4)
; before being executed

This practice is called let-over-lambda. Some modern functional programming languages like OCaml, directly create a closure instead of creating an anonymous function for let statements; that practice is fairly more performant, but the execution result is identical.

Some updates on yasih’s development.

A sketchy complete R5RS implementation is almost done. Some primitives are still missing and I will work on them in the next updates, in which I will also focus on adding additional features like string interpolation, stack tracing, and probably optimization features like a garbage collector.