Posted on October 28, 2016
{-# LANGUAGE FlexibleInstances #-}
{-# LANGUAGE DeriveFunctor #-}
{-# LANGUAGE ScopedTypeVariables #-}
{-# LANGUAGE TypeApplications #-}

module Bidirectional.Serialization.Example.Lookahead where

import Control.Monad
import Data.Char
import Data.Functor
import Data.Maybe
import Prelude hiding (print)
import qualified Prelude

import Bidirectional.Serialization.Classes

Bidirectional lookahead

In the previous example we implemented parsers with no lookahead. A somewhat awkward detail: when a left parenthesis is encountered, the handling of its matching right parenthesis is done in a nested auxiliary parser, obfuscating a little the fact that S-expressions are well-parenthesized strings.

Using lookaheads, we can avoid matching delimiters in auxiliary parsers, so that their structure can be made more explicit in the implementation, improving readability.

We augment the CharStream typeclass with a lookahead method.

class Profunctor p => CharStream p where

  -- Consume one character.
  char :: p Char Char

  -- One-character lookahead.
  -- Also detects end of file.
  lookahead :: p (Maybe Char) (Maybe Char)

Typeclass law

Typeclass laws specify additional constraints that the methods ought to satisfy. By restricting the possible interactions between typeclass methods, they serve as a (semi-)formal form of documentation and help programmers reason about polymorphic code involving such typeclasses.

The following sequence of actions:

should be the same as just consuming the character.

More formally, assuming MonadPlus1 p (i.e., MonadPlus (p a) for all a), then char should be equivalent to:

lmap Just lookahead >>= \c' -> case c' of
  Nothing -> mzero
  Just c -> mfilter (c ==) char

For simplicity’s sake, we will not use MonadPlus here. A more basic law, though less elegant because of partial values1, might be that, requiring just a Monad instance, char should be equivalent to:

lmap Just lookahead >>= \c' -> char >>= \d ->
  if c' == Just d then
    return d

The types


We keep the same parser as in the previous post.

newtype Parser a = Parser { parse :: String -> Maybe (String, a) }
  deriving Functor

runParser :: Parser a -> String -> Maybe a
runParser p s = case parse p s of
  Just ([], a) -> Just a
  _ -> Nothing

instance Applicative Parser where
  pure a = Parser (\s -> Just (s, a))
  Parser f_ <*> Parser a_ = Parser $ \s -> do
    (s', f) <- f_ s
    (s'', a) <- a_ s'
    return (s'', f a)

instance Monad Parser where
  Parser a_ >>= f = Parser $ \s -> do
    (s', a) <- a_ s
    (s'', b) <- parse (f a) s'
    return (s'', b)

parseChar :: Parser Char
parseChar = Parser parseChar'
    parseChar' [] = Nothing
    parseChar' (c : s) = Just (s, c)

An additional function for lookaheads.

-- Lookaheads never fail.
parseLookahead :: Parser (Maybe Char)
parseLookahead = Parser (Just . parseLookahead')
    parseLookahead' [] = ([], Nothing)
    parseLookahead' s@(c : _) = (s, Just c)

instance CharStream (Parsing Parser) where
  char = Parsing parseChar
  lookahead = Parsing parseLookahead


The printer is more surprising. It carries a Lookahead, which “announces” a character to be written (or the end of the stream).

data Lookahead
  = Unknown
  | Expect Char
  | End
  deriving (Eq, Ord, Read, Show)

data Printer a
  = Printer (String, a) Lookahead
  | Failed (String, Lookahead, String, Lookahead)
  -- Reports the point of failure, useful for debugging.
  deriving Functor

runPrinter :: Printer a -> Either String String
runPrinter (Printer (s, _) (Expect c)) =
  Left $ "Incomplete printer: " ++ show (s, c)
runPrinter (Printer (s, _) _) = Right s
runPrinter (Failed u) =
  Left $ "Conflicting printer: " ++ show u

Whenever a character is written, we ensure that it is indeed the character announced by the lookahead (if any).

The matchAhead function takes the current lookahead, a string to write, and a new lookahead and checks their consistency, returning the new lookahead after writing the string. A faster printer would unsafely not perform this check.

Its complexity is largely devoted to the case where nothing is being written, then the lookaheads are expected to match, with Unknown acting as a wildcard.

matchAhead :: Lookahead -> [Char] -> Lookahead -> Maybe Lookahead
matchAhead Unknown _ next' = Just next'
matchAhead next [] Unknown = Just next
matchAhead (Expect c) (c' : _) next' = guard (c == c') $> next'
matchAhead (Expect c) [] next'@(Expect c') = guard (c == c') $> next'
matchAhead (Expect _) [] End = Nothing
matchAhead End (_ : _) _ = Nothing
matchAhead End [] End = Just End
matchAhead End [] (Expect _) = Nothing

instance Applicative Printer where
  pure a = Printer ("", a) Unknown
  Printer (s, f) next <*> Printer (s', a) next'
    | Just next'' <- matchAhead next s' next'
      = Printer (s ++ s', f a) next''
    | otherwise
      = Failed (s, next, s', next')
  Failed u <*> _ = Failed u
  _ <*> Failed v = Failed v

instance Monad Printer where
  Printer (s, a) next >>= f = case f a of
    Printer (s', b) next'
      | Just next'' <- matchAhead next s' next'
        -> Printer (s ++ s', b) next''
      | otherwise -> Failed (s, next, s', next')
    Failed v -> Failed v
  Failed u >>= _ = Failed u

printChar :: Char -> Printer Char
printChar c = Printer ([c], c) Unknown

Whereas pure and printChar set the lookahead field to Unknown, non-trivial lookaheads are created via printLookahead. Note that this does not write anything, it is only a promise to write the given character later.

printLookahead :: Maybe Char -> Printer (Maybe Char)
printLookahead c' = Printer ([], c') (toLookahead c')
    toLookahead Nothing = End
    toLookahead (Just c) = Expect c

instance CharStream (Printing Printer) where
  char = Printing printChar
  lookahead = Printing printLookahead

A simple tokenizer

We will implement a conversion from a character stream to a token stream made of alphanumerical words and parentheses, separated (or not) by spaces.

data Token
  = Atom String -- Alphanumerical string
  | LPar -- '('
  | RPar -- ')'
  deriving Show

For example:

(a(bc()(d))e f)

should be parsed as

[ LPar, Atom "a", LPar, Atom "bc", LPar, RPar
, LPar, Atom "d", RPar, RPar, Atom "e", Atom "f", RPar

We also wish to obtain a pretty-printer at the same time, separating elements of a list (delimited by brackets) with spaces, but leaving no space after an left parenthesis or before a right one.

(a (bc (d)) e f)


A stream of tokens starts with any number of whitespace characters, followed by a sequence of individual tokens (each consuming whitespace after itself).

tokens :: forall p. (CharStream p, Monad1 p) => p [Token] [Token]
tokens = case monad1 @p @[Token] of
  Dict -> noSpace *> many' token

The many' combinator iterates a parser of Maybe until it returns Nothing, cumulating their results. As a printer, the parameter p receives a list and is expected to only print its head, but it is allowed to inspect its tail in order to prettify the output (here, inserting spaces between certain tokens); many' p traverses the list to print it one element at a time using p.

  :: forall p a
  .  (Profunctor p, Monad1 p)
  => p [a] (Maybe a) -> p [a] [a]
many' p = case monad1 @p @[a] of
  Dict -> do
    a' <- p
    case a' of
      Nothing -> return []
      Just a -> (a :) <$> tail =. many' p

Handling spaces

We have two bidirectional parsers to consume spaces. They differ in their behaviour as printers: noSpace prints nothing, postSpace may print one space depending on the next token. Both still need to inspect the next token to feed to lookahead, which allows the parsers not to consume any non-space character.

-- Consume optional spaces. Print no spaces.
-- The printing context looks at the first token to obtain
-- a lookahead.
noSpace :: forall p. (CharStream p, Monad1 p) => p [Token] ()
noSpace = space firstTokenFirst noSpace

-- Consume spaces after a token.
-- As a printer, assumes that the previous token was either an atom
-- or a closing bracket, and looks at the next token to determine
-- whether to print a space or not.
postSpace :: forall p. (CharStream p, Monad1 p) => p [Token] ()
postSpace = space firstTokenPre noSpace

-- Consume spaces. Parameterized by a custom lookahead producer.
  :: forall p a
  .  (CharStream p, Monad1 p)
  => (a -> Maybe Char) -> p a () -> p a ()
space nextChar moreSpaces = case monad1 @p @a of
  Dict -> do
    c' <- nextChar =. lookahead
    case c' of
      Just c | isSpace c -> consume >> moreSpaces
        where consume = lmap (const c) char
      _ -> return ()

-- The first character of the first token.
firstTokenFirst :: [Token] -> Maybe Char
firstTokenFirst (t : _) = Just (tokenFirst t)
firstTokenFirst [] = Nothing

-- The first character of a token.
tokenFirst :: Token -> Char
tokenFirst (Atom a) = head a
tokenFirst LPar = '('
tokenFirst RPar = ')'

-- The first character after some implicit token followed by
-- the given list of tokens.
-- This allows to print spaces prettily between tokens.
firstTokenPre :: [Token] -> Maybe Char
firstTokenPre (RPar : _) = Just ')'
firstTokenPre (_ : _) = Just ' '
firstTokenPre [] = Nothing

One token at a time

We use a lookahead to distinguish the type of the next token, deferring to auxiliary parsers atom and consume to actually consume it, followed by the appropriate whitespace consumer.

-- Parse a token and consume following spaces.
token :: forall p. (CharStream p, Monad1 p) => p [Token] (Maybe Token)
token = case monad1 @p @[Token] of
  Dict -> do
    c' <- firstTokenFirst =. lookahead
    case c' of
      Nothing -> return Nothing
      Just c
        | isAlphaNum c -> atom <* lmap tail postSpace
        | '(' == c -> consume $> Just LPar <* lmap tail noSpace
        | ')' == c -> consume $> Just RPar <* lmap tail postSpace
        | otherwise -> fail $ "Unexpected character: " ++ show c
          consume = lmap (const c) char
    -- The atom parser only needs one character following
    -- the token it is parsing.
    atom = dimap
      (\(Atom a : as) -> a ++ maybeToList (firstTokenPre as))
      (Just . Atom)
      (atom' [])

-- Parse an atom (sequence of alphanumerical characters).
-- This parser carries an accumulator to enable tail calls.
atom' :: forall p. (CharStream p, Monad1 p) => [Char] -> p [Char] [Char]
atom' acc = case monad1 @p @[Char] of
  Dict -> do
    c' <- firstChar =. lookahead
    case c' of
      Just c | isAlphaNum c -> consume >> lmap tail (atom' (c : acc))
        where consume = lmap (const c) char
      _ -> (return . reverse) acc
    firstChar (c : _) = Just c
    firstChar [] = Nothing


main :: IO ()
main = do
    s = "(a(bc()(d))e f)"
    Just e = runParser (parsing tokens) s
  Prelude.print e
  either putStrLn putStrLn . runPrinter $ printing tokens e

  1. This might not be a good simplification because the Monad laws are rather fragile in the presence of partial values.