Replacing the Read Class

Tickets

Problems with the Haskell 98 class

• the backtracking parser is very inefficient

• writing instances is painful, compared to a monadic interface

• on failure, parsers give no indication of what went wrong

Proposed replacements

• parsers based on ReadP, from Parallel Parsing Processes by Koen Classen, JFP 2004.

  • abstract type with monadic interface, invoked with

    readP_to_S :: ReadP a -> String -> [(a, String)]

    (The internal representation is more complex.)

  • choice operator supports full nondeterminism by parsing alternatives in parallel. (There is also a committed choice operator.)

  • mostly deterministic parsers are much more efficient than ReadS

  • returns no information on failure, but could be altered to do so, as it knows the position at which all alternatives are exhausted.

  • requires PolymorphicComponents

  GHC has used a new Read class based on these parsers since 5.04 (July 2002).

• Malcolm's experimental Poly TextParser in the development version of HaXml.

  • monadic interface, invoked with

    runParser :: TextParser a -> String -> (Either String a, String)

    (The internal representation is more general, parameterised on token type.)

  • returns only the first successful alternative, but with internal backtracking, and an explicit commit operator to prevent backtracking where necessary

  • on failure, returns an error string and the unparsed input

  • failure messages can be defined/combined/nested with convenient combinators

• Could also use something based on Daan's Parsec combinators

  • monadic interface, invoked with (simplifying slightly)

    parse :: Parser a -> SourceName -> String -> Either ParseError a

  • mostly deterministic, with explicit 'try' operator for non-determinism

  • on failure, reports an error message including the position at which the error occured.

  • already widely used for other parsing tasks