It would be pretty damn cool if you could create a data type for
generically describing a monadic parser, and then use template haskell to generate a concrete parser from that data type. That would allow you to create your specification in a generic way and then target different parsers like parsec, attoparsec, etc. There is some code coming up in a few paragraphs that should describe this idea more clearly. I would like to suggest that while it would be cool, it is impossible. As proof, I will attempt to create a simple monadic parser that has only one combinator: anyChar :: ParserSpec Char We need the GADTs extension and some imports: > {-# LANGUAGE GADTs, TemplateHaskell #-} > import Control.Monad (join) > import qualified Text.Parsec.Char as P > import Language.Haskell.TH (ExpQ, appE) > import Language.Haskell.TH.Syntax (Lift(lift)) > import Text.Parsec (parseTest) > import qualified Text.Parsec.Char as P > import Text.Parsec.String (Parser) Next we define a type that has a constructor for each of the different combinators we want to support, plus constructors for the functor and monad methods: > data ParserSpec a where > AnyChar :: ParserSpec Char > Return :: a -> ParserSpec a > Join :: ParserSpec (ParserSpec a) -> ParserSpec a > FMap :: (a -> b) -> ParserSpec a -> ParserSpec b > > instance Lift (ParserSpec a) where > lift _ = error "not defined because we are screwed later anyway." In theory, we would extend that type with things like `Many`, `Some`, `Choice`, etc. In Haskell, we are used to seeing a `Monad` defined in terms of `return` and `>>=`. But, we can also define a monad in terms of `fmap`, `return` and `join`. We will do that in `ParserSpec`, because it makes the fatal flaw more obvious. Now we can define the `Functor` and `Monad` instances: > instance Functor ParserSpec where > fmap f p = FMap f p > instance Monad ParserSpec where > return a = Return a > m >>= f = Join ((FMap f) m) and the `anyChar` combinator: > anyChar :: ParserSpec Char > anyChar = AnyChar And now we can define a simple parser that parses two characters and returns them: > charPair :: ParserSpec (Char, Char) > charPair = > do a <- anyChar > b <- anyChar > return (a, b) Now, we just need to define a template haskell function that generates a `Parser` from a `ParserSpec`: > genParsec :: (Lift a) => ParserSpec a -> ExpQ > genParsec AnyChar = [| anyChar |] > genParsec (Return a) = [| return a |] > genParsec (Join p) = genParsec p > -- genParsec (FMap f p) = appE [| f |] (genParsec p) -- uh-oh Looking at the `FMap` case we see the fatal flaw. In order to generate the parser we would need some way to transform any arbitrary Haskell function of type `a -> b` into Template Haskell. Obviously, that is impossible (for some definition of obvious). Therefore, we can assume that it is not possible to use Template Haskell to generate a monadic parser from a monadic specification. We can also assume that `Applicative` is not available either. Seems likely that `Category` based parsers would also be out. Now, we can, of course, do the transformation at runtime: > interpretParsec :: ParserSpec a -> Parser a > interpretParsec AnyChar = P.anyChar > interpretParsec (Return a) = return a > interpretParsec (FMap f a) = fmap f (interpretParsec a) > interpretParsec (Join mm) = join (fmap interpretParsec (interpretParsec mm)) > test = parseTest (interpretParsec charPair) "ab" My fear is that doing that will result in added runtime overhead. One reason for wanting to create a compile-time parser generator is to have the opportunity to generate very fast parsing code. It seems like here we can only be slower than the parser we are targeting? Though.. perhaps not? Perhaps the parser returned by `interpretParsec` has all the interpret stuff removed and is as fast as if we have constructed it by hand? I don't have an intuitive feel for it.. I guess criterion would know.. Any thoughts? - jeremy _______________________________________________ Haskell-Cafe mailing list [hidden email] http://www.haskell.org/mailman/listinfo/haskell-cafe |
On 13-03-12 04:06 PM, Jeremy Shaw wrote:
> It would be pretty damn cool if you could create a data type for > generically describing a monadic parser, and then use template haskell > to generate a concrete parser from that data type. [...] > I would like to suggest that while it would be cool, it is > impossible. Impossibility proofs are notoriously difficult. You showed that this approach: >> data ParserSpec a where >> AnyChar :: ParserSpec Char >> Return :: a -> ParserSpec a >> Join :: ParserSpec (ParserSpec a) -> ParserSpec a >> FMap :: (a -> b) -> ParserSpec a -> ParserSpec b does not work. The flaw is indeed in FMap. It should not take a function as first argument, but rather a *description* of a function (the same way ParserSpec gives you a description of a parser). Then you can make it work, if your 'description' language is adequate. For some strange reason, I am biased towards 'finally tagless' descriptions, but YMMV. Jacques _______________________________________________ Haskell-Cafe mailing list [hidden email] http://www.haskell.org/mailman/listinfo/haskell-cafe |
In reply to this post by Jeremy Shaw-3
Why not the parsers package [1]? Write the parser against the Parsing class and then use trifecta or write instances for attoparsec or parsec. With enough inlining perhaps the overhead of the class gets optimized away? [1] http://hackage.haskell.org/package/parsers On Tue, Mar 12, 2013 at 9:06 PM, Jeremy Shaw <[hidden email]> wrote: It would be pretty damn cool if you could create a data type for _______________________________________________ Haskell-Cafe mailing list [hidden email] http://www.haskell.org/mailman/listinfo/haskell-cafe |
In reply to this post by Jacques Carette
On Tue, Mar 12, 2013 at 3:32 PM, Jacques Carette <[hidden email]> wrote:
> On 13-03-12 04:06 PM, Jeremy Shaw wrote: >>> data ParserSpec a where >>> AnyChar :: ParserSpec Char >>> Return :: a -> ParserSpec a >>> Join :: ParserSpec (ParserSpec a) -> ParserSpec a >>> FMap :: (a -> b) -> ParserSpec a -> ParserSpec b > > > does not work. The flaw is indeed in FMap. It should not take a function > as first argument, but rather a *description* of a function (the same way > ParserSpec gives you a description of a parser). Then you can make it work, > if your 'description' language is adequate. Right. But, then I would not be able to use Haskell's existing do notation -- and I would have to poorly recreate a subset of Haskell. And, I think, ParsecSpec would not be a real monad. But.. that is sort of the conclusion -- if you want to do compile-time generation, then the data-type can not contain any function values -- at least none that would need to be lifted into the generated code. And, there is no way to make a type with a real Monad instance which does not contain such a function. _______________________________________________ Haskell-Cafe mailing list [hidden email] http://www.haskell.org/mailman/listinfo/haskell-cafe |
In reply to this post by Dag Odenhall
Mostly, because I want to do other sorts of compile-time inspections
on the parser. Being able to generate the parser is just the easiest part to get started with. - jeremy On Tue, Mar 12, 2013 at 3:36 PM, [hidden email] <[hidden email]> wrote: > Why not the parsers package [1]? Write the parser against the Parsing class > and then use trifecta or write instances for attoparsec or parsec. With > enough inlining perhaps the overhead of the class gets optimized away? > > [1] http://hackage.haskell.org/package/parsers > > > On Tue, Mar 12, 2013 at 9:06 PM, Jeremy Shaw <[hidden email]> wrote: >> >> It would be pretty damn cool if you could create a data type for >> generically describing a monadic parser, and then use template haskell >> to generate a concrete parser from that data type. That would allow >> you to create your specification in a generic way and then target >> different parsers like parsec, attoparsec, etc. There is some code >> coming up in a few paragraphs that should describe this idea more >> clearly. >> >> I would like to suggest that while it would be cool, it is >> impossible. As proof, I will attempt to create a simple monadic parser >> that has only one combinator: >> >> anyChar :: ParserSpec Char >> >> We need the GADTs extension and some imports: >> >> > {-# LANGUAGE GADTs, TemplateHaskell #-} >> > import Control.Monad (join) >> > import qualified Text.Parsec.Char as P >> > import Language.Haskell.TH (ExpQ, appE) >> > import Language.Haskell.TH.Syntax (Lift(lift)) >> > import Text.Parsec (parseTest) >> > import qualified Text.Parsec.Char as P >> > import Text.Parsec.String (Parser) >> >> Next we define a type that has a constructor for each of the different >> combinators we want to support, plus constructors for the functor and >> monad methods: >> >> > data ParserSpec a where >> > AnyChar :: ParserSpec Char >> > Return :: a -> ParserSpec a >> > Join :: ParserSpec (ParserSpec a) -> ParserSpec a >> > FMap :: (a -> b) -> ParserSpec a -> ParserSpec b >> > >> > instance Lift (ParserSpec a) where >> > lift _ = error "not defined because we are screwed later anyway." >> >> In theory, we would extend that type with things like `Many`, `Some`, >> `Choice`, etc. >> >> In Haskell, we are used to seeing a `Monad` defined in terms of >> `return` and `>>=`. But, we can also define a monad in terms of >> `fmap`, `return` and `join`. We will do that in `ParserSpec`, because >> it makes the fatal flaw more obvious. >> >> Now we can define the `Functor` and `Monad` instances: >> >> > instance Functor ParserSpec where >> > fmap f p = FMap f p >> >> > instance Monad ParserSpec where >> > return a = Return a >> > m >>= f = Join ((FMap f) m) >> >> and the `anyChar` combinator: >> >> > anyChar :: ParserSpec Char >> > anyChar = AnyChar >> >> And now we can define a simple parser that parses two characters and >> returns them: >> >> > charPair :: ParserSpec (Char, Char) >> > charPair = >> > do a <- anyChar >> > b <- anyChar >> > return (a, b) >> >> Now, we just need to define a template haskell function that generates >> a `Parser` from a `ParserSpec`: >> >> > genParsec :: (Lift a) => ParserSpec a -> ExpQ >> > genParsec AnyChar = [| anyChar |] >> > genParsec (Return a) = [| return a |] >> > genParsec (Join p) = genParsec p >> > -- genParsec (FMap f p) = appE [| f |] (genParsec p) -- uh-oh >> >> Looking at the `FMap` case we see the fatal flaw. In order to >> generate the parser we would need some way to transform any arbitrary >> Haskell function of type `a -> b` into Template Haskell. Obviously, >> that is impossible (for some definition of obvious). >> >> Therefore, we can assume that it is not possible to use Template >> Haskell to generate a monadic parser from a monadic specification. >> >> We can also assume that `Applicative` is not available either. Seems >> likely that `Category` based parsers would also be out. >> >> Now, we can, of course, do the transformation at runtime: >> >> > interpretParsec :: ParserSpec a -> Parser a >> > interpretParsec AnyChar = P.anyChar >> > interpretParsec (Return a) = return a >> > interpretParsec (FMap f a) = fmap f (interpretParsec a) >> > interpretParsec (Join mm) = join (fmap interpretParsec (interpretParsec >> > mm)) >> >> > test = parseTest (interpretParsec charPair) "ab" >> >> My fear is that doing that will result in added runtime overhead. One >> reason for wanting to create a compile-time parser generator is to have >> the opportunity to generate very fast parsing code. It seems like here >> we can only be slower than the parser we are targeting? Though.. >> perhaps not? Perhaps the parser returned by `interpretParsec` has all >> the interpret stuff removed and is as fast as if we have constructed >> it by hand? I don't have an intuitive feel for it.. I guess criterion >> would know.. >> >> Any thoughts? >> >> - jeremy >> >> _______________________________________________ >> Haskell-Cafe mailing list >> [hidden email] >> http://www.haskell.org/mailman/listinfo/haskell-cafe > > _______________________________________________ Haskell-Cafe mailing list [hidden email] http://www.haskell.org/mailman/listinfo/haskell-cafe |
In reply to this post by Jeremy Shaw-3
Jeremy Shaw wrote: > It would be pretty damn cool if you could create a data type for > generically describing a monadic parser, and then use template haskell > to generate a concrete parser from that data type. That would allow > you to create your specification in a generic way and then target > different parsers like parsec, attoparsec, etc. There is some code > coming up in a few paragraphs that should describe this idea more > clearly. After rather mild and practical restrictions, the problem is solvable. (BTW, even the problem of lifting arbitrary functional values, let alone handles and other stuff, is solvable in realistic settings -- or even completely, although less practically.) Rather than starting from the top -- implementing monadic or applicative parser, let's start from the bottom and figure out what we really need. It seems that many real-life parsers aren't using the full power of Applicative, let alone monad. So why to pay, a whole lot, for what we don't use. Any parser combinator library has to be able to combine parsers. It seems the applicative rule <*> :: Parser (a->b) -> Parser a -> Parser b is very popular. It is indeed very useful -- although not the only thing possible. One can come up with a set of combinators that are used for realistic parsing. For example, *> :: Parser a -> Parser b -> Parser b for sequential composition, although expressible via <*>, could be defined as primitive. Many other such combinators can be defined as primitives. In other words: the great advantage of Applicative parser combinators is letting the user supply semantic actions, and executing those actions as parsing progresses. There is also a traditional approach: the parser produces an AST or a stream of parsing events, which the user consumes and semantically processes any way they wish. Think of XML parsing: often people parse XML and get a DOM tree, and process it afterwards. An XML parser can be incremental: SAX. Parsers that produce AST need only a small fixed set of combinators. We never need to lift arbitrary functions since those parsers don't accept arbitrary semantic actions from the user. For that reason, these parsers are also much easy to analyze. Let's take the high road however, applicative parsers. The <*> rule is not problematic: it neatly maps to code. Consider newtype Code a = Code Exp which is the type-annotated TH Code. We can easily define app_code :: Code (a->b) -> Code a -> Code b app_code (Code f) (Code x) = Code $ AppE f x So, Code is almost applicative. Almost -- because we only have a restricted pure: pureR :: Lift a => a -> Code a with a Lift constraint. Alas, this is not sufficient for realistic parsers, because often we have to lift functions, as in the example of parsing a pair of characters: pure (\x y -> (x,y)) <*> anyChar <*> anyChar But aren't functions really unliftable? They are unliftable by value, but we can also lift by reference. Here is an example, using tagless final framework, since it is extensible. We define the basic minor Applicative > class Sym repr where > pureR :: Lift a => a -> repr a > app :: repr (a->b) -> repr a -> repr b > > infixl 4 `app` And a primitive parser, with only one primitive parser. > class Sym repr => Parser repr where > anychar :: repr Char For our example, parsing two characters and returning them as a pair, we need pairs. So, we extend our parser with three higher-order _constants_. > class Sym repr => Pair repr where > pair :: repr (a -> b -> (a,b)) > prj1 :: repr ((a,b) -> a) > prj2 :: repr ((a,b) -> b) And here is the example. > test1 = pair `app` anychar `app` anychar One interpretation of Sym is to generate code (another one could analyze the parsers) > data C a = C{unC :: Q Exp} Most interesting is the instance of pairs. Actually, it is not that interesting: we just lift functions by reference. > pair0 x y = (x,y) > > instance Pair C where > pair = C [e| pure pair0 |] > prj1 = C [e| pure fst |] > prj2 = C [e| pure snd |] Because tagless-final is so extensible, any time we need a new functional constant, we can easily introduce it and define its code, either by building a code expression or by referring to a global name that is bound to the desired value. The latter is `lift by reference' (which is what dynamic linking does). The obvious limitation of this approach is that all functions to lift must be named -- because we lift by reference. We can also build anonymous functions, if we just add lambda to our language. If we go this way we obtain something like http://okmij.org/ftp/meta-programming/index.html#meta-haskell (which has lam, let, arrays, loops, etc.) Sample code, for reference {-# LANGUAGE TemplateHaskell #-} {-# LANGUAGE NoMonomorphismRestriction #-} module P where import Language.Haskell.TH import Language.Haskell.TH.Syntax import Language.Haskell.TH.Ppr import Control.Applicative import Text.ParserCombinators.ReadP class Sym repr where pureR :: Lift a => a -> repr a app :: repr (a->b) -> repr a -> repr b infixl 4 `app` class Sym repr => Parser repr where anychar :: repr Char -- Higher-order constants class Sym repr => Pair repr where pair :: repr (a -> b -> (a,b)) prj1 :: repr ((a,b) -> a) prj2 :: repr ((a,b) -> b) -- parse two characters and return them as a pair test1 = pair `app` anychar `app` anychar -- Implementations -- we don't need Q monad actually, neither here -- nor anywhere! -- It's a bummer that lift has the signature t -> Q Exp -- rather than t -> Exp! data C a = C{unC :: Q Exp} instance Sym C where pureR = C . lift app f x = C $ appE (appE (varE '(Control.Applicative.<*>)) (unC f)) (unC x) instance Parser C where anychar = C . varE $ 'get pair0 x y = (x,y) instance Pair C where pair = C [e| pure pair0 |] prj1 = C [e| pure fst |] prj2 = C [e| pure snd |] printC :: C a -> IO String printC m = runQ (fmap pprint $ unC m ) test1C = printC test1 {- "(Control.Applicative.<*>) ((Control.Applicative.<*>) (Control.Applicative.pure P.pair0) Text.ParserCombinators.ReadP.get) Text.ParserCombinators.ReadP.get" -} _______________________________________________ Haskell-Cafe mailing list [hidden email] http://www.haskell.org/mailman/listinfo/haskell-cafe |
All,
2013/3/13 <[hidden email]>: > So, Code is almost applicative. Almost -- because we only have a > restricted pure: > pureR :: Lift a => a -> Code a > with a Lift constraint. Alas, this is not sufficient for realistic > parsers, because often we have to lift functions, as in the example of > parsing a pair of characters: I've previously used an approach like this in the grammar-combinators library. See http://hackage.haskell.org/packages/archive/grammar-combinators/0.2.7/doc/html/Text-GrammarCombinators-Base-ProductionRule.html#t:LiftableProductionRule and http://hackage.haskell.org/packages/archive/grammar-combinators/0.2.7/doc/html/Text-GrammarCombinators-Utils-LiftGrammar.html. The approach uses a restricted pure like this: class ProductionRule p => LiftableProductionRule p where epsilonL :: a -> Q Exp -> p aSource and associated epsilonLS :: (Lift v, LiftableProductionRule p) => v -> p v epsilonLS v = epsilonL v $ lift v There is a function liftGrammar which lifts a grammar that uses the type class to a list of declarations using TH. This allowed me to start from a context-free grammar, transform it to a non-left-recursive grammar, optimize it and then lift it using TH. In some tests, I found that this improved performance significantly over using the transformed grammar directly, even when I try to force the transformation to happen before the benchmark. I assume this is because the lifted grammar is optimised better by the compiler. Regards, Dominique _______________________________________________ Haskell-Cafe mailing list [hidden email] http://www.haskell.org/mailman/listinfo/haskell-cafe |
2013/3/13 Dominique Devriese <[hidden email]>:
> class ProductionRule p => LiftableProductionRule p where > epsilonL :: a -> Q Exp -> p aSource > > and associated > epsilonLS :: (Lift v, LiftableProductionRule p) => v -> p v > epsilonLS v = epsilonL v $ lift v Note that the point of providing epsilonL as primitive and not just epsilonLS is that I can then still lift most functions I use: epsilonL (,) [| (,) |] Even though functions are not necessarily liftable. This is an alternative to Oleg's adding of e.g. pair etc. as DSL primitives. Dominique _______________________________________________ Haskell-Cafe mailing list [hidden email] http://www.haskell.org/mailman/listinfo/haskell-cafe |
In reply to this post by Jeremy Shaw-3
Jeremy,
The problem you're trying to solve might seem tricky but it is in fact quite solvable. In Feldspar[1] we use monads quite frequently and generate code from them, in a similar fashion to what you're trying to do. We've written a paper about how we do it[2] that I welcome you to read. If you have any questions regarding the paper I'd be happy to try to answer them. There are two parts to the trick. One is to use the continuation monad to get a monad instance. The other trick is to restrict any functions you have in your data type (like FMap in your example) so that they can be reified into something that can be compiled, which would be Template Haskell in your case.
To help you along the way I've prepared some code to give you an idea of how this can be done. This code only shows the continuation monad trick but I hope it's useful nonetheless. \begin{code} {-# LANGUAGE GeneralizedNewtypeDeriving #-} {-# LANGUAGE RankNTypes #-} {-# LANGUAGE GADTs #-} module MonadReify where newtype Cont r a = C { unC :: (a -> r) -> r }
instance Monad (Cont r) where return a = C $ \k -> k a m >>= f = C $ \k -> unC m (\a -> unC (f a) k) data ParserSpec a where
AnyChar :: ParserSpec Char
Return :: a -> ParserSpec a Join :: ParserSpec (ParserSpec a) -> ParserSpec a FMap :: (a -> b) -> ParserSpec a -> ParserSpec b bindSpec p f = Join (FMap f p)
newtype Parser a = P { unP :: forall r. Cont (ParserSpec r) a } instance Monad Parser where return a = P (return a) m >>= f = P (unP m >>= \a -> unP (f a))
anyChar :: Parser Char anyChar = P (C $ \k -> bindSpec AnyChar k) reifyParser :: Parser a -> (forall r. ParserSpec r -> b) -> b reifyParser (P (C f)) g = g (f (\a -> Return a))
\end{code} Cheers, Josef [1]https://github.com/Feldspar/feldspar-language [2]http://www.cse.chalmers.se/~josefs/publications/paper21_cameraready.pdf On Tue, Mar 12, 2013 at 9:06 PM, Jeremy Shaw <[hidden email]> wrote: It would be pretty damn cool if you could create a data type for _______________________________________________ Haskell-Cafe mailing list [hidden email] http://www.haskell.org/mailman/listinfo/haskell-cafe |
Free forum by Nabble | Edit this page |