{-# LANGUAGE Rank2Types, GeneralizedNewtypeDeriving, MultiParamTypeClasses, UndecidableInstances, FunctionalDependencies, FlexibleInstances, GADTs #-} -------------------------------------------------------------------------- -- | -- Module : Control.Monad.CC -- Copyright : (c) R. Kent Dybvig, Simon L. Peyton Jones and Amr Sabry -- License : MIT -- -- Maintainer : Dan Doel -- Stability : Experimental -- Portability : Non-portable (rank-2 types, multi-parameter type classes, -- functional dependencies) -- -- A monadic treatment of delimited continuations. -- -- Adapted from the paper -- /A Monadic Framework for Delimited Continuations/, -- by R. Kent Dybvig, Simon Peyton Jones and Amr Sabry -- (<http://www.cs.indiana.edu/~sabry/papers/monadicDC.pdf>) -- -- This module implements the delimited continuation monad and transformer, -- using the sequence-of-frames implementation from the original paper. module Control.Monad.CC ( -- * The CC monad CC(), runCC, -- * The CCT monad transformer CCT(), runCCT, SubCont(), Prompt, MonadDelimitedCont(..), -- * Assorted useful control operators reset, shift, control, shift0, control0, abort -- * Examples -- $Examples ) where import Control.Applicative import Control.Monad.Identity import Control.Monad.State import Control.Monad.Reader import Control.Monad.Trans import Control.Monad.CC.Seq import Control.Monad.CC.Prompt -- newtype Frame m ans a b = Frame (a -> CCT ans m b) data Frame m ans a b = FFrame (a -> b) | MFrame (a -> CCT ans m b) type Cont ans m a = Seq (Frame m) ans a newtype SubCont ans m a b = SC (SubSeq (Frame m) ans a b) -- | The CCT monad transformer allows you to layer delimited control -- effects over an arbitrary monad. -- -- The CCT transformer is parameterized by the following types -- -- * ans : A region parameter, so that prompts and subcontinuations -- may only be used in the same region they are created. -- -- * m : the underlying monad -- -- * a : The contained value. A value of type CCT ans m a can be though -- of as a computation that calls its continuation with a value of -- type 'a' newtype CCT ans m a = CCT { unCCT :: Cont ans m a -> P ans m ans } instance (Monad m) => Functor (CCT ans m) where fmap f (CCT e) = CCT $ \k -> e (PushSeg (FFrame f) k) instance (Monad m) => Applicative (CCT ans m) where pure = return (<*>) = ap instance (Monad m) => Monad (CCT ans m) where return v = CCT $ \k -> appk k v (CCT e1) >>= e2 = CCT $ \k -> e1 (PushSeg (MFrame e2) k) instance MonadTrans (CCT ans) where lift m = CCT $ \k -> lift m >>= appk k instance (MonadReader r m) => MonadReader r (CCT ans m) where ask = lift ask local f m = CCT $ \k -> local f (unCCT m k) instance (MonadState s m) => MonadState s (CCT ans m) where get = lift get put = lift . put instance (MonadIO m) => MonadIO (CCT ans m) where liftIO = lift . liftIO -- Applies a continuation to a value. appk :: Monad m => Cont ans m a -> a -> P ans m ans appk EmptyS a = return a appk (PushP _ k) a = appk k a appk (PushSeg f k) a = appFrame f a k where appFrame (MFrame g) b l = unCCT (g b) l appFrame (FFrame g) b l = appk l (g b) -- | Executes a CCT computation, yielding a value in the underlying monad runCCT :: (Monad m) => (forall ans. CCT ans m a) -> m a runCCT c = runP (unCCT c EmptyS) -- | The CC monad may be used to execute computations with delimited control. newtype CC ans a = CC { unCC :: CCT ans Identity a } deriving (Functor, Monad, Applicative, MonadDelimitedCont (Prompt ans) (SubCont ans Identity)) -- | Executes a CC computation, yielding a resulting value. runCC :: (forall ans. CC ans a) -> a runCC c = runIdentity (runCCT (unCC c)) -- | A typeclass for monads that support delimited control operators. -- The type varibles represent the following: -- -- m : The monad itself -- -- p : The associated type of prompts that may delimit computations in the monad -- -- s : The associated type of sub-continuations that may be captured class (Monad m) => MonadDelimitedCont p s m | m -> p s where -- | Creates a new, unique prompt. newPrompt :: m (p a) -- | Delimits a computation with a given prompt. pushPrompt :: p a -> m a -> m a -- | Abortively capture the sub-continuation delimited by the given -- prompt, and call the given function with it. The prompt does not appear -- delimiting the sub-continuation, nor the resulting computation. withSubCont :: p b -> (s a b -> m b) -> m a -- | Pushes a sub-continuation, reinstating it as part of the continuation. pushSubCont :: s a b -> m a -> m b instance (Monad m) => MonadDelimitedCont (Prompt ans) (SubCont ans m) (CCT ans m) where newPrompt = CCT $ \k -> newPromptName >>= appk k pushPrompt p (CCT e) = CCT $ \k -> e (PushP p k) withSubCont p f = CCT $ \k -> let (subk, k') = splitSeq p k in unCCT (f (SC subk)) k' pushSubCont (SC subk) (CCT e) = CCT $ \k -> e (pushSeq subk k) -- | An approximation of the traditional /reset/ operator. Creates a new prompt, -- calls the given function with it, and delimits the resulting computation -- with said prompt. reset :: (MonadDelimitedCont p s m) => (p a -> m a) -> m a reset e = newPrompt >>= \p -> pushPrompt p (e p) -- ----- -- These originally had types like: -- -- ((a -> m b) -> m b) -> m a -- -- but I came to the conclusion that it would be convenient to be able to pass -- in monadically typed values. -- As a specific example, this makes the difference between -- -- > shift q (\f -> f (dref p)) -- -- and -- -- > join $ shift q (\f -> f (dref p)) -- -- In other words, one can expressed in terms of the other (I think), but -- the fact that one has to insert a 'join' /outside/ the shift, and not -- anywhere near where the sub-continuation is actually used is rather -- odd, and difficult to remember compared to the difference between: -- -- > shift q (\f -> f (return pureValue)) -- -- and -- -- > shift q (\f -> f pureValue) -- ----- -- | The traditional /shift/ counterpart to the above 'reset'. Reifies the -- subcontinuation into a function, keeping both the subcontinuation, and -- the resulting computation delimited by the given prompt. shift :: (MonadDelimitedCont p s m) => p b -> ((m a -> m b) -> m b) -> m a shift p f = withSubCont p $ \sk -> pushPrompt p $ f (\a -> pushPrompt p $ pushSubCont sk a) -- | The /control/ operator, traditionally the counterpart of /prompt/. It does -- not delimit the reified subcontinuation, so control effects therein can -- escape. The corresponding prompt is performed equally well by 'reset' above. control :: (MonadDelimitedCont p s m) => p b -> ((m a -> m b) -> m b) -> m a control p f = withSubCont p $ \sk -> pushPrompt p $ f (\a -> pushSubCont sk a) -- | Abortively captures the current subcontinuation, delimiting it in a reified -- function. The resulting computation, however, is undelimited. shift0 :: (MonadDelimitedCont p s m) => p b -> ((m a -> m b) -> m b) -> m a shift0 p f = withSubCont p $ \sk -> f (\a -> pushPrompt p $ pushSubCont sk a) -- | Abortively captures the current subcontinuation, delimiting neither it nor -- the resulting computation. control0 :: (MonadDelimitedCont p s m) => p b -> ((m a -> m b) -> m b) -> m a control0 p f = withSubCont p $ \sk -> f (\a -> pushSubCont sk a) -- | Aborts the current continuation up to the given prompt. abort :: (MonadDelimitedCont p s m) => p b -> m b -> m a abort p e = withSubCont p (\_ -> e) ------------------------------------------------------------------------------- -- $Examples -- -- This module provides many different control operators, so hopefully the -- examples herein can help in selecting the right ones. The most raw are the -- four contained in the 'MonadDelimitedCont' type class. The first, of course, -- is 'newPrompt', which should be straight forward enough. Next comes -- 'pushPromp't, which is the basic operation that delimits a computation. -- In the absense of other control operators, it's simply a no-op, so -- -- > pushPrompt p (return v) == return v -- -- 'withSubCont' is the primitive that allows the capture of sub-continuations. -- Unlike callCC, 'withSubCont' aborts the delimited continuation it captures, -- so: -- -- > pushPrompt p ((1:) `liftM` (2:) `liftM` withSubCont p (\k -> return [])) -- -- will yield a value of [] on running, not [1, 2]. -- -- The final primitive control operator is 'pushSubCont', which allows the use -- of the sub-continuations captured using 'withSubCont'. So: -- -- > pushPrompt p ((1:) `liftM1 (2:) `liftM` -- > withSubCont p (\k -> pushSubCont k (return []))) -- -- will yield the answer [1, 2]. /However/, Capturing a sub-continuation and -- immediately pusshing it /is not/ a no-op, because the sub-continuation -- does not contain the delimiting prompt (and, of course, pushSubCont does -- not re-instate it, as it doesn't know what prompt was originally used). -- Thus, capturing and pushing a sub-continuation results in the net loss of -- one delimiter, and said delimiter will need to be re-pushed to negate that -- effect, if desired. -- -- Out of these four primitive operators have been built various functional -- abstractions that incorporate one or more operations. On the delimiting -- side is 'reset', which combines both prompt creation and delimiting. In -- some papers on the subject (such as /Shift to Control/), each capture -- operator would be paired with a corresponding delimiter operator (and -- indeed, a separate CPS transform). However, since prompts are explicitly -- passed in this implementation, a single delimiter suffices for supporting -- all capture operators (although 'pushPrompt' will need to be used if one -- wishes to explicitly push a prompt more than once). -- -- The simplest control flow operator is 'abort', which, as its name suggests, -- simply aborts a given sub-continuation. For instance, the second example -- above can be written: -- -- > pushPrompt p ((1:) `liftM` (2:) `liftM` abort p (return [])) -- -- The rest of the functions reify the sub-continuation into a function, -- so that it can be used. The shift/control operators all have similar -- effects in this regard, but differ as to where they delimit various -- parts of the resulting computation. Some names may help a bit for the -- following explanation, so consider: -- -- > shift p (\f -> e) -- -- /p/ is, obviously, the prompt; /f/ is the reified continuation, and /e/ -- is the computation that will be run in the aborted context. With these -- names in mind, the control operators work as follows: -- -- * 'shift' delimits both /e/ and every invocation of /f/. So, effectively, -- when using 'shift', control effects can never escape a delimiter, and -- computations of the form: -- -- > reset (\p -> <computations with shift p>) -- -- /look/ pure from the outside. -- -- * 'control' delimits /e/, but not the sub-continuation in /f/. Thus, if -- the sub-continuation contains other 'control' invocations, the effects -- may escape an enclosing delimiter. So, for example: -- -- > reset (\p -> shift p (\f -> (1:) `liftM` f (return [])) -- >>= \y -> shift p (\_ -> return y)) -- -- yields a value of [1], while replacing the 'shift's with 'control' -- yields a value of []. -- -- * 'shift0' delimits /f/, but not /e/. So: -- -- > reset (\p -> (1:) `liftM` pushPrompt p -- > (shift0 p (\_ -> shift0 p (\_ -> return [])))) -- -- yields [], whereas using 'shift' would yield [1]. -- -- * 'control0' delimits neither /e/ nor /f/, and is, in effect, the reified -- analogue to using withSubCont and pushSubCont directly. -- -- For a more complete and in-depth discussion of these four control operators, -- see /Shift to Control/, by Chung-chieh Shan. -- -- A small example program follows. It uses delimited continuations to reify a -- monadic loop into an iterator object. Saving references to old iterators -- allows one to effecively store references to various points in the traversal. -- Effectively, this is a simple, restricted case of a generalized zipper. -- -- > data Iterator r a = I a (CC r (Iterator r a)) | Done -- > -- > current :: Iterator r a -> Maybe a -- > current (I a _) = Just a -- > current Done = Nothing -- > -- > next :: Iterator r a -> CC r (Iterator r a) -- > next (I _ m) = m -- > next Done = return Done -- > -- > iterator :: ((a -> CC r ()) -> CC r ()) -> CC r (Iterator r a) -- > iterator loop = reset $ \p -> -- > loop (\a -> -- > shift p $ \k -> -- > return $ I a (k $ return ())) >> return Done -- > -- > test = do i <- iterator $ forM_ [1..5] -- > go [] i -- > where -- > go l Done = return l -- > go l i = do let (Just a) = current i -- > l' = replicate a a ++ l -- > i' <- next i -- > go l' i' -- -- The results are what one might expect from such an iterator object: -- -- > *Test> runCC test -- > [5,5,5,5,5,4,4,4,4,3,3,3,2,2,1]