开发者

Haskell: How to write interactive interpreter on top of a State monad?

We're working on a model filesystem that uses a state monad internally. We have a type class with operations like these:

class Monad m => FS m where
  isDirectory  :: Path -> m Bool
  children     :: Path -> m [Path]
  ...

We're working on a little interactive interpreter that will offer commands like cd, ls, cat, and so on. An operation in the interpreter can be written this way:

fsop :: FS m => Operation -> m Response

The definitions of Operation and Respons开发者_JS百科e aren't important; if you like, take them to be strings.

The problem I am trying to solve is to write a top-level loop in the I/O monad that interprets filesystem Operations and prints responses. If IO were an instance of FS (that is, if we were working directly with the IO monad), life would be simple: we could write

loop :: Path -> IO ()
loop currentDir = do
        op <- getLine
        case read op of
          ChangeDir d -> loop d -- should test 'isDirectory d', but let's not
          Ls -> do { files <- children currentDir
                   ; mapM_ putStrLn files
                   ; loop currentDir }
          Exit -> return ()

But that's not what I want. I want to use Control.Monad.State:

newtype Filesystem a = Filesystem (State (Data.Map.Map Path Contents) a)

and to declare

instance Monad Filesystem ...
instance FS Filesystem ...

Using the FS abstraction, I can write a single-step function that should work with any instance, and indeed the following code compiles:

step :: FS fs => Path -> Operation -> fs (Path, Response)
step currentDir op = 
        case op of
          ChangeDir d -> return (d, "")
          Ls -> do { files <- children currentDir
                   ; return (currentDir, unlines files) }

At this point I am totally stuck. What I want to do is write an interactive loop in the IO monad, which can read Operations and print Responses, but which works on a state monad that is not necessarily IO. (One of the reasons for having a model that is not in IO is that so we can test QuickCheck properties.)

I feel like this has to be a standard problem—an interactive read-eval-print loop on top of a stateful abstraction that is not IO—but I must be missing something breathtakingly obvious because I can't seem to figure it out. I've looked online but have not been enlightened.

Any help writing an interactive, IO-performing computation that can call step would be greatly appreciated.


What about using monad transformers? They are more or less standard way to combine monads. Here an simplistic example:

type Foo a = StateT String IO a

replT :: Foo ()
replT = do
  str   <- liftIO getLine
  state <- get
  liftIO $ putStrLn ("current state: " ++ state)
  liftIO $ putStrLn ("setting state: " ++ str)
  put str
  replT

Below are results of running replT from within ghci.

*Main> runStateT replT "Initial state"
asd
current state: Initial state
setting state: asd
zxc
current state: asd
setting state: zxc
asdasd

There are three monad transformers libs. mtl, transformers and monadLib. I cannot recommend any of them since I don't use them much.


Disclaimer: I can't promise the following is a good way to go about it, but working through it sounds fun. Let's take it for a spin, shall we?.


A few obligatory imports

First, let's toss some data types out there. I'm going to fill in some details and tweak things a bit, in order to define a simple "file system" that we can actually interact with.

type Path = String
type Response = Maybe String
type Contents = [String]

data Operation = Cd Path 
               | Ls 
               | MkDir Path
               | Quit
    deriving (Read, Show)

Next, we'll do something a bit edgy... strip out all the monads. What? This is madness! Perhaps, but sometimes all the hidden plumbing that >>= provides hides things just a bit too much.

For the file system itself, we'll just store the current working directory and a map from paths to their children. We'll also need a handful of functions to interact with it.

data Filesystem = Filesystem { wd :: Path, files :: M.Map Path Contents }
    deriving Show

newFS = Filesystem "/" (M.singleton "/" [])

isDirectory p fs = M.member p $ files fs
children p fs = fromMaybe [] . M.lookup p $ files fs
cd p fs = fs { wd = p }
create p fs = let newPath = wd fs ++ p ++ "/"
                  addPath = M.insert newPath [] . M.adjust (p:) (wd fs)
              in (newPath, fs { files = addPath $ files fs })

Now for a monad-less version of the step function. It needs to take an Operation and a Filesystem, and return a Response and a (possibly modified) Filesystem:

step :: Operation -> Filesystem -> (Response, Filesystem)
step (Cd d) fs = (Just "Ok\n", cd d fs)
step (MkDir d) fs = first (\d -> Just $ "Created " ++ d ++ "\n") $ create d fs
step Ls fs = let files = children (wd fs) fs
             in (Just $ unlines files, fs)
step Quit fs = (Nothing, fs)

...hmm, that type signature already looks a lot like the guts of a State monad. Oh well, just ignore it for now, and charge blindly onward.

Now, what we want is a function that will provide a general-purpose interface to a Filesystem interpreter. Particularly, we want the interface to be at least somewhat self-contained so that whatever uses the interface doesn't have to step through manually, yet we want the interface to be sufficiently oblivious to the code using it that we can wire it up to the IO monad, some other Monad, or even no monad at all.

What this tells us primarily is that we'll need to interleave the external code with the interpreter in some fashion, rather than having either part be in control. Now, Haskell is a functional language, so that means that using lots of higher-order functions is good, right? Sounds plausible to me, so here's the strategy we'll use: If a function doesn't know what to do next, we'll hand it another function that we assume does. Repeat until everybody knows what's going on. A flawless plan, no?

The heart of it all is the step function, so we'll start by just calling that.

interp1 :: Operation -> Filesystem -> (Response, Filesystem)
interp1 op fs = step op fs

...well, it's a start. I guess. But wait, where is the Operation coming from? We need the external code to provide that, but we can't just ask for it without getting all mixed up with unsavory characters like IO. So we get another function to do the dirty work for us:

interp2 :: ((Operation -> (Response, Filesystem)) -> t) -> Filesystem -> t
interp2 inp fs = inp (\op -> step op fs)

Of course, now all we have is some stupid t that we don't even know what it is. We know it has to have a Response and a Filesystem in it somewhere, but we can't do anything with it, so we'll hand it back to another function, along with some instructions on how to proceed... which will of course involve passing in yet more functions. It's functions all the way down, you know.

interp3 :: ((Operation -> (Response, Filesystem)) -> a)
           -> (a -> ((Response, Filesystem) -> b) -> c)
           -> (Filesystem -> b)
           -> (String -> Filesystem -> b) 
           -> Filesystem 
           -> c
interp3 inp check done out fs = check (inp (\op -> step op fs)) test
    where test (Nothing, fs) = done fs
          test (Just s, fs)  = out s fs

...well that's pretty ugly. But don't worry, all is going according to plan. We can make a couple observations next:

  • The type a only exists between inp and check, so in hindsight, we might as well combine them ahead of time and just pass the composed function to the interpreter.
  • When we call done, it ought to mean exactly what it says on the tin. So the return type for done should be the same as the whole interpreter, meaning b and c ought to be the same type.

Now, if done ends the whole thing, what's out? As the name none-too-subtly implies, it's providing output to the external code, but where does it go after that? It needs to loop back into the interpreter somehow, and we might note that our interpreter is not yet recursive. The way forward is clear--the interpreter, like Jormungand, thus seizes its own tail; looping back around indefinitely till the interpretation finishes (or until Ragnarök, whichever comes first).

interp4 :: ((Operation -> (Response, Filesystem)) 
               -> ((Response, Filesystem) -> r) -> r)
           -> (Filesystem -> r)
           -> (String -> Filesystem -> (Filesystem -> r) -> r)
           -> Filesystem
           -> r
interp4 checkInp done out fs = checkInp (\op -> step op fs) test
    where loop = interp4 checkInp done out
          test (Nothing, fs) = done fs
          test (Just s, fs)  = out s fs loop

...oh, did I mention that it works now? No, seriously!

Here's some IO code to use the interface:

ioIn f k = putStr "> " >> (k . f =<< readLn)
ioDone fs = putStrLn "Done" >> return fs 
ioOut x fs k = putStr x >> k fs

ioInterp :: IO Filesystem
ioInterp = interp4 ioIn ioDone ioOut newFS

And here's code that runs a list of commands, producing a list of output strings:

scriptIn f k (x:xs) = k (f x) xs
scriptDone fs xs = ["Done\n"]
scriptOut r fs k xs = r : k fs xs

scriptInterp :: [Operation] -> [String]
scriptInterp = interp4 scriptIn scriptDone scriptOut newFS

Examples of running both in GHCi here, if just the code doesn't tickle your imagination sufficiently.


Well, that's that. Or is it? Frankly, that interpreter is code only a mother could love. Is there something that would tie it all together elegantly? Something to reveal the underlying structure of the code?

...okay, so it's pretty obvious where this leads. The overall design of functions tail-calling each other in circles looks an awful lot like continuation-passing style, and not once but twice in the interpreter's type signature can be found the characteristic pattern (foo -> r) -> r, better known as the continuation monad.

Unfortunately, even after all that, continuations make my head hurt and I'm not sure how best to disentangle the very ad-hoc structure of the interpreter into a computation running in a MonadCont.


I can think of two solutions here:

1) Use a monad transformer library. I can't improve on Shimuuar's reply, except in some details on the libraries. Transformers by itself doesn't provide the necessary instances; you would need to use transformers and either monads-tf or monads-fd, which offer implementations based on type families and fundeps, respectively. I prefer monads-tf if you go this route. The api is almost identical to that of mtl. I don't have experience with MonadLib, but it looks quite good also.

2) Write your main loop in IO, and for each loop iteration call runState to evaluate the state monad. Something like the following:

loop path state = do
  op <- readOp
  let ((newpath, resp), newstate) = runState (step path op) state
  print resp
  loop newpath newstate

This should work, but it's far less idiomatic than using monad transformers.


Require your instances of FS to be instance of MonadIO, not just Monad:

class MonadIO m => FS m where ...

Then you will have available the liftIO method to lift FS into IO:

liftIO :: MonadIO m => m a -> IO a

so you can write in the IO monad:

files <- liftIO $ children currentDir

etc. Of course, that means you will need to implement liftIO for each FS before you even write the FS instance, but for this application (without having seen the actual details) it sounds like that should be simple.

0

上一篇:

下一篇:

精彩评论

暂无评论...
验证码 换一张
取 消

最新问答

问答排行榜