monad-unlift: Typeclasses for representing monad transformer unlifting

monad-unlift

Typeclasses for providing for unlifting of monad transformers and stacks.

Note that concrete implementations of common transformers implementing these type classes are provided by the monad-unlift-ref package.

Synopsis

import Control.Concurrent.Async
import Control.Monad.Trans.Unlift
import Control.Monad.Trans.RWS.Ref
import Control.Monad.IO.Class
import Data.Mutable

-- Some artbirary data type for the MonadReader
data SomeEnv = SomeEnv Int

myFunc :: RWSRefT
    -- The WriterT piece is contained by an IORef
    IORef
    -- For efficiency, we store the state in a primitive reference
    (PRef RealWorld)
    SomeEnv   -- Reader
    [String]  -- Writer
    Int       -- State
    IO
    (String, String)
myFunc = do
    -- Get the unlift function. Due to weaknesses in ImpredicativeTypes, we
    -- need to use a newtype wrapper. You can also use askRunBase.
    --
    -- If you want to just unwrap one transformer layer, use
    -- askUnlift/askRun/Unlift.
    UnliftBase run <- askUnliftBase

    -- Note that we can use unlift to turn our transformer actions into IO
    -- actions. Unlike the standard RWST, actions from separate threads are
    -- both retained due to mutable references.
    --
    -- In real life: you shouldn't rely on this working, as RWST is not thread
    -- safe. This example is provided as a good demonstration of the type level
    -- functionality.
    liftIO $ concurrently (run foo) (run bar)
  where
    foo = do
        tell ["starting foo"]
        modify (+ 1)
        tell ["leaving foo"]
        return "foo is done"
    bar = do
        tell ["starting bar"]
        SomeEnv e <- ask
        modify (+ e)
        tell ["leaving bar"]
        return "bar is done"

main :: IO ()
main = do
    ((w, x), y, z) <- runRWSRefT myFunc (SomeEnv 5) 6
    print w -- foo is done
    print x -- bar is done
    print y -- 12 = 6 + 5 + 1
    print z -- starting and leaving statements, order ambiguous

Overview

A common pattern is to have some kind of a monad transformer, and want to pass an action into a function that requires actions in a base monad. That sounds a bit abstract, so let's give a concrete example:

-- From async
concurrently :: IO a -> IO b -> IO (a, b)

func1 :: ReaderT Foo IO String
func2 :: ReaderT Foo IO Double

doBoth :: ReaderT Foo IO (String, Double)
doBoth = _

Doing this manually is possible, but a bit tedious:

doBoth :: ReaderT Foo IO (String, Double)
doBoth = ReaderT $ \foo -> concurrently
    (runReaderT func1 foo)
    (runReaderT func2 foo)

This also doesn't generalize at all; you'll be stuck writing concurrently variants for every monad transformer stack. Fortunately, the monad-control package generalizes this to a large number of transformer stacks. Let's implement our generalized concurrently:

concurrentlyG :: MonadBaseControl IO m
              => m a -> m b -> m (StM m a, StM m b)
concurrentlyG f g = liftBaseWith $ \run ->
    concurrently (run f) (run g)

Notice how, in the signature for concurrentlyG, we no longer return (a, b), but (StM m a, StM m b). This is because there may be additional monadic context for each thread of execution, and we have no way of merging these together in general. Some examples of context are:

• With WriterT, it's the values that you called tell on
• With EitherT, the returned value may not exist at all

In addition to this difficulty, many people find the types in monad-control difficult to navigate, due to their extreme generality (which is in fact the power of that package!).

There is a subset of these transformer stacks that are in fact monad morphisms. Simply stated, these are transformer stacks that are isomorphic to ReaderT. For these monads, there is not context in the returned value. Therefore, there's no need to combine returned states or deal with possibly missing values.

This concept is represented by the monad-unlift package, which provides a pair of typeclasses for these kinds of transformer stacks. Before we dive in, let's see how we solve our concurrentlyG problem with it:

concurrentlyG :: MonadBaseUnlift IO m
              => m a -> m b -> m (a, b)
concurrentlyG f g = do
    UnliftBase run <- askUnliftBase
    liftBase $ concurrently (run f) (run g)

Notice how we get (a, b) in the return type as desired. There's no need to unwrap values or deal with context.

MonadTransUnlift

MonadTransUnlift is a class for any monad transformer which is isomorphic to ReaderT, in the sense that the environment can be captured and applied later. Some interesting cases in this space are:

• IdentityT and things isomorphic to it; in this case, you can think of the environment as being ()
• Transformers which contain a mutable reference in their environment. This allows them to behave like stateful transformers (e.g., StateT or WriterT), but still obey the monad morphism laws. (See below for more details.)

Due to weaknesses in GHC's ImpredicativeTypes, we have a helper datatype to allow for getting polymorphic unlift functions, appropriately named Unlift. For many common cases, you can get away with using askRun instead, e.g.:

bar :: ReaderT Foo IO ()

baz :: ReaderT Foo IO ()
baz = do
    run <- askRun
    liftIO $ void $ forkIO $ run bar

Using Unlift, this would instead be:

    Unlift run <- askUnlift
    liftIO $ void $ forkIO $ run bar

or equivalently:

    u <- askUnlift
    liftIO $ void $ forkIO $ unlift u bar

MonadBaseUnlift

MonadBaseUnlift extends this concept to entire transformer stacks. This is typically the typeclass that people end up using. You can think of these two typeclasses in exactly the same way as MonadTrans and MonadIO, or more precisely MonadTrans and MonadBase.

For the same ImpredicativeTypes reason, there's a helper type UnliftBase. Everything we just discussed should transfer directly to MonadBaseUnlift, so learning something new isn't necessary. For example, you can rewrite the last snippet as:

    u <- askUnliftBase
    liftIO $ void $ forkIO $ unliftBase u bar

Reference transformers

When playing transformer stack games with a transformer like StateT, it's common to accidentally discard state modifications. Additionally, in the case of runtime exceptions, it's usually impossible to retain the state. (Similar statements apply to WriterT and RWST, both in strict and lazy variants.)

Another approach is to use a ReaderT and hold onto a mutable reference. This is problematic since there's no built in support for operations like get, put, or tell. What we want is to have a MonadState and/or MonadWriter instance.

To address this case, this package includes variants of those transformers that use mutable references. These references are generic using the mutable-containers package, which allows you to have highly efficient references like PRef instead of always using boxed references like IORef.

Note that, for generality, the reference transformers take type parameters indicating which mutable reference type to use. Some examples you may use are:

• IORef for boxed references in IO
• STRef s for boxed references in ST
• PRef RealWorld for an unboxed reference in IO

See the synopsis for a complete example.

conduit

The transPipe function in conduit has caused confusion in the past due to its requirement of provided functions to obey monad morphism laws. This package makes a good companion to conduit to simplify that function's usage.

Other notable instances

Both the HandlerT transformer from yesod-core and LoggingT/NoLoggingT are valid monad morphisms. HandlerT is in fact my first example of using the "environment holding a mutable reference" technique to overcome exceptions destroying state.

{-# LANGUAGE FlexibleContexts  #-}
{-# LANGUAGE OverloadedStrings #-}
{-# LANGUAGE TemplateHaskell   #-}
import Control.Concurrent.Async
import Control.Monad.IO.Class
import Control.Monad.Logger
import Control.Monad.Trans.Unlift

main :: IO ()
main = runStdoutLoggingT foo

foo :: (MonadLogger m, MonadBaseUnlift IO m, MonadIO m) => m ()
foo = do
    run <- askRunBase
    a <- liftIO $ async $ run $ $logDebug "Hello World!"
    liftIO $ wait a