How exactly ExceptT differs to EarlyDo

After publishing my early GHC plugin, I saw a lot of response from people saying “Why not ExceptT?”, despite having outlined it in the README of the project. After reading some sincere responses failing to see why I bothered with this at all, it seems I haven’t explained myself well.

I’m happy for people to just disagree because they have different values, but I want to make sure we’re talking about the same trade-offs.

I think people need to see real code to see why I think this is a substantial improvement. You’ll have to use your imagination generously to raise the contrived to the “real world”.

(I’m calling this plugin EarlyDo so I can talk specifically about the syntax extension.)

Setup

Here is a module of three functions (but use your imagination to make a couple dozen), where the IO actions return Either failure a. The proposal I have been repeatedly given is to use ExceptT. I’m typically using RIO, so I define a fake one, and then use a real one later.

Basic ExceptT

To do concurrency or exception handling or any with-like thing, you need the equivalent of MonadUnliftIO to capture the current context, and then re-run the monad further in the new thread. Here we do it manually to demonstrate what’s happening.

import Control.Concurrent.Async
import Control.Monad.Trans.Except
import Control.Monad.Trans.Reader
import Control.Monad.Trans

type RIO r a = ReaderT r IO a

foo :: Int -> RIO () (Either () ())
foo i = runExceptT $ do
  _ <- ExceptT $ bar (i + 1)
  ExceptT $ zot (i + 2)

bar :: Int -> RIO () (Either () ((), ()))
bar i =
  runExceptT $ do
    ExceptT $ zot 0
    r <- ExceptT $ fmap Right ask
    (x, y) <-
      ExceptT $ lift $
      fmap
        (\(x, y) -> (,) <$> x <*> y)
        (concurrently
         (flip runReaderT r $ foo (i + 1))
         (flip runReaderT r $ zot (i + 2)))
    pure (x, y)

zot :: Int -> RIO () (Either () ())
zot i = runExceptT $ do
  _ <- ExceptT $ bar (i + 1)
  ExceptT $ zot (i + 2)

(Why am I not using ExceptT () (RIO ()) .. in my type signatures? Continue reading to the end.)

Using MonadUnliftIO

Here’s a version using MonadUnliftIO. Not that the work of unlifting has been done for me.

import           Control.Monad.Trans.Except
import           RIO (RIO)
import           UnliftIO

foo :: Int -> RIO () (Either () ())
foo i = runExceptT $ do
  _ <- ExceptT $ bar (i + 1)
  ExceptT $ zot (i + 2)

bar :: Int -> RIO () (Either () ((), ()))
bar i =
  runExceptT $ do
    ExceptT $ zot 0
    (x, y) <-
      ExceptT $
      fmap
        (\(x, y) -> (,) <$> x <*> y)
        (concurrently
         (foo (i + 1))
         (zot (i + 2)))
    pure (x, y)

zot :: Int -> RIO () (Either () ())
zot i = runExceptT $ do
  _ <- ExceptT $ bar (i + 1)
  ExceptT $ zot (i + 2)

Evaluating cost and ergonomics

What remains is the runExceptT and ExceptT calls which cannot be eliminated. Here we are unsatisfied, because all the functions in our module are using this “pattern”. A pattern is what we call repeating yourself because your language can’t abstract it for you. We pay an O(1) cost per definition, and an O(n) cost for calls per definition.

Additionally, any traverse call that is just MonadIO m => .. in the middle of the function, or logging function, will also run in ExceptT, causing a performance penalty that we didn’t want. If we want to avoid that penalty, we must use lift. But I haven’t benchmarked this. You are welcome to do so.

You pay to short-circuit and you pay to do normal actions in the base monad, and if you’re not careful you may accidentally pay extra.

Given that pretty much all of my apps have monads that look something like this, this seemed worth improving.1

EarlyDo

With EarlyDo, we have:

{-# OPTIONS -F -pgmF=early #-}
import Control.Early
import RIO (RIO)
import UnliftIO

foo :: Int -> RIO () (Either () ())
foo i = do
 _ <- bar (i + 1)?
 zot (i + 2)

bar :: Int -> RIO () (Either () ((), ()))
bar i = do
  zot 0?
  (x, y) <-
    fmap (\(x, y) -> (,) <$> x <*> y) (concurrently (foo (i + 1)) (zot (i + 2)))?
  pure (Right (x, y))

zot :: Int -> RIO () (Either () ())
zot i = do
  _ <- bar (i + 1)?
  zot (i + 2)

A single character that appears in a not-normally-valid syntax position signals short-circuiting. Any base monad action like mapRIO (runs in RIO r) can be used without lifting or unlifting. I have to explicitly produce Right in bar, which is a cost, but not in zot/foo.

Using our own monad type

Finally, we can use our own wrapper monad:

{-# OPTIONS -F -pgmF=early #-}
{-# LANGUAGE GeneralizedNewtypeDeriving #-}
import Control.Early
import RIO (RIO)
import UnliftIO

data AppEnv
newtype App a = App { runApp :: RIO AppEnv a} deriving (Functor, Applicative, Monad, MonadIO)
instance MonadUnliftIO App where withRunInIO = wrappedWithRunInIO App runApp

foo :: Int -> App (Either () ())
foo i = do
  _ <- bar (i + 1)?
  zot (i + 2)

bar :: Int -> App (Either () ((), ()))
bar i = do
  zot 0?
  (x, y) <-
    fmap (\(x, y) -> (,) <$> x <*> y) (concurrently (foo (i + 1)) (zot (i + 2)))?
  pure (Right (x, y))

zot :: Int -> App (Either () ())
zot i = do
  _ <- bar (i + 1)?
  zot (i + 2)

The App monad has the full power of all the previously mentioned things.

But not with ExceptT. I cannot put ExceptT in this monad stack and retain the MonadUnliftIO instance. It has no valid instance!

But you should still use ExceptT because I like it

Could the mechanics of this syntax use ExceptT underneath? Sure, but then the whole do block would be under ExceptT, and I’d have to lift the rest of the do notation too (even if it’s not necessary, incurring extra cost for no reason). Even if I could tell which statements should be ExceptT’d or lift’d, it would be overkill.

At this level, it doesn’t matter much. All you need the syntax to produce is a case expression. The use of an Early class is unneccessary, it just makes it easy to use either Maybe or Either with the same syntax.


  1. If your answer is “just define really short names for runExceptT and ExceptT and lift”, perhaps you could go back to using >>= and >> instead of do, and ApplicativeDo, and RecursiveDo, and Arrows, and report your experience.

    I’ve never once used Arrow and I’ve used RecursiveDo as a novelty, but mfix was sufficient in the end. ApplicativeDo is a brittle extension, but that doesn’t stop me from using it.