Thank to all who replied. After giving up on continuations (which make everything but forward branches seductively easy), I settled on something with the following interface:

class MonadGoto lbl gm | gm -> lbl where liftG :: Monad m => m () -> gm m () label :: Monad m => gm m lbl gotoIf :: Monad m => m Bool -> lbl -> gm m ()

Note that, unlike a monad transformer, only ()-producing computations can be embedded. This limitation reflects the problem of def-use across basic-block boundaries. Of course, computations in the embedded monad can be done there, then lifted, e.g.:

;top <- label ; ... ; liftG (do v <- m1 ; ... ; m2 v ; ...) ; brPositive r3 top ; ...

I've implemented MonadGoto as

newtype GotoT m a = GotoT { unGotoT :: WriterT [GotoLbldT m ()] LblSupply a }

data GotoLbldT m a = Label !Lbl | GotoIf (m Bool) !Lbl | Inner (m a)

wrL = GotoT . tell . (:[]) instance MonadGoto Lbl GotoT where liftG = wrL . Inner label = GotoT $ lift fromSupply >>= \l -> tell [Label l] >> return l gotoIf p l = wrL $ GotoIf p l

where GotoLbldT gets converted to

data GotoLbldRunT m a = ThenLabel { before :: m (), next :: GotoLbldRunT m a, labelLbl :: !Lbl } | ThenGotoIf { before :: m (), next :: GotoLbldRunT m a , gotoPred :: m Bool, gotoLbl :: !Lbl } | Tail (m a)

which can then be run

type JumpArray m v = [(Lbl, m v)] runGM :: Monad m => JumpArray m v -> GotoLbldRunT m v -> (JumpArray m v, m v)

. My Verilog has delayed branches (because that's the easy thing to do.) So, I implemented a DelayedGotoT which provides for delayed branches by shuffling the "GotoLbldT"s and using a StateT to save the result of the gotoPred evaluation. "Erkok, Levent" writes:

You need a fairly recent version of Hugs to run this example, November 2002 release would do (try with -98). For ghc, you need the CVS version, or wait till the next release. As far as I know, no other implementation supports mdo, nor there are any plans for it.

For the moment I've chosen to do the mdo desugaring manually instead of moving to the CVS ghc (or exclusively using Hugs.) Will mdo be in the next release? (That's 5.04.2, I think -- though I've recently been confused about GHC releases.) The relevant revision to Lex.lhs is pretty recent and doesn't seem to be on ghc-5-04-branch. thanks, mike