{-# LANGUAGE GeneralizedNewtypeDeriving #-}

{-|

This module provides leak-free and referentially transparent
higher-order discrete signals.

-}

module FRP.Elerea.Simple
    (
    -- * The signal abstraction
      Signal
    , SignalGen
    -- * Embedding into I/O
    , start
    , external
    , externalMulti
    -- * Basic building blocks
    , delay
    , snapshot
    , generator
    , memo
    , until
    -- * Derived combinators
    , stateful
    , transfer
    , transfer2
    , transfer3
    , transfer4
    -- * Signals with side effects
    -- $effectful
    , execute
    , effectful
    , effectful1
    , effectful2
    , effectful3
    , effectful4
    -- * A longer example
    -- $example
    ) where

import Control.Applicative
import Control.Concurrent.MVar
import Control.Monad
import Control.Monad.Fix
import Data.IORef
import Data.Maybe
import Prelude hiding (until)
import System.Mem.Weak

-- | A signal represents a value changing over time.  It can be
-- thought of as a function of type @Nat -> a@, where the argument is
-- the sampling time, and the 'Monad' instance agrees with the
-- intuition (bind corresponds to extracting the current sample).
-- Signals and the values they carry are denoted the following way in
-- the documentation:
--
-- > s = <<s0 s1 s2 ...>>
--
-- This says that @s@ is a signal that reads @s0@ during the first
-- sampling, @s1@ during the second and so on.  You can also think of
-- @s@ as the following function:
--
-- > s t_sample = [s0,s1,s2,...] !! t_sample
--
-- Signals are constrained to be sampled sequentially, there is no
-- random access.  The only way to observe their output is through
-- 'start'.
newtype Signal a = S (IO a) deriving (Functor, Applicative, Monad)

-- | A dynamic set of actions to update a network without breaking
-- consistency.
type UpdatePool = [Weak (IO (), IO ())]

-- | A signal generator is the only source of stateful signals.  It
-- can be thought of as a function of type @Nat -> a@, where the
-- result is an arbitrary data structure that can potentially contain
-- new signals, and the argument is the creation time of these new
-- signals.  It exposes the 'MonadFix' interface, which makes it
-- possible to define signals in terms of each other.  The denotation
-- of signal generators happens to be the same as that of signals, but
-- this partly accidental (it does not hold in the other variants), so
-- we will use a separate notation for generators:
--
-- > g = <|g0 g1 g2 ...|>
--
-- Just like signals, generators behave as functions of time:
--
-- > g t_start = [g0,g1,g2,...] !! t_start
--
-- The conceptual difference between the two notions is that signals
-- are passed a sampling time, while generators expect a start time
-- that will be the creation time of all the freshly generated
-- signals in the resulting structure.
newtype SignalGen a = SG { unSG :: IORef UpdatePool -> IO a }

-- | The phases every signal goes through during a superstep.
data Phase a = Ready a | Updated a a

instance Functor SignalGen where
    fmap = liftM

instance Applicative SignalGen where
    pure = return
    (<*>) = ap

instance Monad SignalGen where
    return x = SG $ \_ -> return x
    SG g >>= f = SG $ \p -> g p >>= \x -> unSG (f x) p

instance MonadFix SignalGen where
    mfix f = SG $ \p -> mfix $ \x -> unSG (f x) p

-- | Embedding a signal into an 'IO' environment.  Repeated calls to
-- the computation returned cause the whole network to be updated, and
-- the current sample of the top-level signal is produced as a
-- result. This is the only way to extract a signal generator outside
-- the network, and it is equivalent to passing zero to the function
-- representing the generator.  In general:
--
-- > replicateM n =<< start <|<<x0 x1 x2 x3 ...>> ...|> == take n [x0,x1,x2,x3,...]
--
-- Example:
--
-- > do
-- >     smp <- start (stateful 3 (+2))
-- >     res <- replicateM 5 smp
-- >     print res
--
-- Output:
--
-- > [3,5,7,9,11]
start :: SignalGen (Signal a) -- ^ the generator of the top-level signal
      -> IO (IO a)            -- ^ the computation to sample the signal
start (SG gen) = do
    pool <- newIORef []
    S sample <- gen pool
    return $ do
        res <- sample
        superstep pool
        return res

-- | Performing the two-phase superstep.
superstep :: IORef UpdatePool -> IO ()
superstep pool = loop id []
  where
    deref ptr = (fmap.fmap) ((,) ptr) (deRefWeak ptr)
    loop getPtrs final = do
      (ptrs,acts) <- unzip.catMaybes <$> (mapM deref =<< readIORef pool)
      case acts of
          [] -> do
              sequence_ final
              writeIORef pool (getPtrs [])
          _  -> do
              writeIORef pool []
              mapM_ fst acts
              loop ((ptrs++) . getPtrs) (mapM_ snd acts : final)

-- | Auxiliary function used by all the primitives that create a
-- mutable variable.
addSignal :: (a -> IO a)      -- ^ sampling function
          -> (a -> IO ())     -- ^ aging function
          -> IORef (Phase a)  -- ^ the mutable variable behind the signal
          -> IORef UpdatePool -- ^ the pool of update actions
          -> IO (Signal a)    -- ^ the signal created
addSignal sample update ref pool = do
  let upd = readIORef ref >>= \v -> case v of
              Ready x -> update x
              _       -> return ()

      fin = readIORef ref >>= \v -> case v of
              Updated x _ -> writeIORef ref $! Ready x
              _           -> error "Signal not updated!"

      sig = S $ readIORef ref >>= \v -> case v of
              Ready x     -> sample x
              Updated _ x -> return x

  updateActions <- mkWeak sig (upd,fin) Nothing
  modifyIORef pool (updateActions:)
  return sig

-- | The 'delay' combinator is the elementary building block for
-- adding state to the signal network by constructing delayed versions
-- of a signal that emit a given value at creation time and the
-- previous output of the signal afterwards (@--@ is undefined):
--
-- > delay x0 s = <| <<x0 s0 s1 s2 s3 ...>>
-- >                 <<-- x0 s1 s2 s3 ...>>
-- >                 <<-- -- x0 s2 s3 ...>>
-- >                 <<-- -- -- x0 s3 ...>>
-- >                 ...
-- >              |>
--
-- It can be thought of as the following function (which should also
-- make it clear why the return value is 'SignalGen'):
--
-- > delay x0 s t_start t_sample
-- >   | t_start == t_sample = x0
-- >   | t_start < t_sample  = s (t_sample-1)
-- >   | otherwise           = error \"Premature sample!\"
--
-- The way signal generators are extracted by 'generator' ensures that
-- the error can never happen.
--
-- Example (requires the @DoRec@ extension):
--
-- > do
-- >     smp <- start $ do
-- >         rec let fib'' = liftA2 (+) fib' fib
-- >             fib' <- delay 1 fib''
-- >             fib <- delay 1 fib'
-- >         return fib
-- >     res <- replicateM 7 smp
-- >     print res
--
-- Output:
--
-- > [1,1,2,3,5,8,13]
delay :: a                    -- ^ initial output at creation time
      -> Signal a             -- ^ the signal to delay
      -> SignalGen (Signal a) -- ^ the delayed signal
delay x0 (S s) = SG $ \pool -> do
  ref <- newIORef (Ready x0)

  let update x = s >>= \x' -> x' `seq` writeIORef ref (Updated x' x)

  addSignal return update ref pool

-- | A formal conversion from signals to signal generators, which
-- effectively allows for retrieving the current value of a previously
-- created signal within a generator.  This includes both signals
-- defined in an external scope as well as those created earlier in
-- the same generator.  In the model, it corresponds to the identity
-- function.
snapshot :: Signal a -> SignalGen a
snapshot (S s) = SG $ \_ -> s

-- | Auxiliary function.
memoise :: IORef (Phase a) -> a -> IO a
memoise ref x = writeIORef ref (Updated undefined x) >> return x

-- | A reactive signal that takes the value to output from a signal
-- generator carried by its input with the sampling time provided as
-- the start time for the generated structure.  It is possible to
-- create new signals in the monad, which is the key to defining
-- dynamic data-flow networks.
--
-- > generator << <|x00 x01 x02 ...|>
-- >              <|x10 x11 x12 ...|>
-- >              <|x20 x21 x22 ...|>
-- >              ...
-- >           >> = <| <<x00 x11 x22 ...>>
-- >                   <<x00 x11 x22 ...>>
-- >                   <<x00 x11 x22 ...>>
-- >                   ...
-- >                |>
--
-- It can be thought of as the following function:
--
-- > generator g t_start t_sample = g t_sample t_sample
--
-- It has to live in the 'SignalGen' monad, because it needs to
-- maintain an internal state to be able to cache the current sample
-- for efficiency reasons. However, this state is not carried between
-- samples, therefore start time doesn't matter and can be ignored.
--
-- Refer to the longer example at the bottom to see how it can be
-- used.
generator :: Signal (SignalGen a) -- ^ the signal of generators to run
          -> SignalGen (Signal a) -- ^ the signal of generated structures
generator (S s) = SG $ \pool -> do
  ref <- newIORef (Ready undefined)

  let sample = (s >>= \(SG g) -> g pool) >>= memoise ref

  addSignal (const sample) (const (() <$ sample)) ref pool

-- | Memoising combinator.  It can be used to cache results of
-- applicative combinators in case they are used in several places.
-- It is observationally equivalent to 'return' in the 'SignalGen'
-- monad.
--
-- > memo s = <|s s s s ...|>
--
-- For instance, if @s = f \<$\> s'@, then @f@ will be recalculated
-- once for each sampling of @s@.  This can be avoided by writing @s
-- \<- memo (f \<$\> s')@ instead.  However, 'memo' incurs a small
-- overhead, therefore it should not be used blindly.
--
-- All the functions defined in this module return memoised signals.
memo :: Signal a             -- ^ the signal to cache
     -> SignalGen (Signal a) -- ^ a signal observationally equivalent to the argument
memo (S s) = SG $ \pool -> do
  ref <- newIORef (Ready undefined)

  let sample = s >>= memoise ref

  addSignal (const sample) (const (() <$ sample)) ref pool

-- | A signal that is true exactly once: the first time the input
-- signal is true.  Afterwards, it is constantly false, and it holds
-- no reference to the input signal.  For instance (assuming the rest
-- of the input is constantly @False@):
--
-- > until <<False False True True False True ...>> =
-- >     <| <<False False True  False False False False False False False ...>>
-- >        << ---  False True  False False False False False False False ...>>
-- >        << ---   ---  True  False False False False False False False ...>>
-- >        << ---   ---   ---  True  False False False False False False ...>>
-- >        << ---   ---   ---   ---  False True  False False False False ...>>
-- >        << ---   ---   ---   ---   ---  True  False False False False ...>>
-- >        << ---   ---   ---   ---   ---   ---  False False False False ...>>
-- >        ...
-- >     |>
--
-- It is observationally equivalent to the following expression (which
-- would hold onto @s@ forever):
--
-- > until s = do
-- >     step <- transfer False (||) s
-- >     dstep <- delay False step
-- >     memo (liftA2 (/=) step dstep)
--
-- Example:
--
-- > do
-- >     smp <- start $ do
-- >         cnt <- stateful 0 (+1)
-- >         tick <- until ((>=3) <$> cnt)
-- >         return $ liftA2 (,) cnt tick
-- >     res <- replicateM 6 smp
-- >     print res
--
-- Output:
--
-- > [(0,False),(1,False),(2,False),(3,True),(4,False),(5,False)]
until :: Signal Bool             -- ^ the boolean input signal
      -> SignalGen (Signal Bool) -- ^ a one-shot signal true only the first time the input is true
until (S s) = SG $ \pool -> do
  ref <- newIORef (Ready undefined)

  rsmp <- mfix $ \rs -> newIORef $ do
    x <- s
    writeIORef ref (Updated undefined x)
    when x $ writeIORef rs $ do
      writeIORef ref (Updated undefined False)
      return False
    return x

  let sample = join (readIORef rsmp)

  addSignal (const sample) (const (() <$ sample)) ref pool

-- | A signal that can be directly fed through the sink function
-- returned.  This can be used to attach the network to the outer
-- world.  The signal always yields the value last written to the
-- sink.  In other words, if the sink is written less frequently than
-- the network sampled, the output remains the same during several
-- samples.  If values are pushed in the sink more frequently, only
-- the last one before sampling is visible on the output.
--
-- Example:
--
-- > do
-- >     (sig,snk) <- external 4
-- >     smp <- start (return sig)
-- >     r1 <- smp
-- >     r2 <- smp
-- >     snk 7
-- >     r3 <- smp
-- >     snk 9
-- >     snk 2
-- >     r4 <- smp
-- >     print [r1,r2,r3,r4]
--
-- Output:
--
-- > [4,4,7,2]
external :: a                         -- ^ initial value
         -> IO (Signal a, a -> IO ()) -- ^ the signal and an IO function to feed it
external x = do
  ref <- newIORef x
  return (S (readIORef ref), writeIORef ref)

-- | An event-like signal that can be fed through the sink function
-- returned.  The signal carries a list of values fed in since the
-- last sampling, i.e. it is constantly @[]@ if the sink is never
-- invoked.  The order of elements is reversed, so the last value
-- passed to the sink is the head of the list.  Note that unlike
-- 'external' this function only returns a generator to be used within
-- the expression constructing the top-level stream, and this
-- generator can only be used once.
--
-- Example:
--
-- > do
-- >     (gen,snk) <- externalMulti
-- >     smp <- start gen
-- >     r1 <- smp
-- >     snk 7
-- >     r2 <- smp
-- >     r3 <- smp
-- >     snk 9
-- >     snk 2
-- >     r4 <- smp
-- >     print [r1,r2,r3,r4]
--
-- Output:
--
-- > [[],[7],[],[2,9]]
externalMulti :: IO (SignalGen (Signal [a]), a -> IO ()) -- ^ a generator for the event signal and the associated sink
externalMulti = do
  var <- newMVar []
  return (SG $ \pool -> do
             ref <- newIORef (Ready undefined)
             let sample = modifyMVar var $ \list -> memoise ref list >> return ([], list)
             addSignal (const sample) (const (() <$ sample)) ref pool
         ,\val -> do vals <- takeMVar var
                     putMVar var (val:vals)
         )

-- | A pure stateful signal.  The initial state is the first output,
-- and every subsequent state is derived from the preceding one by
-- applying a pure transformation.
--
-- Example:
--
-- > do
-- >     smp <- start (stateful "x" ('x':))
-- >     res <- replicateM 5 smp
-- >     print res
--
-- Output:
--
-- > ["x","xx","xxx","xxxx","xxxxx"]
stateful :: a                    -- ^ initial state
         -> (a -> a)             -- ^ state transformation
         -> SignalGen (Signal a)
stateful x0 f = mfix $ \sig -> delay x0 (f <$> sig)

-- | A stateful transfer function.  The current input affects the
-- current output, i.e. the initial state given in the first argument
-- is considered to appear before the first output, and can never be
-- observed, and subsequent states are determined by combining the
-- preceding state with the current output of the input signal using
-- the function supplied.
--
-- Example:
--
-- > do
-- >     smp <- start $ do
-- >         cnt <- stateful 1 (+1)
-- >         transfer 10 (+) cnt
-- >     res <- replicateM 5 smp
-- >     print res
--
-- Output:
--
-- > [11,13,16,20,25]
transfer :: a                    -- ^ initial internal state
         -> (t -> a -> a)        -- ^ state updater function
         -> Signal t             -- ^ input signal
         -> SignalGen (Signal a)
transfer x0 f s = mfix $ \sig -> do
    sig' <- delay x0 sig
    memo (liftA2 f s sig')

-- | A variation of 'transfer' with two input signals.
transfer2 :: a                     -- ^ initial internal state
          -> (t1 -> t2 -> a -> a)  -- ^ state updater function
          -> Signal t1             -- ^ input signal 1
          -> Signal t2             -- ^ input signal 2
          -> SignalGen (Signal a)
transfer2 x0 f s1 s2 = mfix $ \sig -> do
    sig' <- delay x0 sig
    memo (liftA3 f s1 s2 sig')

-- | A variation of 'transfer' with three input signals.
transfer3 :: a                           -- ^ initial internal state
          -> (t1 -> t2 -> t3 -> a -> a)  -- ^ state updater function
          -> Signal t1                   -- ^ input signal 1
          -> Signal t2                   -- ^ input signal 2
          -> Signal t3                   -- ^ input signal 3
          -> SignalGen (Signal a)
transfer3 x0 f s1 s2 s3 = mfix $ \sig -> do
    sig' <- delay x0 sig
    memo (liftM4 f s1 s2 s3 sig')

-- | A variation of 'transfer' with four input signals.
transfer4 :: a                                 -- ^ initial internal state
          -> (t1 -> t2 -> t3 -> t4 -> a -> a)  -- ^ state updater function
          -> Signal t1                         -- ^ input signal 1
          -> Signal t2                         -- ^ input signal 2
          -> Signal t3                         -- ^ input signal 3
          -> Signal t4                         -- ^ input signal 4
          -> SignalGen (Signal a)
transfer4 x0 f s1 s2 s3 s4 = mfix $ \sig -> do
    sig' <- delay x0 sig
    memo (liftM5 f s1 s2 s3 s4 sig')

{- $effectful

The following combinators are primarily aimed at library implementors
who wish build abstractions to effectful libraries on top of Elerea.

-}

-- | An IO action executed in the 'SignalGen' monad. Can be used as
-- `liftIO`.
execute :: IO a -> SignalGen a
execute act = SG $ \_ -> act

-- | A signal that executes a given IO action once at every sampling.
--
-- In essence, this combinator provides cooperative multitasking
-- capabilities, and its primary purpose is to assist library writers
-- in wrapping effectful APIs as conceptually pure signals.  If there
-- are several effectful signals in the system, their order of
-- execution is undefined and should not be relied on.
--
-- Example:
--
-- > do
-- >     smp <- start $ do
-- >         ref <- execute $ newIORef 0
-- >         effectful $ do
-- >             x <- readIORef ref
-- >             putStrLn $ "Count: " ++ show x
-- >             writeIORef ref $! x+1
-- >             return ()
-- >     replicateM_ 5 smp
--
-- Output:
--
-- > Count: 0
-- > Count: 1
-- > Count: 2
-- > Count: 3
-- > Count: 4
--
-- Another example (requires mersenne-random):
--
-- > do
-- >     smp <- start $ effectful randomIO :: IO (IO Double)
-- >     res <- replicateM 5 smp
-- >     print res
--
-- Output:
--
-- > [0.12067753390401374,0.8658877349182655,0.7159264443196786,0.1756941896012891,0.9513646060896676]
effectful :: IO a                 -- ^ the action to be executed repeatedly
          -> SignalGen (Signal a)
effectful act = SG $ \pool -> do
  ref <- newIORef (Ready undefined)

  let sample = act >>= memoise ref

  addSignal (const sample) (const (() <$ sample)) ref pool

-- | A signal that executes a parametric IO action once at every
-- sampling.  The parameter is supplied by another signal at every
-- sampling step.
effectful1 :: (t -> IO a)          -- ^ the action to be executed repeatedly
           -> Signal t             -- ^ parameter signal
           -> SignalGen (Signal a)
effectful1 act (S s) = SG $ \pool -> do
  ref <- newIORef (Ready undefined)

  let sample = s >>= act >>= memoise ref

  addSignal (const sample) (const (() <$ sample)) ref pool

-- | Like 'effectful1', but with two parameter signals.
effectful2 :: (t1 -> t2 -> IO a)   -- ^ the action to be executed repeatedly
           -> Signal t1            -- ^ parameter signal 1
           -> Signal t2            -- ^ parameter signal 2
           -> SignalGen (Signal a)
effectful2 act (S s1) (S s2) = SG $ \pool -> do
  ref <- newIORef (Ready undefined)

  let sample = join (liftM2 act s1 s2) >>= memoise ref

  addSignal (const sample) (const (() <$ sample)) ref pool

-- | Like 'effectful1', but with three parameter signals.
effectful3 :: (t1 -> t2 -> t3 -> IO a) -- ^ the action to be executed repeatedly
           -> Signal t1                -- ^ parameter signal 1
           -> Signal t2                -- ^ parameter signal 2
           -> Signal t3                -- ^ parameter signal 3
           -> SignalGen (Signal a)
effectful3 act (S s1) (S s2) (S s3) = SG $ \pool -> do
  ref <- newIORef (Ready undefined)

  let sample = join (liftM3 act s1 s2 s3) >>= memoise ref

  addSignal (const sample) (const (() <$ sample)) ref pool

-- | Like 'effectful1', but with four parameter signals.
effectful4 :: (t1 -> t2 -> t3 -> t4 -> IO a) -- ^ the action to be executed repeatedly
           -> Signal t1                      -- ^ parameter signal 1
           -> Signal t2                      -- ^ parameter signal 2
           -> Signal t3                      -- ^ parameter signal 3
           -> Signal t4                      -- ^ parameter signal 4
           -> SignalGen (Signal a)
effectful4 act (S s1) (S s2) (S s3) (S s4) = SG $ \pool -> do
  ref <- newIORef (Ready undefined)

  let sample = join (liftM4 act s1 s2 s3 s4) >>= memoise ref

  addSignal (const sample) (const (() <$ sample)) ref pool

-- | The Show instance is only defined for the sake of Num...
instance Show (Signal a) where
  showsPrec _ _ s = "<SIGNAL>" ++ s

-- | Equality test is impossible.
instance Eq (Signal a) where
  _ == _ = False

-- | Error message for unimplemented instance functions.
unimp :: String -> a
unimp = error . ("Signal: "++)

instance Ord t => Ord (Signal t) where
  compare = unimp "compare"
  min = liftA2 min
  max = liftA2 max

instance Enum t => Enum (Signal t) where
  succ = fmap succ
  pred = fmap pred
  toEnum = pure . toEnum
  fromEnum = unimp "fromEnum"
  enumFrom = unimp "enumFrom"
  enumFromThen = unimp "enumFromThen"
  enumFromTo = unimp "enumFromTo"
  enumFromThenTo = unimp "enumFromThenTo"

instance Bounded t => Bounded (Signal t) where
  minBound = pure minBound
  maxBound = pure maxBound

instance Num t => Num (Signal t) where
  (+) = liftA2 (+)
  (-) = liftA2 (-)
  (*) = liftA2 (*)
  signum = fmap signum
  abs = fmap abs
  negate = fmap negate
  fromInteger = pure . fromInteger

instance Real t => Real (Signal t) where
  toRational = unimp "toRational"

instance Integral t => Integral (Signal t) where
  quot = liftA2 quot
  rem = liftA2 rem
  div = liftA2 div
  mod = liftA2 mod
  quotRem a b = (fst <$> qrab,snd <$> qrab)
    where qrab = quotRem <$> a <*> b
  divMod a b = (fst <$> dmab,snd <$> dmab)
    where dmab = divMod <$> a <*> b
  toInteger = unimp "toInteger"

instance Fractional t => Fractional (Signal t) where
  (/) = liftA2 (/)
  recip = fmap recip
  fromRational = pure . fromRational

instance Floating t => Floating (Signal t) where
  pi = pure pi
  exp = fmap exp
  sqrt = fmap sqrt
  log = fmap log
  (**) = liftA2 (**)
  logBase = liftA2 logBase
  sin = fmap sin
  tan = fmap tan
  cos = fmap cos
  asin = fmap asin
  atan = fmap atan
  acos = fmap acos
  sinh = fmap sinh
  tanh = fmap tanh
  cosh = fmap cosh
  asinh = fmap asinh
  atanh = fmap atanh
  acosh = fmap acosh

{- $example

For a not entirely trivial example, let's create a dynamic collection
of countdown timers, where each expired timer is removed from the
collection.  First of all, we'll need a simple tester function:

@
 sigtest gen = 'replicateM' 15 '=<<' 'start' gen
@

We can try it with a trivial example:

@
 \> sigtest $ 'stateful' 2 (+3)
 [2,5,8,11,14,17,20,23,26,29,32,35,38,41,44,47]
@

Our first definition will be a signal representing a simple named
timer:

@
 countdown :: String -\> Int -\> SignalGen (Signal (String,Maybe Int))
 countdown name t = do
   let tick prev = do { t \<- prev ; 'guard' (t \> 0) ; 'return' (t-1) }
   timer \<- 'stateful' (Just t) tick
   'return' ((,) name '<$>' timer)
@

Let's see if it works:

@
 \> sigtest $ countdown \"foo\" 4
 [(\"foo\",Just 4),(\"foo\",Just 3),(\"foo\",Just 2),(\"foo\",Just 1),(\"foo\",Just 0),
  (\"foo\",Nothing),(\"foo\",Nothing),(\"foo\",Nothing),...]
@

Next, we will define a timer source that takes a list of timer names,
starting values and start times and creates a signal that delivers the
list of new timers at every point:

@
 timerSource :: [(String, Int, Int)] -\> SignalGen (Signal [Signal (String, Maybe Int)])
 timerSource ts = do
   let gen t = 'mapM' ('uncurry' countdown) newTimers
           where newTimers = [(n,v) | (n,v,st) \<- ts, st == t]
   cnt \<- 'stateful' 0 (+1)
   'generator' (gen '<$>' cnt)
@

Now we need to encapsulate the timer source signal in another signal
expression that takes care of maintaining the list of live timers.
Since working with dynamic collections is a recurring task, let's
define a generic combinator that maintains a dynamic list of signals
given a source and a test that tells from the output of each signal
whether it should be kept.  We can use @mdo@ expressions (a variant of
@do@ expressions allowing forward references) as syntactic sugar for
'mfix' to make life easier:

@
 collection :: Signal [Signal a] -\> (a -\> Bool) -\> SignalGen (Signal [a])
 collection source isAlive = mdo
   sig \<- 'delay' [] ('map' 'snd' '<$>' collWithVals')
   coll \<- 'memo' ('liftA2' (++) source sig)
   let collWithVals = 'zip' '<$>' ('sequence' '=<<' coll) '<*>' coll
   collWithVals' \<- 'memo' ('filter' (isAlive . 'fst') '<$>' collWithVals)
   'return' $ 'map' 'fst' '<$>' collWithVals'
@

We need recursion to define the @coll@ signal as a delayed version of
its continuation, which does not contain signals that need to be
removed in the current sample.  At every point of time the running
collection is concatenated with the source.  We define @collWithVals@,
which simply pairs up every signal with its current output.  The
output is obtained by extracting the current value of the signal
container and sampling each element with 'sequence'.  We can then
derive @collWithVals'@, which contains only the signals that must be
kept for the next round along with their output.  Both @coll@ and
@collWithVals'@ have to be memoised, because they are used more than
once (the program would work without that, but it would recalculate
both signals each time they are used).  By throwing out the respective
parts, we can get both the final output and the collection for the
next step (@coll'@).

Now we can easily finish the original task:

@
 timers :: [(String, Int, Int)] -\> SignalGen (Signal [(String, Int)])
 timers timerData = do
   src \<- timerSource timerData
   getOutput '<$>' collection src ('isJust' . 'snd')
     where getOutput = 'fmap' ('map' (\\(name,Just val) -> (name,val)))
@

As a test, we can start four timers: /a/ at t=0 with value 3, /b/ and
/c/ at t=1 with values 5 and 3, and /d/ at t=3 with value 4:

@
 \> sigtest $ timers [(\"a\",3,0),(\"b\",5,1),(\"c\",3,1),(\"d\",4,3)]
 [[(\"a\",3)],[(\"b\",5),(\"c\",3),(\"a\",2)],[(\"b\",4),(\"c\",2),(\"a\",1)],
  [(\"d\",4),(\"b\",3),(\"c\",1),(\"a\",0)],[(\"d\",3),(\"b\",2),(\"c\",0)],
  [(\"d\",2),(\"b\",1)],[(\"d\",1),(\"b\",0)],[(\"d\",0)],[],[],[],[],[],[],[]]
@

If the noise of the applicative lifting operators feels annoying, she
(<http://personal.cis.strath.ac.uk/~conor/pub/she/>) comes to the
save.  Among other features it provides idiom brackets, which can
substitute the explicit lifting.  For instance, it allows us to define
@collection@ this way:

@
 collection :: Stream [Stream a] -> (a -> Bool) -> StreamGen (Stream [a])
 collection source isAlive = mdo
   sig \<- 'delay' [] (|'map' ~'snd' collWithVals'|)
   coll \<- 'memo' (|source ++ sig|)
   collWithVals' \<- 'memo' (|'filter' ~(isAlive . 'fst') (|'zip' ('sequence' '=<<' coll) coll|)|)
   'return' (|'map' ~'fst' collWithVals'|)
@

-}