kleene-0: src/Kleene/DFA.hs
{-# LANGUAGE BangPatterns #-}
{-# LANGUAGE FlexibleInstances #-}
{-# LANGUAGE FunctionalDependencies #-}
{-# LANGUAGE GADTs #-}
{-# LANGUAGE Safe #-}
{-# LANGUAGE ScopedTypeVariables #-}
module Kleene.DFA (
DFA (..),
-- * Conversions
fromRE,
toRE,
fromERE,
toERE,
fromTM,
fromTMEquiv,
toKleene,
) where
import Prelude ()
import Prelude.Compat
import Algebra.Lattice ((\/))
import Data.IntMap (IntMap)
import Data.IntSet (IntSet)
import Data.List (intercalate)
import Data.Map (Map)
import Data.Maybe (fromMaybe)
import Data.RangeSet.Map (RSet)
import qualified Data.Function.Step.Discrete.Closed as SF
import qualified Data.IntMap as IM
import qualified Data.IntSet as IS
import qualified Data.Map as Map
import qualified Data.MemoTrie as MT
import qualified Data.RangeSet.Map as RSet
import Kleene.Classes
import qualified Kleene.ERE as ERE
import Kleene.Internal.Pretty
import qualified Kleene.RE as RE
-- | Deterministic finite automaton.
--
-- A deterministic finite automaton (DFA) over an alphabet \(\Sigma\) (type
-- variable @c@) is 4-tuple \(Q\), \(q_0\) , \(F\), \(\delta\), where
--
-- * \(Q\) is a finite set of states (subset of 'Int'),
-- * \(q_0 \in Q\) is the distinguised start state (@0@),
-- * \(F \subset Q\) is a set of final (or accepting) states ('dfaAcceptable'), and
-- * \(\delta : Q \times \Sigma \to Q\) is a function called the state
-- transition function ('dfaTransition').
--
data DFA c = DFA
{ dfaTransition :: !(IntMap (SF.SF c Int))
-- ^ transition function
, dfaAcceptable :: !IntSet
-- ^ accept states
, dfaBlackholes :: !IntSet
-- ^ states we cannot escape
}
deriving Show
-------------------------------------------------------------------------------
-- Construction
-------------------------------------------------------------------------------
-- | Convert 'RE.RE' to 'DFA'.
--
-- >>> putPretty $ fromRE $ RE.star "abc"
-- 0+ -> \x -> if
-- | x <= '`' -> 3
-- | x <= 'a' -> 2
-- | otherwise -> 3
-- 1 -> \x -> if
-- | x <= 'b' -> 3
-- | x <= 'c' -> 0
-- | otherwise -> 3
-- 2 -> \x -> if
-- | x <= 'a' -> 3
-- | x <= 'b' -> 1
-- | otherwise -> 3
-- 3 -> \_ -> 3 -- black hole
--
-- Everything and nothing result in blackholes:
--
-- >>> traverse_ (putPretty . fromRE) [RE.empty, RE.star RE.anyChar]
-- 0 -> \_ -> 0 -- black hole
-- 0+ -> \_ -> 0 -- black hole
--
-- Character ranges are effecient:
--
-- >>> putPretty $ fromRE $ RE.charRange 'a' 'z'
-- 0 -> \x -> if
-- | x <= '`' -> 2
-- | x <= 'z' -> 1
-- | otherwise -> 2
-- 1+ -> \_ -> 2
-- 2 -> \_ -> 2 -- black hole
--
-- An example with two blackholes:
--
-- >>> putPretty $ fromRE $ "c" <> RE.star RE.anyChar
-- 0 -> \x -> if
-- | x <= 'b' -> 2
-- | x <= 'c' -> 1
-- | otherwise -> 2
-- 1+ -> \_ -> 1 -- black hole
-- 2 -> \_ -> 2 -- black hole
--
fromRE :: forall c. (Ord c, Enum c, Bounded c) => RE.RE c -> DFA c
fromRE = fromTM
-- | Convert 'ERE.ERE' to 'DFA'.
--
-- We don't always generate minimal automata:
--
-- >>> putPretty $ fromERE $ "a" /\ "b"
-- 0 -> \_ -> 1
-- 1 -> \_ -> 1 -- black hole
--
-- Compare this to an @complement@ example
--
-- Using 'fromTMEquiv', we can get minimal automaton, for the cost of higher
-- complexity (slow!).
--
-- >>> putPretty $ fromTMEquiv $ ("a" /\ "b" :: ERE.ERE Char)
-- 0 -> \_ -> 0 -- black hole
--
-- >>> putPretty $ fromERE $ complement $ star "abc"
-- 0 -> \x -> if
-- | x <= '`' -> 3
-- | x <= 'a' -> 2
-- | otherwise -> 3
-- 1+ -> \x -> if
-- | x <= 'b' -> 3
-- | x <= 'c' -> 0
-- | otherwise -> 3
-- 2+ -> \x -> if
-- | x <= 'a' -> 3
-- | x <= 'b' -> 1
-- | otherwise -> 3
-- 3+ -> \_ -> 3 -- black hole
--
fromERE :: forall c. (Ord c, Enum c, Bounded c) => ERE.ERE c -> DFA c
fromERE = fromTM
-- | Create from 'TransitionMap'.
--
-- See 'fromRE' for a specific example.
fromTM :: forall k c. (Ord k, Ord c, TransitionMap c k) => k -> DFA c
fromTM = fromTMImpl Nothing
-- | Create from 'TransitonMap' minimising states with 'Equivalent'.
--
-- See 'fromERE' for an example.
--
fromTMEquiv :: forall k c. (Ord k, Ord c, TransitionMap c k, Equivalent c k) => k -> DFA c
fromTMEquiv = fromTMImpl (Just equivalent)
fromTMImpl :: forall k c. (Ord k, Ord c, TransitionMap c k)
=> Maybe (k -> k -> Bool)
-> k
-> DFA c
fromTMImpl mequiv re = DFA
{ dfaTransition = transition
, dfaAcceptable = IS.fromList
[ i
| (re', i) <- Map.toList lookupMap
, nullable re'
]
, dfaBlackholes = blackholes
}
where
transition = IM.fromList
[ (i, js)
| (re', pm) <- Map.toList tm
, let i = fromMaybe 0 $ Map.lookup re' lookupMap
, let js = SF.normalise $ fmap (\re'' -> fromMaybe 0 $ Map.lookup re'' lookupMap) pm
]
blackholes = IS.fromList
[ i
| (i, sf) <- IM.toList transition
, sf == pure i
]
tm = transitionMap re
-- reversing makes error state go last, usually
lookupMap :: Map k Int
lookupMap = makeLookup 1 lookupMap' (reverse $ Map.toList $ Map.delete re tm)
lookupMap' :: Map k Int
lookupMap' = case Map.lookup re tm of
Nothing -> Map.empty
Just _ -> Map.singleton re 0
makeLookup :: Int -> Map k Int -> [(k, b)] -> Map k Int
makeLookup = maybe makeLookupEq makeLookupEquiv mequiv
makeLookupEq :: Int -> Map k Int -> [(k, b)] -> Map k Int
makeLookupEq !_ !acc [] = acc
makeLookupEq !n acc ((x, _) : xs) = makeLookup (n + 1) (Map.insert x n acc) xs
-- this differs from makeLookupEq. We don't insert new states right away,
-- but check whether equivalent state is already in the map.
--
-- This causes n^2 of exp m operations, where n = number of states and
-- m size of @k@.
makeLookupEquiv :: (k -> k -> Bool) -> Int -> Map k Int -> [(k, b)] -> Map k Int
makeLookupEquiv _ !_ !acc [] = acc
makeLookupEquiv eq !n acc ((x, _) : xs) = case ys of
[] -> makeLookup (n + 1) (Map.insert x n acc) xs
((_, i) : _) -> makeLookup n (Map.insert x i acc) xs
where
ys = [ p | p@(y, _) <- Map.toList acc, eq x y ]
-------------------------------------------------------------------------------
-- Destruction
-------------------------------------------------------------------------------
-- | Convert 'DFA' to 'RE.RE'.
--
-- >>> putPretty $ toRE $ fromRE "foobar"
-- ^foobar$
--
-- For 'RE.string' regular expressions, @'toRE' . 'fromRE' = 'id'@:
--
-- prop> let s = take 5 s' in RE.string (s :: String) === toRE (fromRE (RE.string s))
--
-- But in general it isn't:
--
-- >>> let aToZ = RE.star $ RE.charRange 'a' 'z'
-- >>> traverse_ putPretty [aToZ, toRE $ fromRE aToZ]
-- ^[a-z]*$
-- ^([a-z]|[a-z]?[a-z]*[a-z]?)?$
--
-- @
-- not-prop> (re :: RE.RE Char) === toRE (fromRE re)
-- @
--
-- However, they are 'RE.equivalent':
--
-- >>> RE.equivalent aToZ (toRE (fromRE aToZ))
-- True
--
-- And so are others
--
-- >>> all (\re -> RE.equivalent re (toRE (fromRE re))) [RE.star "a", RE.star "ab"]
-- True
--
-- @
-- expensive-prop> RE.equivalent re (toRE (fromRE (re :: RE.RE Char)))
-- @
--
-- Note, that @'toRE' . 'fromRE'@ can, and usually makes regexp unrecognisable:
--
-- >>> putPretty $ toRE $ fromRE $ RE.star "ab"
-- ^(a(ba)*b)?$
--
-- We can 'complement' DFA, therefore we can complement 'RE.RE'.
-- For example. regular expression matching string containing an @a@:
--
-- >>> let withA = RE.star RE.anyChar <> "a" <> RE.star RE.anyChar
-- >>> let withoutA = toRE $ complement $ fromRE withA
-- >>> putPretty withoutA
-- ^([^a]|[^a]?[^a]*[^a]?)?$
--
-- >>> let withoutA' = RE.star $ RE.REChars $ RSet.complement $ RSet.singleton 'a'
-- >>> putPretty withoutA'
-- ^[^a]*$
--
-- >>> RE.equivalent withoutA withoutA'
-- True
--
-- Quite small, for example 2 state DFAs can result in big regular expressions:
--
-- >>> putPretty $ toRE $ complement $ fromRE $ star "ab"
-- ^([^]|a(ba)*(ba)?|a(ba)*([^b]|b[^a])|([^a]|a(ba)*([^b]|b[^a]))[^]*[^]?)$
--
-- We can use @'toRE' . 'fromERE'@ to convert 'ERE.ERE' to 'RE.RE':
--
-- >>> putPretty $ toRE $ fromERE $ complement $ star "ab"
-- ^([^]|a(ba)*(ba)?|a(ba)*([^b]|b[^a])|([^a]|a(ba)*([^b]|b[^a]))[^]*[^]?)$
--
-- >>> putPretty $ toRE $ fromERE $ "a" /\ "b"
-- ^[]$
--
-- See <https://mathoverflow.net/questions/45149/can-regular-expressions-be-made-unambiguous>
-- for the description of the algorithm used.
--
toRE :: forall c. (Ord c, Enum c, Bounded c) => DFA c -> RE.RE c
toRE = toKleene
-- | Convert 'DFA' to 'ERE.ERE'.
toERE :: forall c. (Ord c, Enum c, Bounded c) => DFA c -> ERE.ERE c
toERE = toKleene
-- | Convert to any 'Kleene'.
--
-- See 'toRE' for a specific example.
--
toKleene :: forall k c. (Ord c, Enum c, Bounded c, FiniteKleene c k) => DFA c -> k
toKleene (DFA tr acc _) = unions
[ re 0 j maxN
| j <- IS.toList acc
]
where
maxN | IM.null tr = 1
| otherwise = succ $ fst $ IM.findMax tr
{-
-- this is useful for debug
table =
[ show i ++ " " ++ show j ++ " " ++ show k ++ " = " ++ pretty (re i j k)
| k <- [0..pred maxN]
, i <- [0..pred maxN]
, j <- [0..pred maxN]
]
-}
re i j k = MT.memo re' (i, j, k)
re' (i, j, k)
| k <= 0 = if i == j then eps \/ r else r
| otherwise = re i j k' \/ (re i k' k' <> star (re k' k' k') <> re k' j k')
where
r = maybe empty fromRSet $ Map.lookup (i, j) re0map
k' = k - 1
re0map :: Map (Int, Int) (RSet c)
re0map = Map.fromListWith RSet.union
[ ((i, j), RSet.singletonRange (lo, hi))
| (i, tr') <- IM.toList tr
, (lo, hi, j) <- toPieces tr'
]
toPieces :: (Enum a, Bounded a, Ord a) => SF.SF a b -> [(a, a, b)]
toPieces (SF.SF m v)
| maxBound `Map.member` m = toPieces' m
| otherwise = toPieces' (Map.insert maxBound v m)
toPieces' :: (Enum a, Bounded a) => Map a b -> [(a, a, b)]
toPieces' = go minBound . Map.toList where
go _lo [] = []
go lo ((k, v) : kv) = (lo, k, v) : go (succ k) kv
-------------------------------------------------------------------------------
-- Operations
-------------------------------------------------------------------------------
-- | Run 'DFA' on the input.
--
-- Because we have analysed a language, in some cases we can determine an input
-- without traversing all of the input.
-- That's not the cases with 'RE.RE' 'match'.
--
-- >>> let dfa = fromRE $ RE.star "abc"
-- >>> map (match dfa) ["", "abc", "abcabc", "aa", 'a' : 'a' : undefined]
-- [True,True,True,False,False]
--
-- Holds:
--
-- @
-- 'match' ('fromRE' re) xs == 'match' re xs
-- @
--
-- prop> all (match (fromRE r)) $ take 10 $ RE.generate (curry QC.choose) 42 (r :: RE.RE Char)
--
instance Ord c => Match c (DFA c) where
match (DFA tr acc bh) = go (0 :: Int) where
go s _ | IS.member s bh = IS.member s acc
go s [] = IS.member s acc
go s (c : cs) = case IM.lookup s tr of
Nothing -> False
Just sf -> go (sf SF.! c) cs
-- | Complement DFA.
--
-- Complement of 'DFA' is way easier than of 'RE.RE': complement accept states.
--
-- >>> let dfa = complement $ fromRE $ RE.star "abc"
-- >>> putPretty dfa
-- 0 -> \x -> if
-- | x <= '`' -> 3
-- | x <= 'a' -> 2
-- | otherwise -> 3
-- 1+ -> \x -> if
-- | x <= 'b' -> 3
-- | x <= 'c' -> 0
-- | otherwise -> 3
-- 2+ -> \x -> if
-- | x <= 'a' -> 3
-- | x <= 'b' -> 1
-- | otherwise -> 3
-- 3+ -> \_ -> 3 -- black hole
--
-- >>> map (match dfa) ["", "abc", "abcabc", "aa","abca", 'a' : 'a' : undefined]
-- [False,False,False,True,True,True]
--
instance Complement c (DFA c) where
complement (DFA tr acc err) = DFA tr acc' err where
acc' = IS.difference (IM.keysSet tr) acc
-------------------------------------------------------------------------------
-- Debug
-------------------------------------------------------------------------------
instance Show c => Pretty (DFA c) where
pretty dfa = intercalate "\n"
[ show i ++ acc ++ " -> " ++ SF.showSF sf ++ bh
| (i, sf) <- IM.toList (dfaTransition dfa)
, let acc = if IS.member i (dfaAcceptable dfa) then "+" else ""
, let bh = if IS.member i $ dfaBlackholes dfa then " -- black hole" else ""
]
-- $setup
-- >>> :set -XOverloadedStrings
-- >>> import Data.Foldable (traverse_)
-- >>> import Algebra.Lattice ((/\))
--
-- >>> import Test.QuickCheck ((===))
-- >>> import qualified Test.QuickCheck as QC
--
-- >>> newtype Smaller a = Smaller a deriving (Show)
-- >>> let intLog2 = (`div` 10)
-- >>> instance QC.Arbitrary a => QC.Arbitrary (Smaller a) where arbitrary = QC.scale intLog2 QC.arbitrary; shrink (Smaller a) = map Smaller (QC.shrink a)