{-# LANGUAGE OverloadedStrings, DeriveDataTypeable #-} {-# OPTIONS_GHC -Wall #-} {-| This module contains the core calculus for the Morte language. This language is a minimalist implementation of the calculus of constructions, which is in turn a specific kind of pure type system. If you are new to pure type systems you may wish to read \"Henk: a typed intermediate language\". Morte is a strongly normalizing language, meaning that: * Every expression has a unique normal form computed by `normalize` * You test expressions for equality of their normal forms using `==` * Equational reasoning preserves normal forms Strong normalization comes at a price: Morte forbids recursion. Instead, you must translate all recursion to F-algebras and translate all corecursion to F-coalgebras. If you are new to F-(co)algebras then you may wish to read "Morte.Tutorial" or read \"Recursive types for free!\": Morte is designed to be a super-optimizing intermediate language with a simple optimization scheme. You optimize a Morte expression by just normalizing the expression. If you normalize a long-lived program encoded as an F-coalgebra you typically get a state machine, and if you normalize a long-lived program encoded as an F-algebra you typically get an unrolled loop. Strong normalization guarantees that all abstractions encodable in Morte are \"free\", meaning that they may increase your program's compile times but they will never increase your program's run time because they will normalize to the same code. -} module Morte.Core ( -- * Syntax Var(..), Const(..), Expr(..), Context, -- * Core functions typeWith, typeOf, normalize, -- * Utilities used, shift, prettyExpr, prettyTypeError, -- * Errors TypeError(..), TypeMessage(..), -- * Builders buildConst, buildVar, buildExpr, buildTypeMessage, buildTypeError, ) where import Control.Applicative ((<$>), (<*>)) import Control.DeepSeq import Control.Exception (Exception) import Control.Monad.Trans.State (State, evalState) import qualified Control.Monad.Trans.State as State import Data.Binary (Binary(get, put), Get, Put) import Data.Binary.Get (getWord64le) import Data.Binary.Put (putWord64le) import Data.Monoid (mempty, (<>)) import Data.String (IsString(fromString)) import Data.Text () -- For the `IsString` instance import Data.Text.Lazy (Text, unpack) import qualified Data.Text.Encoding as Text import qualified Data.Text.Lazy as Text import Data.Text.Lazy.Builder (Builder, toLazyText, fromLazyText) import Data.Text.Lazy.Builder.Int (decimal) import Data.Typeable (Typeable) import Data.Word (Word8) {-| Label for a bound variable The `Text` field is the variable's name (i.e. \"@x@\"). The `Int` field disambiguates variables with the same name if there are multiple bound variables of the same name in scope. Zero refers to the nearest bound variable and the index increases by one for each bound variable of the same name going outward. The following diagram may help: > +-refers to-+ > | | > v | > \(x : *) -> \(y : *) -> \(x : *) -> x@0 > > +-------------refers to-------------+ > | | > v | > \(x : *) -> \(y : *) -> \(x : *) -> x@1 This `Int` behaves like a De Bruijn index in the special case where all variables have the same name. You can optionally omit the index if it is @0@: > +refers to+ > | | > v | > \(x : *) -> \(y : *) -> \(x : *) -> x Zero indices are omitted when pretty-printing `Var`s and non-zero indices appear as a numeric suffix. -} data Var = V Text Int deriving (Eq, Show) putUtf8 :: Text -> Put putUtf8 txt = put (Text.encodeUtf8 (Text.toStrict txt)) getUtf8 :: Get Text getUtf8 = do bs <- get case Text.decodeUtf8' bs of Left e -> fail (show e) Right txt -> return (Text.fromStrict txt) instance Binary Var where put (V x n) = do putUtf8 x putWord64le (fromIntegral n) get = V <$> getUtf8 <*> fmap fromIntegral getWord64le instance IsString Var where fromString str = V (Text.pack str) 0 instance NFData Var where rnf (V n p) = rnf n `seq` rnf p {-| Constants for the calculus of constructions The only axiom is: > ⊦ * : □ ... and all four rule pairs are valid: > ⊦ * ↝ * : * > ⊦ □ ↝ * : * > ⊦ * ↝ □ : □ > ⊦ □ ↝ □ : □ -} data Const = Star | Box deriving (Eq, Show, Bounded, Enum) instance Binary Const where put c = case c of Star -> put (0 :: Word8) Box -> put (1 :: Word8) get = do n <- get :: Get Word8 case n of 0 -> return Star 1 -> return Box _ -> fail "get Const: Invalid tag byte" instance NFData Const where rnf c = seq c () axiom :: Const -> Either TypeError Const axiom Star = return Box axiom Box = Left (TypeError [] (Const Box) (Untyped Box)) rule :: Const -> Const -> Either TypeError Const rule Star Box = return Box rule Star Star = return Star rule Box Box = return Box rule Box Star = return Star -- | Syntax tree for expressions data Expr -- | > Const c ~ c = Const Const -- | > Var (V x 0) ~ x -- > Var (V x n) ~ x@n | Var Var -- | > Lam x A b ~ λ(x : A) → b | Lam Text Expr Expr -- | > Pi x A B ~ ∀(x : A) → B -- > Pi unused A B ~ A → B | Pi Text Expr Expr -- | > App f a ~ f a | App Expr Expr deriving (Show) lookupN :: Eq a => a -> [(a, b)] -> Int -> Maybe b lookupN a ((a', b'):abs') n | a /= a' = lookupN a abs' n | n > 0 = lookupN a abs' $! n - 1 | n == 0 = Just b' | otherwise = Nothing lookupN _ [] _ = Nothing lookupCtx :: Var -> Context -> Maybe Expr lookupCtx (V x n) ctx = lookupN x ctx n instance Eq Expr where eL0 == eR0 = evalState (go (normalize eL0) (normalize eR0)) [] where go :: Expr -> Expr -> State [(Text, Text)] Bool go (Const cL) (Const cR) = return (cL == cR) go (Var (V xL nL)) (Var (V xR nR)) = do ctx <- State.get return (nL == nR && case lookupN xL ctx nL of Nothing -> xL == xR Just xR' -> xR' == xR ) go (Lam xL tL bL) (Lam xR tR bR) = do ctx <- State.get State.put ((xL, xR):ctx) eq1 <- go tL tR eq2 <- go bL bR State.put ctx return (eq1 && eq2) go (Pi xL tL bL) (Pi xR tR bR) = do ctx <- State.get State.put ((xL, xR):ctx) eq1 <- go tL tR eq2 <- go bL bR State.put ctx return (eq1 && eq2) go (App fL aL) (App fR aR) = do b1 <- go fL fR b2 <- go aL aR return (b1 && b2) go _ _ = return False instance Binary Expr where put e = case e of Const c -> do put (0 :: Word8) put c Var x -> do put (1 :: Word8) put x Lam x _A b -> do put (2 :: Word8) putUtf8 x put _A put b Pi x _A _B -> do put (3 :: Word8) putUtf8 x put _A put _B App f a -> do put (4 :: Word8) put f put a get = do n <- get :: Get Word8 case n of 0 -> Const <$> get 1 -> Var <$> get 2 -> Lam <$> getUtf8 <*> get <*> get 3 -> Pi <$> getUtf8 <*> get <*> get 4 -> App <$> get <*> get _ -> fail "get Expr: Invalid tag byte" instance IsString Expr where fromString str = Var (fromString str) instance NFData Expr where rnf e = case e of Const c -> rnf c Var v -> rnf v Lam x _A b -> rnf x `seq` rnf _A `seq` rnf b Pi x _A _B -> rnf x `seq` rnf _A `seq` rnf _B App f a -> rnf f `seq` rnf a {-| Bound variable names and their types Variable names may appear more than once in the `Context`. The `Var` @x\@n@ refers to the @n@th occurrence of @x@ in the `Context` (using 0-based numbering). -} type Context = [(Text, Expr)] -- | The specific type error data TypeMessage = UnboundVariable | InvalidInputType Expr | InvalidOutputType Expr | NotAFunction | TypeMismatch Expr Expr | Untyped Const deriving (Show) instance NFData TypeMessage where rnf tm = case tm of UnboundVariable -> () InvalidInputType e -> rnf e InvalidOutputType e -> rnf e NotAFunction -> () TypeMismatch e1 e2 -> rnf e1 `seq` rnf e2 Untyped c -> rnf c -- | A structured type error that includes context data TypeError = TypeError { context :: Context , current :: Expr , typeMessage :: TypeMessage } deriving (Typeable) instance Show TypeError where show = unpack . prettyTypeError instance Exception TypeError instance NFData TypeError where rnf (TypeError ctx crr tym) = rnf ctx `seq` rnf crr `seq` rnf tym -- | Render a pretty-printed `Const` as a `Builder` buildConst :: Const -> Builder buildConst c = case c of Star -> "*" Box -> "□" -- | Render a pretty-printed `Var` as a `Builder` buildVar :: Var -> Builder buildVar (V txt n) = fromLazyText txt <> if n == 0 then mempty else "@" <> decimal n -- | Render a pretty-printed `Expr` as a `Builder` buildExpr :: Expr -> Builder buildExpr = go False False where go :: Bool -> Bool -> Expr -> Builder go parenBind parenApp e = case e of Const c -> buildConst c Var x -> buildVar x Lam x _A b -> (if parenBind then "(" else "") <> "λ(" <> fromLazyText x <> " : " <> go False False _A <> ") → " <> go False False b <> (if parenBind then ")" else "") Pi x _A b -> (if parenBind then "(" else "") <> (if used x b then "∀(" <> fromLazyText x <> " : " <> go False False _A <> ")" else go True False _A ) <> " → " <> go False False b <> (if parenBind then ")" else "") App f a -> (if parenApp then "(" else "") <> go True False f <> " " <> go True True a <> (if parenApp then ")" else "") {-| Determine whether a `Pi`-bound variable should be displayed Notice that if any variable within the body of a `Pi` shares the same name and an equal or greater DeBruijn index we display the `Pi`-bound variable. To illustrate why we don't just check for equality, consider this type: > forall (a : *) -> forall (a : *) -> a@1 The @a\@1@ refers to the outer @a@ (i.e. the left one), but if we hid the inner @a@ (the right one), the type would make no sense: > forall (a : *) -> * -> a@1 ... because the @a\@1@ would misleadingly appear to be an unbound variable. -} used :: Text -> Expr -> Bool used x e0 = go e0 0 where go e n = case e of Var (V x' n') | x == x' && n' >= n -> True | otherwise -> False Lam x' _A b -> go _A n || (go b $! n') where n' = if x == x' then n + 1 else n Pi x' _A _B -> go _A n || (go _B $! n') where n' = if x == x' then n + 1 else n App f a -> go f n || go a n Const _ -> False -- | Render a pretty-printed `TypeMessage` as a `Builder` buildTypeMessage :: TypeMessage -> Builder buildTypeMessage msg = case msg of UnboundVariable -> "Error: Unbound variable\n" InvalidInputType expr -> "Error: Invalid input type\n" <> "\n" <> "Type: " <> buildExpr expr <> "\n" InvalidOutputType expr -> "Error: Invalid output type\n" <> "\n" <> "Type: " <> buildExpr expr <> "\n" NotAFunction -> "Error: Only functions may be applied to values\n" TypeMismatch expr1 expr2 -> "Error: Function applied to argument of the wrong type\n" <> "\n" <> "Expected type: " <> buildExpr expr1 <> "\n" <> "Argument type: " <> buildExpr expr2 <> "\n" Untyped c -> "Error: " <> buildConst c <> " has no type\n" -- | Render a pretty-printed `TypeError` as a `Builder` buildTypeError :: TypeError -> Builder buildTypeError (TypeError ctx expr msg) = "\n" <> ( if Text.null (toLazyText buildContext ) then mempty else "Context:\n" <> buildContext <> "\n" ) <> "Expression: " <> buildExpr expr <> "\n" <> "\n" <> buildTypeMessage msg where buildKV (key, val) = fromLazyText key <> " : " <> buildExpr val buildContext = (fromLazyText . Text.unlines . map (toLazyText . buildKV) . reverse) ctx {-| Substitute all occurrences of a variable with an expression > subst x n C B ~ B[x@n := C] -} subst :: Text -> Int -> Expr -> Expr -> Expr subst x n e' e = case e of Lam x' _A b -> Lam x' (subst x n e' _A) b' where n' = if x == x' then n + 1 else n b' = n' `seq` subst x n' (shift 1 x' e') b Pi x' _A _B -> Pi x' (subst x n e' _A) _B' where n' = if x == x' then n + 1 else n _B' = n' `seq` subst x n' (shift 1 x' e') _B App f a -> App (subst x n e' f) (subst x n e' a) Var (V x' n') -> if x == x' && n == n' then e' else e Const k -> Const k {-| @shift n x@ adds @n@ to the index of all free variables named @x@ within an `Expr` -} shift :: Int -> Text -> Expr -> Expr shift d x0 e0 = go e0 0 where go e c = case e of Lam x _A b -> Lam x (go _A c) (go b $! c') where c' = if x == x0 then c + 1 else c Pi x _A _B -> Pi x (go _A c) (go _B $! c') where c' = if x == x0 then c + 1 else c App f a -> App (go f c) (go a c) Var (V x n) -> n' `seq` Var (V x n') where n' = if x == x0 && n >= c then n + d else n Const k -> Const k {-| Type-check an expression and return the expression's type if type-checking suceeds or an error if type-checking fails `typeWith` does not necessarily normalize the type since full normalization is not necessary for just type-checking. If you actually care about the returned type then you may want to `normalize` it afterwards. -} typeWith :: Context -> Expr -> Either TypeError Expr typeWith ctx e = case e of Const c -> fmap Const (axiom c) Var x -> case lookupCtx x ctx of Nothing -> Left (TypeError ctx e UnboundVariable) Just a -> return a Lam x _A b -> do let ctx' = [ (x', shift 1 x _A') | (x', _A') <- (x, _A):ctx ] _B <- typeWith ctx' b let p = Pi x _A _B _t <- typeWith ctx p return p Pi x _A _B -> do eS <- fmap whnf (typeWith ctx _A) s <- case eS of Const s -> return s _ -> Left (TypeError ctx e (InvalidInputType _A)) let ctx' = [ (x', shift 1 x _A') | (x', _A') <- (x, _A):ctx ] eT <- fmap whnf (typeWith ctx' _B) t <- case eT of Const t -> return t _ -> Left (TypeError ctx' e (InvalidOutputType _B)) fmap Const (rule s t) App f a -> do e' <- fmap whnf (typeWith ctx f) (x, _A, _B) <- case e' of Pi x _A _B -> return (x, _A, _B) _ -> Left (TypeError ctx e NotAFunction) _A' <- typeWith ctx a if _A == _A' then do let a' = shift 1 x a _B' = subst x 0 a' _B return (shift (-1) x _B') else do let nf_A = normalize _A nf_A' = normalize _A' Left (TypeError ctx e (TypeMismatch nf_A nf_A')) {-| `typeOf` is the same as `typeWith` with an empty context, meaning that the expression must be closed (i.e. no free variables), otherwise type-checking will fail. -} typeOf :: Expr -> Either TypeError Expr typeOf = typeWith [] -- | Reduce an expression to weak-head normal form whnf :: Expr -> Expr whnf e = case e of App f a -> case whnf f of Lam x _A b -> whnf (shift (-1) x b') -- Beta reduce where a' = shift 1 x a b' = subst x 0 a' b _ -> e _ -> e -- | Returns whether a variable is free in an expression freeIn :: Var -> Expr -> Bool freeIn v@(V x n) = go where go e = case e of Lam x' _A b -> n' `seq` (go _A || if x == x' then freeIn (V x n') b else go b) where n' = n + 1 Pi x' _A _B -> n' `seq` (go _A || if x == x' then freeIn (V x n') _B else go _B) where n' = n + 1 Var v' -> v == v' App f a -> go f || go a Const _ -> False {-| Reduce an expression to its normal form, performing both beta reduction and eta reduction `normalize` does not type-check the expression. You may want to type-check expressions before normalizing them since normalization can convert an ill-typed expression into a well-typed expression. -} normalize :: Expr -> Expr normalize e = case e of Lam x _A b -> case b' of App f a -> case a of Var v' | v == v' && not (v `freeIn` f) -> shift (-1) x f -- Eta reduce | otherwise -> e' where v = V x 0 _ -> e' _ -> e' where b' = normalize b e' = Lam x (normalize _A) b' Pi x _A _B -> Pi x (normalize _A) (normalize _B) App f a -> case normalize f of Lam x _A b -> normalize (shift (-1) x b') -- Beta reduce where a' = shift 1 x (normalize a) b' = subst x 0 a' b f' -> App f' (normalize a) Var _ -> e Const _ -> e {-| Pretty-print an expression The result is a syntactically valid Morte program -} prettyExpr :: Expr -> Text prettyExpr = toLazyText . buildExpr -- | Pretty-print a type error prettyTypeError :: TypeError -> Text prettyTypeError = toLazyText . buildTypeError