Sound Haskell

Dana N. Xu

University of Cambridge

Joint work with Simon Peyton Jones

Microsoft Research CambridgeKoen Claessen

Chalmers University of Technology

Module UserPgm where

f :: [Int]->Intf xs = head xs `max` 0

: … f [] …

Program Errors Give Headache!

Glasgow Haskell Compiler (GHC) gives at run-time

Exception: Prelude.head: empty list

Module Prelude where

head :: [a] -> ahead (x:xs) = xhead [] = error “empty list”

Types>> > > > > Contracts >> > > > >

head (x:xs) = x

head :: [Int] -> Int

…(head 1)…

head :: {x | not (null xs)} -> {r | True}

…(head [])…

Bug!

Bug!Contract

(original Haskell boolean expression)

Type not :: Bool -> Boolnot True = Falsenot False = True

null :: [a] -> Boolnull [] = Truenull (x:xs) = False

Preconditionshead :: {xs | not (null xs)} -> {r | True}head (x:xs’) = x

f xs = head xs `max` 0

Warning: f [] calls head which may fail head’s precondition!

f_ok xs = if null xs then 0 else head xs `max` 0

No more warnings from compiler!

Expressiveness of the Specification Language

data T = T1 Bool | T2 Int | T3 T T

sumT :: T -> IntsumT :: {x | noT1 x} -> {r | True}sumT (T2 a) = asumT (T3 t1 t2) = sumT t1 + sumT t2

noT1 :: T -> BoolnoT1 (T1 _) = FalsenoT1 (T2 _) = TruenoT1 (T3 t1 t2) = noT1 t1 && noT1 t2

Expressiveness of the Specification LanguagesumT :: T -> IntsumT :: {x | noT1 x} -> {r | True}sumT (T2 a) = asumT (T3 t1 t2) = sumT t1 + sumT t2

rmT1 :: T -> TrmT1 :: {x | True} -> {r | noT1 r}rmT1 (T1 a) = if a then T2 1 else T2 0rmT1 (T2 a) = T2 armT1 (T3 t1 t2) = T3 (rmT1 t1) (rmT1 t2)

For all crash-free t::T, sumT (rmT1 t) will not crash.

Functions without Annotationsdata T = T1 Bool | T2 Int | T3 T T

noT1 :: T -> BoolnoT1 (T1 _) = FalsenoT1 (T2 _) = TruenoT1 (T3 t1 t2) = noT1 t1 && noT1 t2

(&&) True x = x(&&) False x = False

No abstraction is more compact than the function definition itself!

Higher Order Functions

all :: (a -> Bool) -> [a] -> Boolall f [] = Trueall f (x:xs) = f x && all f xs

filter :: (a -> Bool) -> [a] -> [a]filter :: {p | True} -> {xs | True} -> {r | all p r}filter p [] = []filter p (x:xs’) = case (p x) of True -> x : filter p xs’ False -> filter p xs’

Contracts for higher-order function’s parameterf1 :: (Int -> Int) -> Int

f1 :: ({x | True} -> {y | y >= 0}) -> {r | r >= 0}

f1 g = (g 1) - 1

f2:: {r | True}

f2 = f1 (\x -> x – 1)

Error: f1’s postcondition fails because (g 1) >= 0 does not imply (g 1) – 1 >= 0Error: f2 calls f1 which fails f1’s precondition

Lazinessfst (a,b) = a

Option 1 ()fst :: {x | True} -> {r | True}

Option 2 ()fst :: ({x | True}, Any) -> {r | True} …fst (5, error “f”)…

fstN :: (Int, Int) -> IntfstN :: ({x | True}, Any) -> {r | True}fstN (a, b) n = if n > 0 then fstN (a+1, b) (n-1)

else a

g2 = fstN (5, error “fstN”) 100

Every expression satisfies Any

Various Exampleszip :: [a] -> [b] -> [(a,b)]zip :: {xs | True} -> {ys | sameLen xs ys} -> {rs | sameLen rs xs }

sameLen [] [] = TruesameLen (x:xs) (y:ys) = sameLen xs yssameLen _ _ = False

f91 :: Int -> Intf91 :: { n <= 101 } -> {r | r == 91 }f91 n = case (n <= 100) of

True -> f91 (f91 (n + 11)) False -> n – 10

Sortingsorted [] = Truesorted (x:[]) = Truesorted (x:y:xs) = x <= y && sorted (y : xs)

insert :: {i | True} -> {xs | sorted xs} -> {r | sorted r}

merge :: [Int] -> [Int] -> [Int]merge :: {xs | sorted xs} -> {ys | sorted ys} -> {r | sorted r}

bubbleHelper :: [Int] -> ([Int], Bool)bubbleHelper :: {xs | True} -> {r | not (snd r) ==> sorted (fst r)}

Insertsort, mergesort, bubblesort :: {xs | True} -> {r | sorted r}

(==>) True x = x(==>) False x = True

Contract Synonymtype Ok = {x | True}

type NonNull = {x | not (null x)}

head :: [Int] -> Int

head :: NonNull -> Ok

head (x:xs) = x

{-# type Ok = {x | True} -#}

{-# type NonNull = {x | not (null x)} #-}

{-# contract head :: NonNull -> Ok #-}

Actual Syntax

What we can’t dog1, g2 :: Ok -> Okg1 x = case (prime x > square x) of True -> x False -> error “urk”

g2 xs ys = case (rev (xs ++ ys) == rev ys ++ rev xs) of True -> xs False -> error “urk”

Hence, three possible outcomes: (1) Definitely Safe (no crash, but may loop)(2) Definite Bug (definitely crashes)(3) Possible Bug

Crash!

Crash!

LanguageSyntax

following Haskell’slazy semantics

Two special constructors BAD is an expression that crashes.error :: String -> aerror s = BAD

head (x:xs) = xhead [] = BAD

UNR (short for “unreachable”) is an expression that gets stuck. This is not a crash, although execution comes to a halt without delivering a result. (identifiable infinite loop)

Crashing

Definition (Crash).

A closed term e crashes iff e !* BAD

Definition (Crash-free Expression)

An expression e is crash-free iff

8 C. BAD 2 C, ` C[[e]] :: (), C[[e]] !* BAD

What to Check?Does function f satisfies its contract t (written f2 t)?

Goal: main 2 {x | True}

At the definition of each function f,assuming the given precondition holds,we check1. No pattern matching failure2. Precondition of all calls in the body of f holds3. Postcondition holds for f itself.

How to Check?Given e and t We construct a term (eBt) (pronounced e “ensures” t) Prove (eBt) is crash-free.

simplify the term (eBt) to e’ and e’ is syntactically safe, then we are done.

Theorem 1eBt is crash-free , e 2 t

Theorem 2simpl (eBt) is syntactically safe ) eBt is crash-free

This Talk

ESC/Haskell (HW’06)

Syntax of Contracts

Full version: x’:{x | x >0} -> {r | r > x’}Short hand: {x | x > 0} -> {r | r > x}

k:({x | x > 0} -> {y | y > 0}) -> {r | r > k 5}

Contract Satisfaction

e" means e diverges or e !* UNR

B pronounced ensuresC pronounced requires

e B {x | p} = case p[e/x] of

True -> eFalse -> BAD

e C {x | p} = case p[e/x] of

True -> e

False -> UNR

Example:5 2 {x | x > 0} 5 B {x | x > 0}= case (5 > 0) of True -> 5 False -> BAD

e B x:t1 ! t2 = v. (e (vC t1)) B t2[vCt1/x]

e C x:t1 ! t2 = v. (e (vB t1)) C t2[vBt1/x]

e B (t1, t2) = case e of (e1, e2) -> (e1 B t1, e2 B t2)

e C (t1, t2) = case e of (e1, e2) -> (e1 C t1, e2 C t2)

Function Contract and Tuple Contract

f :: {x | x > 0} -> {r | True}f x = x

f B {x | x > 0} -> {r | True}=(x.x) B {x | x > 0} -> {r | True}= v. (x.x (v C {x | x > 0})) B {r | True}= v. (case v > 0 of True -> v False -> UNR)

g = … f…

f C {x | x > 0} -> {r | True}= v. (case v > 0 of True -> v False -> BAD)

Higher-Order Function

f1 B ({x | True} -> {y | y >= 0}) -> {r | r >= 0}= … B C B= v1. case (v1 1) >= 0 of True -> case (v1 1) - 1 >= 0 of True -> (v1 1) -1 False -> BAD False -> UNR

f1 :: (Int -> Int) -> Intf1 :: ({x | True} -> {y | y >= 0}) -> {r | r >= 0}f1 g = (g 1) - 1

f2:: {r | True}f2 = f1 (\x -> x – 1)

e B Any = UNRe C Any = BAD

f5 :: Any -> {r | True}f5 x = 5

error :: String -> a -- HM Typeerror :: {x | True} -> Any -- Contract

Any Contract

Properties of B and CLemma1:For all closed, crash-free e, and closed t,e C t 2 t

Lemma2:For all e and t, if e2 t, then (a) e v e B t(b) e C t v e

Definition (Crashes-More-Often):e1 v e2 iff for all C, ` C[[ei]] :: () for i=1,2 and(1) C[[e2]] !* BAD ) C[[e1]] !* BAD

About 30 Lemmas Lemma [Monotonicity of Satisfaction ]:If e1 2 t and e1 v e2, then e22 tLemma [Congruence of v]:e1 v e2 ) 8 C. C[[e1]] v C[[e2]]Lemma [Idempotence of Projection]:8 e, t. e B t B t ≡ e B t 8 e, t. e C t C t ≡ eC t Lemma [A Projection Pair]:8 e, t. e B t C t v eLemma [A Closure Pair]:8 e, t. e v eC t B t

Contributions Automatic static contract checking instead of dynamic contract checking. Compared with ESC/Haskell

Allow pre/post specification for higher-order function’s parameter through contracts.

Reduce more false alarms caused by Laziness and in an efficient way. Allow user-defined data constructors to be used to define contracts Specifications are more type-like, so we can define contract synonym.

We develop a concise notation (B and C) for contract checking, which enjoys many properties. We give a new and relatively simpler proof of the soundness and completeness of dynamic contract checking, this proof is much trickier than it looks.

We implement the idea in GHC Accept full Haskell Support separate compilation as the verification is modular. Can check functions without contract through CEG unrolling.