The H2 Wiki


accidentally-exponential

Accidentally exponential in a compiler

The introduction

I accidentally introduced exponential run time into a compiler I was writing and did not realise for over three years.

The language

Once upon a time I was writing an AST to represent expressions in a programming language. In reality the language was SQL, but for the sake accessibility let’s pretend it was a simple arithmetic language that allows us to write zero, one, and (non-empty) sums and products.

data Expr a = Zero a
            | One
            | Sum (NonEmpty (Expr a))
            | Prod (NonEmpty (Expr a))

The reason why Zero has an argument will be explained in due course. First, let’s implement wrapper functions to provide a nice API and write an example expression.

zero :: Expr ()
zero = Zero ()

(.+) :: Expr a -> Expr a -> Expr a
e1 .+ e2 = Sum (pure e1 <> pure e2)

(.*) :: Expr a -> Expr a -> Expr a
e1 .* e2 = Prod (pure e1 <> pure e2)

example :: Expr ()
example = (One .+ zero .+ zero .+ One) .* (zero .+ One .+ One)

We expect the example expression to evaluate to 2 * 2, i.e. 4. Let’s write an evaluator and check. The evaluator will be recursive on the structure of terms of type Expr. We are good functional programmers so we will write a data type and general fold function to encapsulate the recursion pattern. They basically write themselves, given the definition of Expr.

data ExprFold a r = ExprFold
  { zeroF :: a -> r
  , oneF  :: r
  , sumF  :: NonEmpty r -> r
  , prodF :: NonEmpty r -> r
  }

exprFold :: ExprFold a r -> Expr a -> r
exprFold f = \case
  Zero a  -> zeroF f a
  One     -> oneF f
  Prod es -> prodF f (fmap (exprFold f) es)
  Sum es  -> sumF f (fmap (exprFold f) es)

Now it’s simple, almost tautological, to write an evaluator to interpret terms in the integers.

int :: Expr a -> Int
int = exprFold ExprFold
  { zeroF = const 0
  , oneF  = 1
  , sumF  = sum
  , prodF = product
  }
> int example
4

The task

Now I’ll explain the reason for the argument to the Zero constructor. The backend that we are targeting doesn’t have a very good concept of Zero! We can happily send One to it but sending Zero is quite tricky. In reality the backend is SQL and it’s not terribly straightforward to generically represent an empty query in SQL.

Instead of dealing with that problem directly, my approach was to give Expr a type parameter. Then expressions of type Expr () can contain Zero and expressions of type Expr Void cannot.

Expressions that might contain Zero need to have it removed before the query is sent to the backend. The following function removes all occurrences of Zero from an expression. If the expression itself was actually zero it returns Nothing, otherwise it returns a Just. Because the type parameter in the return type is polymorphic we have a static guarantee that the result doesn’t contain any Zeros!

-- Just an expression, or Nothing.  Nothing means that the value
-- of the expression is zero.
removeZero :: Expr unit -> Maybe (Expr void)
removeZero = exprFold ExprFold
  { zeroF   = const Nothing
    -- ^ Zero is converted to Nothing
  , oneF    = pure One
    -- ^ One does not need to change
  , sumF    = \es ->
       Sum <$> nonEmpty (catMaybes (toList es))
  -- ^ We filter all zeros out of the Sum.  Terms that are zero
  -- (Nothing) are removed by catMaybes. If they were *all* zero
  -- then the argument to nonEmpty is [] so it returns Nothing.
  -- If some were not zero then we end up with a NonEmpty list of
  -- non-zeroes.
  , prodF   = \es ->
       Prod <$> (traverse removeZero =<< sequence es)
  -- ^ If any of the terms in a Prod is Nothing then the entire
  -- product is Nothing (because the product of a list
  -- containing zero is itself zero).
  }
>  fmap int (removeZero example)
Just 4
> fmap int (removeZero (zero .* One))
Nothing

We’re only ever going to use removeZero at type

removeZero :: Expr () -> Maybe (Expr Void)

but the more general type signature

removeZero :: Expr unit -> Maybe (Expr void)

is valid, so let’s use that instead. After all, more polymorphism gives fewer places for bugs to hide, right?

The bug

removeZero calculates the correct result but regrettably it has exponential run time! How can this be? The point of expressing removeZero in terms of a fold function was to make it brainlessly simple to write it efficiently and correctly.

The problem is that I called removeZero in the body of its own definition. I wasn’t supposed to do that! exprFold was supposed to encapsulate all the recursion so that I need not use explicit recursion myself. For the prodF field I wrote

prodF = \es -> Prod <$> (traverse removeZero =<< sequence es)

when I should have written

prodF = \es -> Prod <$> sequence es

The argument to prodF, es, has already had the recursive call of removeZero applied to each element by exprFold. I didn’t need to call removeZero again. Calling it again doesn’t change the result (because there are no longer any zeros) but it doubles the amount of work at each Prod node, making the run time exponential.

The stars aligned

Ironically, if removeZero were less polymorphic, taking an argument of type Expr () instead of Expr unit, then the type checker would have caught this bug. The recursive call is at type Expr void which Haskell’s polymorphic recursion very happily allows to unify with Expr unit. Furthermore, if removeZero were not idempotent then this bug would have been quickly discovered by the test suite. The bug could only occur because the following three conditions coincided:

The reality

This bug occurred in real life in Opaleye, my PostgreSQL EDSL. The bug was added in Feb 2016 and only removed nearly three and a half years later, after being reported by Christopher Sasarak. (Despite the commit message the slowdown is exponential, not quadratic. Suppose that each level of the tree has only one branch. We’re running removeEmpty twice for each branch, so cost (Prod es) = 2 * sum (cost es).)

The conclusion

Opaleye makes good use of Haskell’s type system to check correctness. It has a comprehensive set of unit tests and property tests. However, because the buggy function returns the correct result there is no way that Haskell’s type system can spot the error. Nor can unit tests, property tests, Liquid Haskell, or any other system which only checks the calculated value. Some form of performance testing is required. If you have a library without a performance testing suite perhaps you will be lucky enough, like me, that your users act as performance testers for you!