sbv-8.10: SMT Based Verification: Symbolic Haskell theorem prover using SMT solving.
Copyright(c) Levent Erkok
LicenseBSD3
Maintainererkokl@gmail.com
Stabilityexperimental
Safe HaskellNone
LanguageHaskell2010

Data.SBV.Tools.BoundedFix

Description

Bounded fixed-point unrolling.

Synopsis

Documentation

bfix :: (SymVal a, Uninterpreted (SBV a -> r)) => Int -> String -> ((SBV a -> r) -> SBV a -> r) -> SBV a -> r Source #

Bounded fixed-point operation. The call bfix bnd nm f unrolls the recursion in f at most bnd times, and uninterprets the function (with the name nm) after the bound is reached.

This combinator is handy for dealing with recursive definitions that are not symbolically terminating and when the property we are interested in does not require an infinite unrolling, or when we are happy with a bounded proof. In particular, this operator can be used as a basis of software-bounded model checking algorithms built on top of SBV. The bound can be successively refined in a CEGAR like loop as necessary, by analyzing the counter-examples and rejecting them if they are false-negatives.

For instance, we can define the factorial function using the bounded fixed-point operator like this:

    bfac :: SInteger -> SInteger
    bfac = bfix 10 "fac" fact
      where fact f n = ite (n .== 0) 1 (n * f (n-1))

This definition unrolls the recursion in factorial at most 10 times before uninterpreting the result. We can now prove:

>>> prove $ \n -> n .>= 1 .&& n .<= 9 .=> bfac n .== n * bfac (n-1)
Q.E.D.

And we would get a bogus counter-example if the proof of our property needs a larger bound:

>>> prove $ \n -> n .== 10 .=> bfac n .== 3628800
Falsifiable. Counter-example:
  s0 = 10 :: Integer

  fac :: Integer -> Integer
  fac _ = 2

The counter-example is telling us how it instantiated the function fac when the recursion bottomed out: It simply made it return 2 for all arguments at that point, which provides the (unintended) counter-example.

By design, if a function defined via bfix is given a concrete argument, it will unroll the recursion as much as necessary to complete the call (which can of course diverge). The bound only applies if the given argument is symbolic. This fact can be used to observe concrete values to see where the bounded-model-checking approach fails:

>>> prove $ \n -> n .== 10 .=> observe "bfac_n" (bfac n) .== observe "bfac_10" (bfac 10)
Falsifiable. Counter-example:
  bfac_10 = 3628800 :: Integer
  bfac_n  = 7257600 :: Integer
  s0      =      10 :: Integer

  fac :: Integer -> Integer
  fac _ = 2

Here, we see further evidence that the SMT solver must have decided to assign the value 2 in the final call just as it was reaching the base case, and thus got the final result incorrect. (Note that 7257600 = 2 * 3628800.) A wrapper algorithm can then assert the actual value of bfac 10 here as an extra constraint and can search for "deeper bugs."