====== Lecture 3 (Skeleton) ======

Summary of what we are doing in today's class:

{{vcg-big-picture.png}}


===== Verification condition generation: converting programs into formulas =====

==== Context ====

Recall that we can
  * represent programs using guarded command language, e.g. desugaring of 'if' into non-deterministic choice and assume
  * give meaning to guarded command language statements as relations
  * we can represent relations using set comprehensions; if our program c has two state components, we can represent its meaning R( c ) as $\{((x_0,y_0),(x,y)) \mid F  \}$, where F is some formula that has x,y,x_0,y_0 as free variables.

  * simple values: variables are integers.  Later we will talk about modeling pointers and arrays, but what we say now applies

Our goal is to find rules for computing R( c ) that are
  * correct
  * efficient
  * create formulas that we can effectively prove later


What exactly do we prove about the formula R( c ) ?

We prove that this formula is **valid**:

  R( c ) -> error=false



==== Formulas for basic statements ====

In our simple language, basic statements are assignment, havoc, assume, assert.

  R(x=t) = (x=t & y=y_0 & error=error_0)

**Note**: all our statements will have the property that if error_0 = true, then error=true.  That is, you can never recover from an error state.  This is convenient: if we prove no errors at the end, then there were never errors in between.

**Note**: the condition y=y_0 & error=error_0 is called <b>frame condition</b>.  There are as many conjuncts as there are components of the state.  This can be annoying to write, so let us use shorthand frame(x) for it.  The shorthand frame(x) denotes a conjunction of v=v_0 for all v that are distinct from x (in this case y and error).  We can have zero or more variables as arguments of frame, so frame() means that nothing changes.

  R(havoc x) = frame(x)
  R(assume F) = F[x:=x_0, y:=y_0, error:=error_0] & frame()
  R(assert F) = (F -> frame)

**Note**:

  x=t  is same as  havoc(x);assume(x=t)

  assert false = crash  (stops with error)

  assume true  = skip   (does nothing)



==== Composing formulas using relation composition ====

This is perhaps the most direct way of transforming programs to formulas.  It creates formulas that are linear in the size of the program.

Non-deterministic choice is union of relations, that is, disjunction of formulas:

  CR(c1 [] c2) = CR(c1) | CR(c2)

In sequential composition we follow the rule for composition of relations.  We want to get again formula with free variables x_0,y_0,x,y.  So we need to do renaming.  Let x_1,y_1,error_1 be fresh variables.

  CR(c1 ; c2) = exists x_1,y_1,error_1.  CR(c1)[x:=x_1,y:=y_1,error:=error_1] & CR(c2)[x:=x_1,y:=y_1,error:=error_1]

The base case is

  CR(c)=R(c)

when c is a basic command.





==== Avoiding accumulation of equalities ====

This approach generates many variables and many frame conditions.  

Ignoring error for the moment, we have, for example:

  R(x=3) = (x=3 & y=y_0)
  R(y=x+2) = (y=x_0 + 2 & x=x_0)

  CR(x=3;y=x+2) = x_1=3 & y_1 = y_0 & y = x_1 + 2 & x = x_1

But if a variable is equal to another, it can be substituted using the substitution rules

  (exists x_1. x_1=t & F(x_1))     <->    F(t)
  (forall x_1. x_1=t -> F(x_1)     <->    F(t)

We can apply these rules to reduce the size of formulas.


==== Approximation ====

If (F -> G) is value, we say that F is stronger than F and we say G is weaker than F.

When a formula would be too complicated, we can instead create a simpler approximate formula.  To be sound, if our goal is to prove a property, we need to generate a *larger* relation, which corresponds to a weaker formula describing a relation, and a stronger verification condition.  (If we were trying to identify counterexamples, we would do the opposite).

We can replace "assume F" with "assume F1" where F1 is weaker.  Consequences:
  * omtiting complex if conditionals (assuming both branches can happen - as in most type systems)
  * replacing complex assignments with arbitrary change to variable: because x=t is havoc(x);assume(x=t) and we drop the assume

This idea is important in static analysis.








==== Symbolic execution ====

Symbolic execution converts programs into formulas by going forward.  It is therefore somewhat analogous to the way an [[interpreter]] for the language would work.  

Avoid renaming all the time.

  SE(F,k, c1; c2) = SE(F & R(c1), k+1, c2)             (update formula)

  SE(F,k,(c1 [] c2); c2) = SE(F, k, c1) | SE(F,k,c2)   (explore both branches)

Note: how many branches do we get?

Strongest postcondition:
\begin{equation*}
  sp(P,r) = \{ s_2 \mid \exists s_1.\ s_1 \in P \land (s_1,s_2) \in r \}
\end{equation*}
Like composition of a set with a relation.  It's called ''relational image'' of set $P$ under relation $r$.

Note: when proving our verification condition, instead of proving that semantics of relation implies error=false, it's same as proving that the formula for set sp(U,r) implies error=false, where U is the universal relation, or, in terms of formulas, computing the strongest postcondition of formula 'true'.

==== Weakest preconditions ====

While symbolic execution computes formula by going forward along the program syntax tree, weakest precondition computes formula by going backward.

  wp(Q, x=t) =
  wp(Q, assume F) =
  wp(Q, assert F) =
  wp(Q, c1 [] c2) = 
  wp(Q, c1 ; c2) = 

==== Inferring Loop Invariants ====

Suppose we compute strongest postcondition in a program where we unroll loop k times.
  * What does it denote?  
  * What is its relationship to loop invariant?

Weakening strategies
  * maintain a conjunction
  * drop conjuncts that do not remain true

Alternative:
  * decide that you will only loop for formulas of restricted form, as in abstract interpretation and data flow analysis (next week)





===== One useful decision procedure: Proving quantifier-free linear arithmetic formulas =====

Suppose that we obtain (one or more) verification conditions of the form
\begin{equation*}
 F\ \rightarrow\ (\mbox{error}=\mbox{false})
\end{equation*}

whose validity we need to prove.  We here assume that F contains only linear arithmetic.  Note: we can check satisfiability of $F\ \land\ (\mbox{error}=\mbox{true})$.  We show an algorithm to check this satisfiability.

==== Quantifier Presburger arithmetic ====

Here is the grammar:

  var = x | y | z | ...                    (variables)
  K = ... | -2 | -1 | 0 | 1 | 2 | ...      (integer constants)
  T ::= var | T + T | K * T                (terms)
  A ::= T=T | T <= T                       (atomic formulas)
  F ::= A  |  F & F |  F|F  |  ~F          (formulas)

To get full Presburger arithmetic, allow existential and universal quantifiers in formula as well.

Note: we can assume we have boolean variables (such as 'error') as well, because we can represent them as 0/1 integers.

Satisfiability of quantifier-free Presburger arithmetic is decidable.

Proof: small model theorem.

==== Small model theorem for Quantifier-Free Presburger Arithmetic (QFPA) ====

First step: transform to disjunctive normal form.

Next: reduce to integer linear programming:
\begin{equation*}
  A\vec x = \vec b, \qquad \vec x \geq \vec 0
\end{equation*}
where $A \in {\cal Z}^{m,n}$ and $x \in {\cal Z}^n$.

Then solve integer linear programming (ILP) problem
  * [[wk>Integer Linear Programming]]
  * online book chapter on ILP
  * [[http://www.gnu.org/software/glpk/|GLPK]] tool

We can prove small model theorem for ILP - gives bound on search.

Short proof by {{papadimitriou81complexityintegerprogramming.pdf|Papadimitriou}}:
  * solution of Ax=b (A regular) has as components rationals of form p/q with bounded p,q
  * duality of linear programming
  * obtains bound $M = n(ma)^{2m+1}$, which needs $\log n + (2m+1)\log(ma)$ bits
  * we could encode the problem into SAT: use circuits for addition, comparison etc.

Note: if small model theorem applies to conjunctions, it also applies to arbitrary QFPA formulas.  

Moreover, one can improve these bounds.  One tool based on these ideas is [[http://www.cs.cmu.edu/~uclid/|UCLID]].

Alternative: enumerate disjuncts of DNF on demand, each disjunct is a conjunction, then use ILP techniques (often first solve the underlying linear programming problem over reals).  Many SMT tools are based on this idea (along with Nelson-Oppen combination: next class).
  * [[http://www.cs.nyu.edu/acsys/cvc3/download.html|CVC3]] (successor of CVC Lite)
  * [[http://combination.cs.uiowa.edu/smtlib/|SMT-LIB]] Standard for formulas, competition


==== Full Presburger arithmetic ====

Full Presburger arithmetic is also decidable.

Approaches:
  * Quantifier-Elimination (Omega tool from Maryland) - see homework
  * Automata Theoretic approaches: LASH, MONA (as a special case)

===== Papers =====

  * Verification condition generation in Spec#: http://research.microsoft.com/~leino/papers/krml157.pdf
  * Loop invariant inference for set algebra formulas: {{hob-tcs.pdf}}
  * Induction-iteration method for machine code checking: http://www.cs.wisc.edu/wpis/papers/pldi00.ps

  * Presburger Arithmetic (PA) bounds: {{papadimitriou81complexityintegerprogramming.pdf}}
  * Specializing PA bounds: http://www.lmcs-online.org/ojs/viewarticle.php?id=43&layout=abstract