CSE503: Software Engineering
Lecture 11  (January 29, 1999)

David Notkin

 

v     An alternative formulation to Hoare triples is to use Dijkstra's weakest precondition predicate transformers that, in essence, provide a functional approach to handling the predicates

Ø      wp(Q,S) is defined to be the weakest logical predicate that must be true before executing S in order to prove that postcondition Q is true afterwards

Ø      Ex:

§        wp(x > 10,x:=x+1) = x > 9

§        wp(Q,skip) = Q

§        wp(Q,S1;S2) = wp(wp(Q,S2),S1)

§        wp(Q, if C then S1 else S2) = (wp(Q,S1) and C) or (wp(Q,S2) and not C)

Ø      In general, there is a direct analogy between Hoare triples and wp proofs, although on the whole, wp proofs seem a little more straightforward, as they seem to capture the intuitive notion of backwards reasoning for proofs more effectively

v     The last topic in this area of program verification that I want to cover is the basics of how to verify abstract data types

Ø      The basic idea here is that you have an abstract data type A (representing a stack, for instance) and an implementation C of that data type. 

Ø      How do you prove that the C satisfies the specification provided by A?

Ø      The first step is to have a precise definition of the ADT A

ADT Stack is
    private s == <>
                const maxSize;
    Push(x)       { precondition == !Full()
                         postcondition == s' = s || <x> }
    Pop()           { precondition == !Empty()
                         postcondition == s' = <s1..sN-1> }
    Top(): int    { precondition == !Empty()
                          postcondition == Top = sN and s' = s }
     Empty(): bool  { postcondition == Empty = (s = <>) and s' = s }
     Full(): bool  { postcondition == Full = (len(s) = maxSize and s' = s }
end Stack

Ø      We also define an invariant over the abstraction representations

§        In this case, we might say that the abstract invariant is defined as
len(s) <= maxSize

Ø      We also have a concrete implementation of the stack, perhaps using arrays to represent the sequence

package stackImpl is
     private int sArray[1..maxSize];
     private int sTop := 0;
     procedure sPush(int x) is
           if !sFull then
                  sTop := sTop +1;
                  sArray[sTop] := x;
           fi
      end
      procedure…

Ø      We also need to define a concrete invariant, which in this case is
sTop <= maxSize

Ø      In essence, the abstraction, with its specification, describes how to map an abstract stack to another abstract stack (perhaps with a return value of the operation)

Ø      And similarly, the implementation, with its specification (which can be written in Hoare triples or wp's), describes how to map a concrete stack to another concrete stack

Ø      However, to prove anything about the relationship between the abstraction and the implementation requires an explicit abstraction function (AF), which maps a state in the concrete implementation to a state in the abstract representation.

§        That is, it describes how to interpret a state in the concrete machine as a state in the abstract machine

§        For the above example, the abstraction function maps the [1..sTop] elements in sArray to a sequence of elements in the stack's s variable

§        Note that many different concrete states map to the same abstract state

·        sTop = 3 and sArray = [1,2,3,4,5,6] and
sTop = 3 and sArray = [1,2,3,6,5,4] map to the same abstract state:

s = <1,2,3>

§        In general, abstraction functions are many-to-one; this is one reason that the inverse relation that maps abstract to concrete states is less useful in proofs (although it seems sensible on the face of it)

Ø      This allows us to think about the relationship between the two levels of abstraction, with the basics captured in the figure at the end of the notes

Ø      Sketch of proof technique

§        For each initial concrete state i, show that AF(i) satisfies the abstract invariant

·        That is, make sure that any initial concrete state is a legal abstract state

§        For each operation of the concrete implementation and its associated abstract operation, given a legal concrete state c, we need to show
opabs(AF(c)) = AF(opconc(c))

§        Essentially, this is a proof by induction that shows that the concrete operations satisfy the abstract operations

Ø      The basic technique can be used to show satisfaction over any pairs of representations, including state machines, different levels of specifications, etc.

v     OK, let's get back to the question you asked last lecture: who cares about this?

Ø      I'd claim that there are at least two reasons to understand at least the basics of proving programs correct

§        It really does provide a sound basis for understanding the relationship between specifications and implementations, even if it is not realizable for sizeable systems

§        When you are writing or debugging your own code, or perhaps even TAing or helping someone else with their code, it gives you an intellectual tool (which you can and will and probably should use informally most of the time) to help

·        I don't think about loop invariants in 99% of the loops I write, but when I am having a problem, I often informally write down or at least determine what the loop invariant should be, and this often helps

·       

When debugging a system that uses ADTs, the idea of separating out potential problems in (a) the abstract definition, (b) the concrete implementation, and (c) the representation function between them can be useful, too