Presentation is loading. Please wait.

Presentation is loading. Please wait.

Checking Validity of Quantifier-Free Formulas in Combinations of First-Order Theories Clark W. Barrett Ph.D. Dissertation Defense Department of Computer.

Similar presentations


Presentation on theme: "Checking Validity of Quantifier-Free Formulas in Combinations of First-Order Theories Clark W. Barrett Ph.D. Dissertation Defense Department of Computer."— Presentation transcript:

1 Checking Validity of Quantifier-Free Formulas in Combinations of First-Order Theories Clark W. Barrett Ph.D. Dissertation Defense Department of Computer Science Stanford University August 2001

2 The Problem: First-Order Logic l First-Order Logic is a mathematical system for making precise statements. l Statements in first-order logic are made up of the following pieces: Variables x, y Constants 0, John, Functions f ( x ), x + y Predicates p ( x ), x > y, x = y Boolean connectives,,, Quantifiers, l Example: Every rectangle is a square x. ( Rectangle ( x ) Square ( x ))

3 The Problem: First-Order Theories l A first-order theory is a set of first-order statements about a related set of constants, functions, and predicates. A theory of arithmetic might include the following statements about 0 and + : x. ( x + 0 = x ) x,y. ( x + y = y + x )

4 The Problem: Validity l An expression is valid if every possible way of interpreting it results in a true statement. x = x p ( x ) x = y f ( x ) = f ( y ) f ( x ) = f ( y ) x = y Valid Invalid l An expression is valid in a theory if every possible way of interpreting it in that theory results in a true statement. x 0 l An expression is valid in a theory if every possible way of interpreting it in that theory results in a true statement. x 0 Invalid in the theory of real arithmetic l An expression is valid in a theory if every possible way of interpreting it in that theory results in a true statement. x 0 Valid in positive real arithmetic

5 The Problem: Validity Checking Suppose T is a first-order theory and is a first-order formula We write T = as an abbreviation for is valid in T A classical result in Computer Science states that in general, the question of whether T = is undecidable. It is impossible to write a program that can always figure out whether T = However, given appropriate restrictions on T and, a program can automatically decide T = We consider theories T such that T = is decidable when is quantifier-free.

6 Motivation l Many interesting and practical problems can be solved by checking the validity of a formula in some theory. l As evidence of this claim, consider the following widely- used tools tools which include decision procedures for checking validity u PVS [Owre et al. 92] u STeP [Manna et al. 96, Bjørner 99] u ESC [Detlefs et al. 98] u Mona [Klarlund and Møller 98] u SVC [Barrett et al. 96]

7 The SVC Story l Roots in processor verification u [Burch and Dill 94] u [Jones et al. 95] l Internal use at Stanford u Symbolic simulation [Su et al. 98] u Software specification checking [Park et al. 98] u Infinite-state model checking [Das and Dill 01] l External use since public release in 1998 u Model Checking [Boppana et al. 99] u Theorem prover proof assistance [Heilmann 99] u Integration into programming languages [Day et al. 99] u Many others

8 The SVC Story l Despite its success, SVC has many limitations u Gaps in theoretical understanding u Outgrown its original software architecture u Unnecessarily slow performance in some cases l This thesis is the result of ongoing efforts to address these limitations. u New contributions to underlying theory u A flexible and efficient implementation u Techniques for faster and more robust performance

9 Outline l Validity Checking Overview u The Problem u Motivation u The SVC Story u Top-Level Algorithm l Methods for Combining Theories l Implementation l Adapting Techniques from Propositional Satisfiability l Contributions and Conclusions

10 Top-Level Algorithm l Consider the following formula in the theory of arithmetic x > y y > x x = y l Step 1: Choose an atomic formula l Step 2: Consider two cases: Replace the atomic formula with true Replace the atomic formula is with false l Step 3: Simplify true y > x x = yfalse y > x x = y true y > x x = y

11 Top-Level Algorithm l Consider the following formula in the theory of arithmetic x > y y > x x = y true y > x x = yfalse y > x x = y true y > x x = y true x = y true false x y y x x y This formula is unsatisfiable

12 Validity Checking Overview l A literal is an atomic formula or its negation The validity checker is built on top of a core decision procedure for satisfiability in T of a set of literals. l The method for checking satisfiability will vary greatly depending on the theory in question l The most powerful technique for producing a satisfiability procedure is by combining other satisfiability procedures

13 Outline l Validity Checking Overview l Methods for Combining Theories u The Problem u Shostaks Method u The Nelson-Oppen Method u A Combined Method l Implementation l Adapting Techniques from Propositional Satisfiability l Contributions and Conclusions

14 The Problem l Consider the following theories: Real linear arithmetic: +, -,0,1,…, Arrays: s [ i ], update ( s, i, v ) Uninterpreted functions and predicates: f ( x ), p ( x ),… l And the following set of literals in the combined theory: p ( y ) s = update ( t, i, 0 ) x - y - z = 0 z + s [ i ] = f ( x - y ) p ( x - f ( f ( z ) ) ) l Consider the following theories: Real linear arithmetic: +, -,0,1,…, Arrays: s [ i ], update ( s, i, v ) Uninterpreted functions and predicates: f ( x ), p ( x ),… l And the following set of literals in the combined theory: p ( y ) s = update ( t, i, 0 ) x - y - z = 0 z + s [ i ] = f ( x - y ) p ( x - f ( f ( z ) ) ) l Consider the following theories: Real linear arithmetic: +, -,0,1,…, Arrays: s [ i ], update ( s, i, v ) Uninterpreted functions and predicates: f ( x ), p ( x ),… l And the following set of literals in the combined theory: p ( y ) s = update ( t, i, 0 ) x - y - z = 0 z + s [ i ] = f ( x - y ) p ( x - f ( f ( z ) ) ) l Consider the following theories: Real linear arithmetic: +, -,0,1,…, Arrays: s [ i ], update ( s, i, v ) Uninterpreted functions and predicates: f ( x ), p ( x ),… l And the following set of literals in the combined theory: p ( y ) s = update ( t, i, 0 ) x - y - z = 0 z + s [ i ] = f ( x - y ) p ( x - f ( f ( z ) ) ) l Consider the following theories: Real linear arithmetic: +, -,0,1,…, Arrays: s [ i ], update ( s, i, v ) Uninterpreted functions and predicates: f ( x ), p ( x ),… l And the following set of literals in the combined theory: p ( y ) s = update ( t, i, 0 ) x - y - z = 0 z + s [ i ] = f ( x - y ) p ( x - f ( f ( z ) ) ) l Question: Given a method to decide satisfiability of literals in each theory, how do we decide the satisfiability of literals in the combined theory? l Two main approaches, each with advantages and disadvantages u Shostak [Shostak 84] u Nelson-Oppen [Nelson and Oppen 79]

15 Shostaks Method l Has formed an ongoing strand of research u Originally published in 1984 [Shostak 84] u Several clarifying papers since then u [Cyrluk et al. 96] u [Ruess and Shankar 01] l Used in several automated deduction systems u PVS, STeP, SVC l Unfortunately, remains difficult to understand u Details are nonintuitive u Simple proof of correctness has been especially elusive l Contribution : A new presentation of a key subset of Shostaks original algorithm.

16 Shostaks Method: Canonizer l There are two main components in a Shostak satisfiability procedure: the canonizer and the solver. l The canonizer rewrites terms into a unique form T = a = b canon ( a ) = canon ( b ) l Example: canonizer for linear arithmetic u Combines like terms canon ( x + x ) = 2 x u Imposes an ordering on the variables canon ( y + x ) = x + y

17 Shostaks Method: Solver A set of equations E is said to be in solved form if the left- hand side of each equation is a variable which appears only once in E in solved formnot in solved form x = y + zx = y + z w = z - aw = z + x v = 3 y + b 2 v = 3 y + b S means replace each left-hand side variable occurring in S with its corresponding right-hand side E ( w + x + y + z ) = z - a + y + z + y + z

18 Shostaks Method: Solver l The solver transforms an equation into an equisatisfiable set of equations in solved form If T = a b, then solve ( a = b ) = { false } u Otherwise: solve ( a = b ) = a set of equations E in solved form T = ( a = b x. E ) x is a set of fresh variables appearing in E, but not in a or b. l Example: solver for real linear arithmetic solve ( x - y - z = 0 ) = { x = y + z } solve ( x + 1 = x - 1 ) = { false }

19 The Simplified Algorithm Given a set of equations and disequations Step 1: Use the solver to convert into an equisatisfiable set of equations E in solved form u Use a generalization of Gaussian elimination with back substitution

20 Choose matrix row The Simplified Algorithm Given a set of equations and disequations Step 1: Use the solver to convert into an equisatisfiable set of equations E in solved form Select an equation from Apply E as a substitution to Solve to get E Apply E as a substitution to E Add E to E - x - 3 y + 2 z = - 1 x - y - 6 z = 1 2 x + y - 10 z = 3 E

21 The Simplified Algorithm Given a set of equations and disequations Step 1: Use the solver to convert into an equisatisfiable set of equations E in solved form Select an equation from Apply E as a substitution to Solve to get E Apply E as a substitution to E Add E to E - x - 3 y + 2 z = - 1 x - y - 6 z = 1 2 x + y - 10 z = 3 E Apply previous rows Choose matrix row

22 The Simplified Algorithm Given a set of equations and disequations Step 1: Use the solver to convert into an equisatisfiable set of equations E in solved form Select an equation from Apply E as a substitution to Solve to get E Apply E as a substitution to E Add E to E x = - 3 y + 2 z + 1 x - y - 6 z = 1 2 x + y - 10 z = 3 E Apply previous rows Make pivot 1 Choose matrix row Apply to previous rows

23 The Simplified Algorithm Given a set of equations and disequations Step 1: Use the solver to convert into an equisatisfiable set of equations E in solved form Select an equation from Apply E as a substitution to Solve to get E Apply E as a substitution to E Add E to E x - y - 6 z = 1 2 x + y - 10 z = 3 E Apply previous rows Make pivot 1 Choose matrix row Apply to previous rows x = - 3 y + 2 z + 1

24 The Simplified Algorithm Given a set of equations and disequations Step 1: Use the solver to convert into an equisatisfiable set of equations E in solved form Select an equation from Apply E as a substitution to Solve to get E Apply E as a substitution to E Add E to E - 3 y + 2 z + 1 - y - 6 z = 1 2 x + y - 10 z = 3 E Apply previous rows Make pivot 1 Choose matrix row Apply to previous rows x = - 3 y + 2 z + 1

25 The Simplified Algorithm Given a set of equations and disequations Step 1: Use the solver to convert into an equisatisfiable set of equations E in solved form Select an equation from Apply E as a substitution to Solve to get E Apply E as a substitution to E Add E to E y = - z 2 x + y - 10 z = 3 E Apply previous rows Make pivot 1 Choose matrix row Apply to previous rows x = - 3 y + 2 z + 1

26 The Simplified Algorithm Given a set of equations and disequations Step 1: Use the solver to convert into an equisatisfiable set of equations E in solved form Select an equation from Apply E as a substitution to Solve to get E Apply E as a substitution to E Add E to E y = - z 2 x + y - 10 z = 3 E Apply previous rows Make pivot 1 Choose matrix row Apply to previous rows x = - 3( - z ) + 2 z + 1

27 The Simplified Algorithm Given a set of equations and disequations Step 1: Use the solver to convert into an equisatisfiable set of equations E in solved form Select an equation from Apply E as a substitution to Solve to get E Apply E as a substitution to E Add E to E 2 x + y - 10 z = 3 E Apply previous rows Make pivot 1 Choose matrix row Apply to previous rows x = 5 z + 1 y = - z

28 The Simplified Algorithm Given a set of equations and disequations Step 1: Use the solver to convert into an equisatisfiable set of equations E in solved form Select an equation from Apply E as a substitution to Solve to get E Apply E as a substitution to E Add E to E 2(5 z + 1) + ( - z ) - 10 z = 3 E Apply previous rows Make pivot 1 Choose matrix row Apply to previous rows x = 5 z + 1 y = - z

29 The Simplified Algorithm Given a set of equations and disequations Step 1: Use the solver to convert into an equisatisfiable set of equations E in solved form Select an equation from Apply E as a substitution to Solve to get E Apply E as a substitution to E Add E to E z = - 1 E Apply previous rows Make pivot 1 Choose matrix row Apply to previous rows x = 5 z + 1 y = - z

30 The Simplified Algorithm Given a set of equations and disequations Step 1: Use the solver to convert into an equisatisfiable set of equations E in solved form Select an equation from Apply E as a substitution to Solve to get E Apply E as a substitution to E Add E to E z = - 1 E Apply previous rows Make pivot 1 Choose matrix row Apply to previous rows x = 5( - 1) + 1 y = - ( - 1)

31 The Simplified Algorithm Given a set of equations and disequations Step 1: Use the solver to convert into an equisatisfiable set of equations E in solved form Select an equation from Apply E as a substitution to Solve to get E Apply E as a substitution to E Add E to E E Apply previous rows Make pivot 1 Choose matrix row Apply to previous rows x = - 4 y = 1 z = - 1

32 The Simplified Algorithm Given a set of equations and disequations Step 1: Use the solver to convert into an equisatisfiable set of equations E in solved form u Step 2: Use this set of equations together with the canonizer to check if any disequality is violated For each a b Check if canon ( E ( a ) ) = canon ( E ( b ) ) E x = - 4 y = 1 z = - 1 2 y - 10 x 6( z - 2 x ) 2(1)-10(-4) 6(-1-2(-4)) 42 42

33 The Simplified Algorithm Given a set of equations and disequations Step 1: Use the solver to convert into an equisatisfiable set of equations E in solved form u Step 2: Use this set of equations together with the canonizer to check if any disequality is violated For each a b Check if canon ( E ( a ) ) = canon ( E ( b ) ) E x = 5 z + 1 y = - z 1 - 4 y x - z 1 - 4( - z ) (5 z + 1) - z 4 z + 1

34 The Simplified Algorithm Given a set of equations and disequations Step 1: Use the solver to convert into an equisatisfiable set of equations E in solved form u Step 2: Use this set of equations together with the canonizer to check if any disequality is violated For each a b Check if canon ( E ( a ) ) = canon ( E ( b ) ) l Technical detail: u If there is more than one disequality, the theory must be convex

35 Shostaks Method: Combining Theories l In what sense is this algorithm a method for combining theories? Two Shostak theories T 1 and T 2 can often be combined to form a new Shostak theory T = T 2 T 2 Compose canonizers: canon = canon 1 o canon 2 u Often, solvers can also be combined u Treat terms from other theory as variables u Repeatedly apply solvers from each theory until resulting set of equations is in solved form

36 Shostaks Method: Contributions l Shostaks original algorithm is much more complicated because it includes a decision procedure for the theory of pure equality with uninterpreted functions l Why is the simplified version a contribution? u Can be applied directly to produce decision procedures, even combinations of decision procedures u Much easier to understand and prove correct u Provides intuition for understanding the original algorithm u Provides the foundation for a generalization of the original Shostak method based on a variation of Nelson-Oppen

37 Nelson-Oppen l Developed for the Stanford Pascal Verifier u [Nelson and Oppen 79] u [Nelson 80, Oppen 80] l Tinelli and Harandi discovered a new (simpler) proof and an important optimization u [Tinelli and Harandi 96] l Used in real systems u ESC u EHDM [von Henke et al. 88] u Vampyre [ http://www-cad.eecs.berkeley.edu/~rupak/Vampyre ]

38 Nelson-Oppen l Unlike Shostak, Nelson-Oppen does not impose a specific strategy on individual theories u Instead of a solver and canonizer, u Each theory provides a complete satisfiability procedure u Technical detail: Each theory must be stably infinite l There are two phases in the version of Nelson-Oppen presented by Tinelli and Harandi u Purification phase u Check phase

39 Nelson-Oppen: Purification Phase l Transform a set of literals in a combined theory to an equisatisfiable set of literals such that each literal is pure : u A pure literal contains symbols from only a single theory l Consider again the following set of literals in a combined theory of arithmetic, arrays, and uninterpreted functions p ( y ) s = update ( t, i, 0 ) x - y - z = 0 z + s [ i ] = f ( x - y ) p ( x - f ( f ( z ) ) ) j = 0

40 Nelson-Oppen: Purification Phase l Transform a set of literals in a combined theory to an equisatisfiable set of literals such that each literal is pure : u A pure literal contains symbols from only a single theory l Consider again the following set of literals in a combined theory of arithmetic, arrays, and uninterpreted functions p ( y ) s = update ( t, i, j ) x - y - z = j z + s [ i ] = f ( x - y ) p ( x - f ( f ( z ) ) ) j = 0 k = s [ i ]

41 Nelson-Oppen: Purification Phase l Transform a set of literals in a combined theory to an equisatisfiable set of literals such that each literal is pure : u A pure literal contains symbols from only a single theory l Consider again the following set of literals in a combined theory of arithmetic, arrays, and uninterpreted functions p ( y ) s = update ( t, i, j ) x - y - z = j z + k = f ( x - y ) p ( x - f ( f ( z ) ) ) j = 0 k = s [ i ] j = 0 k = s [ i ] l = x - y m = z + k

42 Nelson-Oppen: Purification Phase l Transform a set of literals in a combined theory to an equisatisfiable set of literals such that each literal is pure : u A pure literal contains symbols from only a single theory l Consider again the following set of literals in a combined theory of arithmetic, arrays, and uninterpreted functions p ( y ) s = update ( t, i, j ) l - z = j m = f ( l ) p ( x - f ( f ( z ) ) ) j = 0 k = s [ i ] l = x - y m = z + k

43 Nelson-Oppen: Purification Phase l Transform a set of literals in a combined theory to an equisatisfiable set of literals such that each literal is pure : u A pure literal contains symbols from only a single theory l Consider again the following set of literals in a combined theory of arithmetic, arrays, and uninterpreted functions j = 0 k = s [ i ] l = x - y m = z + k n = f ( f ( z ) ) ) v = x - n p ( y ) s = update ( t, i, j ) l - z = j m = f ( l ) p ( v )

44 Nelson-Oppen: Purification Phase l Transform a set of literals in a combined theory to an equisatisfiable set of literals such that each literal is pure : u A pure literal contains symbols from only a single theory l Consider again the following set of literals in a combined theory of arithmetic, arrays, and uninterpreted functions s = update ( t, i, j ) k = s [ i ] p ( y ) m = f ( l ) p ( v ) n = f ( f ( z ) ) ) l - z = j j = 0 l = x - y m = z + k v = x - n

45 Nelson-Oppen: Check Phase Definitions l Shared variables are variables that appear in literals from more than one theory Shared: l, z, j, y, m, k, v, n Unshared: x, s, t, i s = update ( t, i, j ) k = s [ i ] p ( y ) m = f ( l ) p ( v ) n = f ( f ( z ) ) ) l - z = j j = 0 l = x - y m = z + k v = x - n l An arrangement of a set is a set of equalities that partitions the set into equivalence classes Suppose S = { a, b, c } Some arrangements of S u { a b, a c, b c }{ { a }, { b }, { c } } { a = b, a c, b c }{ { a, b }, { c } } { a = b, a = c, b = c }{ { a, b, c } }

46 Nelson-Oppen: Check Phase Choose an arrangement A of the shared variables For each theory, check if the set of literals pure in that theory together with the arrangement A is satisfiable l If an arrangement exists that is compatible with each set of literals, then the original set of literals is satisfiable in the combined theory Arrays s = update ( t, i, j ) k = s [ i ] Uninterpreted p ( y ) m = f ( l ) p ( v ) n = f ( f ( z ) ) ) Arithmetic l - z = j j = 0 l = x - y m = z + k v = x - n A ( l, z, j, y, m, k, v, n )

47 Nelson-Oppen: A Variation l Contribution : A Variation of Nelson-Oppen u The purification phase can be eliminated u Instead, simply partition the formulas according to the outer-most symbol p ( y ) s = update ( t, i, 0 ) x - y - z = 0 z + s [ i ] = f ( x - y ) p ( x - f ( f ( z ) ) ) Arithmetic x - y - z = 0 z + s [ i ] = f ( x - y ) Arrays s = update ( t, i, 0 ) Uninterpreted p ( y ) p ( x - f ( f ( z ) ) )

48 Nelson-Oppen: A Variation l Contribution : A Variation of Nelson-Oppen u The purification phase can be eliminated u Instead, simply partition the formulas according to the outer-most symbol Choose an arrangement A of the shared terms which appear in a term or formula belonging to another theory u For each theory, check if the set of literals assigned to that theory together with the arrangement is satisfiable u Terms with foreign symbols are treated as variables Arithmetic x - y - z = 0 z + s [ i ] = f ( x - y ) Arrays s = update ( t, i, 0 ) Uninterpreted p ( y ) p ( x - f ( f ( z ) ) ) A ( s [ i ], x - y, f ( x - y ), 0, y, z, f ( f ( z ) ), x - f ( f ( z ) ) )

49 Nelson-Oppen: A Variation l Contribution : A Variation of Nelson-Oppen u The purification phase can be eliminated u Instead, simply partition the formulas according to the outer-most symbol Choose an arrangement A of the shared terms which appear in a term or formula belonging to another theory u For each theory, check if the set of literals assigned to that theory together with the arrangement is satisfiable u Terms with foreign symbols are treated as variables l Contributions of this variation u Fewer formulas given to each theory u Easier to implement u Easier to combine with Shostak

50 Combining Shostak and Nelson-Oppen l Theory requirements u Shostak requires convexity u Nelson-Oppen requires stable-infiniteness u Contribution : The following theorem relates the two Every convex first-order theory with no trivial models is stably-infinite u The proof is based on first-order compactness u Note: if a convex theory does admit trivial models, it can usually be modified to include the non-triviality axiom: x,y. x y

51 Combining Shostak and Nelson-Oppen l Contribution : An algorithm for combining the two methods Equalities are processed according to the Shostak algorithm to get a set of equalities E in solved form l All literals are partitioned as in the Nelson-Oppen variation The key idea is to consider the partial arrangement induced on the shared terms S by canon and E : A = : { a = b a, b S canon ( E ( a )) = canon ( E ( b )) } An arrangement A is chosen as in the Nelson-Oppen variation, but this arrangement must include A = This arrangement is automatically consistent with E u The non-Shostak theories are checked for consistency with the arrangement as before

52 Outline l Validity Checking Overview l Methods for Combining Theories l Implementation l Adapting Techniques from Propositional Satisfiability l Contributions and Conclusions

53 Implementation: Approach l Based on Nelson-Oppen and Shostak combination l Online algorithm l Optimizations A Union-Find data structure and an Update List are used to efficiently keep track of both E and A simultaneously u Simplify phase added u Each new formula is simplified u Enables rewrites that can reduce the number of shared terms l Flexible theory interface u Accommodates Nelson-Oppen theories, Shostak theories, and more

54 Implementation: Interface l Recall the top-level algorithm x > y y > x x = y l Choose an atomic formula l Consider two cases: u Add to the set of choices made and simplify Add to the set of choices made and simplify l Repeat until formula is true or set of choices is unsatisfiable Interface from top-level : AddFact, Simplify, Satisfiable true y > x x = yfalse y > x x = y true y > x x = y

55 AddFact Simplify Theory-specific code Top-level code Assert Equalities Rewrite Assert Formula Setup Term Satisfiable RewriteSolveUpdate AssertSetup AddSharedTerm CheckSat Satisfiable

56 AddFact Simplify Top-level code Assert Equalities Rewrite Assert Formula Setup Term Satisfiable p ( y ), s = update ( t, i, 0), x - y - z = 0, z + s [ i ] = f ( x - y ), p ( x - f ( f ( z ))) Uninterpreted Arrays Arithmetic (Shostak) E p ( y ) y y Update List

57 AddFact Simplify Top-level code Assert Equalities Rewrite Assert Formula Setup Term Satisfiable p ( y ), s = update ( t, i, 0), x - y - z = 0, z + s [ i ] = f ( x - y ), p ( x - f ( f ( z ))) Uninterpreted Arrays Arithmetic (Shostak) E s = update ( t, i, 0) 0 y p ( y ) 0 s = update ( t, i, 0) s =... Update List

58 AddFact Simplify Top-level code Assert Equalities Rewrite Assert Formula Setup Term Satisfiable p ( y ), s = update ( t, i, 0), x - y - z = 0, z + s [ i ] = f ( x - y ), p ( x - f ( f ( z ))) Uninterpreted Arrays Arithmetic (Shostak) E Update List x - y - z = 0 y p ( y ) 0 s = update ( t, i, 0) x = y + z s = update ( t, i, 0) x - y - z = 0 x = y + z y + z x =... y + z

59 AddFact Simplify Top-level code Assert Equalities Rewrite Assert Formula Setup Term Satisfiable p ( y ), s = update ( t, i, 0), x - y - z = 0, z + s [ i ] = f ( x - y ), p ( x - f ( f ( z ))) Uninterpreted Arrays Arithmetic (Shostak) E Update List z + s [ i ] = f ( x - y ) y p ( y ) 0 s = update ( t, i, 0) z = f ( z ) s = update ( t, i, 0) z = f ( z ) x = y + z y + z z=f (z)z=f (z) f (z)f (z) z + s [ i ] =... s[i]s[i] s[i]s[i] 0 0 z = f ( z ) z f ( z ) z = f ( z ) x - y z z z = f ( z )

60 AddFact Simplify Top-level code Assert Equalities Rewrite Assert Formula Setup Term Satisfiable p ( y ), s = update ( t, i, 0), x - y - z = 0, z + s [ i ] = f ( x - y ), p ( x - f ( f ( z ))) Uninterpreted Arrays Arithmetic (Shostak) E Update List p ( x - f ( f ( z ))) y p ( y ) 0 s = update ( t, i, 0) x = y + f ( z ) y + z = y + f ( z ) p ( x - …) z z f (z)f (z) f (z)f (z) p ( y ) z f ( z ) z = f ( z ) f ( f ( z )) f ( z ) z = f ( z ) x - f ( z ) y y p ( y ) Satisfiable

61 Implementation: Contributions l Better implementation of Nelson-Oppen u Online algorithm u Each theory only needs to consider a subset of the shared terms u Simplify phase u Can reduce number of shared terms u Equality reasoning is only done once u Simple algorithm with detailed proof u Flexible theory interface l Combined with Shostak u Generalizes original Shostak algorithm Efficient: same data structure for E and A

62 Outline l Validity Checking Overview l Methods for Combining Theories l Implementation l Adapting Techniques from Propositional Satisfiability u The Problem u Combining with SAT u Results l Contributions and Conclusions

63 The Problem l Recall the top-level algorithm x > y y > x x = y l Choose an atomic formula l Consider two cases: u Add to the set of choices made and simplify Add to the set of choices made and simplify l Repeat until formula is true or set of choices is unsatisfiable true y > x x = yfalse y > x x = y true y > x x = y

64 The Problem l The choice of which atomic formula to try next can make a dramatic difference in performance l SVC includes clever heuristics that improve performance significantly l We are convinced that better performance is possible u Equivalent formulas can vary significantly in performance u Research in a related area, Boolean satisfiability (SAT), has advanced significantly l Strategy : Find a way to apply SAT techniques to first-order validity checking

65 Combining with SAT: Approach l Generate SAT problem from validity-checking problem u Negate the formula whose validity is in question u Extract Boolean structure from resulting formula u Convert to CNF [Larabee 92] u Run SAT on converted formula l If SAT reports unsatisfiabile, the formula is valid l The inverse is not true u A satisfying assignment must be checked for first-order consistency

66 Combining with SAT: Initial Results l Implementation u GRASP SAT engine [Silva 96] u SVC2 l Initial results were disappointing u Examples of interest could not be proved by just considering Boolean structure u SAT techniques do not compensate for the loss of information resulting from translation to SAT l Idea : u Incrementally give SAT more information

67 Combining with SAT: Conflict Clauses l A conflict clause captures a minimal set of decisions that lead to a conflict and keeps SAT from ever making the same set of choices true y > x x yfalse y > x x y true x y f ( x ) = f ( y ) y > x x y true false true y > x x y Unsatisfiable f ( x ) f ( y ) y x x = y

68 Combining with SAT: Conflict Clauses l How do we get a conflict clause from the first-order satisfiability algorithm u Using all decisions too slow u Black-box minimization methods too slow l Solution : Use proof-production! u Aaron Stump has extended several SVC decision procedures to produce a proof for every result deduced u By looking at what assumptions are used in a proof of inconsistency, a conflict clause can be obtained

69 Results

70 Results: Preliminary Conclusions l Naïve approach does not work well l Adding conflict clauses results in dramatic speed-ups on several examples l Most helpful on formulas with more Boolean structure l Still more work to be done u Find out source of performance problems u Compare to related work u [Goel et al. 98] u [Bryant et al. 99]

71 Outline l Validity Checking Overview l Methods for Combining Theories l Implementation l Adapting Techniques from Propositional Satisfiability l Contributions and Conclusions

72 Thesis Contributions l A new presentation of the core of Shostaks algorithm u Easier to understand and prove correct u Can be applied directly to produce decision procedures u Forms the foundation of a generalization l A new variation of Nelson-Oppen u Eliminates purification phase u Fewer formulas given to each theory u Easier to implement u Easier to combine with Shostak l A new algorithm combining Shostak and Nelson-Oppen u Theoretical result relating convex and stable-infinite u Generalization of Shostaks original method

73 Thesis Contributions l A detailed and provably correct implementation u Online u Optimized to eliminate redundant equality reasoning u Optimized to reduce number of shared terms u Flexible theory API l Faster search by combining with SAT u Methodology and implementation for extracting CNF u Better performance via conflict clauses u Conflict clauses from proofs (with Aaron Stump) u Dramatic improvements on several examples

74 Future Work l Relaxing restrictions on theories and formulas u Non-disjoint signatures u Non-stably-infinite theories u Formulas with quantifiers l Individual Theories u Efficient implementation for Presburger arithmetic u Better techniques for accommodating third-party decision procedures l SAT u Understand cases where combination with SAT fails

75 Acknowledgements l Advisor: David Dill l Orals Committee: John Gill, Zohar Manna, John Mitchell, Natarajan Shankar l Stanford Associates: Aaron Stump, Jeremy Levitt, Satyaki Das, Jeffrey Xsu, Robert Jones, Vijay Ganesh, Kanna Shimizu, Husam Abu-Haimed, Jens Skakkebæk, David Park, Shankar Govindaraju, Madan Musuvathi, Chris Wilson l Others: Cesare Tinelli l SVC Users l Personal: Friends and family

76 Validity Checking Overview l Top-level Algorithm CheckValid(h,c) IF c = true THEN RETURN TRUE; IF !Satisfiable(h) THEN RETURN FALSE; IF c = false THEN RETURN FALSE; subgoals := ApplyTactic(h,c); FOREACH (h,c) in subgoals DO IF !CheckValid(h,c) THEN RETURN FALSE; RETURN TRUE; CheckValid(h,c) IF c = true THEN RETURN TRUE; IF !Satisfiable(h) THEN RETURN FALSE; IF c = false THEN RETURN FALSE; subgoals := ApplyTactic(h,c); FOREACH (h,c) in subgoals DO IF !CheckValid(h,c) THEN RETURN FALSE; RETURN TRUE; ApplyTactic(h,c) Let e be an atomic formula appearing in c; h1 := AddFact(h,e); c1 := Simplify(h1,c); h2 := AddFact(h,!e); c2 := Simplify(h2,c); RETURN {(h1,c1),(h2,c2)}; If CheckValid( T, ) = TRUE, then T =

77 Shostaks Method: Convexity l A set of literals S is convex in a theory T if T S does not entail any disjunction of equalities without entailing one of the equalities itself l A theory T is convex if every set of literals in the language of T is convex in T

78 Shostaks Method: Requirements on T l Shostak Theory T u Signature of T contains no predicate symbols u T is convex Canonizer such that a,b. T = a =b iff a = b Solver such that if T = a b, then a =b { false } u Otherwise: a =b = a set of equations E in solved form u T = a =b x. E, where x is the set of variables appearing in E, but not in a or b. u The variables in x are guaranteed to be fresh.

79 The Simplified Algorithm Given a set of equations and disequations Step 1: Use the solver to convert into an equisatisfiable set of equations E in solved form u Step 2: Use this set of equations together with the canonizer to check if any disequality is violated Suppose a b canon ( E ( a ) ) = canon ( E ( b ) ) T = E ( a ) = E ( b ) T E = a = b T E { a b } is unsatisfiable l Technical detail: The method is complete only for convex theories

80 Shostaks Method: The Algorithm Shostak,,, := ; WHILE DO BEGIN Remove some equality a = b from ; Let a:= a and b:= b ; Let := a= b ; IF = false THEN RETURN FALSE; Let := U ; END IF a = b for some a b in THEN RETURN FALSE; ELSE RETURN TRUE; Shostak(,,, ) = TRUE iff is satisfiable in T

81 Nelson-Oppen: Definitions l Theories must be stably-infinite u A theory T is stably-infinite if every quantifier-free formula is satisfiable in T iff it is satisfiable in an infinite model of T l Terminology for combinations of theories u Theories T 1, T 2, … T n with signatures 1, 2, … n u As with Shostak, signatures must be disjoint u Members of i are called i-symbols u An expression containing only i-symbols is called pure u An i-term is a constant i-symbol, an application of a functional i-symbol, or an i-variable u Each variable is associated arbitrarily with a theory

82 Nelson-Oppen: Definitions l Terminology for combinations of theories (continued) u An i-predicate is the application of a predicate i-symbol u An atomic i-formula is an i-predicate or an equation whose left-hand side is an i-term u An i-literal is an atomic i-formula or its negation u An occurrence of a term is i-alien if it is a j-term (i j) and all its super-terms are i-terms l If S is a set of terms, then an arrangement of S is a set of equations and disequations induced by a partition of S u S = { a, b, c } u Partition P = { { a, b }, { c } } u Arrangement : { a = b, a c, b c }

83 Nelson-Oppen: Purification Phase NO-Purify( ) WHILE != DO BEGIN Let be some i-literal in ; IF is pure THEN Remove from ; i := i U { }; ELSE Let t be an i-alien j-term in ; Replace every occurrence of t in with a new j-variable z; := U { j = t }; ENDIF END RETURN 1^…^ n; is satisfiable in T iff 1 ^ 2 ^ … n is satisfiable in T

84 Nelson-Oppen: Check Phase NO-Check( 1,... n,Sat1,…,Satn) Let S be the set of variables which appear in more than one i; Let A be an arrangement of S; sat := TRUE; FOREACH i DO BEGIN sat := sat ^ Sati( i^A); END RETURN sat; l The second step is non-deterministic 1 ^ 2 ^ … n is satisfiable in T iff it is possible for NO-Check to return TRUE l If the theories are convex, the algorithm can be determinized inexpensively

85 Nelson-Oppen: A Variation l The purification phase can be eliminated l S is a set of terms rather than a set of variables l In calls to Sat i, i-alien terms are treated as variables NO-Check(,Sat 1,…,Sat n ) Let S be the set of terms which are i-alien in either an i-literal or an i-term in ; Let A be an arrangement of S; sat := TRUE; FOREACH set of i-literals i in DO BEGIN sat := sat ^ Sat i ( i^A); END RETURN sat;

86 Combining Shostak and Nelson-Oppen NO-Shostak(,,,Sat NO ) Let S be the set of shared terms; Let be the 1-equalities, the 1-disequalities, and NO the 2-literals in ; := ; LOOP BEGIN IF !Sat NO ( NO ^A = ) THEN RETURN FALSE; ELSE IF !Sat NO ( NO ^A) THEN Choose a,b from S such that T 2 NO A |= a=b, but a=b A = ELSE IF = THEN BREAK; ELSE Remove some equality a = b from ; Let a:= (a) and b:= (b); Let := (a= b); IF = {false} THEN RETURN FALSE; Let := ( ) U ; END IF A THEN RETURN TRUE; ELSE RETURN FALSE;

87 Combining Shostak and Nelson-Oppen NO-Shostak(,, ) := ;S := ; LOOP BEGIN IF t1=f(x 1,…,x n ), t 2 =f(y1,…,yn) with t1,t2 in S and norm(xi)=norm(yi) but norm(t1) != norm(t2) THEN a := t1, b := t2; ELSE IF = THEN RETURN TRUE; ELSE Remove some equality a = b from ; Let a:= can(a) and b:= can(b); Add each sub-term of a,b to S; Let := (a= b); IF = {false} THEN RETURN FALSE; Let := ( ) U ; END RETURN TRUE;

88 Individual Theories l SVC contains decision procedures for a number of individual theories u Pure equality with uninterpreted functions u Real linear arithmetic u Arrays u Bit-vectors u Records l In our efforts to revisit and improve these decision procedures, a number of interesting issues were uncovered u Finite domains u Strategies for arithmetic

89 Finite Domains l Theoretical technicalitiy u Cannot directly combine a theory with only finite models u Not stably-infinite u Union of theories likely to actually be inconsistent u Solution: Form an extended theory whose relativized reduct with respect to a new predicate P is the theory with a finite domain. l Implementation strategy for nonconvexity u Keep track of the terms for which P holds u Use graph coloring to determine satisfiability

90 Arithmetic l Suppose we want to handle linear arithmetic formulas with mixed variable types: some real and some integer. l One approach is the following: u Split weak inequalities into the disjunction of an equation and a strong inequality u Use Shostak-style solver to eliminate all equations that can be solved for a real variable u Use Fourier-Motzkin techniques to eliminate all real variables from inequalities u Eliminate disequalities which can be solved for a real variable u Whats left can be done with Presburger decision procedures

91 Math symbols ()


Download ppt "Checking Validity of Quantifier-Free Formulas in Combinations of First-Order Theories Clark W. Barrett Ph.D. Dissertation Defense Department of Computer."

Similar presentations


Ads by Google