220 likes | 423 Views
Rahul Sharma and Alex Aiken (Stanford University). From Invariant Checking to Invariant Inference Using Randomized Search. Randomized Search. x = i ; y = j; while y!=0 do x = x-1; y = y-1; if( i ==j ) assert x==0. No! Yes!. Invariants. Heap. Numerical. Arrays. Strings.
E N D
Rahul Sharma and Alex Aiken (Stanford University) From Invariant Checking to Invariant Inference Using Randomized Search
Randomized Search x = i; y = j; while y!=0 do x = x-1; y = y-1; if( i==j ) assert x==0 • No! • Yes!
Invariants Heap Numerical Arrays Strings
Loop Invariant Inference assume P while B do S assert Q • Find that satisfies Find a valuation of unknown predicates that makes the verification conditions (VCs) valid
Inference as Search • Given a set of candidates • Goal is to find a candidate that satisfies the VCs • This problem is hard! • Effective domain specific approaches • Numerical, arrays, linked lists, etc. Is it possible to have a general search procedure?
C2I • (Domain-specific) Checker + (General) Search = Inference • To obtain an invariant inference engine • Instantiate the search with a search space • An SMT solver to check
Main Contribution • A generally applicable randomized search • Numerical, array, linked lists, and strings • Competitive performance with specialized approaches
MCMC Search • Markov Chain Monte Carlo (MCMC) sampling • The only known tractable solution method for high dimensional irregular search spaces [andrieu 03][chenney 00]
MCMC Overview 37 73 47 17 29 42 23
Search • while( ) • Propose a random modification to • if cost decreased then accept • if cost increased then • with some probability accept anyway • return
Cost Function (First Attempt) • Problems • Throughput < 1000 iterations per second • No incremental feedback
Cost Function (Fixed) • Given sets of concrete states • G: some reachable states • B: some bad states • Z: some implications I s b g t Incremental feedback Efficient to evaluate
Concrete States assume P while B do S assert Q • Reachable state • Pair , • Bad state ,
Overall Algorithm • Given G, Z, and B, for the cost function • Run search until a 0-cost candidate is found • SMT solver checks that satisfies all the VCs • If yes, then done • Update G, Z, or B and repeat • SMT solvers can generate counterexamples • If not then generate from executions
LIA Invariants • Program has integral variables • Search space: • Transformations for MCMC: • Update a • Update a • Update all and of a single inequality
Arrays • Fluid updates abstraction of DDA (ESOP’10) • Z3 fails to generate counterexamples • MCMC on this search space times out on ~30% • Restrict search space: handle each in under a second
Lists • Search space: Boolean combinations of atoms • Atoms are relations • Reachability relations • Use EPR (CAV’13) for check
Strings • Operations that intermix strings and integers • length(s), indexOf(s1, s2), substr(s1, i1, i2), … • Search space: Boolean combinations of predicates • Z3-Str (FSE’13) for check
Conclusion • Static invariant inference is a hard problem, made easier by separating search and check • Search based techniques can work • Competitive with other methods • Easier to retarget to new domains • Future work, scale MCMC to full program proofs
Related Work • Pranav Garg, ChristofLöding, P. Madhusudan, Daniel Neider: ICE: A Robust Framework for Learning Invariants. CAV 2014 • ShacharItzhaky, NikolajBjørner, Thomas W. Reps, MoolySagiv, Aditya V. Thakur: Property-Directed Shape Analysis. CAV 2014 • Rajeev Alur, RastislavBodík, GarvitJuniwal, Milo M. K. Martin, MukundRaghothaman, Sanjit A. Seshia, Rishabh Singh, Armando Solar-Lezama, EminaTorlak, Abhishek Udupa: Syntax-guided synthesis. FMCAD 2013 • Ashutosh Gupta, RupakMajumdar, Andrey Rybalchenko: From tests to proofs. STTT 15(4) (2013) • Yungbum Jung, Soonho Kong, Bow-Yaw Wang, Kwangkeun Yi: Deriving Invariants by Algorithmic Learning, Decision Procedures, and Predicate Abstraction. VMCAI 2010 • SumitGulwani, NebojsaJojic: Program verification as probabilistic inference. POPL 2007: 277-289