170 likes | 445 Views
Boosting Concolic Testing via Interpolation. Joxan Jaffar – NUS, Singapore Vijayaraghavan Murali – NUS, Singapore Jorge Navas – Univ. of Melbourne, Australia. Concolic Testing. Our method. Interpolation to attack path-explosion Generate at each point an “interpolant”
E N D
Boosting Concolic Testing via Interpolation JoxanJaffar– NUS, Singapore VijayaraghavanMurali – NUS, Singapore Jorge Navas – Univ. of Melbourne, Australia
Our method • Interpolation to attack path-explosion • Generate at each point an “interpolant” • Succinctly captures why any paththrough the program point is not buggy • When interpolant is implied along a different path, subsume it – guaranteed to not hit the bug • (Potentially) exponential savings!
Example Half interpolant Full interpolant s=0 skip s=s+1 skip skip s=s+2 s=s+2 skip
Points to note • Assumption: “bug conditions” e.g., if (C) throw error • Only nodes with full interpolants capable of subsumption • Half interpolant at a node becomes full interpolant when all paths arising from it have been explored • Catch: We assumed DFS-order • There exist many other heuristics for concolic testing!
Example: non-DFS Half interpolant Full interpolant s=0 skip s=s+1 skip skip s=s+2 s=s+2 Cannot subsume because of half interpolant! skip
Goal: Accelerate full interpolant formation Explore with restriction Tree full of full interpolants Give each PP one chance to be subsumed Demand subsumption! Greedy Confirmation (GC)
Example: with GC Invoke GC at: s=0 skip s=s+1 skip skip s=s+2 s=s+2 skip
Experiments • Evaluate using Directed Automated Random Testing (DART) technique • Interpolation algorithm: TRACER framework for symbolic execution • Concolic tester: Crest, modified to “consult” TRACER for subsumption • Different search heuristics – CFG, Uniform Random Search (URS) and DFS • Benchmarks: device driver programs from SV-COMP
Related work • Boonstoppel et al. TACAS’08 “KLEE”, Dawei et al. FSE’11 • Use analysis information – live/dead, dependencies etc. – to discard certain tests • We use logical information (interpolants) to discard paths • Burnim et al. ASE’08 • Use branch coverage as a measure of quality of tests and try to improve it • Branch coverage typically does not suffer from path explosion, so our focus is path coverage
Conclusion • Mitigate path explosion using interpolation – full and half interpolant framework • Challenge: non-DFS search heuristics interfere with full interpolant formation • Greedy confirmation to accelerate formation of full interpolants • Experimental results prove that our method works in practice