ALICe: A Framework to Improve Affine Loop Invariant Computation

International audienceA crucial point in program analysis is the computation of loop invariants. Accurate invariants are required to prove properties on a program but they are difficult to compute. Extensive research has been carried out but, to the best of our knowledge, no benchmark has ever been developed to compare algorithms and tools. We present ALICe, a toolset to compare automatic computation techniques of affine loop scalar invariants. It comes with a benchmark that we built using 102 test cases which we found in the loop invariant bibliography, and interfaces with three analysis programs, that rely on different techniques: Aspic, ISL and PIPS. Conversion tools are provided to handle format heterogeneity of these programs. Experimental results show the importance of model coding and the poor performances of PIPS on concurrent loops. To tackle these issues, we use two model restructurations techniques whose correctness is proved in Coq, and discuss the improvements realized