EN FR
EN FR


Section: Research Program

Numerical Programs

In recent years, we demonstrated our capability towards specifying and proving properties of floating-point programs, properties which are both complex and precise about the behavior of those programs: see the publications [70] , [119] , [66] , [114] , [69] , [65] , [106] , [104] but also the web galleries of certified programs at our Web page (http://toccata.lri.fr/gallery/index.en.html ), the Hisseo project (http://hisseo.saclay.inria.fr/ ), S. Boldo's page (http://www.lri.fr/~sboldo/research.html ), and industrial case studies in the U3CAT ANR project. The ability to express such complex properties comes from models developed in Coq [6] . The ability to combine proof by reasoning and proof by computation is a key aspect when dealing with floating-point programs. Such a modeling provides a safe basis when dealing with C source code [5] . However, the proofs can get difficult even on short programs, and to achieve them some automation is needed, and obtained by combining SMT solvers and Gappa [67] , [83] , [51] [10] . Finally, the precision of the verification is obtained thanks to precise models of floating-point computations, taking into account the peculiarities of the architecture (e.g. x87 80-bit floating-point unit) and also the compiler optimizations [71] , [111] .

The directions of research concerning floating-point programs that we pursue are the following.

Making Formal Verification of Floating-point Programs Easier

A first goal is to ease the formal verification of floating-point programs: the primary objective is still to improve the scope and efficiency of our methods, so as to ease further the verification of numerical programs. The on-going development of the Flocq library continues towards the formalization of bit-level manipulations and also of exceptional values (e.g. infinities). We believe that good candidates for applications of our techniques are smart algorithms to compute efficiently with floats, which operate at the bit-level. The formalization of real numbers is being revamped too: higher-level numerical algorithms are usually built on some mathematical properties (e.g. computable approximations of ideal approximations), which then have to be proved during the formal verification of these algorithms.

Easing the verification of numerical programs also implies more automation. SMT solvers are generic provers well-suited for automatically discharging verification conditions, but they tend to be confused by floating-point arithmetic [77] . Our goal is to improve the arithmetic theories of Alt-Ergo, so that they support floating-point arithmetic along their other theories, if possible by reusing the heuristics developed for Gappa.

Continuous Quantities, Numerical Analysis

The goal is to handle floating-point programs that are related to continuous quantities. This includes numerical analysis programs we have already worked on [15] [66] [4] . But our work is only a beginning: we were able to solve the difficulties to prove one particular scheme for one particular partial differential equation. We need to be able to easily prove this kind of programs. This requires new results that handle generic schemes and many partial differential equations. The idea is to design a toolbox to prove these programs with as much automation as possible. We wish this could be used by numerical analysts that are not or hardly familiar with formal methods, but are interested in the formal correctness of their schemes and their programs.

Another very interesting kind of programs (especially for industrial developers) are those based on hybrid systems, that is where both discrete and continuous quantities are involved. This is a longer term goal, but we may try to go towards this direction. A first problem is to be able to specify hybrid systems: what are they exactly expected to do? Correctness usually means not going into a forbidden state but we may want additional behavioral properties. A second problem is the interface with continuous systems, such as sensors. How can we describe their behavior? Can we be sure that the formal specification fits? We may think about Ariane V where one piece of code was shamelessly reused from Ariane IV. Ensuring that such a reuse is allowed requires to correctly specify the input ranges and bandwidths of physical sensors.

Studying hybrid systems is among the goals of the new ANR project Cafein.

Certification of Floating-point Analyses

In coordination with our second theme, another objective is to port the kernel of Gappa into either Coq or Why3, and then extract a certified executable. Rather than verifying the results of the tool a posteriori with a proof checker, they would then be certified a priori. This would simplify the inner workings of Gappa, help to support new features (e.g. linear arithmetic, elementary functions), and make it scale better to larger formulas, since the tool would no longer need to carry certificates along its computations. Overall the tool would then be able to tackle a wider range of verification conditions.

An ultimate goal would be to develop the decision procedure for floating-point computations, for SMT context, that is mentioned in Section  3.2.2 , directly as a certified program in Coq or Why3.