Des nombreuses applications de méthodes formelles se fondent sur les solveurs SMT pour valider automatiquement les conditions à vérifier et fournissent des certificats de leurs résultats. Nous visons à la fois à améliorer l'efficacité des solveurs SMT et à accroître leur fiabilité. Notre première contribution est un cadre uniforme pour le raisonnement avec des formules quantifiées dans les solveurs SMT, dans lequel généralement diverses techniques d'instanciation sont utilisées. Nous montrons que les principales techniques d'instanciation peuvent être jetées dans ce cadre. Le cadre repose sur le problème de l'E-ground (dis)unification. Nous présentons une procédure de décision pour résoudre ce problème en pratique: Fermeture de congruence avec variables libres (CCFV}). Nous mesurons l'impact de CCFV dans les solveurs SMT veriT et CVC4. Nous montrons que nos implémentations présentent des améliorations par rapport aux approches à la fine pointe de la technologie. Notre deuxième contribution est un cadre pour le traitement des formules tout en produisant des preuves détaillées. Les principaux composants de notre cadre de production de preuve sont un algorithme de récurrence contextuelle générique et un ensemble extensible de règles d'inférence. Avec des structures de données appropriées, la génération des preuves ne crée que des frais généraux linéaires et les vérifications peuvent être vérifiées en temps linéaire. Nous avons également mis en œuvre l'approche en veriT. Cela nous a permis de simplifier considérablement la base du code tout en augmentant le nombre de problèmes pour lesquels des preuves détaillées peuvent être produites / In many formal methods applications it is common to rely on SMT solvers to automatically discharge conditions that need to be checked and provide certificates of their results. In this thesis we aim both to improve their efficiency of and to increase their reliability. Our first contribution is a uniform framework for reasoning with quantified formulas in SMT solvers, in which generally various instantiation techniques are employed. We show that the major instantiation techniques can be all cast in this unifying framework. Its basis is the problem of E-ground (dis)unification, a variation of the classic rigid E-unification problem. We introduce a decision procedure to solve this problem in practice: Congruence Closure with Free Variables (CCFV). We measure the impact of optimizations and instantiation techniques based on CCFV in the SMT solvers veriT and CVC4, showing that our implementations exhibit improvements over state-of-the-art approaches in several benchmark libraries stemming from real world applications. Our second contribution is a framework for processing formulas while producing detailed proofs. The main components of our proof producing framework are a generic contextual recursion algorithm and an extensible set of inference rules. With suitable data structures, proof generation creates only a linear-time overhead, and proofs can be checked in linear time. We also implemented the approach in veriT. This allowed us to dramatically simplify the code base while increasing the number of problems for which detailed proofs can be produced

Dependent type theories are a kind of mathematical foundations investigated both for the formalisation of mathematics and for reasoning about programs. They are implemented as the kernel of many proof assistants and programming languages with proofs (Coq, Agda, Idris, Dedukti, Matita, etc). Dependent types allow to encode elegantly and constructively the universal and existential quantifications of higher-order logics and are therefore adapted for writing logical propositions and proofs. However, their usage is not limited to the area of pure logic. Indeed, some recent work has shown that they can also be powerful for driving the construction of programs. Using more precise types not only helps to gain confidence about the program built, but it can also help its construction, giving rise to a new style of programming called Type-Driven Development. However, one difficulty with reasoning and programming with dependent types is that proof obligations arise naturally once programs become even moderately sized. For example, implementing an adder for binary numbers indexed over their natural number equivalents naturally leads to proof obligations for equalities of expressions over natural numbers. The need for these equality proofs comes, in intensional type theories (like CIC and ML) from the fact that in a non-empty context, the propositional equality allows us to prove as equal (with the induction principles) terms that are not judgementally equal, which implies that the typechecker can't always obtain equality proofs by reduction. As far as possible, we would like to solve such proof obligations automatically, and we absolutely need it if we want dependent types to be use more broadly, and perhaps one day to become the standard in functional programming. In this thesis, we show one way to automate these proofs by reflection in the dependently typed programming language Idris. However, the method that we follow is independent from the language being used, and this work could be reproduced in any dependently-typed language. We present an original type-safe reflection mechanism, where reflected terms are indexed by the original Idris expression that they represent, and show how it allows us to easily construct and manipulate proofs. We build a hierarchy of correct-by-construction tactics for proving equivalences in semi-groups, monoids, commutative monoids, groups, commutative groups, semi-rings and rings. We also show how each tactic reuses those from simpler structures, thus avoiding duplication of code and proofs. Finally, and as a conclusion, we discuss the trust we can have in such machine-checked proofs.