Contributions to code optimization and high performance library generation

The number of transistors as well as the frequency of processors have followed Moore's law for the past decades, at the expense of an increase in architecture complexity. As improvements in processor clock frequencies have levelled out, a major shift to parallelism is taking place. Parallelism is now the only way to enhance performance, and compilers and operating systems are essential to improve performance automatically. Due to the complexity of hardware mechanisms, hard to model, compilers are still far from being able to generate high performance application codes, even on a single core. The work presented in this thesis focuses on two aspects of this issue: we rst describe results obtained for the optimization and generation of high performance libraries and their automatic reuse in applications, and then describe performance evaluation techniques and tools to guide the optimization process. Results of this work is followed by research perspectives.Le nombre de transistors des processeurs, ainsi que leur frequence, ont suivi la loi de Moore pendant plusieurs decennies, au prix d'une complexite croissante des architectures. La recente de l'accroissement en frequence a notamment deux consequences: le parallelisme est desormais un des seuls vecteurs de gain de performances, et la chaine de compilation ainsi que le systeme d'exploitation sont indispensables pour l'obtention automatique de ces performances. Du a la complexite des mecanismes architecturaux difficiles a modeliser de facon realiste, les compilateurs restent cependant loin de pouvoir generer automatiquement des applications hautes performances, meme pour un seul core. Le travail que nous presentons se focalise sur d'une part l'optimisation et la generation de bibliotheques hautes performances et leur reutilisation automatique dans un contexte applicatif, d'autre part, sur l'evaluation et la modelisation des performances an de guider l'optimisation. Les resultats de ces travaux sont suivis de perspectives de recherche