Durant ces dernières années, Il est de plus en plus compliqué d'écrire du code qui utilise une architecture au mieux de ses capacités. Certaines opérations clefs ont soit un accélérateur dédié, ou admettent une implémentation finement optimisée qui délivre les meilleurs performances. Ainsi, il est intéressant d'identifier ces opérations pendant la compilation d'un programme, et de faire appel à une implémentation optimisée.Nous nous intéressons dans cette thèse au problème de détection de ces opérations. Nous proposons un procédé qui détecte des sous-calculs correspondant à des opérations d'algèbre linéaire à l'intérieur de programmes polyédriques. L'idée principale de ce procédé est de découper le programme en sous-calculs isolés, et essayer de reconnaître chaque sous-calculs comme une combinaison d'opérateurs d'algèbre linéaire.Le découpage du calcul est effectué en utilisant une transformation de programme appelée tuilage monoparamétrique. Cette transformation partitionne le calcul en tuiles dont la forme est un agrandissement paramétrique d'une tuile de taille constante. Nous montrons que le programme tuilé reste polyédrique tout en permettant une paramétrisation limitée des tailles de tuile. Les travaux précédents sur le tuilage nous forçaient à choisir l'une de ces deux propriétés.Ensuite, afin d'identifier les opérateurs, nous introduisons un algorithme de reconnaissance de template, qui est une extension d'un algorithme d'équivalence de programme. Nous proposons plusieurs extensions afin de tenir compte des propriétés sémantiques communément rencontrées en algèbre linéaire.Enfin, nous combinons les deux contributions précédentes en un procédé qui détecte les sous-calculs correspondant à des opérateurs d'algèbre linéaire. Une de ses composantes est une librairie de template, inspirée de la spécification BLAS. Nous démontrons l'efficacité de notre procédé sur plusieurs applications. / Writing a code which uses an architecture at its full capability has become an increasingly difficult problem over the last years. For some key operations, a dedicated accelerator or a finely tuned implementation exists and delivers the best performance. Thus, when compiling a code, identifying these operations and issuing calls to their high-performance implementation is attractive. In this dissertation, we focus on the problem of detection of these operations. We propose a framework which detects linear algebra subcomputations within a polyhedral program. The main idea of this framework is to partition the computation in order to isolate different subcomputations in a regular manner, then we consider each portion of the computation and try to recognize it as a combination of linear algebra operations.We perform the partitioning of the computation by using a program transformation called monoparametric tiling. This transformation partitions the computation into blocks, whose shape is some homothetic scaling of a fixed-size partitioning. We show that the tiled program remains polyhedral while allowing a limited amount of parametrization: a single size parameter. This is an improvement compared to the previous work on tiling, that forced us to choose between these two properties.Then, in order to recognize computations, we introduce a template recognition algorithm. This template recognition algorithm is built on a state-of-the-art program equivalence algorithm. We also propose several extensions in order to manage some semantic properties.Finally, we combine these two previous contributions into a framework which detects linear algebra subcomputations. A part of this framework is a library of template, based on the BLAS specification. We demonstrate our framework on several applications.

abstract: Dynamic software update (DSU) enables a program to update while it is running. DSU aims to minimize the loss due to program downtime for updates. Usually DSU is done in three steps: suspending the execution of an old program, mapping the execution state from the old program to a new one, and resuming execution of the new program with the mapped state. The semantic correctness of DSU depends largely on the state mapping which is mostly composed by developers manually nowadays. However, the manual construction of a state mapping does not necessarily ensure sound and dependable state mapping. This dissertation presents a methodology to assist developers by automating the construction of a partial state mapping with a guarantee of correctness. This dissertation includes a detailed study of DSU correctness and automatic state mapping for server programs with an established user base. At first, the dissertation presents the formal treatment of DSU correctness and the state mapping problem. Then the dissertation presents an argument that for programs with an established user base, dynamic updates must be backward compatible. The dissertation next presents a general definition of backward compatibility that specifies the allowed changes in program interaction between an old version and a new version and identified patterns of code evolution that results in backward compatible behavior. Thereafter the dissertation presents formal definitions of these patterns together with proof that any changes to programs in these patterns will result in backward compatible update. To show the applicability of the results, the dissertation presents SitBack, a program analysis tool that has an old version program and a new one as input and computes a partial state mapping under the assumption that the new version is backward compatible with the old version. SitBack does not handle all kinds of changes and it reports to the user in incomplete part of a state mapping. The dissertation presents a detailed evaluation of SitBack which shows that the methodology of automatic state mapping is promising in deal with real world program updates. For example, SitBack produces state mappings for 17-75% of the changed functions. Furthermore, SitBack generates automatic state mapping that leads to successful DSU. In conclusion, the study presented in this dissertation does assist developers in developing state mappings for DSU by automating the construction of state mappings with a correctness guarantee, which helps the adoption of DSU ultimately. / Dissertation/Thesis / Doctoral Dissertation Computer Science 2015