Thèse soutenue

Accélération matérielle de la compilation à la volée pour les systèmes embarqués

FR
Auteur / Autrice : Alexandre Carbon
Direction : Henri-Pierre Charles
Type : Thèse de doctorat
Discipline(s) : Informatique
Date : Soutenance en 2013
Etablissement(s) : Paris 6

Résumé

FR  |  
EN

Developed since the 60s, JIT compilation is widely used since 15 years. This is the consequence of two main phenomena: the increasing dynamism of applications and the increasing demand concerning virtualization. The transfer of these issues to the embedded domain leads to experience JIT compilation on small and sparse resources. However, the management of JIT compilation algorithms’ complexity and irregularity on small resources (in-order processors, limited speculation, limited memory hierarchies) leads to important scaling-down problems in terms of performance. As a consequence, JIT compilation solutions are less attractive in this domain. While several software optimizations have been already proposed in the literature, we propose in this thesis the development of hardware accelerations coupled to the processor in charge of the JIT compilation. The final aim is to propose a more efficient solution in terms of performance with respect to embedded constraints. Based on the LLVM framework compiler (LLC), our experiments highlight two critical points in terms of performance: the associative array and dynamic memory allocation management and the instruction graph handling for instructions to compile and optimize. Two accelerators have been proposed in this way. Concerning the management of associative arrays, we obtain gains up to 25 % on LLC with an area overhead under 1. 4 % of the associated processor.