Compilation hybride guidée pour profilage

The end of chip frequency scaling capacity, due heat dissipation limitations, made manufacturers search for an alternative to sustain the processing capacity growth. The chosen solution was to increase the hardware parallelism, by packing multiple independent processors in a single chip, in a Multiple-Instruction Multiple-Data (MIMD) fashion, each with special instructions to operate over a vector of data, in a Single-Instruction Multiple-Data (SIMD) manner. Such paradigm change, brought to software developer the convoluted task of producing efficient and scalable applications. Programming languages and associated tools evolved to aid such task for new developed applications. But automated optimizations capable of coping with such a new complex hardware, from legacy, single threaded applications, is still lacking.To apply code transformations, either developers or compilers, require to assert that, by doing so, they are not changing the expected comportment of the application producing unexpected results. But syntactically poor codes, such as use of pointer parameters with multiple possible indirections, complex loop structures, or incomplete codes, make very hard to extract application behavior solely from the source code in what is called a static analyses. To cope with the lack of information extracted from the source code, many tools and research has been done in, how to use dynamic analyses, that does application profiling based on run-time information, to fill the missing information. The combination of static and dynamic information to characterize an application are called hybrid analyses. This works advocates for the use of hybrid analyses to be able to optimizations on loops, regions where most of computations are done. It proposes a framework capable of statically applying some complex loop transformations, that previously would be considered unsafe, by assuring their safe use during run-time with a lightweight test.The proposed framework uses application execution profiling to help the static loop optimizer to: 1) identify and classify program hot-spots, so as to focus only on regions vital for the execution time; 2) guide the optimizer in understanding the overall loop behavior, so as to reduce the valid loop transformations search space; 3) using instruction's memory access functions, it statically builds a lightweight run-time test that determine, based on the program parameters values, if a given optimization is safe to be used or not. It's applicability is shown by performing complex loop transformations into a variety of loops, obtained from applications of different fields, and demonstrating that the run-time overhead is insignificant compared to the loop execution time or gained performance, in the vast majority of cases.L'auteur n'a pas fourni de résumé en françai