Abstract
Numerical accuracy and execution time of programs using the floating-point arithmetic are major challenges in many computer science applications. The improvement of these criteria is the subject of many research works. However we notice that the accuracy improvement decrease the performances and conversely. Indeed, improvement techniques of numerical accuracy, such as expansions or compensations, increase the number of computations that a program will have to execute. The more the number of computations added is, the more the performances decrease. This thesis work presents a method of accuracy improvement which take into account the negative effect on the performances. So we automatize the error-free transformations of elementary floating-point operations because they present a high potential of parallelism. Moreover we propose some transformation strategies allowing partial improvement of programs to control more precisely the impact on execution time. Then, tradeoffs between accuracy and performances are assured by code synthesis. We present also several experimental results with the help of tools implementing all the contributions of our works.