Adaptive loop transformations for scientific programs

To facilitate the eecient execution of scientiic programs , parallelizing compilers apply a wide range of loop transformations. In some situations it may not be possible to determine the applicability and usefulness of the transformations at compile-time. In this paper we develop adaptive versions of various loop transformations which lend themselves to eecient application at run-time. Instead of explicitly applying a transformation at compile-time, we generate adaptive code which is able to behave like the transformed code, if so desired, at run-time. Therefore an adaptive program can be viewed as expressing multiple ways of executing a program. The selection of the particular execution is based upon run-time information. Adaptive programs ooer a more practical and powerful alternative to mul-tiversion programs.