Abstract

In recent years, several approaches have been proposed to use profile information in compiler optimization. This profile information can be used at the source level to guide loop transformations as well as in the backend to guide low level optimizations. At the same time, profile guided library generators have been proposed also, like Atlas, Spiral, or FFTW, that tune their routines for the underlying hardware. These approaches have led to excellent performance improvements. However, a possible drawback of these approaches is that applications are optimized using a single or a limited set of data inputs. It is well known that programs can exhibit vastly differing behaviors for different inputs. Therefore, it is not clear whether the performance numbers reported are still valid for other input than the input used to optimize the program. In this paper, we address this problem for a specific statistical compiler tuning method. We use three different platforms and several SPECint2000 benchmarks. We show that when we tune the compiler using train data, we obtain a compiler setting that still performs well for reference data. These results suggest that profile guided optimization may be more stable than is sometimes believed and that a limited number of train data sets is sufficient to obtain a well optimized program for all inputs. 1

Citations