Automatic High-Level Profile Directed Code Optimization for Parallel Architectures

Code optimization techniques improve the performance of code at the IR (intermediate representation) level of the compiler. Parallel code optimization and execution has always been better than sequential because parallel applications can fully utilize the power of modern multicore processor architectures. A compilation framework based on LLVM featuring parallel task compilation and automatic parallel code generation is developed in this research. Good features of polyhedral model are used for optimization along with profile directed feedback technique to provide good performance. The impact of this framework on programmer's productivity is also discussed. The results reveal that the new technique outperforms current compilers for many applications taken from Polybench/C-4.1 benchmark suite.