tobias christian grosser Diploma

Sustained growth in high performance computing and the availability of advanced mobile devices increase the use of computation intensive applications. To ensure fast execution and consequently low power usage modern hardware provides multi-level caches, multiple cores, SIMD instructions or even dedicated vector accelerators. Taking advantage of those manually is difficult and often not possible in a portable way. Fortunately, advanced techniques that increase data-locality and parallelism with the help of polyhedral abstractions are known to be effective in exploiting hardware capabilities. Yet, their automatic use is currently limited. They are mostly implemented in language specific source-to-source compilers which can only optimize manually annotated code that matches a certain canonical structure. Furthermore, polyhedral optimizers often target C or CUDA code, which limits efficient communication with compiler internals and can lead to missed optimization opportunities. In this thesis we present Polly, a project to enable polyhedral optimizations in LLVM. LLVM is an infrastructure project used in compilers for a large set of different programming languages. It is built around a low-level intermediate representation (LLVMIR) that allows language independent optimizations. We present how Polly can apply polyhedral transformations on this representation. This includes the detection of static control parts, their translation into a Z-polyhedra based representation, optimizations on this representation and finally, the generation of optimized LLVM-IR. We also define an interface to connect external optimizers and show a novel approach to detect parallelism which is used to generate SIMD and OpenMP code. Finally, we show in some experiments how Polly can be used to automatically apply optimizations for data locality and parallelism.