Providing new parallel programming models/abstractions as a set of library functions has the huge advantage that it allows for an relatively easy incremental porting path for legacy HPC applications, in contrast to the huge effort needed when novel concepts are only provided in new programming languages or language extensions. However, performance issues are to be expected with fine granular usage of library functions. In previous work, we argued that binary rewriting can bridge the gap by tightly coupling application and library functions at runtime. We showed that runtime specialization at the binary level, starting from a compiled, generic stencil code can help in approaching performance of manually written, statically compiled version. In this paper, we analyze the benefits of post-processing the re-written binary code using standard compiler optimizations as provided by LLVM. To this end, we present our approach for efficiently converting x86-64 binary code to LLVM-IR. Using the mentioned generic code for arbitrary 2d stencils, we present performance numbers with and without LLVM postprocessing. We find that we can now achieve the performance of variants specialized by hand.
展开▼
