The TRegion Interface and Compiler Optimizations for OPENMP Target Regions

OPENMP is a well established, single-source programming language extension to introduce parallelism into (historically) sequential base languages, namely C/C++ and Fortran. To program not only multi-core CPUs but also many-cores and heavily parallel accelerators, OPENMP 4.0 adopted a flexible offloading scheme inspired by the hierarchy in many GPU designs. The flexible design of the offloading scheme allows to use it in various application scenarios. However, it may also result in a significant performance loss, especially because OPENMP semantics is traditionally interpreted solely in the language front-end as a way to avoid problems with the "sequential-execution-minded" optimization pipeline. Given the limited analysis and transformation capabilities in a modern compiler front-end, the actual syntax used for OPENMP offloading can substantially impact the observed performance. The compiler front-end will always have to favor correct but overly conservative code, if certain facts are not syntactically obvious. \ In this work, we investigate how we can delay (target specific) implementation decisions currently taken early during the compilation of OPENMP offloading code. We prototyped our solution in LLVM/Clang, an industrial strength OPENMP compiler, to show that we can use semantic source code analyses as a rational instead of relying on the user provided syntax. Our preliminary results on the rather simple Rodinia benchmarks already show speedups of up to 1.55x.