Description
Compilers are the cornerstone of software development. However, its outdated, purely sequential architecture scales only insufficiently with the size of today's software projects. The compiler’s runtime, which is unproductive for the development process, is taking up an increasingly large share of the development cycle.
The majority of the runtime is spent on data flow analyses. The computed information enables to apply the optimizations, e.g., to reduce the runtime of a program.
This dissertation describes the framework ParCan, which makes it possible to execute fixpoint-based data flow analyses in a data-parallel fashion on a graphics card (GPGPU).
By integrating ParCan into the LLVM compiler framework, its runtime could be reduced by up to 31%.
Within the scope of this thesis, further issues such as the efficiency of graph structures as well as the efficient, deadlock-free synchronization of threads on the GPU were addressed.
Reviews
There are no reviews yet.