A nanopass framework for commercial compiler development

Contemporary commercial compilers typically handle sophisticated high-level source languages, generate efficient assembly or machine code for multiple hardware architectures, run under and generate code to run under multiple operating systems, and support source-level debugging, profiling, and other program development tools. As a result, commercial compilers tend to be among the most complex of software systems. Nanopass frameworks are designed to help make this complexity manageable. A nanopass framework is a domain-specific language, embedded in a general purpose programming language, to aid in compiler development. A nanopass compiler is comprised of many small passes, each of which performs a single task and specifies only the interesting transformations to be performed by the pass. Intermediate languages are formally specified by the compiler writer, which allows the infrastructure both to verify that the output of each pass is well-formed and to fill in the uninteresting boilerplate parts of each pass. Prior nanopass frameworks were prototype systems aimed at educational use, but we believe that a suitable nanopass framework can be used to support the development of commercial compilers. We have created such a framework and have demonstrated its effectiveness by using the framework to create a new commercial compiler that is a “plug replacement” for an existing commercial compiler. The new compiler uses a more sophisticated, although slower, register allocator and implements nearly all of the optimizations of the original compiler, along with several “new and improved” optimizations. When compared to the original compiler on a set of benchmarks, code generated by the new compiler runs, on average, 21.5% faster. The average compile time for these benchmarks is less than twice as long as with the original compiler. This dissertation provides a description of the new framework, the new compiler, and several experiments that demonstrate the performance and effectiveness of both, as well as a presentation of several optimizations performed by the new compiler and facilitated by the infrastructure.