ITR: Compiler Technology for Automatic Parallelization via Thread-Level Speculation

点击翻译按钮获取中文摘要

For many practical reasons (e.g., hardware development cost, powerconsumption, etc.), it would be attractive if we could increase theeffective performance of a computer system by simply adding more processorsto it. While it is reasonably straightforward to build such systems thatsupport "parallel processing", relatively few of today's programs wouldenjoy any performance gain on such machines because they were not writtenwith parallel processing in mind. The ideal solution for harnessing thepotential benefits of parallel processing without placing a large burden onthe programmer would be for the compiler to automatically transform asequential program into an efficient parallel program. While there hasbeen progress on using compilers to automatically "parallelize" regularnumeric programs (typically written in FORTRAN), there has been littleprogress in automatically parallelizing broader classes of programs. Thekey stumbling block has been that compilers have traditionally createdparallelism by proving that potential threads are always independent. Tocircumvent this limitation, recent hardware prooposals enable the compilerto optimistically create parallel threads without proving independence. Inthis project, we plan to develop the compiler technology necessary to fullyexploit this new potential for harnessing parallel processors.