HW-SW co-design techniques for modern programming languages

Computer Architecture;Compiler;Inline Caching;Garbage Collection

Choi, Jiho

关键词: Computer Architecture;    Compiler;    Inline Caching;    Garbage Collection;   

Others  :  https://www.ideals.illinois.edu/bitstream/handle/2142/102482/CHOI-DISSERTATION-2018.pdf?sequence=1&isAllowed=y

美国|英语

来源: The Illinois Digital Environment for Access to Learning and Scholarship

【 摘 要 】

Modern programming languages raise the level of abstraction, hide the details of computer systems from programmers, and provide many convenient features. Such strong abstraction from the details of computer systems with runtime support of many convenient features increases the productivity of programmers.Such benefits, however, come with performance overheads. First, many of modern programming languages use a dynamic type system which incurs overheads of profiling program execution and generating specialized codes in the middle of execution. Second, such specialized codes constantly add overheads of dynamic type checks. Third, most of modern programming languages use automatic memory management which incurs memory overheads due to metadata and delayed reclamation as well as execution time overheads due to garbage collection operations.This thesis makes three contributions to address the overheads of modern programming languages. First, it describes the enhancements to the compiler of dynamic scripting languages necessary to enable sharing of compilation results across executions. These compilers have been developed with little consideration for reusing optimization efforts across executions since it is considered difficult due to dynamic nature of the languages. As a first step toward enabling the reuse of compilation results of dynamic scripting languages, it focuses on inline caching (IC) which is one of the fundamental optimization techniques for dynamic type systems. Second, it describes a HW-SW co-design technique to further improve IC operations. While the first proposal focuses on expensive IC miss handling during JavaScript initialization, the second proposal accelerates IC hit operations to improve the overall performance. Lastly, it describes how to exploit common sharing patterns of programs to reduce overheads of reference counting for garbage collection. It minimizes atomic operations in reference counting by biasing each object to a specific thread.