Scalable Size Inliner for Mobile Applications (WIP)

International Conference on Languages, Compilers, and Tools for Embedded Systems (LCTES) at PLDI

Abstract

Inlining is critical for both performance and size in mobile apps. When building large mobile apps, ThinLTO, a scalable link-time optimization is imperative in order to achieve both optimal size and build scalability. However, inlining with ThinLTO is not tuned to reduce the code size because each module inliner works independently without modeling the size cost across modules, and functions are often not eligible to import due to private references, appearing in Objective-C or Swift for iOS. This paper extends the bitcode summary to perform a global inlining analysis to find inline candidates for saving the code size. Using this summary information, a pre-inliner eagerly inlines the candidates that are proven to shrink the size. When the inline candidates are not eligible to import, a pre-merger combines their bitcode modules to remove inline restrictions. Our work improves the size of real-world mobile apps when compared to the MinSize (-Oz) optimization level. We reduced the code size by 2.8% for SocialApp and 4.0% for ChatApp.

Latest Publications

Boosted Dense Retriever

Patrick Lewis, Barlas Oğuz, Wenhan Xiong, Fabio Petroni, Wen-tau Yih, Sebastian Riedel

NAACL - 2022