[MPLR'25] Fast and Compact: Reducing Size of AOT-Compiled Java Code Without Sacrificing Performance скачать в хорошем качестве

[MPLR'25] Fast and Compact: Reducing Size of AOT-Compiled Java Code Without Sacrificing Performance

Fast and Compact: Reducing Size of AOT-Compiled Java Code without Sacrificing Performance (Video, MPLR 2025) Christoph Pichler, Bernhard Urban-Forster, Paley Li, Roland Schatz, and Hanspeter Mössenböck (JKU Linz, Austria; Oracle Labs, Austria; Oracle Labs, Czechia; Oracle Labs, Austria; JKU Linz, Austria) Abstract: GraalVM Native Image is a technology that compiles Java applications to native executables ahead of time (AOT). Due to highly aggressive optimizations, it is possible to reach a peak performance similar to that of a regular Java application executed on a Java Virtual Machine (JVM) and its Just-in-Time (JIT) compiler. The advantage of Native Image is much faster start-up when compared to the JVM and the JIT compiler. However, binary files resulting from Native Image tend to be quite large. In this work, we present an approach to shrink the executables produced by Native Image while maintaining high peak performance: Instead of compiling everything to machine code, we exclude rarely or never executed methods and keep them as bytecode instead. When called, these methods can be run with a lightweight interpreter. If an interpreted method turns out to cause high overhead, its execution continues in native mode by simulating a JIT compiler. For the DaCapo benchmark suite, we are able to reduce the sizes of Native Image executables by up to 36.9% without significantly reducing peak performance or increasing start-up times. Article: https://doi.org/10.1145/3759426.3760976 ORCID: https://orcid.org/0009-0008-7223-6692, https://orcid.org/0009-0007-2769-7388, https://orcid.org/0009-0001-7461-3331, https://orcid.org/0009-0006-3100-8816, https://orcid.org/0000-0001-7706-7308 Video Tags: AOT compilation, interpretation, performance, size reduction, GraalVM Native Image, splashws25mplrmain-p27-p, doi:10.1145/3759426.3760976, orcid:0009-0008-7223-6692, orcid:0009-0007-2769-7388, orcid:0009-0001-7461-3331, orcid:0009-0006-3100-8816, orcid:0000-0001-7706-7308 Presentation at the MPLR 2025 conference, October 12–18, 2025, https://conf.researchr.org/home/icfp-... Sponsored by ACM SIGPLAN, ACM SIGAda,