Open Source Summit North America 2022

Technology based on open source evolves every year with novel ideas that break paradigms. This evolution requires the reinvention of toolchains constantly. The performance and security of a SW project could be improved (or affected) by choosing the precise configuration in the compiler, linker, or binary tools available to developers. This presentation aims to show some of the new GNU toolchain features for 2022. Among those new features to discover is the enablement of vectorized half-precision floating-point format (FP16) instructions, heavily used in Deep Learning applications. New optimizations for memory transactions using wider vector registers and the enablement of DT_RELR for x86 platforms give the developers the tools to improve memory-bound applications as well as file size. At the same time, new enhancements on the integrated static analyzer give the developer the capability to detect security bugs easily. Having a better understanding of the toolchains allows developers to showcase the best of new platform architecture technology for users’ applications as well as boost the innovation and security of incoming projects.

The kernel receives thousands of patches per month, with recent major releases containing well over ten thousand. Testing a patch means first building the affected source code. But using an arbitrary configuration file provides no guarantee of build coverage. defconfig only builds around 15% of the patches contained in v5.13. allyesconfig provides high coverage, but incurs higher build times (three hours for allyesconfig vs. 20 minutes for defconfig on a commodity desktop). Moreover, allyesconfig kernels are less suited for run-time testing due to their size, and they preclude combinatorial configuration testing, e.g., with randconfig. The speakers' approach, krepair, automatically repairs a configuration file to ensure high patch coverage while retaining most settings from the original configuration file. This allows testers to build the patch without sacrificing tests for a variety of hardware and software configurations. Our evaluation on thousands of real patches shows that krepair finds repairs to allnoconfig and defconfig that change only a couple hundred options on average, achieving patch coverage comparable to allyesconfig which, in contrast, requires enabling over ten thousand additional options. The talk will also demonstrate using the free and open-source tool.