Workshop: LLVM-HPC2026: The Twelfth Workshop on the LLVM Compiler Infrastructure in HPC
Authors: Michael Kuron (Siemens Industry Software GmbH)
Abstract: An ahead-of-time compiler's ability to optimize strongly affects the runtime performance of an application, yet limited knowledge of runtime parameters restricts which optimizations it can apply. Just-in-time compilers can make more informed optimization decisions, but have seen limited adoption in high-performance computing due to build toolchain requirements and programming language restrictions. In this work, we use LLVM as a just-in-time compiler that can be tightly integrated into any application. We automatically specialize compute kernels by inspecting LLVM's intermediate representation to find simple loop-invariant conditionals, which we remove, and virtual function calls, which we inline. This suits C++ lambda functions better than previous work, which requires explicit annotation. We apply this method to a multiphysics simulation package that runs on general-purpose graphics processing units. This reduces iteration time by up to 35% by significantly simplifying control flow complexity of the compiled kernels and improves developer productivity through lower build times.
Back to LLVM-HPC2026: The Twelfth Workshop on the LLVM Compiler Infrastructure in HPC Archive Listing
Back to Full Workshop Archive Listing