Workshop: LLVM-HPC2026: The Twelfth Workshop on the LLVM Compiler Infrastructure in HPC
Authors: Marek Rouven Beckmann, Peter Arzt, Tim Heldmann, and Felix Wolf (Technical University of Darmstadt)
Abstract: Fortran retains continued relevance in high-performance and scientific computing, yet its tooling ecosystem lags behind that of C and C++. Call graphs are a common foundation for analysis tools, and constructing them for modern Fortran requires resolving language features that elude simple source analysis. We present a static call-graph generator for Fortran implemented as a Flang plugin. Building on LLVM Flang's parse tree and semantic information, it produces whole-program call graphs from source code and resolves type-bound procedures, polymorphic dispatch, finalizers, operator overloading, generic interfaces, and C interoperability. We evaluate it on the production codes ICON, NEKO, and CP2K, and on the Fortran benchmarks of SPEChpc 2021, comparing against static LLVM-IR and dynamic Score-P call graphs. Across all applications, our analysis produces a reliable and detailed representation of procedures and call relationships. A case study demonstrates that the resulting graphs are well suited to guide instrumentation for performance analysis.
Back to LLVM-HPC2026: The Twelfth Workshop on the LLVM Compiler Infrastructure in HPC Archive Listing
Back to Full Workshop Archive Listing