Compiler-Assisted Floating-Point Error Analysis and Profiling with FPChecker

FPChecker is an open-source tool that uses compiler instrumentation to detect and profile floating-point errors in HPC (high-performance computing) and scientific code. Floating-point arithmetic (e.g., IEEE 754) is inherently imprecise because numbers like 0.1 cannot be represented exactly in binary, and rounding errors accumulate during computation. FPChecker automatically identifies where precision is lost — e.g., cancellation (subtracting nearly equal numbers) or catastrophic rounding — without requiring the programmer to manually rewrite code or run extended-precision tests. It belongs to a family of tools (alongside Verrou, CADNA, and Herbgrind) used by computational scientists and engineers who need to verify the numerical reliability of simulations in weather forecasting, physics, ML, and finance. The talk at ISC 2026 presents FPChecker's profiling capabilities for large-scale parallel applications.