When working with optimizing compilers, it is sometimes beneficial to inspect intermediate results to figure out whether an optimization has fired or not. While taking a dive into LLVM’s fast-math semantics I recently encountered a subtle quirk that may affect others as well.