AnICA: analyzing inconsistencies in microarchitectural code analyzers

Microarchitectural code analyzers, i.e., tools that estimate the throughput of machine basic blocks, are important utensils in tool belt performance engineers. Recent like llvm-mca, uiCA, and Ithemal use a variety techniques different models for their predictions. When put to test, it is common see these state-of-the-art give very results. These inconsistencies either errors, or they point rarely documented assumptions made by designers. In this paper, we present AnICA, taking inspiration from differential testing abstract interpretation systematically analyze among analyzers. Our evaluation shows AnICA can summarize thousands few dozen descriptions directly lead high-level insights into behavior tools. several case studies, further demonstrate how automatically finds characterizes known unknown bugs as well quirk AMD's Zen microarchitectures.