Detection of Optimizations Missed by the Compiler

Abstract

With the increasing significance of compilers in software development, identifying optimization bugs and enhancing their performance has become a significant challenge. In recent years, many studies have targeted only specific analyses to identify missed optimizations (e.g., in data flow analyses). While there have been some general approaches, most rely on differential testing between different compilers, which makes it difficult to identify optimization bugs that are common to multiple compilers. This paper tackles these challenges by introducing an effective and general approach called MOD. MOD works by using a manually-curated mapping between optimizations, ensuring that code should be consistently optimized: if one optimization triggers but the other does not, that indicates a bug in either of the two optimizations. We implemented MOD as a tool to detect missed optimizations in the available expressions of the LLVM. Experimental results show that MOD can report 20 correct alerts in one hour of detection of random test programs.