Systems and methods for automatically testing one or more versions of a compiler of are disclosed. A compiler is instrumented to generated data exposing various internal decisions and/or actions made by the compiler. Subsequently, multiple distinct versions of the compiler are executed to compile a code corpus associated with a particular programming language. Output (including instrumentation output) from the compilation of the code corpus for each version of the compiler is obtained and compared to identify behavioral changes that may exist between the various versions of the compiler.