Compiler backends are critical for translating high-level code into efficient machine instructions, yet they remain relatively underexplored in compiler testing. Effective backend testing requires programs that expose low-level backend behaviors, yet such features are difficult to generate and are often eliminated by earlier optimization passes. As a result, existing testing approaches frequently fail to exercise diverse backend behaviors and therefore miss deep crashes and miscompilations.

We present BackSmith, a black-box approach for testing compiler backends across compilers and architectures. BackSmith generates code snippets with two complementary properties: backend-oriented features that directly stress backend mechanisms such as instruction selection and register allocation, and optimization-resistant features that preserve program diversity by resisting excessive middle-end canonicalization. To further increase coverage of rare but critical backend behaviors, BackSmith also generates code snippets whose compiled assembly exhibits unique patterns. It then integrates all three kinds of features into seed programs for backend testing.

We evaluated BackSmith on 16 mature GCC and LLVM backends. Over five months of testing, BackSmith uncovered 104 previously unknown backend bugs, 88 of which have been confirmed or fixed, demonstrating the effectiveness of our approach in systematically exposing deep backend defects.