This is about the conflict between compiler optimizations and sanitizers. Compiler optimizations exploit UB behavior to optimize code and, in the process, they can transform UB code to well-defined code by picking an arbitrary interpretation for the code. Sanitizers introduce additional runtime checks into code to detect UB behavior. Obviously, sanitizers cannot detect UB that has been elided by a compiler optimization.
This paper uses differential testing to understand where this happens in LLVM.