There's more discussion about this on this reddit thread: http://www.reddit.com/r/programming/comments/x1ql4/undefined_behavior_in_c_contest_winners/ http://www

There's more discussion about this on this reddit thread: http://www.reddit.com/r/programming/comments/x1ql4/undefined_behavior_in_c_contest_winners/ http://www.reddit.com/r/programming/comments/x1ql4/undefined... The problem is that realloc builtin's return type is internally flagged as noalias, making clang think it was safe to use constant propagation for the printf. I guess LLVM doesn't take the equality check into account. > 1. Breaking reflexivity is not OK. Ever. a==a should hold. This reminds me of the "proof" that 1=2. Once an undefined operation is introduced, the end result means nothing. > 2. The compiler knows that this is undefined behavior, and uses this to "optimize" to a wrong result. Maybe, maybe not. It's not a requirement that a compiler even recognizes that undefined behavior happens. In many cases behavior is undefined simply because it would be too expensive or impractical for the compiler to deal with that behavior.