They are using the LLVM to optimize the code before it is run by the computer. They say no code runs faster than the one optimized by llvm. Sometimes it doesn't work after llvm, llvm breaks it.
LLVM reinterprets your entire code and sometimes it interprets wrongly and breaks your ~~undefined behavior~~ smart tricks.
Especially O3 they say does this a lot.
Now that openAI has released O4, hope is that will improve.