Wouldn't LLVM adjust the models if it is beneficial to its code generation, even if the result less accurately reflects the processor? (I think GCC does that.)