The advantage is that you can reason about the correctness of code. Sure, if you write an entire code base depending on its size you may be able to reason about all the side effects in your head. But when moving to a larger project, or one developed by a team - all bets are off. It is really important to have the effects embedded in the type system. Even with a large test suite.
On top of this, having professionally written C/C++ for years in the embedded systems space I can say that Monads for IO is not awkward. In fact, it feels far more powerful because you are not limited to the awful semicolon operator for chaining effects.
Lastly FP != slower (even on single core). I agree Haskell can be hard to reason about performance wise at times due to choosing lazy by default. However, have you ever heard of Sisal? Many strict by default FP languages are quite fast, however they sacrifice purity. But to be honest, with experience Haskell can be damn fast also.