All non-toy languages and compilers only add constant overhead; unfortunately that's still a problem, the goal is to not have any overhead at all.
malloc/free is fast with a moving GC, but moving is not, and if an object is never moved it's likely that it should not have been allocated on the heap in the first place.
The cmpxchg problem is solvable by epoch-based reclamation libraries (RCU, crossbeam-epoch, etc.), and while a traditional GC also solves it, it's not necessary.
Even more, the performance limitations of most GC languages have more to do with the lack of good ways of writing code which simply doesn't allocate, rather then the problem of collection. GCs are often faster than malloc/free, but not as fast as simply not allocating/freeing anything.
Finally, it's important to remember that there are algorithms that are significantly more difficult to implement without a GC tahn with a GC. Even simple compare-and-swap atomic sets can require many times more code and care to implement if you have to handle cleanup of temporaries as well.