What abstractions are you referring to? If you're using traits and writing your functions with generics there shouldn't be any overhead, other than the increased number of instructions, but this is no different than C++ generics. There are trait objects, but these come up somewhat less often then you would think, and have the same performance as using abstract classes/polymorphism in C++. But maybe you're referring to something else?
I am referring to trait objects, and I use those pretty extensively (although I limit my usage because I am making games and want to avoid a pointer dereference). For libraries you might not need them as much, but I definitely find their convenience indespensible for application side programming.
Personally, I've found that most cases you can avoid using trait objects by just using generics, but it obviously depends on your use case. You especially need trait objects to have things like heterogeneous arrays. Trait objects are pretty much the same as C++ classes with virtual methods, with one tweak in the representation. A C++ object is referred to by a pointer to (vtable_ptr, class_fields). A trait object "&Foo" is a fat pointer (vtable_ptr, ptr_to_fields). So a trait object reference is two pointers, but doesn't have a pointer stuck to the front of all of its structs. Method invocation requires you to dereference the vtable, get the function pointer, and call that, passing the data pointer. In C++ you would dereference your pointer, dereference the vtable, then pass the pointer to the function. If anything, I think the Rust method may be faster because you avoid the double de-reference, but it's probably so small as to be unmeasurable in all but pathological uses.
