[microtonal]

The potential danger in computationally intensive code is that you have a number of wrappers around the actual data (e.g. in the article AttributeSet wraps a hash table), each requiring a pointer dereference and a vtable lookup.

So, IMO it depends on the language features available and the scenario whether it is a code smell. For instance, Haskell allows you to introduce a new type whose data representation is the same as the original type using newtype [1]. In contrast to a type alias, newtype makes a different type, but it still allows you to define functions on that type in terms of the original type (using the type's constructor). If you want to hide the underlying representation, you simply do not export the constructors from the enclosing module.

tl;dr: in Haskell you do get a better abstraction without the overhead. I am pretty sure many other language can do as well (private inheritance in C++?).

[1] http://www.haskell.org/onlinereport/decls.html#datatype-rena...