Comparing to other generics implementations - they do add a lot of code, much more than is really needed to achieve good performance. Sure, some of this code can be reduced at linking stage, but for this to work you need to statically link everything which has its own set of problems (overhead of repeated code in every binary).

IMHO the approach is a great example of premature optimisation: generate 2^N versions then try to fold some similar ones to save space. Much better approach is: generate 1 version and specialize small parts when really needed. Technically, assuming sufficiently smart compiler, both would end up with exactly the same code, performing exactly the same, but the latter solution is preferable for other non-performance wise reasons (like better type checking, ability to distribute binary libraries, faster compile times, etc.).

BTW: I still often find Scala containers to be often much faster than C++ STL. But this is probably for much different reasons than template/generics implementation.