The ML family of languages is my favorite (second only to C++, natch.) but their brand of parametric polymorphism is equivalent in C++ to MyTemplate<T*> (or, if you like boost, MyTemplate<reference_wrapper<T> > would be more accurate.) They are not the same as MyTemplate<T>. This is by choice of ML language implementors, as they see it as a defect of C++ that every template instantiation has to be compiled independently. Instead, they turn static compile-time polymorphism into run-time dynamism using indirection and function pointers. (Maybe MLton doesn't do this, but I know GHC does.)

It's the difference between C's qsort and C++'s std::sort. The latter blows the former out of the water when sorting small things. This brand of polymorphism is unique to C++ and until someone does it better (which no one seems to have any interest in, except maybe Walter Bright) then C++ will always be "the best C."

This is why I said "pretty fast", not necessarily faster than C++. Using ML means trading a small percentage of C++'s raw speed for code that can be debugged, adapted, etc. with much less trouble. I have found this to be a more than reasonable trade-off in virtually all cases.