The existing compilation model of C++ templates---namely
inlining---suffers from a problem that each object code and the
resulting executable tend to become huge, and also, that
implementations of templates cannot be hidden.  In this paper, we
present a method of separately compiling C++ templates.  This method
is based on source-to-source transformation of C++ code.  At runtime,
the transformed code (1) performs boxing and unboxing on polymorphic
data (as many ML compilers do) in order to deal with the parametric
polymorphism of templates and (2) passes around overloaded functions
(as the Glasgow Haskell Compiler does) in order to treat the ad hoc
polymorphism of C++.  Our method also exploits the multiple
inheritance and abstract interface mechanisms of C++ to handle method
invocations on polymorphic objects.