No      X3J17/96-0065
                                        WG21/N0883
                                Date:   March 12, 1996
                                Author: Michael S. Ball
 
 
 
 
 
Observations on the Template Compilation Model
 
Of all of the arguments made for and against the "Separation
model" for template compilation, there are three that I find
persuasive.
 
Name Access
 
Though the "separation" model does provide some advantages in
limiting name access, I believe that this is the same sort of
short-term advantage provided by the C preprocessor, and that it
has many of the same longer-term liabilities that we have seen
from the preprocessor. The advantage is that you can pay less
attention to the structure of your interfaces and
implementations. The liability is that the code so structured is
less reusable, harder to change, and significantly more
difficult to manipulate with advanced program development tools.
We have tools to control interfaces and name access within the
language, and using these tools will provide considerably more
reusable code than an ad hoc structuring into files. File
structuring and the preprocessor are reasonable compromises for
dealing with limited resources, but the need to maintain
compatibility with them has considerably slowed the development
of more advanced tools.
 
Costs
 
The cost factor, though not overwhelming, is persuasive. The
only examples that we have of the separation model have
considerably higher costs than the inclusion model. Though we
all have clever schemes that we think can bring the cost down,
we have no evidence that these will work in practice. We
estimate that the additional cost for separation will ultimately
be small, but we have no evidence of this, and in existing
practice the cost is quite high. Even the proponents of the
separation model recognize this, and introduced the explicit
instantiation syntax to allow the programmer to optimize the
instantiation process. One justification for this was the large
reduction in compilation time that judicious use of explicit
instantiation could produce.
 
Lack of Experience
 
My major concern, however, is that the separation model has
technical problems. I do not know what these are, but I know, as
well as I know that the sun will rise tomorrow, that they are
there. We will not discover them until we have a number of
implementations in use by a large number of customers. We have a
great deal of experience with the inclusion model, and we have a
much better idea of its problems. If we issue the standard with
such a large and complex area untested, we will guarantee
incompatible implementation, exactly the situation that the
standard is trying to avoid.