egcs-ss-970825: template function linking problem

[Alexandre Oliva]

Jonathan Larmour writes:

> At 21:14 29/08/97 -0300, Alexandre Oliva wrote:
>> Jonathan Larmour writes:
>> 
>>> template <class T> void problem_fn(T &obj);
>> 
>>> //  printf("%p\n", &problem_fn); // Uncomment this line and it stops
> linking

>> Besides gcc failing to instantiate the template it decided to use,
>> there's no way to find out which template function it should have
>> used, so an error should have been flagged at compile-time.

> Is that true? Remember the linking error I got _did_ have the correct type

What do you mean correct type?  Why should it have preferred
problem_fn<int> over problem_fn<void> or
problem_fn<::foo::bar::anytype_at_all*> ? :-)

> which implies gcc knew exactly which type to use

which was wrong.

>> The correct syntax would be to refer to &problem_fn<int>, but we still
>> can't do that in gcc, AFAIK.

> Specifying the type exactly slightly removes the point of templates though!

Not in this case.  Given the above declaration of problem_fn, you may
either:

// know in advance which type to instantiate:
void foo() { printf("%p\n", problem_fn<int>); }

// or not:
template <class T>
void bar() { printf("%p\n", problem_fn<T>); }

But the compiler must not be able to guess for you unless it has
enough context to decide.  In the example I gave below:

>> void (*f)(int&) = problem_fn;

The compiler is able (and required) to deduce the appropriate type for
problem_fn, so it can guess which specialization you're talking about.
However, in an ellipsis argument such as the one for printf, there's
no context information for the appropriate specialization to be
selected.

> This didn't work either when I tried it :-(. (Or do you mean, if the
> instantiation problem was fixed?).

That's what I meant.

-- 
Alexandre Oliva
mailto:oliva@dcc.unicamp.br mailto:aoliva@acm.org
Universidade Estadual de Campinas, SP, Brasil