Logo Questions Linux Laravel Mysql Ubuntu Git Menu
 

Semantic correctness of non-instantiated C++ template functions

The following C++ code does not compile e.g. with g++-4.7 or clang++-3.2:

struct Bar {};

template<typename T>
void foo(T t, Bar bar) {
    t.compiler_does_not_care();
    bar.nonexistent_method();
}

int main() {}

Why do compilers check the code of the template function foo for semantic correctness (where they can) even though it is never instantiated? Is this standard compliant?

like image 366
user1225999 Avatar asked Aug 20 '14 13:08

user1225999


People also ask

Is it necessary to instantiate a template?

In order for any code to appear, a template must be instantiated: the template arguments must be provided so that the compiler can generate an actual class (or function, from a function template).

What is typename in template?

In template definitions, typename provides a hint to the compiler that an unknown identifier is a type. In template parameter lists, it's used to specify a type parameter.

What is the difference between typename and class in template?

There is no difference between using <typename T> OR <class T> ; i.e. it is a convention used by C++ programmers.

Which one is suitable syntax for function template?

Which one is suitable syntax for function template? Explanation: Both class and typename keywords can be used alternatively for specifying a generic type in a template.


1 Answers

Bar is a non dependent name (i.e. its type does not depend on T), so the compiler is required to verify the correctness of the code during the first phase of name-lookup (see the note below).

Since Bar has no nonexistent_method() method, the compiler is required to issue a diagnosis.

If you change your template to:

template<typename T>
void foo(T t, T bar) {
    t.compiler_does_not_care();
    bar.nonexistent_method();
}

No non-dependent names are involved, so no error is emitted since the template is never instantiated (phase 2 of the lookup)


Notes:

  • Comprehensible description of two-phase name lookup from LLVM :

1) Template definition time: when the template is initially parsed, long before it is instantiated, the compiler parses the template and looks up any "non-dependent" names. A name is "non-dependent" if the results of name lookup do not depend on any template parameters, and therefore will be the same from one template instantiation to another.

2) Template instantiation time: when the template is instantiated, the compiler looks up any "dependent" names, now that it has the full set of template arguments to perform lookup. The results of this lookup can (and often do!) vary from one template instantiation to another.

  • As for the why non-dependent name lookup can't be deferred to the second stage, see this other SO post; it seems that it is mostly for historical reasons.
like image 65
quantdev Avatar answered Sep 17 '22 20:09

quantdev