ADL insanity
Consider the following program (hat tip to Agustín Bergé):
#include <stdio.h>
namespace A {
struct A {};
void call(void (*f)()) {
f();
}
}
void f() {
puts("Hello world");
}
int main() {
call(f);
}
Right now, this code doesn’t compile. Clang will helpfully tell you:
error: use of undeclared identifier 'call'; did you mean 'A::call'?
call(f);
^~~~
A::call
But we can make it compile by adding a completely unused overload of f
!
#include <stdio.h>
namespace A {
struct A {};
void call(void (*f)()) {
f();
}
}
void f() {
puts("Hello world");
}
void f(A::A); // UNUSED
int main() {
call(f);
}
This program compiles just fine and prints Hello world
on GCC, Clang, and ICC (but not MSVC).
And apparently it’s right to do so!
From [basic.lookup.argdep]/2:
[I]f the argument is the name or address of a set of overloaded functions and/or function templates, its associated entities and namespaces are the union of those associated with each of the members of the set, i.e., the entities and namespaces associated with its parameter types and return type.
And the standard continues:
Additionally, if the aforementioned set of overloaded functions is named with a template-id, its associated entities and namespaces also include those of its type template-arguments and its template template-arguments.
That is to say, GCC is technically correct to accept (and Clang, ICC, and MSVC are technically wrong to reject) the following program:
#include <stdio.h>
namespace B {
struct B {};
void call(void (*f)()) {
f();
}
}
template<class T>
void f() {
puts("Hello world");
}
int main() {
call(f<B::B>);
}
In both of these cases, we have an “argument” with no actual type. f
and f<B::B>
are names of overload sets, and an overload set doesn’t have a type. In order to collapse
the overload set down to a single function, we need to know what kind of function pointer type
is acceptable to the best-matching overload of call
… and that means we need to build
a candidate set for call
, which means we need to look up the name call
. Which happens via
ADL. Which normally requires that we know the types of our arguments!
So f
and f<B::B>
are “arguments without types,” but they are also arguments which
nevertheless contribute to argument-dependent lookup!
I suspect that there is no C++ code in the wild that relies on this little-known and frankly insane feature of ADL. It would be interesting to learn otherwise.