with the following code (a boiled-down version of my original code)
#include
#include
template class A;
As I have figured out by now (see also the edits to the question) what the problem was. It turns out that all answers (sofar) didn't solve the (full) problem.
Essentially, the evaluation of the decltype
expression for the global function required eventually the private member A<X>::a
(through its occurrence in the decltype
expression of A<x>::diff()
). This opens the next question about what the standard says about this.
auto diff(A<Y> const&y) const -> decltype(a-y.a)
is the problem; regardless of anything else, if X
and Y
are different types, A<X>
and A<Y>
are different types and cannot peek at each other's privates. Templates are not covariant!
The specific error here might be an eccentricity of GCC (in that it doesn't spot that X
is the same type as Y
) but the more general case where you might be trying to diff two different types (and why else would you have a separate template type in your diff function?) will never work, regardless of compiler.
TL;DR: Gcc appears to have a bug where trailing return types on template member functions are not treated as within the class's scope.
This bug causes gcc to fail when instantiating the template member function auto diff(A<Y> const&y) const -> decltype(a-y.a)
because a
is private and gcc thinks private members are inaccessible here.
The code builds fine with clang and VC++, and I don't see anything trying to access A<double>::a
outside A<double>
, so it looks like a gcc bug to me.
Others have mentioned that A<X>
and A<Y>
are different classes, but that's not the case here, both are A<double>
. I believe that means that friendship is not necessary in this case, although to work in the general case A<X>
does need to be friends with other specializations of A.
Specifically, a
in y.a
is a dependent name so it cannot be looked up until A<Y>
is known. At that point lookup is done, the accessibility is tested and it should be found that A<double>
does have access to A<double>::a
.
Here's the exact code I compiled in both clang (svn-3.2) and VC++11 (Since I'm using clang on Windows I can't #include <iostream>
)
#include <cmath>
template<typename X> class A {
X a;
public:
A(X x) : a(x) {}
template<typename Y>
auto diff(A<Y> const&y) const -> decltype(a-y.a)
{ return a-y.a; }
};
template<typename X, typename Y>
inline auto dist(A<X> const&x, A<Y> const&y) -> decltype(std::abs(x.diff(y)))
{ return std::abs(x.diff(y)); }
int main()
{
A<double> x(2.0), y(4.5);
return (int) dist(x,y);
}
This code results in build errors on gcc 4.5 similar to what you describe.
Replacing
auto diff(A<Y> const&y) const -> decltype(a-y.a)
with
auto diff(A<Y> const&y) const -> typename std::common_type<X,Y>::type
causes the code to work on gcc 4.5.
This indicates to me a bug where gcc is failing to treat trailing return types as inside the class's scope. Some testing reveals that the trailing return type must be on a template function to trigger the bug.
There is an error with your code:
template<typename Y>
auto diff(A<Y> const&y) const -> decltype(a-y.a)
{ return a-y.a; }
Here, A<Y>
is a different type, so A<X>
cannot see it's a
data member. Only A<Y>
can see A<Y>::a
.
Edit: that said, in your particular case, X
and Y
are both double
, so I would naively expect that to compile. Note that in the best of cases, this construction should only compile when X
and Y
are the same, which may not be what you want.
Your function auto diff(A<Y> const& y)
access the private
variable A::a
from outside the class in your decltype
statement.
Your A::a
should be public
if you want to use it the way you do in your diff
function.
EDIT: it seems the friendship solution is much better suited for this problem, then just making it public.