In this Q&A I wrote a little wrapper class that provides reverse iterator access to a range, relying on the c++1z language feature template argument deduction for class templates (p0091r3, p0512r0)
#include <iostream>
#include <iterator>
#include <vector>
template<class Rng>
class Reverse
{
Rng const& rng;
public:
Reverse(Rng const& r) noexcept
:
rng(r)
{}
auto begin() const noexcept { using std::end; return std::make_reverse_iterator(end(rng)); }
auto end() const noexcept { using std::begin; return std::make_reverse_iterator(begin(rng)); }
};
int main()
{
std::vector<int> my_stack;
my_stack.push_back(1);
my_stack.push_back(2);
my_stack.puhs_back(3);
// prints 3,2,1
for (auto const& elem : Reverse(my_stack)) {
std::cout << elem << ',';
}
}
However, doing a nested application of Reverse
does not yield the original iteration order
// still prints 3,2,1 instead of 1,2,3
for (auto const& elem : Reverse(Reverse(my_stack))) {
std::cout << elem << ',';
}
Live Example (same output for g++ 7.0 SVN and clang 5.0 SVN)
The culprit seems to be the template argument deduction for class templates because the usual wrapper function does allow for correct nesting
template<class Rng>
auto MakeReverse(Rng const& rng) { return Reverse<Rng>(rng); }
// prints 1,2,3
for (auto const& elem : MakeReverse(MakeReverse(my_stack))) {
std::cout << elem << ',';
}
Live Example (same output for g++ and clang)
Question: is nested template argument deduction for class templates supposed to work only "one level" deep, or is this a bug in the current implementations of both g++ and clang?
Piotr's answer correctly explains what is happening - the move constructor is a better match than your constructor template.
But (h/t T.C. as usual) there's a better fix than just writing a factory anyway: you can add an explicit deduction guide to handle the wrapping:
The point of this is to override the copy deduction candidate, thanks to the newly added preference in [over.match.best] for this:
Hence, we'd have four generated functions, borrowing again from Piotr's naming:
Before,
#3
was preferred as being more specialized. Now,#4
is preferred as being a deduction guide. So, we can still write:and that works.
This may be explained in [over.match.class.deduct]/p1:
My understanding is that the compiler invents the following two functions (two - including a copy constructor that is implicitly generated for this class):
and then tries to select the best overload based on the call:
which resolves to #2 because this one is more specialized. The conclusion is that:
involves a copy constructor to construct the outer
Reverse
instance.