Is Boost.Phoenix inherently slower than the equivalent C++11 lambdas (does it use virtual calls, 'volatile' usage, etc.)?

I always thought Boost.Phoenix used type-inference to infer everything statically, until I tried this code:

#include <vector>
#include <boost/phoenix/phoenix.hpp>

using namespace boost::phoenix;
using namespace boost::phoenix::placeholders;

struct Foo { int x; };

int main()
{
    std::vector<int> bar;
    bind(&Foo::x, ref(bar)[_1])("invalid index");   // oops
    return 0;
}

and got the warning:

warning C4239: nonstandard extension used : 'argument' : conversion from const char [3] to volatile const boost::proto::detail::anyns::any &
A non-const reference may only be bound to an lvalue

That surprised me. I didn't expect to see any anywhere, much less volatile!

Does that mean Boost.Phoenix is therefore actually inherently slower than its equivalent C++11 lambdas (ignoring the particular compiler I'm using here)?