std::result_of and SFINAE

ISO/IEC JTC1 SC22 WG21 N3462=12-0152 - 2012-10-18

Eric Niebler (eric.niebler@gmail.com)
Daniel Walker (daniel.j.walker@gmail.com)
Joel de Guzman (djowel@gmail.com)

This paper deals with the use of result_of in contexts where SFINAE is a consideration; i.e., in function signatures. It is an update to N3436. result_of is intended to be a simple way to compute the result type of some callable when provided with arguments of specified types. It is a shorthand for a longer type computation using decltype and declval. However, there are cases where the use of result_of leads to a hard error whereas the equivalent use of decltype and declval would not. This happens in function signatures, when the direct use of decltype of an ill-formed call expression causes the function to be dropped due to SFINAE.

This paper addresses the issue by making result_of SFINAE-friendly. It does this by conditioning the presence of the nested type typedef on whether the call expression is ill-formed or not.

This paper will do the following:

Problem
    Trivial Example
    Reports from the Field
Solution
    Proposed Resolution
    Options Considered and Dismissed
    Implementation Experience
    Differences from the Pre-Portland Mailing
Acknowledgements
Appendix A: Example of an Error in a Non-Immediate Context
Appendix B: Example of Why Fn Shouldn't Be Required to Be a Callable Type

Problem

The direct use of decltype in a function signature would cause SFINAE to kick in when the expression in the decltype is invalid. In contrast, the use of result_of in a function signature, as it is currently specified, causes a hard error because the call expression appears in a non-SFINAE context: within the definition of the nested type typedef.

Trivial Example

Below is a program that compiles when the return type is declared using decltype, but that fails to compile when using result_of.


//#define BUGBUG
#include <type_traits>
#include <utility>
#include <string>

struct eat { template<typename T> eat(T const &) {} };
struct not_incrementable {};

struct inc {
    template<typename T>
    auto operator()(T t) const -> decltype(t++)
        { return t++; }
};

template<typename A>
#ifdef BUGBUG
  typename std::result_of<inc(A)>::type            // HARD ERROR HERE
#else
  decltype(std::declval<inc>()(std::declval<A>())) // SFINAE HERE
#endif
try_inc(A a) {
    return inc()(a);
}

not_incrementable try_inc(eat) {
    return not_incrementable();
}

int main() {
    int x = try_inc(1); // OK
    not_incrementable y = try_inc(std::string("foo")); // OK, not_incrementable
}

Reports from the Field

You might wonder if this is really a problem that needs to be fixed, or whether instead it is simply a case of people misusing result_of. Even if result_of were never intended to be used this way, there is evidence that users find the current behavior unsatisfactory:

Boost
When boost::result_of was switched to use decltype, some of Boost's tests started failing. After investigation, we discovered that the TR1-style result_of had been implemented such that it was SFINAE-friendly, and that folks had discovered this and come to rely on it. The amount of affected code within Boost was small, but Boost became concerned about its downstream users and are now in the process of making their decltype-based result_of SFINAE-friendly to prevent downstream breakage. (Aside: the problems in Boost have already been found by users and reported both on StackOverflow and in Boost's bug tracker.)
GCC bug report
The linked bug report, entitled "substitution failure reports error with result_of", describes this problem. Says Paolo Carlini:

"Unfortunately, I don't think you can use std::result_of for sfinae purposes, if I understand correctly it's a well know annoyance which you have to overcome by open coding with std::declval."

LWG DR1225
The submitter of DR1225 wants std::result_of to be SFINAE-friendly (among other things) and offers wording. He is under the impression that this change would require compiler support (perhaps in order to correctly handle errors in the non-immediate instantiation context; see CWG DR1227). This part of the DR was rejected, the rationale given being: "The wish to change result_of into a compiler-support trait was beyond the actual intention of the submitter Sebastian." Note that a SFINAE-friendly result_of can be implemented in standard C++11, at least insofar as it be allowed to fail when the first error appears in a non-immediate context (see Appendix A).
Complaints from Boost's Users

From: Ian McCulloch
Subject: Re: result_of tutorial
Date: 2005-03-17 08:22:29 GMT

Thanks Peter, I think I understand how it works now. Unfortunately, the lack of SFINAE is a real showstopper for me. I want to write functions like

template <typename T>
typename result_of<negate(T)>::type
operator-(T const& x)
{
   return negate()(x);
}

but the lack of SFINAE here makes boost::result_of essentially useless for this. But I imagine it won't be difficult to make a new version based on boost that would work."

A Blog Post
Although written in Japanese, the author shows an invoke function defined with result_of that fails to compile because it (result_of) is not SFINAE-friendly. It also calls out std::common_type as another trait that would benefit from being SFINAE-friendly.

Solution

Notes About the Proposed Wording Change

The intention of the proposed change is to condition the presence of the nested type typedef of result_of on whether the call expression used in the decltype-specifier is well-formed. The changes occur in Table 57 (section 20.9.7.6, [meta.trans.other]). An attempt is made at the standardese to permit result_of to issue a hard error if and only if the first error is encountered in a non-immediate context. This bit of language borrows heavily from the specification of is_assignable in Table 49 (section 20.9.4.3, [meta.unary.prop]). (See Appendix A for an example of an error in a non-immediate context.)

Proposed Resolution

Options Considered and Dismissed

Eric Niebler started a discussion on the -lib reflector about result_of and SFINAE in c++std-lib-32993. Below are some of the considerations brought up in the ensuing thread, and the authors' thoughts.

Should result_of simply be deprecated?

In c++std-lib-32999, Nikolay Ivchenkov floats the idea that result_of be deprecated, thereby solving the SFINAE problem by encouraging users to use decltype and declval directly. Others (namely Howard Hinnant and Pete Becker) echoed Nikolay's sentiment that result_of's unconventional use of function types is confusing to some users. In c++std-lib-33010, Anthony Williams points out that result_of<X(Y)>::type is simpler than the equivalent type calculation involving decltype and declval. The authors of this paper agree and do not recommend deprecating result_of.

Should result_of be implemented in terms of more general and idiomatic traits?

In c++std-lib-33011, Howard Hinnant described an interface that he has come to prefer: std::invoke_of and std::invokable. Rather than taking one template argument of function type, invoke_of would take N parameters, the first of which is the callable type and the rest are the types of the arguments. invokable is a Boolean trait that takes the same parameters and computes whether the specified call expression is well-formed or not. result_of is trivially implemented in terms of such traits.

The authors of this paper are not opposed to this change. It is, however, a larger change than necessary. As such, the authors are not proposing it here. Note that adopting the resolution suggested in this paper would not block any future proposal that recommends such a change.

Shouldn't we also address the other traits that could be made SFINAE-friendly?

As noted by some users and by Marc Glisse in c++std-lib-32994, result_of is far from the only trait that could benefit from the SFINAE treatment. iterator_traits and common_type are obvious candidates. The authors of this paper have chosen to narrowly focus on result_of for lack of time.

Implementation Experience

Has this been implemented yet? Yes. In c++std-lib-33002, Howard Hinnant claims that his libc++ library has already shipped a SFINAE-friendly result_of. In c++std-lib-33013, he links to his implementation. It can be found at http://llvm.org/svn/llvm-project/libcxx/trunk/include/type_traits.

Daniel Krügler claimed in a private exchange to have recently implemented a SFINAE-friendly result_of for gcc.

Also, Boost recently switched to a SFINAE-friendly result_of, although it hasn't shipped it yet at the time of writing (2012-09-20).

Differences from Pre-Portland Mailing

Since the pre-Portland mailing went out, Daniel Krügler brought two issues to the authors' attention. Both have been addressed by changing the proposed wording. The changes are as follows:

  1. The requirement on Fn to be a "callable type (20.8.1), reference to function, or reference to callable type" has been struck.
  2. An additional requirement for type completeness of Fn and all the types in the parameter pack ArgTypes has been added.

In support of (1), Daniel Krügler observes that simply removing that requirement allows for SFINAE to drop overloads based on the non-callablility of the Fn template parameter. "Not ever callable" is not that logically different from "not callable with certain arguments," and would in fact be handled naturally by the most naive implementation. Appendix B shows an example where this distinction makes a difference.

The change described in bullet (2) is to make all the types in the template parameter complete. Daniel Krügler describes his reasoning as follows in a private email:

The other point (I mentioned it already in my other follow-up) is that your suggested wording omits an important constraint, namely type completeness. The reason for this can be directly deduced from the definition of INVOKE in 20.8.2 p1: Bullets 1 and 3 require to determine whether T1 is a of the same or of a derived type of the class type of the pointer of member. Also, we need a complete type of Fn, if this is a class type with a potentially existing function call operator. As a conservative start I suggest to impose the very same precondition as we have for is_constructible or for common_type [...]

Acknowledgements

Many thanks to those on the committee reflector who provided feedback about this issue; namely, Marc Glisse, Nikolay Ivchenkov, David Abrahams, Anthony Williams, Pete Becker, Peter Dimov, Doug Gregor and especially Daniel Krügler, who suggested wording and pointed out a number of ways to improve the resolution. Thanks are also due to the Boost users and developers who pushed for this change and made suggestions, especially Michel Morin and Jeffrey Lee Hellrung, Jr.

Appendix A: Example of an Error in a Non-Immediate Context

The current proposal allows that while testing an expression for well-formed-ness, a hard error can occur. No attempt is made to make result_of usable with such "indirectly" ill-formed expressions. Below is an example of a callable expression that will trigger a hard error, regardless of whether result_of is "SFINAE-friendly" or not.

template<class X>
struct Fail
{
  static_assert(sizeof(X)==0,"duh");
  typedef int type;
};

struct Fun
{
  template<class T>
  typename Fail<T>::type operator()(T)
      { return 0; }
};

template<class T>
typename std::result_of<Fun(T)>::type foo(T)
   { return 0; }

template<class T>
int foo(...)
   { return 0; }

int main()
{
   foo(0);
}

Appendix B: Example of Why Fn Shouldn't Be Required to Be a Callable Type

This example comes from Daniel Krügler. It demonstrates some legitimate code where you'd like SFINAE to exclude the use of result_of when the callable is not, in fact, callable. If the original requirement on the callability of Fn is retained, the code below becomes undefined, and possibly a hard error.

#include <type_traits>

template<class Fn>
typename std::result_of<Fn(int)>::type f(Fn& t);

template<class T, class U>
int f(U u);

struct S {};

int main() {
  f<S>(0);
}