bind_front should not unwrap reference_wrapper

1. Introduction

This paper proposes a change in the behaviour of the std::bind_front in regards to the bound arguments of the std::reference_wrapper<T> type — the arguments should not be unwrapped (passing rw.get() to underlying callable), and should be propagated unchanged instead (passing rw unmodified to underlying callable).

This change is the result of analysis of use case provided by the Abseil team, that I have not considered during the initial design. The main motivation of the change is to reduce the extent of damage that may be implied by code that assumes "incorrect" behavior.

Changes proposed in this paper need to be considered in the C++20 timeline, as they would constitute breaking change after the publication of standard in the current form.

2. Revision history

2.1. Revision 0

Initial revision.

3. Motivation and Scope

This paper discusses the behavior of the code that uses std::bind_front in conjunction with std::reference_wrapper instantations. To illustrate lets consider the following code:

Thingy thingy;
auto boundFunctor = std::bind_front(func, std::ref(thingy));

There are two possible and conflicting behaviors of the boundFunctor() invocation (for purpose of presentation we refer to bound reference_wrapper as thingyRef):

• unwrapping: func(thingyRef.get()), i.e. passing Thingy & — mimics behaviour of std::bind_front w.r.t. reference_wrapper,
• propagation: func(thingyRef), i.e. passing std::reference_wrapper<Thingy> — std::reference_wrapper is treated as any other class.

3.1. Unwrapping use case: Double conversion

One of the situations when the difference between unwrapping and propagation semantic is visible, is when the functor is bound with the reference to the object (e.g. std::reference_wrapper<std::string>, and the code accepts an view to it (e.g. std::string_view). To illustrate lets consider:

void functionAcceptingStringView(std::string_view);
void functionAcceptingSpanOfIntegers(std::span<int>);

std::string s; 
auto fs = std::bind_front(&functionAcceptingStringView, std::ref(s));
std::vector<int> v;
auto fv = std::bind_front(&functionAcceptingSpanOfIntegers, std::ref(v));

With the unwrapping (current) semantic, both fs() and fv() compiles correctly, as std::string_view can be implicitly constructed from std::string& and std::span<int> can be implicitly constructed from std::vector<int>&. In case if propagation semantic was provided, both invocation would not compile, as they would require two user defined conversion to be performed:

• std::reference_wrapper<std::string> → std::string& → std::string_view for fs(), and
• std::reference_wrapper<std::vector<int>> → std::vector<int>& → std::span<int> for fv().

auto fs = std::bind_front(&functionAcceptingStringView, std::string_view(s));
auto fv = std::bind_front(&functionAcceptingSpanOfIntegers, std::span(v));

3.2. Propagation use case: Function currying

The difference between the propagation is unwrapping semantic, is that the former support rebinding of the argument. This can be best illustrated with the implementation of the functor PartialApply that implements function currying in C++:

template <typename F>
struct PartialApply {
    PartialApply(F f) : f(f) {}
    F f;

    template <typename... A> auto operator()(A const&... a) const {
        if constexpr (std::is_invocable<F const&, A const&...>::value) {
            return f(a...);
        } else {
            return bind_front(*this, a...);
        }
    }
};

The intent of the above code is that the expression PartialApply(func)(a)(b) is either:

• calling func(a, b) if that is well-formed,
• returning callback that binds func with a and b and accepts consecutive arguments (rebounding).

The above implementation works flawlessly in case of the propagation semantic, however, it fails in the case of unwrapping is used — in the case when rebinding is performed, copy of arguments that were originally passed via std::ref is made.

In the most optimistic scenario the above issue will manifest as compilation error. This happens when move only type is passed by reference:

std::unique_ptr<Thingy> thingy;
auto func = [](std::unique_ptr<Thingy>&, int) {};

PartialApply(func)(std::ref(thingy))(10);

In the case of copyable types, silently copy of the object will be created. This, of course, may have negative performance implication, but in the worst case can lead to dangling references:

std::string str;
auto func = [](std::string& s, int) -> std::string& { return s; };

std::string& sref = PartialApply(func)(std::ref(s))(10); // sref refers to copy of str stored in PartialApply
sref.push_back('a');                                     // use of dangling reference

Finally, the author is not aware of the way of fixing PartialApply than reimplementing custom bind_front alternative that has propagation semantic.

3.3. Historical design note on LWG 2219

The prominent use case for partial function application functions (like std::bind_front and std::bind), is to provide a helper to compose given method on the class, with a specific object. In case if invocation should be performed on existing object instance (not a copy), a std::reference_wrapper was used as follows:

auto bound = std::bind_front(&Object::method, std::ref(instance));

In the time when the std::bind_front was originally proposed (also applies to std::bind), the above code to work property required unwrapping semantic (at least the for first argument). This was changed with the resolution of LWG 2219: INVOKE-ing a pointer to member with a reference_wrapper as the object expression, that have introduced dedicated handling for std::reference_wrapper in INVOKE, and unwrapping of std::reference_wrapper is no longer required to achieve this functionality.

4. Design Decisions

4.1. Preference for compilation errors

As indicated in the motivation section, the difference between unwrapping and propagation semantic manifest itself in a very specific scenario. As a consequence, the user, unaware of this specific behavior, may accidentally create an erroneous code that depends on semantic other that one supplied by the standard. In that case, it would be desired that error is detected as early as possible, preferably at compile time.

In light of the above, this paper proposes to switch to propagation semantic, as code that assumes unwrapping becomes ill-formed. With the current behavior (unwrapping), the code that assumes propagation may lead to runtime bugs.

4.2. Preserving other behavior

The change proposed in this paper has no impact on user-visible behaviour of std::bind_front , despite the fact that functors returned by this functor will now store std::reference_wrapper<T> instead of T&, as:

• functors produced by std::bind_front are not assignable, so the difference in assignment semantic is not visible to the user,
• reference_wrapper, like references, is required to be trivially copyable,
• bind_front is not currently usable in constexpr context.

In addition the P1065: constexpr INVOKE applies constexpr to std::reference_wrapper.

5. Proposed Wording

The proposed wording changes refer to N4810 (C++ Working Draft, 2019-03-15).

Apply following changes to section [func.bind_front] Function template bind_front:

  template <class F, class... Args>
     unspecified bind_front(F&& f, Args&&... args);
  

In the text that follows:

• g is a value of the result of a bind_front invocation,
• FD is the type decay_t<F>,
• fd is a target object of g ([func.def]) of type FD initialized with initializer (std::forward<F>(f)),
• BoundArgs is a pack that denotes std::unwrap_­ref_­decay_­t<Args>...,
• bound_args is a pack of bound argument entities of g ([func.def]) of types BoundArgs... initialized with initializers (std::forward<Args>(args))... respectively, and
• call_args is an argument pack used in a function call expression ([expr.call]) of g,

Update the value of the __cpp_lib_bind_front in table "Standard library feature-test macros" of [support.limits.general] to reflect the date of approval of this proposal.

6. Implementability

This paper can be implemented by simply replacing unwrap_ref_decay_t with decay_t in example implementation from "Implementability" section of the P0365R5: Simplified partial function application paper.

7. Acknowledgements

Titus Winters and Abseil team for providing the feedback on the bind_front and code example that motivated creation of this paper.

Samuel Benzaquen for providing feedback and corrections for this paper.

Special thanks and recognition goes to Sabre (http://www.sabre.com) for supporting the production of this proposal and author's participation in standardization committee.

8. References

1. Jonathan Wakely, Issue 2219. INVOKE-ing a pointer to member with a reference_wrapper as the object expression, (LWG2219, https://wg21.link/lwg2219)
2. Barry Revzin, constexpr INVOKE, (P1065, https://wg21.link/p1065)
3. Richard Smith, "Working Draft, Standard for Programming Language C++" (N4810, https://wg21.link/n4810)
4. Tomasz Kamiński, Simplified partial function application, (P0356R5, https://wg21.link/p0356r5)