auto(x): decay-copy in the language

Changes Since R6

• compare decltype(auto(...)) with decay_t
• drop wording for LWG 3491; the author will propose its resolution separately
• discuss other papers, reference_constructs_from_temporary and “deducing this”

Changes Since R5

• discuss a few points that came across during EWG review
• add a table to summarize the library wording

Changes Since R4

• record LWG feedback
• rebase and refine the wording
• demo examples

Changes Since R3

• complete library wording

Changes Since R2

• dropped decltype(auto)(x) in comply with EWG’s opinion

Changes Since R1

• propose decltype(auto)(x) as well

Changes Since R0

• updated examples
• discussed decltype(auto)(x)
• added library wording

Introduction

This paper proposes auto(x) and auto{x} for casting x into a prvalue as if passing x as a function argument by value. The functionality appears as the decay-copy function in the standard for exposition only.

Motivation

Obtaining a prvalue copy is necessary

A generic way to obtain a copy of an object in C++ is auto a = x; but such a copy is an lvalue. We could often convey the purpose in code more accurately if we can obtain the copy as a prvalue. In the following example, let Container be a concept,

void pop_front_alike(Container auto& x) {
    std::erase(x.begin(), x.end(), auto(x.front()));
}

If we write

void pop_front_alike(Container auto& x) {
    auto a = x.front();
    std::erase(x.begin(), x.end(), a);
}

, questions arise – why this is not equivalent to

void pop_front_alike(Container auto& x) {
    std::erase(x.begin(), x.end(), x.front());
}

The problem is, the statement to obtain an lvalue copy is a declaration:

    auto a = x.front();

The declaration’s primary purpose is to declare a variable, while the variable being a copy is the declaration’s property. In contrast, the expression to obtain an rvalue copy is a clear command to perform a copy:

    auto(x.front())

One might argue that the above is indifferent from

    T(x.front())

However, there are plenty of situations that the T is nontrivial to get. We probably don’t want to write the original example as

void pop_front_alike(Container auto& x) {
    using T = std::decay_t<decltype(x.front())>;
    std::erase(x.begin(), x.end(), T(x.front()));
}

Obtaining a prvalue copy with auto(x) works always

In standard library specification, we use the following exposition only function to fulfill auto(x)'s role:

template<class T>
constexpr decay_t<T> decay_copy(T&& v) noexcept(
    is_nothrow_convertible_v<T, decay_t<T>>) {
    return std::forward<T>(v);
}

This definition involves templates, dependent constexpr, forwarding reference, noexcept, and two traits, and still has caveats if people want to use it in practice. An obvious issue is that decay_copy(x.front()) copies x.front() even if x.front() is a prvalue, in other words, a copy.

There is a less obvious issue that needs a code snippet to reproduce:

class A {
    int x;

public:
    A();

    auto run() {
        f(A(*this));           // ok
        f(auto(*this));        // ok as proposed
        f(decay_copy(*this));  // ill-formed
    }

protected:
    A(const A&);
};

The problem is that decay_copy is nobody’s friend. We can use A directly in this specific example. However, in a more general setting, where a type has access to a set of type T's private or protected copy/move constructors, decay-copy an object of T fails inside that type’s class scope, but auto(x) continues to work.

Discussion

auto(x) is a missing piece

Replacing the char in char('a') with auto, we obtain auto('a'), which is a function-style cast. Such a formula also supports injected-class-names and class template argument deduction in C++17. Introducing auto(x) and auto{x} significantly improves the language consistency:

** The type of x is a specialization of ClassTemplate.

With this proposal, all the cells in the table copy construct form x (due to CTAD’s default behavior) to obtain lvalues, prvalues, and pointers to objects, categorized by their columns. Defining auto(x) as a library^[1] facility loses orthogonality.

Introducing auto(x) into the language even improves the library consistency:

Note that auto(expr) does not turn a move-only lvalue expr into a prvalue regardless whether auto(expr) appears in an immediate context, so there is some difference between decltype(auto(expr)) and decay_t<decltype(expr)>.

Do we also miss decltype(auto){x}?

decltype(auto){arg} can forward arg without computing arg's type. It is equivalent to static_cast<decltype(arg)>(arg) . If arg is a variable of type T&&, arg is an lvalue but static_cast<T&&>(arg) is an xvalue.

EWG discussed this idea, disliked its expert-friendly nature, and concluded that adding this facility would cause the teaching effort to add up.

Does auto works in place of decay-copy in the library specification?

Not as a simple find-and-replace, but can be made to improve the quality of the library specification.

The background is that, despite being exposition-only, decay-copy always materializes its argument and produces a copy. auto(expr) is a no-op if the expr is a prvalue.

In the library specification where uses decay-copy, some do not mean to materialize the expressions; some want a new copy; some do not care. However, with auto(x) semantics, we should distinguish the different needs and explicitly say so when a copy is needed.

What’s the difference between decltype(auto(expr)) and decay_t?

decay_t<decltype(expr)> does not look at expr's value category, so the following can hold:

auto p = make_unique<char>();
static_assert(is_same_v<decay_t<decltype(p)>, unique_ptr<char>>);

decay_t<U&> will also simply strips & even if U is a move-only type.

As a result, the decay-copy exposition-only function is, strictly speaking, insufficiently constrained:

auto p2 = decay_copy(p);  // fails inside definition

The situation where decltype(decay_copy(expr)) does not fail in an immediate context does not imply that the call decay_copy(expr) is well-formed.

decltype(auto(expr)) has no such problem. But this does not necessarily mean that decltype(auto(expr)) is better than decay_t or something. Sometimes you do want to compute a “decayed-type,” and sometimes, you may find it helpful to exclude the expressions that cannot be passed as arguments to a parameter declared auto.

May a different spelling work better, such as prvalue_cast?

First, auto(x) casts to prvalue, but may not be how the name prvalue_cast suggests. If there is a prvalue_cast keyword, you may expect the following code to pass an array prvalue to the function foo:

double v[] = { 1.2, 3.5, .8 };
foo(prvalue_cast(v));

But

foo(auto(v));

will pass a double*. Therefore, we are talking about a different facility.

On the other hand, the term “prvalue” is unfamiliar to ordinary C++ users. While some people may expect a prvalue-casting without decaying, some other people may not have any expectation when seeing such a term.

In contrast, the use of the auto keyword implies that this expression decays. It decays in the same way

auto p = v;

does and in the same way

void bar(auto v);
bar(v);

does. This consistency makes the proposed spelling a lot more teachable comparing to the others.

Can we build a decay_copy with a metaprogram that forwards prvalue?

P2237 Metaprogramming means to replace macros, and the answer is yes. Barry kindly provides the following snippet:

consteval void prvalue_cast(meta::info expr) {
    -> decay_t<|type_of(expr)|>(|expr|);
}

However, we should note that we added auto type-specifier in declarations and new expressions not because we cannot meta-program in those contexts. We added them because they are useful, so does auto in a function-style cast, as shown in this paper.

As a side note, it may be a fun thing to see if someone comes up with the following code in the future:

auto p = new auto(prvalue_cast(v));

Should we also make static_cast<auto>(x) work?

The motivation is that because T(x) is indifferent from (T)x – a C-style cast that may try every possible route to T, some codebases may ban auto(x) altogether with T(x) even though auto is a type-placeholder rather than a type. static_cast<auto>(x) can be a way to evacuate from such a check.

In terms of styles, without loss of generality, char{x} means direct-list-initialization from x. It allows no narrowing and is not a cast. I believe that no codebase would ban T{x}, so auto{x} can be an alternative spelling if auto(x) is accidentally banned.

If we ignore the motivation for a moment, I think:

1. Whether static_cast<auto>(x) should work should also take static_cast<template-id>(x) (CTAD) into account. The latter is currently forbidden.
2. If we have such a paper, we may also want to decide whether static_cast<auto&&>(x) should work. This expression is a std::forward without a need to supply the type, but EWG has already turned down a similar idea (decltype(auto){x}).

Relation with other papers

P2255R0^[2] “A type trait to detect reference binding to temporary” guards std::tuple and std::pair constructors from forming references that bind to temporaries. The change is a good thing because it can make the following code fail to compile:

tuple<const string&> x(auto(obj.get_str()));

auto(x) can make copies if necessary, but this does not imply that “it is safer” without a context.

P0847R6^[3] “Deducing this” introduced by-value member functions. With those, a call to a member function may always copy the object:

foo(vec.sorted());

Combining it with auto(obj) has no cost:

foo(auto(vec).sorted());

You can also use this style when calling some member functions that are unwillingly modifying the implicit object parameters:

foo(auto(vec).sorted_in_place());

Demo

Prevent algorithm from modifying through aliases: https://godbolt.miator.net/z/hhcvbc

Using auto(x) in rvalue fluent interface: https://godbolt.miator.net/z/TY8sxr

How auto(x) assists in defining concepts: https://godbolt.miator.net/z/GTaaeE

Compare diagnosis to new auto(x): https://godbolt.miator.net/z/Ks43an

Wording

The wording is relative to N4878.

Part 1

Modify 7.6.1.4 [expr.type.conv]/1 as indicated:

A simple-type-specifier (9.2.9.3) or typename-specifier (13.8) followed by a parenthesized optional expression-list or by a braced-init-list (the initializer) constructs a value of the specified type given the initializer. If the type is a placeholder for a deduced class type, it is replaced by the return type of the function selected by overload resolution for class template deduction (12.2.2.9) for the remainder of this section. Otherwise, if the type is auto, it is replaced by the type deduced for the variable x in the invented declaration ([dcl.spec.auto]):

    auto x init;

, where init is the initializer.

Modify 9.2.9.6 [dcl.spec.auto]/5 as indicated:

A placeholder type can also be used in the type-specifier-seq in the new-type-id or type-id of a new-expression (7.6.2.8) and as a decl-specifier of the parameter-declaration’s decl-specifier-seq in a template-parameter (13.2). The auto type-specifier can also be used as the simple-type-specifier in an explicit type conversion (functional notation) ([expr.type.conv]).

Part 2

Modify 24.3.2 [range.access.begin]/2 as indicated:

Given a subexpression E with type T, let t be an lvalue that denotes the reified object for E. Then:

• If E is an rvalue and enable_borrowed_range<remove_cv_t<T>> is false, ranges::begin(E) is ill-formed.
• Otherwise, if T is an array type (6.8.3) and remove_all_extents_t<T> is an incomplete type, ranges::begin(E) is ill-formed with no diagnostic required.
• Otherwise, if T is an array type, ranges::begin(E) is expression-equivalent to t + 0.
• Otherwise, if decay-copyauto(t.begin()) is a valid expression whose type models input_or_output_iterator, ranges::begin(E) is expression-equivalent to decay-copyauto(t.begin()).
• Otherwise, if T is a class or enumeration type and decay-copyauto(begin(t)) is a valid expression whose type models input_or_output_iterator with overload resolution performed in a context in which unqualified lookup for begin finds only the declarations
    void begin(auto&) = delete;
  void begin(const auto&) = delete;
  then ranges::begin(E) is expression-equivalent to decay-copyauto(begin(t)) with overload resolution performed in the above context.
• Otherwise, ranges::begin(E) is ill-formed.

Modify 24.3.3 [range.access.end]/2 as indicated:

Given a subexpression E with type T, let t be an lvalue that denotes the reified object for E. Then:

• If E is an rvalue and enable_borrowed_range<remove_cv_t<T>> is false, ranges::end(E) is ill-formed.
• Otherwise, if T is an array type (6.8.3) and remove_all_extents_t<T> is an incomplete type, ranges::end(E) is ill-formed with no diagnostic required.
• Otherwise, if T is an array of unknown bound, ranges::end(E) is ill-formed.
• Otherwise, if T is an array, ranges::end(E) is expression-equivalent to t + extent_v<T>.
• Otherwise, if decay-copyauto(t.end()) is a valid expression whose type models sentinel_for<iterator_t<T>> then ranges::end(E) is expression-equivalent to decay-copyauto(t.end()).
• Otherwise, if T is a class or enumeration type and decay-copyauto(end(t)) is a valid expression whose type models sentinel_for<iterator_t<T>> with overload resolution performed in a context in which unqualified lookup for end finds only the declarations
    void end(auto&) = delete;
  void end(const auto&) = delete;
  then ranges::end(E) is expression-equivalent to decay-copyauto(end(t)) with overload resolution performed in the above context.
• Otherwise, ranges::end(E) is ill-formed.

Modify 24.3.6 [range.access.rbegin]/2 as indicated:

Given a subexpression E with type T, let t be an lvalue that denotes the reified object for E. Then:

• If E is an rvalue and enable_borrowed_range<remove_cv_t<T>> is false, ranges::rbegin(E) is ill-formed.
• Otherwise, if T is an array type (6.8.3) and remove_all_extents_t<T> is an incomplete type, ranges::rbegin(E) is ill-formed with no diagnostic required.
• Otherwise, if decay-copyauto(t.rbegin()) is a valid expression whose type models input_or_output_iterator, ranges::rbegin(E) is expression-equivalent to decay-copyauto(t.rbegin()).
• Otherwise, if T is a class or enumeration type and decay-copyauto(rbegin(t)) is a valid expression whose type models input_or_output_iterator with overload resolution performed in a context in which unqualified lookup for rbegin finds only the declarations
    void rbegin(auto&) = delete;
  void rbegin(const auto&) = delete;
  then ranges::rbegin(E) is expression-equivalent to decay-copyauto(rbegin(t)) with overload resolution performed in the above context.
• […]

Modify 24.3.7 [range.access.rend]/2 as indicated:

Given a subexpression E with type T, let t be an lvalue that denotes the reified object for E. Then:

• If E is an rvalue and enable_borrowed_range<remove_cv_t<T>> is false, ranges::rend(E) is ill-formed.
• Otherwise, if T is an array type (6.8.3) and remove_all_extents_t<T> is an incomplete type, ranges::rend(E) is ill-formed with no diagnostic required.
• Otherwise, if decay-copyauto(t.rend()) is a valid expression whose type models sentinel_for<decltype(ranges::rbegin(E)> then ranges::rend(E) is expression-equivalent to decay-copyauto(t.rend()).
• Otherwise, if T is a class or enumeration type and decay-copyauto(rend(t)) is a valid expression whose type models sentinel_for<decltype(ranges::rbegin(E)> with overload resolution performed in a context in which unqualified lookup for rend finds only the declarations
    void rend(auto&) = delete;
  void rend(const auto&) = delete;
  then ranges::rend(E) is expression-equivalent to decay-copyauto(rend(t)) with overload resolution performed in the above context.
• […]

Modify 24.3.10 [range.prim.size]/2 as indicated:

Given a subexpression E with type T, let t be an lvalue that denotes the reified object for E. Then:

• If T is an array of unknown bound (9.3.4.5), ranges::size(E) is ill-formed.
• Otherwise, if T is an array type, ranges::size(E) is expression-equivalent to decay-copyauto(extent_v<T>).
• Otherwise, if disable_sized_range<remove_cv_t<T>> (24.4.3) is false and decay-copyauto(t.size()) is a valid expression of integer-like type (23.3.4.4), ranges::size(E) is expression-equivalent to decay-copyauto(t.size()).
• Otherwise, if T is a class or enumeration type, disable_sized_range<remove_cv_t<T>> is false and decay-copyauto(size(t)) is a valid expression of integer-like type with overload resolution performed in a context in which unqualified lookup for size finds only the declarations
    void size(auto&) = delete;
  void size(const auto&) = delete;
  then ranges::size(E) is expression-equivalent to decay-copyauto(size(t)) with overload resolution performed in the above context.
• […]

Modify 24.3.13 [range.prim.data]/2 as indicated:

Given a subexpression E with type T, let t be an lvalue that denotes the reified object for E. Then:

• If E is an rvalue and enable_borrowed_range<remove_cv_t<T>> is false, ranges::data(E) is ill-formed.
• Otherwise, if T is an array type (6.8.3) and remove_all_extents_t<T> is an incomplete type, ranges::data(E) is ill-formed with no diagnostic required.
• Otherwise, if decay-copyauto(t.data()) is a valid expression of pointer to object type, ranges::data(E) is expression-equivalent to decay-copyauto(t.data()).
• […]

Modify 32.4.3.3 [thread.thread.constr]/5 as indicated:

Effects: The new thread of execution executes

  invoke(decay-copyauto(std::forward<F>(f)),
   decay-copyauto(std::forward<Args>>(args))…)

with the calls to decay-copy being evaluatedvalues produced by auto being materialized ([conv.rval]) in the constructing thread. Any return value from this invocation is ignored. […]

Modify 32.4.4.2 [thread.jthread.cons]/5 as indicated:

Effects: Initializes ssource. The new thread of execution executes

  invoke(decay-copyauto(std::forward<F>(f)), get_stop_token(),
   decay-copyauto(std::forward<Args>>(args))…)

if that expression is well-formed, otherwise

  invoke(decay-copyauto(std::forward<F>(f)),
   decay-copyauto(std::forward<Args>>(args))…)

with the calls to decay-copy being evaluatedvalues produced by auto being materialized ([conv.rval]) in the constructing thread. Any return value from this invocation is ignored. […]

Modify 32.9.9 [futures.async]/3 as indicated:

Effects: The first function behaves the same as a call to the second function with a policy argument of launch::async | launch::deferred […]:

• If launch::async is set in policy, calls invoke(decay-copyauto(std::forward<F>(f)), decay-copyauto(std::forward<Args>>(args))…) (20.14.4, 32.4.3.3) as if in a new thread of execution represented by a thread object with the calls to decay-copy being evaluatedvalues produced by auto being materialized ([conv.rval]) in the thread that called async. Any return value is stored as the result in the shared state. Any exception propagated from the execution of invoke(decay-copyauto(std::forward<F>(f)), decay-copyauto(std::forward<Args>>(args))…) is stored as the exceptional result in the shared state. The thread object is stored in the shared state and affects the behavior of any asynchronous return objects that reference that state.
• If launch::deferred is set in policy, stores decay-copyauto(std::forward<F>(f)) and decay-copyauto(std::forward<Args>(args))... in the shared state. These copies of f and args constitute a deferred function. Invocation of the deferred function evaluates invoke(std::move(g), std::move(xyz)) where g is the stored value of decay-copyauto(std::forward<F>(f)) and xyz is the stored copy of decay-copyauto(std::forward<Args>(args)).... Any return value is stored as the result in the shared state. Any exception propagated from the execution of the deferred function is stored as the exceptional result in the shared state. […]

Acknowledgments

Thank Alisdair Meredith, Arthur O’Dwyer, and Billy O’Neal for providing examples and feedback for this paper. Thank James Touton for presenting the paper and bringing it forward. Thank Jens Maurer and Casey Carter for reviewing the wording.

References