Document number: P1141R1

Ville Voutilainen
Thomas Köppe
Andrew Sutton
Herb Sutter
Gabriel Dos Reis
Bjarne Stroustrup
Jason Merrill
Hubert Tong
Eric Niebler
Casey Carter
Tom Honermann
Erich Keane
Walter E. Brown
Michael Spertus
2018-10-08

Yet another approach for constrained declarations

Abstract

We propose a short syntax for the constrained declaration of function parameters, function return types and variables. The new syntax is a “constrained auto”, e.g. void sort(Sortable auto& c);.

Contents

  1. Revision history
  2. Proposal summary
  3. Part 1: “Constrained auto
  4. Part 2: Relaxed “constrained auto
  5. Part 3: Meaning of “template <Concept T>
  6. Part 4: Meaning of “template <Concept... T>” and its friends
  7. Part 5: Meaning of “-> Concept auto” and its friends

Revision history

Proposal summary

This paper proposes three things:

  1. A syntax for constrained declarations that is practically a “constrained auto”; the principle being “wherever auto goes, a Constraint auto can also (non-recursively) go”. The semantics are to deduce like auto and additionally check a constraint. In a nutshell,
    void f(Sortable auto x);
    Sortable auto f();      // #1
    Sortable auto x = f();  // #2
    template <Sortable auto N> void f();
    and all combined:
    template <Sortable auto N> Sortable auto f(Sortable auto x)
    {
        Sortable auto y = init;
    }
    An unconstrained version of that is:
    template <auto N> auto f(auto x)
    {
        auto y = init;
    }
    So, this proposal includes auto-typed parameters for functions, which we already allow for lambdas.
  2. An additional relaxation where the auto is optional for the cases #1 and #2 illustrated above:
    Sortable f();
    Sortable x = f();
  3. Simplifying (and thus restricting) the rules in [temp.param]/10, so that template <Sortable S> always means that S is a type parameter, and template <Sortable auto S> always means that S is a non-type parameter. Template template-parameters are no longer supported in this short form. Moreover, Sortable is restricted to be a concept that takes a type parameter or type parameter pack; non-type and template concepts are no longer supported in this short form.

Sortable is a “type concept” in all the examples of this summary.

This paper specifically does not propose

The idea of this approach is to provide a syntax that

Part 1: “Constrained auto

The approach proposed here borrows a subset of P0807R0 An Adjective Syntax for Concepts. The idea is that we don’t try to come up with a notation that does everything that P0807 does; in particular, there is no proposal for a new syntax to introduce a type name.

Function templates

The approach is simple: allow auto parameters to produce function templates (as they produce polymorphic lambdas), and allow the auto to be preceded by a concept name. In every case, such a parameter is a deduced parameter, and we can see which parameters are deduced and which ones are not:

[](auto a, auto& b, const auto& c, auto&& d) {...}; // unconstrained
[](Constraint auto a, Constraint auto& b, const Constraint auto& c, Constraint auto&& d) {...}; // constrained

void f1(auto a, auto& b, const auto& c, auto&& d) {...}; // unconstrained
void f2(Constraint auto a, Constraint auto& b, const Constraint auto& c, Constraint auto&& d) {...}; // constrained

[](Constraint auto&& a, SomethingElse&& b) {...}; // a constrained deduced forwarding reference and a concrete rvalue reference
void f3(Constraint auto&& a, SomethingElse&& b) {...}; // a constrained deduced forwarding reference and a concrete rvalue reference

The appearance of auto (including Constraint auto) in a parameter list tells us that we are dealing with a function template. For each parameter, we know whether it is deduced or not. We can tell apart concepts from types: concepts precede auto, types do not.

Return types and variable declarations

Constrained return types work the same way:

auto f4();                  // unconstrained, deduced.

Constraint auto f5();       // constrained, deduced.

Whatever f6();              // See part 2. If Whatever is a type, not deduced.
                            // If Whatever is a concept, constrained and deduced.

Note that f4, f5 and f6 are not templates (whereas the previous f1, f2 and f3 are templates). Here, there is no mention of auto in the parameter list. Users have the choice of adopting a style where it is explicit as to whether the return type is deduced.

Constrained types for variables work the same way:

auto x1 = f1();             // unconstrained, deduced.

Constraint auto x2 = f2();  // constrained, deduced.

Whatever x3 = f3();         // See part 2. If Whatever is a type, not deduced.
                            // If Whatever is a concept, constrained and deduced.

Again, users can make it so that it is easy to see when deduction occurs.

Since non-type template parameters can be deduced via auto (as in template <auto N> void f();), we also allow a constraint there:

template <Constraint auto N> void f7();

Note, however, that this can only be a type constraint; non-type concepts (including auto concepts) are not allowed in this form.

Other uses of auto

In concert with the general approach that “Constraint auto goes wherever auto goes”, new-expressions and conversion operators work:

auto alloc_next() { return new Sortable auto(this->next_val()); }

operator Sortable auto() { }

A “Constraint auto” cannot be used to indicate that a function declarator has a trailing return type:

Constraint auto f() -> auto; // ill-formed; shall be the single type-specifier auto

decltype(auto) can also be constrained:

auto f() -> Constraint decltype(auto);
Constraint decltype(auto) x = f();

Structured bindings do deduce auto in some cases; however, the auto is deduced from the whole (and not from the individual components). It is somewhat doubtful that applying the constraint to the whole, as opposed to (for example) applying separately to each component, is the correct semantic. Therefore, we propose to defer enabling the application of constraints to structured bindings to separate papers.

General rules

The constraint applies directly to the deduced type. It does not apply to the possibly cv-qualified type described by the type specifiers, nor does it apply to the type declared for the variable:

const Assignable<int> auto&& c = *static_cast<int *>(p); // Assignable<int &, int>

Naturally, if the deduced type is cv-qualified (or a reference), the constraint applies to that type.

To keep things simple, an auto (or decltype(auto)) being constrained is always immediately preceded by the constraint. So, cv-qualifiers and concept-identifiers cannot be freely mixed:

const Contraint auto x = foo(); // ok
Constraint const auto x = foo(); // ill-formed
Constraint auto const y = foo(); // ok

We propose only the ability to apply one single constraint for a parameter, return type, or non-type template parameter. Any proposal to consider multiple constraints should happen separately after C++20.

Partial concept identifiers also work. Given a concept template <typename T, typename... Args> concept Constructible = /* ... */;, we can say:

void f(Constructible<int> auto x);   // Constructible<decltype(x), int> is satisfied

Constructible<int> auto f();

Constructible<int> auto x = f();

template <Constructible<int> auto N> void f();

Part 2: Relaxed “constrained auto

In the return type of a function declaration, we can leave out the auto. So, in addition to

Constraint auto f1();

we can write

Constraint f2();

Neither f1 nor f2 are templates. It seems fairly reasonable to allow omitting the auto, but that is intended to be a relaxation of the general rule, not a replacement for Constraint auto.

Trailing return types in function declarations behave the same way:

auto f1() -> Constraint; // okay

Trailing return types are also used for return-type-requirements; however, auto does not (currently) go in such a context, whereas constraints do. A minimal change we can make as part of this proposal is to take auto as being implicitly modified by a least constrained concept where a constraint is currently allowed in a return-type-requirement. At the same time, Constraint auto would also be allowed. Thus, a plain Constraint in such a position can be seen as a case of relaxed “constrained auto”. The interaction with P1084R0 Today’s return-type-requirements Are Insufficient is discussed later in this paper.

For example:

requires(T t) {
  {t} -> const auto*; // t needs to be a pointer to const
}

requires(T t) {
  {t} -> Constraint auto; // t needs to be of a type that satisfies Constraint
}

In variable declarations, omitting the auto also seems reasonable:

Constraint x = f2();

Note, in particular, that we already have a syntax that does (partial) deduction but doesn’t make that explicit in the syntax:

std::tuple x = foo();

The variable case in particular seems reasonable, considering the already existing deduction syntaxes that don’t call attention to deduction. The user always has a choice to use a more explicit syntax. The return type case might well have a weaker rationale for being allowed. It should be noted, though, that this relaxation in general was present in the TS; this paper is merely not proposing it for parameters.

Certain disambiguation details need to be handled:

bool b(Constructible<int> && bar());         // variable definition

void foo() {
    Constructible<int> * f2();               // disambiguation/type-name interpretation rule required
    Constructible<int> * selector = f2();    // (same?) disambiguation/type-name interpretation rule required
}

Note that the variable definition is unambiguously a variable definition. This proposal proposes no function declaration syntax that would clash with it; in particular, Constructible<int> is not considered a type-name that is short for Constructible<int> auto except in limited contexts. It’s merely something to be aware of, and a demonstration of how the presence of auto avoids ambiguity. For each remaining case above, Constructible<int> does appear within such a limited context when parsing the statements as prospective declaration statements; we propose to generalize the rule for disambiguating in favor of declaration statements to cover this case.

Part 3: Meaning of “template <Concept T>

In [temp.param]/10 we have:

A constrained-parameter declares a template parameter whose kind (type, non-type, template) and type match that of the prototype parameter (17.6.8) of the concept designated by the qualified-concept-name in the constrained-parameter. Let X be the prototype parameter of the designated concept. The declared template parameter is determined by the kind of X (type, non-type, template) and the optional ellipsis in the constrained-parameter as follows.

[Example:

template<typename T> concept C1 = true;
template<template<typename> class X> concept C2 = true;
template<int N> concept C3 = true;
template<typename... Ts> concept C4 = true;
template<char... Cs> concept C5 = true;

template<C1 T> void f1();       // OK, T is a type template-parameter
template<C2 X> void f2();       // OK, X is a template with one type-parameter
template<C3 N> void f3();       // OK, N has type int
template<C4... Ts> void f4();   // OK, Ts is a template parameter pack of types
template<C4 T> void f5();       // OK, T is a type template-parameter
template<C5... Cs> void f6();   // OK, Cs is a template parameter pack of chars

end example]

Does that seem like a mouthful?

That’s because it is. In template <Constraint T>, the kind of T depends on the kind of the prototype parameter of Constraint.

We instead propose that, for such a constrained-parameter syntax:

To be clear, we are not proposing that concepts in general should not have non-type or template template parameters. We are merely proposing for it to be the case that the constrained parameter shortcut is not provided for concepts with such prototype parameters; such concepts would need to be used with a requires-clause. The constrained parameter syntax should mean just one thing. Note that the same syntax template <A T> is still a non-type parameter when A is a type name rather than a concept. We are willing to tolerate this small potential for ambiguity.

The rationale for this part is as follows:

  1. It seems desirable to have the constrained template parameter syntax.
  2. It would be nice if that syntax covered the most common case.
  3. It would further be nice if that syntax covered only the most common case.
  4. The other cases are expected to be so rare that there’s no need to provide a shortcut for them, and they are certainly rare enough that they shouldn’t use the same syntax.

So, to clarify:

Other use cases can be done with requires-clauses.

Part 4: Meaning of “template <Concept... T>” and its friends

In [temp.param]/11 we have:

template<C2... T> struct s3; // associates C2<T...>

This seems to be doing an unexpected thing, which is having the constraint apply to more than one type in a pack at a time. We propose that, regardless of whether the prototype parameter of the named concept is a pack:

In other words,

Part 5: Meaning of “-> Concept auto” and its friends

Constraint auto, by virtue of being allowed where auto may appear, may now appear (at least syntactically) in a trailing return type. Following the general rule that we deduce like auto (and additionally check the constraint), a meaning is given for “-> Concept auto” when it appears as the trailing return type in a function definition. The meaning being that we get plain return type deduction (as with auto) plus a check of the constraint.

In Part 2, we note that trailing return types are also used for return-type-requirements and we propose for a meaning to be given for use of plain auto in such a context. Further building along that direction, the form of return-type-requirement involving a concept to be satisfied becomes merely a generalization of trailing-return-types with placeholder types. The deduction as described in Part 2 is based on the status quo of the post-Rapperswil working draft, and it is similar to that of the TS and to the deduction that occurs for return type deduction.

P1084R0 explains that the existing deduction rules associated with return-type-requirements is lacking in terms of expressing certain desired semantics. As a solution, it proposed making -> Constraint more of a special case in the context of a return-type-requirement. Expressed in terms of deduction, the semantics proposed by P1084R0 is more akin to deduction for a theoretical decltype((auto)) than to deduction for auto. Which is to say that the semantics for

-> Constraint

as proposed in P1084R0 is incongruous with how we would describe the semantics of

-> Constraint auto

in the context of this paper.

This paper raises the possibility of expressing the semantics requested in P1084R0 in a different manner. {(E)} -> Constraint decltype(auto); produces the effect that P1084R0 requests for {E} -> Constraint; in terms of applying Constraint with decltype((E)) as the argument to its prototype parameter. There is a further consideration as to whether convertibility, as opposed to type identity, should be enforced; in particular, P1084R0 leaves the possibility that {A()} -> std::Same<A> would be satisfied, but {A()} -> A would not be.

Three different kinds of convertibility requirement have been applied in the various iterations of return-type-requirement thus far. With WD signifying N4762, TS signifying N4674, and return representing return type deduction (whether under the TS or under this proposal); the table below summarizes:

Concrete type Placeholder type
WD TS return P1084 WD TS return P1084
Deduction only X X
Implicit conversion sequence X
Well-formed copy initialization X X X X X

The deduction-only situation with the WD may well be a wording defect, since {0} -> std::Same<int>& would surprisingly not fail. Similarly, for the TS, the difference between requiring an implicit conversion sequence and requiring a well-formed copy initialization is somewhat subtle. Nevertheless, the treatment for concrete types differ from that of types involving placeholders for every iteration of return-type-requirement thus far (given that return type deduction is present for comparison only). This seems undesirable.

The WD situation shows that some sort of requirement would be needed since deduction is not sufficient. A convertibility requirement is an option; however, P1084 does not impose a convertibility requirement for the placeholder case. Under P1084, the placeholder case would trivially pass a value category-aware type identity check. Requiring such a check of decltype((E)) being the same as the type in the return-type-requirement (after substitution of the deduced placeholders, if any) is also an option. P1084 proposes a generalization of a same-type constraint, applying a change only to the placeholder case; this latter option is an application of a generalized same-type constraint on the non-placeholder case (as well as to placeholder cases that P1084 drops).

For a parenthesized lvalue of type int, the following table describes the value category-aware type identity check:

Type deduced/passed to prototype parameter Type to match Type matches
auto int int no
C int int no
C& int int& yes
C&& (forwarding reference) int& int& yes
C decltype(auto) int& int& yes
const C& int const int& no
int N/A int no
int& N/A int& yes
int&& N/A int&& no

In brief, auto means “is a prvalue”, const C& means “is an lvalue of const T requiring C<T>”, int&& means “is an int xvalue”, and C decltype(auto) has the semantics requested in P1084 for all parenthesized expressions.