Remove the requirement for constrained-type-specifiers to be deduced to the same type from the Concepts TS

Introduction

The requirement that each occurrence of a constrained-type-specifier appearing in the type of a variable declaration, return type of a function declaration, type of a parameter in an abbreviated function declaration or generic lambda expression, or the type of a deduction constraint be deduced to the same type has been controversial as evidenced by various discussions within the committee reflectors, positions stated by committee members in advance of the Toronto 2017 meeting, and the P0464R2 [P0464R2] and P0691R0 [P0691R0] paper submissions. This paper proposes removing the same type deduction requirement from the Concepts TS with the goal of increasing consensus on adopting the remaining Concepts TS functionality into the current working paper.

Removal of support for the same type deduction requirement means that the following examples that are currently ill-formed according to the Concepts TS will now be well-formed.

template<typename> concept bool C = true;
template<typename, typename> class ct {};

ct<C,C> v1 = ct<int,char>{};     // Previously ill-formed; each occurence of C is
                                 // not deduced to the same type.  Now ok.

ct<C,C> f1() {
  return ct<int,char>{};         // Previously ill-formed; each occurence of C is
                                 // not deduced to the same type.  Now ok.
}

ct<C,C> af1(C) {
  return ct<int,char>{};
}
auto v2 = af1(0L);               // Previously results in an instantiation error due
                                 // to C being deduced to long from the function
                                 // argument and ct<int,char> not being convertible
                                 // to ct<long,long>.  Now ok; the return type is
                                 // deduced to ct<int,char>.

auto gl1 = [](C,C){};
auto v3 = gl1(0,'c');            // Previously results in an instantiation error due
                                 // to different types being deduced for each occurence
                                 // of C in the generic lambda expression's parameter
                                 // list.  Now ok.

template<typename T> concept bool X =
  requires (T t) {
    { t } -> ct<C,C>;
  };
static_assert(X<ct<int,char>>);  // Previously fails; ct<int,char> doesn't satisfy
                                 // the deduction constraint.  Now ok.

This change would affect, perhaps siliently, the behavior of existing code. Consider the following example:

C af2(int i, C) {             // Previously, the return type was deduced from the function
  return i;                   // argument.  Now, it is deduced from the return statement.
}
auto v3 = af2(0, 0L);         // Previously resulted in v3 having type long.
                              // Now v3 is deduced to have type int.

Additionally, the following code would become ill-formed due to the return type no longer being deducible.

C af3(C) {                    // Previously ok; the return type matches the deduced type
  return {};                  // for the function argument.  Now ill-formed since the
}                             // return type is not deducible from the return statement.

This paper does not propose changing the requirement that the same type be deduced when multiple declarators are present [ Note: gcc 7.1.0 and earlier fail to diagnose differently deduced types for each declarator when the declared type includes constrained-type-specifiers. See gcc bug 81270 [GccBug81270]end note ]:

ct<C,C> v4 = ct<int,char>{},  // Previously ill-formed; different types are deduced
        v5 = ct<int,long>{};  // for the types of v4 and v5.  Remains ill-formed.

Terminology

Some of these terms have been adopted from P0694R0 [P0694R0].

Motivation

P0464R2 [P0464R2] provides motivation for removing the same-type deduction requirement (consistent resolution) for abbreviated functions in favor of allowing different-type deduction (independent resolution). That motivation is acknowledged and little further motivation with respect to abbreviated functions is provided here. The following discusses additional motivation relative to the types of variable declarations, return types of non-abbreviated functions, and deduction constraints.

The requirement for types specified by constrained-type-specifiers to be deduced to the same type within a declaration (consistent resolution) was introduced based on an expectation that such resolution would be frequently desired and that non-abbreviated declarations could be used when not desired. While certainly true that consistent resolution is often desired, it is also true that is is often not desired. Use of non-abbreviated declarations does allow precise specification, but only for function template declarations; there is currently only one syntax available for variable declarations, non-template function declarations, and deduction constraints.

Consider the following variable declaration copied from the introduction. The declaration is currently ill-formed because the deduced types for both occurrences of C do not match.

ct<C,C> v1 = ct<int,char>{}; // Currently ill-formed.
There are several solutions available for making this well-formed such that each of the arguments to ct are individually deduced, but still required to satisfy C. For example:
template<C T1, C T2> void ctCC_with_independent_resolution(ct<T1,T2>);
template<typename T> concept bool ctCC =
  requires (T t) {
    ctCC_with_independent_resolution(t);
  };

ctCC v1 = ct<int,char>{}; // Ok; ctCC allows independent resolution.
P0694R0 [P0694R0] suggests introducing additional names for the same concept to enable independent resolution. For example:
template<typename T> concept bool C2 = C<T>; // A duplicated concept definition.

ct<C,C2> v1 = ct<int,char>{}; // Ok; independent resolution
Neither of these solutions is satisfactory. Of course, should the behavior be changed to independent resolution as proposed in this paper, then enabling consistent resolution to preserve the current behavior is equally unsatisfactory (though necessary only if it is desired to explicitly reject deduction of different types):
template<C T> void ctCC_with_consistent_resolution(ct<T,T>);
template<typename T> concept bool ctCC =
  requires (T t) {
    ctCC_with_consistent_resolution(t);
  };

ctCC v1 = ct<int,char>{}; // Ill-formed (as intended); ctCC requires consistent resolution.

The extensions proposed in N3878 [N3878] provide additional flexibility, but do not provide the full range of expression needed. For example:

ct<C{T},T> v1 = ct<int,char>{};   // Not proposed; syntax from N3878.
                                  // Ill-formed (as intended); C is deduced to be int for the
                                  // first template argument, the second argument is declared
                                  // to match the deduced type for the first argument,
                                  // and ct<int,char> is not convertible to ct<int,int>.
For full genericity, it appears extensions like those proposed in both N3878 [N3878] and P0694R0 [P0694R0] would be needed to enable expressing:

The above examples are illustrated with variable declarations, but the same concerns apply to constrained-type-specifiers in function return types and deduction constraints as well. Consider the following example constraint written to require an expression to have a type requiring std::tuple with a specific arity, but constrained element types. With the changes proposed in this paper, independent resolution will allow the element types to be deduced to the same or different types. If consistent resolution is is desired, then that additional requirement can be added with the (admittedly unsatisfactory) techniques discussed earlier.

template<typename T> concept bool V =
  requires (T t) {
    { t } -> std::tuple<C,C,C>;
  };

Abbreviated functions have the behavior that constrained-type-specifiers specified within the return type of the function that match a constrained-type-specifier used in a parameter type are not deduced from non-discarded return statements from the body of the function (as is the case when the return type is auto or decltype(auto)), but are rather replaced by the deduced type for the matching parameter(s). This behavior introduces the possibility that a change to the function's parameter list may alter the meaning of the return type. For example, if this function:

C af4(auto p1);
is modified to add a parameter p2 with the constrained-type-specifier C:
C af4(auto p1, C p2);
then the change to the parameter list will result in the return type matching the deduced type for C from the parameter list rather than being deduced from return statements in the function definition. Such an effect is subtle and may be made unintentionally. Extensions like those discussed above would enable making these relationships explicit when desired:
auto af4(auto p1, C{T} p2) -> T; // The return type matches the type deduced
                                 // for parameter p2.

P0694R0 [P0694R0] argues that consistent resolution should be the default for variables declared with equivalent constrained-type-specifiers within the same scope and provides the following example (where Number names a concept):

Number operator+(Number,Number);

template<typename T> concept bool Number2 = requires Number<T>;

void f()
{
  Number   x = 1+2;     // OK
  Number   y = 1.0+2.0; // error: different type bound to same concept in a single scope
  Number2  y = 1.0+2.0; // OK
}
However, consistent reslution as portrayed in this example seems unworkable as a default for variable declarations at namespace scope since the first variable declaration would lock in the deduced type associated with the concept-type-specifier for the remainder of the translation unit. An exception could, of course, be made for variables declared at namespace scope, but such a variance in the rules would be surprising and potentially problematic for code refactoring. Consider replacing several namespace scope variables with static local variables.

The examples above illustrate cases where the author feels consistent resolution results in rejection of code that should be well-formed. Further, when the desire for consistent resolution is present, the author feels such constraints should be explicitly specified separately from general satisfaction of a concept. Finally, since the need to specify all of consistent resolution, independent resolution, and matched resolution requirements exists, the author feels more work is needed. In the mean-time, relaxing the requirement for consistent resolution will enable more code to be well-formed without having to resort to the current unstatisfactory workarounds.

Wording

Hide deleted text

These changes are relative to N4674 [N4674]. If P0696R0 [P0696R0] is adopted, then additional changes relative to the wording adoped in that paper will be required.

Remove the added paragraph 4 inserted in 10.1.7.4 [dcl.spec.auto]:

Add the following after paragraph 3 to describe when constrained-type-specifiers in the return type refer to template parameters.
A constrained-type-specifier C1 within the declared return type of an abbreviated function template declaration does not designate a placeholder if its introduced constraint-expression (10.1.7.4.2) is determined to be equivalent, using the rules in 17.6.5.1 for comparing expressions, to the introduced constraint-expression for a constrained-type-specifier C2 in the parameter-declaration-clause of that function declaration. Instead, C1 is replaced by the template parameter invented for C2 (11.3.5). [ Example:

template<typename T> concept bool C = true;

template<typename... T> struct Tuple;

C const& f1(C); // has one template parameter and no deduced return type
Tuple<C...> f2(C); // has one template parameter and a deduced return type
In the declaration f1, the constraint-expression introduced by the constrained-type-specifiers in the parameter-declaration-clause and return type are equivalent ; they would both introduce the expression C<T>, for some invented template parameter T. In f2, the use of C in the return type would introduce the constraint-expression (C<T> &l&l ...), which is distinct from the constraint-expression C<T> introduced by the invented constrained-parameter (17.1) for the constrained-type-specifier in the parameter-declaration-clause according to the rules in 11.3.5. — end example

Renumber the subsequent paragraphs 5-16 in 10.1.7.4 [dcl.spec.auto] to reflect the removal of paragraph 4 and to align the paragraph numbers with the C++ standard.

In 10.1.7.4.1 [dcl.spec.auto.deduct] paragraph 3 (corresponding to paragraph 4 in the C++ standard), modify the last example to indicate that each constrained-type-specifier corresponds to a different invented template parameter:

[…]
[ Example:
auto cf(int) -> Pair<C, C> { return expr; }
The return type of cf is deduced from the parameter p2 in the call f2(expr) of the following invented function:
template<C T> void f2(Pair<T, T>);
template<C T1, C T2> void f2(Pair<T1, T2>);
Both constrained-type-specifiers in the return type of cf correspond to the samedifferent invented template parameters.

Change the text added as 11.3.5 [dcl.fct] paragraph 19:

Each template parameter is invented as follows. All placeholders designated by constrained-type-specifiers whose corresponding constrained-parameters would introduce equivalent constraint-expressions (17.1), using the rules for comparing expressions in 17.6.5.1, have the sameunique invented template parameters
[ Example:
template<typename T> concept bool C = true;
template<typename, typename> class ct {};

void f1(C a, C b);
void f2(ct<C,C>);
C f3(C c);
The types of a, b, c, and the template arguments to ct each correspond to unique invented template type parameters. The return type of f3 is deduced from non-discarded return statements, if any, in the body of the function (9.4.1). — end example ]
[ Example:
namespace N {
  template<typename T> concept bool C = true;
}
template<typename T> concept bool C = true;
template<typename T, int> concept bool D = true;
template<typename, int = 0> concept bool E = true;
void abbr(C, D<0>);
The constrained-type-specifiers C and D<0> correspond to distinct invented template parameters in the declaration of abbr.
void f0(C a, C b);
The types of a and b are the same invented template type parameter.
void f1(C& a, C* b);
The type of a is a reference to an invented template type parameter T, and the type of b is a pointer to T.
void f2(N::C a, C b);
void f3(D<0> a, D<1> b);
In both functions, the parameters a and b have different invented template type parameters.
void f4(E a, E<> b, E<0> c);
The types of a, b, and c are the same because the constrained-type-specifiers E, E<>, and E<0> all associate the constraint-expression E<T, 0>, where T is an invented template type parameter.
void f5(C head, C... tail);
The types of head and tail are different. Their respective introduced constraint-expressions are C<T> and (C<U> && ...), where T is the template parameter invented for head and U is the template parameter invented for tail (17.1).

Acknowledgements

The author would like to thank Andrew Sutton for his continued dedication to progressing the adoption of Concepts into the C++ working paper.

Thanks to Botond Ballo for initial draft review and feedback.

References

[N3878] "Extensions to the Concept Introduction Syntax in Concepts Lite"
http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2014/n3878.pdf
[N4674] "Working Draft, C++ Extensions for concepts"
http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2017/n4674.pdf
[P0464R2] "Revisiting the meaning of foo(ConceptName,ConceptName)"
http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2017/p0464r2.html
[P0691R0] "Integrating Concepts: “Open” items for consideration"
http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2017/p0691r0.pdf
[P0694R0] "Function declarations using concepts"
http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2017/p0694r0.pdf
[P0696R0] "Remove abbreviated functions and template-introduction syntax from the Concepts TS"
http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2017/p0696r0.html
[GccBug81270] "[concepts] ill-formed code with a constrained variable declaration with multiple declarators with different deduced types not rejected"
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81270