ISO/IEC JTC 1/SC 22/WG 21 P1013R1
Date: 2018-06-10
Audience: none (for historical record)
Thomas Köppe <tkoeppe@google.com>We propose to remove from the working paper the fact that mere application of arguments
to concepts cause evaluation “everywhere” as boolean prvalue expressions and
instead introduce a mildly more verbose “simple-requires-expression”.
Instead of bool b = Sortable<MyType>; we propose
bool b = requires Sortable<MyType>;
This change removes unfortunate corner cases and paves the way for future extensions.
| Before the proposal | With the proposal |
|---|---|
template <typename T> concept Foo = /* ... */;
template <typename T> void f(T) {
static_assert(Foo<T>);
if constexpr (Foo<T>) { /* ... */ }
} |
template <typename T> concept Foo = /* ... */;
template <typename T> void f(T) {
static_assert(requires Foo<T>);
if constexpr (requires Foo<T>) { /* ... */ }
} |
Consider a simple type concept:
For a given type, say int, the current working paper makes Foo<int>
a boolean prvalue. This creates a curious corner case if we consider future syntax extensions for
abbreviated function templates. Consider the following code:
There are two plausible meanings this code could have (assuming further work in the direction of P0745R0):
This is kind of an inverse of the “most vexing parse”. In the status quo of
the working paper, the meaning is that of case #1 (a variable declaration); whereas at
least some have expressed that the preferable meaning would be that of case #2 (a function
declaration). This latter case is the one that is consistent with dropping constraints:
template <typename T> bool x(T) is already a function declaration in
the status quo.
We can avoid this problem if we make it so that applying template argument lists to plain concept names do not form prvalue expressions, thus spuriously occupying a “privileged” syntax. If we retain the status quo and later discover that this problem is serious, it will be a breaking change to remove or change the behaviour. On the other hand, if we remove the behaviour now and later discover that we actually do need it, we can easily add it back in without breaking code written to a level of C++ specified by an International Standard.
We propose that an id-expression ([expr.prim.id]) that denotes the specialization of a concept no longer results in a prvalue except in specific contexts where we can expect normalization to apply the specific rule for normalization of an id-expression that names a specialization of a concept, i.e., it is not a subexpression of an atomic constraint (and in, for example, a constraint-logical-or-expression). To make up for the lost functionality, we introduce a new kind of expression: a “simple-requires-expression”, which turns a specialization of a concept into the boolean value indicating its constraint satisfaction.
Instead of making concepts not be expressions, we could make them expressions precisely when they are “named fully” or “fully specialized”, but treat them as type placeholders when they are only “partially bound”. A hypothetical short function template syntax might then fit in like this:
In this approach, it is necessary to know the declaration of Foo in order to
know whether bool y(Foo<char>) declares a variable or a
function, with the usual concerns about maintenance, default arguments, etc.
This approach presupposes a detail of a future, not-yet-made design. It is a plausible approach, but at the same time this direction is compatible with our main proposal: we can first remove the evaluation-for-satisfaction interpretation of supplying arguments to concept names, and later bring it back with the semantics described here.
The proposed new simple-requires-expression is not strictly required,
since a boolean value can also be obtained from an (ordinary) requires-expression
such as “requires { requires Foo<T>; }”.
We feel that such an expression would be somewhat unwieldy in contexts such as those of
constexpr if and static assertion, and that the new simple-requires-expression
will fit in more naturally.
The issue mentioned by this paper appears in the context of the Concepts TS as an ambiguity between constrained-type-specifiers whose constrained-type-name is a partial-concept-id and simple-template-ids whose template-name names a concept; the Concepts TS does not contain wording to add this ambiguity in either [stmt.ambig] or [dcl.ambig.res].
The Concepts TS has a process for concept resolution (N4674 [temp.constr.resolve]), the primary purpose
of which is to handle overloading of function concepts, that could come into play as part of ambiguity
resolution through determining whether the potential partial-concept-id is or is not a type-name.
As an example, concept resolution will fail when considering C<int> (below) as a placeholder type,
because there is a mismatch between the int type argument and the non-type template parameter. The
alternate interpretation of C<int> as an expression is considered viable.
It is observed that semantic aspects in the details of the definition and of the use weigh into disambiguation with such a strategy.
We move beyond the traditional type-versus-expression ambiguity if we further pursue the direction, indicated to be intended by the design of the TS on the core reflector [core-28404], of allowing concept placeholders with other-than-type prototype parameters. On the reflector, it was observed that the TS admits constrained-type-specifiers (N4674 [dcl.type.simple]) to be placeholders for non-type and template template arguments, but does not extend the grammar to allow constrained-type-specifiers in the corresponding contexts; the latter situation appears to be considered a wording problem.
On that aspect of the design space, in addition to being short on wording to be found already in the Concepts TS, the GCC implementation behaves oddly when presented with such constrained-type-specifiers in contexts where they may have been parsed as a type [GccBug85846][GccBug85991].
Additional design guidance would also be required on the value/value ambiguity between evaluation for substitution versus being a placeholder, and the interpretation as to whether an argument list forms part of a partial-concept-id or not when used with concepts having template prototype parameters.
For example, consider the following:
We may find that these aspects hint towards needing extra syntax to engage the placeholder interpretation as opposed to requiring extra syntax for the evaluation-for-satisfaction; however, the future is hard to predict.
There is no impact on the Standard, since the proposal modifies a feature that has not yet been standardized.
Removing concept id-expressions now means that we do not standardize a feature that we may later regret and cannot change without a break. On the other hand, adding the removed feature back in later, if and when we do need it, is straightforward.
[to be fleshed out]
requires requirement-parameter-listopt requirement-bodyrequires nested-name-specifieropt simple-template-id
An id-expression thatsimple-requires-expression
whose simple-template-id denotes the specialization of a concept
[temp.concept] results in a prvalue of type bool. The expression
is true if the concept’s normalized constraint-expression
[temp.constr.decl] is satisfied [temp.constr.constr] by the specified template arguments
and false otherwise.
[Note: A concept’s constraints are also considered when using a template name [temp.names] and during overload resolution [over], and they are compared during the the partial ordering of constraints [temp.constr.order]. – end note]
Many thanks to Andrew Sutton for valuable discussion, Botond Ballo for necessary poking, and Richard Smith for his pre-Rapperswil review and feedback.
This proposal was presented to EWG in Rapperswil 2018 but did not achieve consensus (SF: 4 | F: 15 | N: 11 | A: 6 | SA: 8). The main points of opposition were: