Using [[nodiscard]] should be Recommended Practice

• • Revision History
  • Introduction
  • Rationale
  • Proposed LEWG Policy
  • Proposed Standard Wording
  • Acknowledgments
  • References

Revision History

• R1 adds more rationale, proposed LEWG policy wording, and an additional bullet in the proposed standard wording to cover cases like std::remove_if.

Introduction

Instead of sprinkling [[nodiscard]] in hundreds of places in the library clauses, there should be a single normative recommendation to use it wherever it's appropriate. LEWG policy should be to avoid spending time discussing it.

This paper proposes a policy for LEWG and new wording for the standard.

Rationale

It's not a good use of time to discuss adding [[nodiscard]] to individual function declarations. That requires LEWG time, LWG time, WG21 votes, editorial changes, and reviews to ensure those changes were correctly applied. For example, P2377R0 [[nodiscard]] in the Standard Library: Clause 23 Iterators Library only covers a single clause and is still nearly ten pages of wording. And it has a mistake in the very first change it proposes to a declaration, which would need to be reviewed and corrected.

What we want is for real implementations to give real warnings to users writing real code. Whether the attribute is literally present on individual declarations in the PDF standard is not important.

P2377R0 argued that it's a good use of committee time to add it to the library because it reminds implementors to use the attribute and improve the quality of their implementations. I feel that a single normative recommendation to use it is a far better use of time than micro-managing how implementations declare individual functions.

Implementers are best placed to decide on which functions to add the attribute for their specific implementation. If a compiler has built-in warnings for side-effect-free expressions, e.g. it can determine that it == end; doesn't modify its arguments and so issues a warning, then there is no need for that implementation to explicitly add the [[nodiscard]] attribute to every inline operator== definition. We should aim for good quality diagnostics, not worry about how that is achieved.

Implementors are also more able to respond promptly to changing needs or to user feedback regarding nodiscard diagnostics. New compiler releases happen more frequently than revisions of the C++ standard, and compiler vendors can trial new nodiscard warnings under an optional setting to gather usage experience. This is more flexible than spending WG21 time to discuss an addition that is then enshrined in the standard for at least three years until the next revision (or until more committee time is spent to revise it via a defect report).

The wording below only talks about functions. The valid uses for nodiscard on types are less common, and it might be worth discussing individual cases, and adding the attribute to declarations in the standard if implementors need extra guidance there. Alternatively, maybe blanket wording can be produced that accurately describes the cases that should produce warnings.

The wording does not cover functions which return a pointer to allocated storage, such as operator new and std::allocator<T>::allocate. All such functions are marked nodiscard already. If the group prefers it, we could remove those existing nodiscard attributes and add a new bullet like "Functions that return a pointer to allocated storage. [Example: operator new returns memory which could not be deallocated if discarded. — end example]"

The wording does not cover functions that are considered likely to be confused with another function, due to ambiguous meaning of English nouns vs verbs, such as std::vector<T, A>::empty() which could be misunderstood to empty the container (which is what the clear function does). It is not easy to specify generic wording that clearly defines what might or might not be confusing, and the particular example of vector::clear is already covered because it's a pure function that is only called for its return value. The more general policy already solves the problem, without a special case based on potential confusion around the name.

Proposed LEWG Policy

Policy: Implementors are encourage to diagnose discarded-value expressions via blanket wording in the library introduction clause, so the explicit use of nodiscard in library wording should be avoided. Subclause [lib.nodiscard] covers the cases that implementors are encouraged to diagnose. Anything covered there does not need to be marked nodiscard in library wording. New proposals should not use nodiscard in proposed wording unless it's a case not covered by [lib.nodiscard], and then the paper should provide rationale for not following the policy.
Rational: Implementers are best placed to decide on which functions to add the attribute for their specific implementation. They do not need the standard to specify where it must be used.
Policy paper: P3122

Proposed Standard Wording

This wording is relative to N4971.

Add a new subclause after 16.4.6.15 [lib.types.movedfrom]:

16.4.6.?? Discarded calls [lib.nodiscard]

Except where shown otherwise, it is unspecified which functions defined in the C++ standard library are marked with a nodiscard attribute.

Recommended practice: Implementations should issue a warning for potentially-evaluated discarded-value expressions ([expr.context]) where the expression is a call to any of the following:

• operator== and operator<=> functions.
• const member functions and explicit object member functions that do not directly or indirectly modify objects accessible via their arguments.
  [Example: vector::empty() const and unordered_map::bucket_size(size_type) const do not modify any objects. — end example]
• Non-const member functions that return pointers, references, or iterators, but do not modify the object parameter or the object that the return value refers to.
  [Example: vector::begin() and set::find(const key_type&) are not const so that they can return a mutable iterator, but they do not modify the container or its elements. basic_string::operator[](size_type) does not modify the string or its contents. — end example]
  
• Non-member functions that do not directly or indirectly modify objects accessible via their arguments, other than potentially modifying their non-reference arguments.
  [Example: as_const(T&) behaves like a cast, it has no side effects. back_inserter(Container&) takes a reference to a non-const object but does not modify it. ranges::next(I) increments its by-value parameter, which does not modify any other objects. — end example]
• The function call operators of customization point objects ([customization.point.object]), unless the return type is void.
  
• Mutating sequence operations ([alg.copy]) which return an iterator such that the position the iterator refers to depends on a predicate or the content of the input range, not just on the size of the input range. [Example: copy_if returns an iterator pointing to the end of the result range, which depends on how many elements of the input range match the predicate. remove and unique return an iterator pointing to the end of the result range, which depends on how many elements from the input range were not included in the output range. — end example]

Acknowledgments

Thanks to Nicolai Josuttis, Hana Dusíková, and Christopher Di Bella for the past work identifying where nodiscard is useful.

References

P0600R1, [[nodiscard]] in the Library, Nicolai Josuttis, 2017.
P2351R0, Mark all library static cast wrappers as [[nodiscard]], Hana Dusíková, 2020.
P2377R0, [[nodiscard]] in the Standard Library: Clause 23 Iterators Library, Christopher Di Bella, 2021.
N4971, Working Draft - Programming Languages -- C++, Thomas Köppe, 2023.