Document Number: | P3311R0 |
---|---|
Date: | 2024-05-22 |
Audience: | SG21 |
Reply-to: | Tom Honermann <tom@honermann.net> |
SG21 attained strong concensus to forward P2900R6 (Contracts for C++) [P2900R6] to EWG during the 2024-02-29 SG21 telecon as its design for a minimally viable solution for contracts in C++. This design does not integrate the traditional C assert facility, nor does it expose interfaces to enable other traditional or ad hoc assert facilities to integrate with the new contracts features.
The P2900R6 design includes a new contract_assert statement that avoids many of the well-known pitfalls inherent in macro based traditional assert facilities. Despite considerable desire and effort to enable use of the assert identifier in the contracts design, no method to do so was identified that both avoided backward compatibility problems and used a syntax that gained consensus. P2884R0 (assert Should Be A Keyword In C++26) [P2884R0] attempted to promote assert to a keyword, but failed to get consensus. P2961R2 (A natural syntax for Contracts) [P2961R2], was approved by SG21 during the Kona, 2023 meeting. Section 5.2, "The assert name clash", details the challenges faced with trying to repurpose that identifier.
As a core language feature, contract_assert does not provoke ODR violations due to NDEBUG being inconsistently defined across translation units, is not subject to being redefined during translation as is permissible for macros (note that contract_assert is proposed as a keyword and is therefore not available for use as an identifier), and is not confused by the presence of a comma that is not contained within a pair of balanced parenthesis (an issue addressed for C assert by P2264R7 (Make assert() macro user friendly for C and C++) [P2264R7] for C++26). contract_assert also guards against some cases of possibly unintended behavior that may otherwise occur when its associated conditional-expression is evaluated. It guards against possibly unintended mutation of some objects by implicitly applying a const qualifier to the result of id-expressions that name a parameter or variable with automatic storage duration and to the pointee type of this expressions. It guards against unintended dependence on side effects by allowing an implementation to elide evaluation when the result of the evaluation can be predetermined or to perform the evaluation multiple times (with an expectation that multiple evaluations will make such side effects more noticeable).The differences between contract_assert and traditional assert facilities make it impossible to directly reimplement traditional assert facilities using contract_assert as currently proposed without imposing potentially extensive changes to code that uses those facilities. Further, the latter two behaviors (implicit application of a const qualifier and evaluation elision or duplication) have become contentious due to a lack of confidence that contract_assert, as currently proposed, suffices as a modern alternative to traditional assert facilities like C assert.
Proposed are extensions to the P2900R6 design that are intended to enable existing assert facilities, including C assert, to be optionally integrated with C++ contracts, with minimal changes, such that assertion failures that occur through use of those facilities are also reported through the C++ contract violation handler. In doing so, the author hopes to reduce the current contention over the design of contract_assert by allowing those that prefer the traditional evaluation and side effect guarantees to use interfaces that behave as they desire without compromising on the protections that contract_assert, as currently proposed, provides.
Vast amounts of code have been written over the lifetime of the C and C++ languages that utilize C assert or other traditional or ad hoc assertion facilities. It is not realistic to expect or require existing uses of traditional assert facilities to be rewritten to use C++ contracts when such code is migrated to C++26. The limitations and pitfalls of traditional assert facilities provides good motivation for a new modern facility. However, greater motivation for use of a new facility is produced if it enables integration with existing facilities (such integration also helps to reduce the negative consequences of the infamous "there are N+1 competing standards" situation). Specifically, the consequences of use of multiple assertion facilities, a persistent reality for many programmers, is reduced if assertion failures can be handled in a uniform way.
While it is not feasible to unconditionally claim the assert identifier for use with C++ contracts without unacceptable backward compatibility impact, it is feasible to offer the ability for projects to opt-in to a mode that handles assertion failures in traditional assert() expressions by invoking the C++ contract violation handler. In the long term, perhaps with assistance from code modernization tools, such integration could become standard practice and eventually allow assert to be more fully integrated into C++ contracts.
There are two basic approaches to enable a traditional assert facility to invoke the contract violation handler.
Enabling user construction of std::contracts::contract_violation objects would require exposing a suitable constructor or factory function to specify property values to be returned by the various property access member functions. std::contracts::contract_violation is designed to allow these property values to be accessed from statically allocated memory associated with each contract assertion present in the program that has a potentially checked semantic. Support for user construction would require specifying an ownership model for user provided values. Further, it is currently implementation-defined whether the class is polymorphic; it could be an abstract class, but the property access member functions are non-virtual. Enabling user construction could provide flexibility for integration of traditional assert facilities with C++ contracts, but would also impose responsibility for appropriate selection of the contract_kind, contract_semantic, and detection_mode property values.
contract_assert is a core language feature that is already capable of evaluating an expression and invoking the contract violation handler. The inability to use it to directly implement a traditional assert interface is superimposed by the express desire to avoid well known problems with those traditional interfaces. However, it is possible to relax those imposed restrictions.
The grammar for the contract_assert statement was designed with future extension in mind.
assertion-statement :contract_assert attribute-specifier-seqopt ( conditional-expression ) ;
While use of an attribute to opt-in to alternative semantics for the contract_assert statement would not require syntactic extension, such use would violate the WG21 ignorability policy for standard attributes. The ignorability policy prohibits a standard attribute from selecting a behavior that is not one of the permissible behaviors an implementation would be allowed to select in the absence of the attribute; an attribute cannot be used to relax constraints.
Fortunately, there is room for context sensitive keywords to be used in between the contract_assert keyword and the opening parenthesis (either before or after the optional attribute-specifier-seq) or between the closing parenthesis and the semicolon.
C assert supports two translation modes. When the NDEBUG macro is defined, arguments to assert expressions are not evaluated; this is the traditional form of the ignore semantic (and a form that leads to ODR violations). When NDEBUG is not defined, assert exhibits behavior like that of the enforce semantic, but with more stringent requirements on how termination is performed; abort() is called and any further processing requires the installation of a signal handler for the SIGABRT signal.
Since NDEBUG may be used to conditionally remove code that an assert expression depends on, it isn't possible, in general, to parse arguments passed to assert when NDEBUG is defined. It is therefore impossible to use NDEBUG as a proxy for a contract semantic; when NDEBUG is defined, no contract specification is present. However, when NDEBUG is not defined, then all contract semantics are applicable, though maintaining conformance with C standard requirements for the behavior of assert requires selecting the enforce semantic, writing the required information to the standard error stream, and calling abort().
Add a new traditional_assert enumerator to std::contracts::contract_kind.
Extend the contract_assert syntax to allow an optional context sensitive traditional keyword after contract_assert.
assertion-statement :contract_assert traditionalopt attribute-specifier-seqopt ( conditional-expression ) ;
Modify the contract_assert behavior as follows when traditional is specified.
Extend the recommended practice for the default contract violation handler to implement the behavior required by the C standard for a failed assertion when invoked with a contract_violation object whose kind() member returns traditional_assert. This includes writing select information to the standard error stream and then calling abort().
Note that the above calls for the violation handler to be invoked with the enforce contract semantic in effect. This is required because there is no mechanism for the contract_assert statement to indicate whether the conditional-expression was evaluated, what its result was, or whether the violation handler was invoked and returned. Code following the contract_assert statement therefore has no information available to determine what action should be taken next. This limitation can be worked around by evaluating the assertion predicate separately and using the result value for the conditional-expression of the contract_assert statement, but this effectively hides the expression from the contracts implementation and prevents exposing it via the comment() member of std::contracts::contract_violation. Should motivation arise, additional extensions to enable use of an alternative contract semantic can be added at a later time. Selection of an alternate semantic could be enabled by parameterizing traditional with an argument enclosed in angle brackets (e.g., traditional<observe>; use of parenthesis would be ambiguous due to the parenthesis that surround the conditional-expression).
Upon inclusion of the assert.h or cassert header, if the NDEBUG macro is not defined and the ASSERT_USES_CONTRACTS macro is defined, define the assert macro as follows.
#define assert(...) \ [&] { contract_assert traditional (__VA_ARGS__); }()
Note that it is not an error to define both NDEBUG and ASSERT_USES_CONTRACTS; when NDEBUG is defined, ASSERT_USES_CONTRACTS has no effect. The effects of the NDEBUG macro must take precedence to preserve backward compatibility with existing code.
None.
Thanks to all SG21 participants for the tremendous amount of time, energy, and expertise that has been contributed towards advancing contracts for C++.
[P2264R7] |
"Make assert() macro user friendly for C and C++", P2264R7, 2023. https://wg21.link/p2264r7 |
[P2884R0] |
"assert Should Be A Keyword In C++26", P2884R0, 2023. https://wg21.link/p2884r0 |
[P2900R6] |
"Contracts for C++", P2900R6, 2024. https://wg21.link/p2900r6 |
[P2961R2] |
"A natural syntax for Contracts", P2961R2, 2023. https://wg21.link/p2961r2 |
These changes are relative to P2900R6 [P2900R6].
Hide inserted textWording will be provided in a future revision contingent on SG21 encouraging further work for the proposed direction.