Document number: D0824R1
Date: 2018-02-05
Project: ISO JTC1/SC22/WG21, Programming Language C++
Audience: Library Evolution Working Group
Reply to: Arthur O'Dwyer <arthur.j.odwyer@gmail.com>, Charley Bay <charleyb123@gmail.com>, Odin Holmes <holmes@auto-intern.de>, Michael Wong <fraggamuffin@gmail.com>, Niall Douglas <s_sourceforge@nedprod.com>

Summary of SG14 discussion on <system_error>

1. Introduction
2. Description of C++11's <system_error> facilities
3. Proposed best practices for using C++11's <system_error> facilities
    3.1. Proposed by Arthur O'Dwyer
    3.2. Proposed by Charley Bay
4. Issues with the <system_error> facilities
    4.1. Use of std::string
    4.2. Proliferation of "two-API" libraries
    4.3. No wording sets aside the 0 enumerator
    4.4. Reliance on singletons
    4.5. No error_category subclass can be a literal type
    4.6. No guidance on attaching extra information to error_code
    4.7. Reliance on a surprising overload of operator==
    4.8. error_category should properly be named error_domain
    4.9. Standard error_code-yielding functions can throw exceptions anyway
    4.10. Underspecified error_code comparison semantics

1. Introduction

This paper summarizes a discussion that took place during the SG14 telecon of 2017-10-11. The discussion concerned the facilities of <system_error> (introduced in C++11), which include errc, error_code, error_condition, error_category, and system_error.

The discussion naturally also concerns the current idioms for exceptionless "disappointment handling" (P157R0, Lawrence Crowl), as seen in <filesystem> (C++17) and std::from_chars/to_chars (also C++17); and it concerns future idioms for exceptionless disappointment handling, as proposed in status_value<E,T> (P0262R0, Lawrence Crowl), in expected<T,E> (P0323R2, Vicente J. Botet Escribá), and in result<T,EC> (Outcome, Niall Douglas).

In all of those future idioms involving expected<T,E>-style result types, SG14 expects that the E type parameter will default to std::error_code — which is right and good. This means that the coming years will see much increased use of std::error_code in both new and existing codebases. This means that SG14 is very interested in correcting deficiencies in std::error_code sooner, rather than later.

On SG14's 2017-10-11 teleconference call, various people contributed to a miscellaneous "laundry list" of perceived deficiencies with <system_error>. Several "best practices" were also proposed for how we expect <system_error> facilities to be used in the best codebases. (Unfortunately, the Standard Library does not follow anyone's proposed "best practice"!) This paper is a summary of that discussion.

2. Description of C++11's <system_error> facilities

<system_error> provides the following patterns for dealing with error codes:

A std::error_code object wraps an integer "error enumerator" and a pointer to an "error category" (that is, a handle to the specific error domain associated with this integer enumerator). The addition of the error domain handle is what allows us to distinguish, say, error=5 "no leader for Kafka partition" (in the rdkafka domain) from error=5 "I/O error" (in the POSIX domain). Two error_code instances compare equal if and only if they represent the same error enumerator in the same domain.

Notice that std::error_condition is not the same type as std::error_code!

    namespace std {
    class error_code {
        int value_;
        std::error_category *cat_;
    public:
        error_code() noexcept : error_code(0, std::system_category()) {}
        error_code(int e, const std::error_category& c) noexcept : value_(e), cat_(&c) {}
        template<class E> error_code(E e) noexcept requires(std::is_error_code_enum_v<E>) { *this = make_error_code(e); }  // intentional ADL
        template<class E> error_code& operator=(E e) noexcept requires(std::is_error_code_enum_v<E>) { *this = make_error_code(e); }  // intentional ADL
        void assign(int e, const std::error_category& c) noexcept { value_ = e; cat_ = &c; }
        void clear() noexcept { *this = std::error_code(); }

        explicit operator bool() const noexcept { return value_ != 0; }
        int value() const noexcept { return value_; }
        const std::error_category& category() const noexcept { return cat_; }

        std::string message() const { return cat_->message(value_); }
        std::error_condition default_error_condition() const noexcept { return cat_->default_error_condition(value_); }
    };
    bool operator==(const std::error_code& a, const std::error_code& b) noexcept { return a.value() == b.value() && &a.category() == &b.category(); }
    } // namespace std

Each "error domain" is represented in code by an object of type std::error_category. The standard effectively implies that these must be singletons, because error_code equality-comparison is implemented in terms of pointer comparison (see above). So two error_category objects located at different memory addresses represent different error domains, by definition, as far as the currently standardized scheme is concerned.

    namespace std {
    class error_category {
    public:
        constexpr error_category() noexcept = default;
        error_category(const error_category&) = delete;
        virtual ~error_category() {}

        virtual const char *name() const noexcept = 0;

        virtual std::error_condition default_error_condition(int e) const noexcept { return std::error_condition(e, *this); }
        virtual std::string message(int e) const = 0;

        virtual bool equivalent(int e, const std::error_condition& condition) const noexcept { return this->default_error_condition(e) == condition; }
        virtual bool equivalent(const std::error_code& code, int e) const noexcept { return code == std::error_code(e, *this); }
    };
    bool operator==(const std::error_category& a, const std::error_category& b) noexcept { return &a == &b; }
    } // namespace std

It is intended that the programmer should inherit publicly from std::error_category and override its pure virtual methods (and optionally override its non-pure virtual methods) to create new library-specific error domains. An error domain encompasses a set of error enumerators with their associated human meanings (for example, error=5 meaning "no leader for Kafka partition"). So for example we can expect that rdkafka_category().message(5) would return std::string("no leader for partition").

The standard also provides a class std::error_condition which is almost identical in implementation to std::error_code.

    namespace std {
    class error_condition {
        int value_;
        std::error_category *cat_;
    public:
        error_condition() noexcept : error_code(0, std::generic_category()) {}
        error_condition(int e, const std::error_category& c) noexcept : value_(e), cat_(&c) {}
        template<class E> error_condition(E e) noexcept requires(std::is_error_condition_enum_v<E>) { *this = make_error_condition(e); }  // intentional ADL
        template<class E> error_condition& operator=(E e) noexcept requires(std::is_error_condition_enum_v<E>) { *this = make_error_condition(e); }  // intentional ADL
        void assign(int e, const std::error_category& c) noexcept { value_ = e; cat_ = &c; }
        void clear() noexcept { *this = std::error_condition(); }

        explicit operator bool() const noexcept { return value_ != 0; }
        int value() const noexcept { return value_; }
        const std::error_category& category() const noexcept { return cat_; }

        std::string message() const { return cat_->message(value_); }
    };
    bool operator==(const std::error_condition& a, const std::error_condition& b) noexcept { return a.value() == b.value() && &a.category() == &b.category(); }
    } // namespace std

The best practice for using std::error_condition is the subject of some debate in SG14; see the rest of this paper. However, clearly the vague intent of std::error_condition is to represent in some way a high-level "condition" which a low-level "error code" (thrown up from the bowels of the system) might or might not "match" in some high-level semantic sense. Notice that the error domain of a default-constructed error_code is system_category(), whereas the error domain of a default-constructed error_condition is generic_category().

The standard provides a highly customizable codepath for comparing an error_code against an error_condition with operator==. (Both error_code and error_condition are value types — in fact they are trivial types — which means that their own operator==(A,A) are just bitwise comparisons. We are now speaking of operator==(A,B).)

    namespace std {
    bool operator==(const std::error_code& a, const std::error_condition& b) noexcept {
        return a.category().equivalent(a.value(), b) || b.category().equivalent(a, b.value());
    }
    bool operator==(const std::error_condition& a, const std::error_code& b) noexcept {
        return a.category().equivalent(a.value(), b) || b.category().equivalent(a, b.value());
    }
    } // namespace std

Recall that the base-class implementation of error_category::equivalent() is just to compare for strict equality; but the programmer's own error_category-derived classes can override equivalent() to have different behavior.

Lastly, the <system_error> header provides an exception class that wraps an error_code (but not an error_condition):

    namespace std {
    class system_error : public std::runtime_error {
        std::error_code code_;
        std::string what_;
    public:
        system_error(std::error_code ec) : code_(ec), what_(ec.message()) {}
        system_error(std::error_code ec, const std::string& w) : code_(ec), what_(ec.message() + ": " + w) {}
        system_error(std::error_code ec, const char *w) : system_error(ec, std::string(w)) {}
        system_error(int e, const std::error_category& cat) : system_error(std::error_code(e, cat)) {}
        system_error(int e, const std::error_category& cat, const std::string& w) : system_error(std::error_code(e, cat), w) {}
        system_error(int e, const std::error_category& cat, const char *w) : system_error(std::error_code(e, cat), w) {}
        const std::error_code& code() const noexcept { return code_; }
        const char *what() const noexcept override { return what_.c_str(); }
    };
    } // namespace std

Note that std::filesystem::filesystem_error inherits from std::system_error.

3. Proposed best practices for using C++11's <system_error> facilities

3.1. Proposed by Arthur O'Dwyer

In the example that follows, we have an application appA, which calls into library libB, which calls into library libC, which calls into library libD.

Arthur O'Dwyer proposes the following general rules:

  1. No enumeration type E should ever satisfy both is_error_code_enum_v<E> and is_error_condition_enum_v<E> simultaneously. (To do so would be to represent a low-level error code value and a high-level abstract condition simultaneously, which is impossible.)
  2. For any enumeration type E, the ADL function make_error_code(E) should exist if-and-only-if is_error_code_enum_v<E>; and the ADL function make_error_condition(E) should exist if-and-only-if is_error_condition_enum_v<E>.
  3. In any enumeration type E satisfying either is_error_code_enum_v<E> or is_error_condition_enum_v<E>, the enumerator value 0 must be set aside as a "success" value, and never allotted to any mode of failure. In fact, the enumeration E should have an enumerator success = 0 or none = 0 to ensure that this invariant is never broken accidentally by a maintainer.
  4. Your library should have exactly as many error_category subclasses as it has enumeration types satisfying either is_error_code_enum_v<E> or is_error_condition_enum_v<E>, in a one-to-one correspondence. No error_category subclass should be "shared" between two enumeration types; and no error_category subclass should exist that is not associated with a specific enumeration type.
  5. Each error_category subclass should be a singleton; that is, it should have a single instance across the entire program.
  6. When your library detects a failure, it should construct an std::error_code representing the failure. This can be done using ADL make_error_code(LibD::ErrCode::failure_mode) or simply using LibD::ErrCode::failure_mode (which works because of std::error_code's implicit constructor from error-code-enum types).
  7. This std::error_code is passed up the stack using out-parameters (as <filesystem>) or using Expected<T, std::error_code>.
  8. When libB receives a std::error_code code that must be checked for failure versus success, it should use if (code) or if (!code).
  9. When libB receives a std::error_code code that must be checked for a particular source-specific error (such as "rdkafka partition lacks a leader"), it should use if (code == LibD::ErrCode::failure_mode). This performs exact equality, and is useful if you know the exact source of the error you're looking for (such as "rdkafka").
  10. When libB receives a std::error_code code that must be checked for a high-level condition (such as "file not found"), which might correspond to any of several source-specific errors across different domains, it may use if (code == LibC::ErrCondition::failure_mode), where LibC::ErrCondition is an error-condition-enum type provided by the topmost library (the one whose API we're calling — not any lower-level library). This will perform semantic classification.
  11. Your library might perhaps provide semantic classification by providing an error-condition-enum type LibB::ErrCondition (and its associated error_category subclass LibB::ErrConditionCategory), which encodes knowledge about the kinds of error values reported by LibC. Ideally, LibB::ErrConditionCategory::equivalent() should defer to LibC::ErrConditionCategory::equivalent() in any case where LibB is unsure of the meaning of a particular error code (for example, if it comes from an unrecognized error domain).
  12. Most likely, std::error_condition and error-condition-enum types should simply not be used. libB should not expect its own callers to write if (code == LibB::ErrCondition::oom_failure); instead libB should expect its callers to write if (LibB::is_oom_failure(code)), where bool LibB::is_oom_failure(std::error_code) is a free function provided by libB. This successfully accomplishes semantic classification, and does it without any operator overloading, and therefore does it without the need for the std::error_condition type.

3.2. Proposed by Charley Bay

In the example that follows, we have an application appA, which calls into library libB, which calls into library libC, which calls into library libD.

Charley Bay proposes (something like) the following general rules (paraphrased here by Arthur O'Dwyer):

  1. No enumeration type E should ever satisfy both is_error_code_enum_v<E> and is_error_condition_enum_v<E> simultaneously. (To do so would be to represent a low-level error code value and a high-level abstract condition simultaneously, which is impossible.)
  2. For any enumeration type E, the ADL function make_error_code(E) should exist if-and-only-if is_error_code_enum_v<E>; and the ADL function make_error_condition(E) should exist if-and-only-if is_error_condition_enum_v<E>.
  3. In any enumeration type E satisfying either is_error_code_enum_v<E> or is_error_condition_enum_v<E>, the enumerator value 0 must be set aside as a "success" value, and never allotted to any mode of failure. In fact, the enumeration E should have an enumerator success = 0 or none = 0 to ensure that this invariant is never broken accidentally by a maintainer.
  4. error_category subclasses are not necessarily singletons. It is conceivable that multiple instances of the same error_category subclass type could exist within the same program.
  5. When your library detects a failure, it should construct an std::error_code representing the failure. This should be done using LibD::ErrCode::failure_mode (which works because of std::error_code's implicit constructor from error-code-enum types).
  6. This std::error_code is passed up the stack using out-parameters (as <filesystem>) or using Expected<T, std::error_code>.
  7. When libB receives a std::error_code code, it must never be checked for a particular source-specific error (such as "rdkafka partition lacks a leader"). Every test should be done on the basis of semantic classification — whether at the coarse granularity of "failure versus success" or at the fine granularity of "partition lacks a leader." It is always conceivable that libC might change out its implementation so that it no longer uses libD; therefore, all testing of error codes returned by a libC API must be expressed in terms of the specific set of abstract failure modes exposed by that same libC API.
  8. When libB receives a std::error_code code, it should not be checked for failure versus success via if (code) or if (!code). Instead, "failure" should be a semantic classification as described in the preceding point.
  9. As explained in the preceding point, exact-equality comparisons should never be used. But with the standard syntax, there is a significant risk that the programmer will accidentally write if (code == LibB::ErrCode::oom_failure) (exact-equality comparison) instead of the intended if (code == make_error_condition(LibB::ErrCode::oom_failure)) (semantic classification). Therefore, under the current standard library design, std::error_condition and error-condition-enum types should not be used. libB should expect its callers to write if (LibB::is_oom_failure(code)), where bool LibB::is_oom_failure(std::error_code) is a free function provided by libB. This successfully accomplishes semantic classification, and does it without any operator overloading, and therefore does it without the need for the std::error_condition type.

4. Issues with the <system_error> facilities

On the 2017-10-11 teleconference, the following issues were discussed.

4.1. Use of std::string

std::error_category's method virtual std::string message(int) const converts an error-code-enumerator into a human-readable message. This functionality is apparently useful only for presentation to humans; i.e., we do not expect that anyone should be treating the result of message() as a unique key, nor scanning into its elements with strstr. However, as a pure virtual method, this method must be implemented by each error_category subclass.

Its return type is std::string, i.e., it introduces a hard-coded dependency on std::allocator<char>. This seems to imply that if you are on a computer system without new and delete, without std::allocator, without std::string, then you cannot implement your own error_category subclasses, which effectively means that you cannot use std::error_code to deal with disappointment. SG14 sees any hard-coded dependency on std::allocator as unfortunate. For a supposedly "fundamental" library like <system_error> it is extremely unfortunate. (LWG issue 2955 is related: std::from_chars used to depend on std::error_code and thence on std::string. It was resolved by the adoption of P0682R1, i.e., std::from_chars simply stopped trying to use std::error_code at all.)

During the SG14 telecon, Charley Bay commented that returning ownership of a dynamically allocated string allows the error_category subclass to return a message that differs based on the current locale. However, nobody on the call claimed that this functionality was important to them. Furthermore, if locale-awareness were desirable, then branching on the current (global) locale would be the wrong way to go about it, because that mechanism would not be usable by multi-threaded programs. The right way to introduce locale-awareness into error_category would be to provide a virtual method std::string message(int, const std::locale&).

SG14 seems to agree that eliminating the dependency on std::allocator would be nice.

SG14 seems to agree that dynamically allocated message strings are not an important feature.

Two ways of removing std::string were proposed: return const char*, or return std::string_view. Arthur O'Dwyer commented that he strongly prefers const char* for simplicity (no new library dependencies) and for consistency with std::error_category::name() and std::exception::what(). Niall Douglas commented that he prefers std::string_view over raw null-terminated const char* whenever possible.

Both ways of removing std::string alter the return type of a pure virtual method and thus inevitably break every subclass of std::error_category ever. SG14 has no way out of this dilemma other than to suggest "wait for std2 and do it correctly there."

4.2. Proliferation of "two-API" libraries

<filesystem> is the poster child for this issue. Every function and method in <filesystem> comes in two flavors: throwing and non-throwing. The throwing version gets the "natural" signature (as is right and expected in C++); and the non-throwing version gets a signature with an extra out-parameter of type std::error_code&. The expectation is apparently that <system_error> users will be willing to write "C-style" code:

        namespace fs = std::filesystem;
        void truncate_if_large(const fs::path& p) noexcept
        {
            std::error_code ec;  // declare an uninitialized variable
            uintmax_t oldsize = fs::file_size(p, ec);
            if (ec) { report_error(ec); return; }
            if (oldsize > 1000) {
                fs::resize_file(p, 1000, ec);
                if (ec) { report_error(ec); return; }
            }
        }
    

It would be nicer if the non-throwing API had exactly the same signatures as the throwing API, except that it should return expected<T> or result<T> instead of T. On the telecon, Arthur O'Dwyer commented that this can't easily be done because you cannot have two functions with the same name and the same signature, differing only in return type.

It is possible to segregate the free functions into a separate namespace, say namespace std::filesystem::nothrow, so that the above code could be written as

        namespace fs = std::filesystem::nothrow;  // hypothetical
        void truncate_if_large(const fs::path& p) noexcept
        {
            auto oldsize = fs::file_size(p);
            if (!oldsize.has_value()) { report_error(oldsize.error()); return; }
            if (oldsize.value() > 1000) {
                auto failure = fs::resize_file(p, 1000);
                if (failure) { report_error(failure.error()); return; }
            }
        }
    
However, this doesn't help with member functions, such as directory_entry::is_symlink().

Having two APIs (throwing and non-throwing) side by side in the same namespace has another disadvantage. There is a significant risk that the programmer might accidentally leave off the out-parameter that signifies "non-throwing-ness", resulting in a call to the throwing version when a call to the non-throwing version was intended.

        namespace fs = std::filesystem;
        void truncate_if_large(const fs::path& p) noexcept
        {
            std::error_code ec;  // declare an uninitialized variable
            uintmax_t oldsize = fs::file_size(p, ec);
            if (ec) { report_error(ec); return; }
            if (oldsize > 1000) {
                fs::resize_file(p, 1000);  // Oops! Bug goes undetected by all major vendors.
                if (ec) { report_error(ec); return; }
            }
        }
    
The LLVM project has already observed this failure mode happening in the wild (bug identified, bugfix commit). It is desirable to clearly segregate throwing from non-throwing functions. But we don't know how to make segregation work for member functions. Therefore perhaps the best outcome would be to stick with a single (non-throwing) API for each library.

If we had a single (non-throwing) API that returned something like Expected<T>, and if Expected<T> had a member function T or_throw() that returned the ex.value() if possible or else threw a system_error initialized from ex.error(), then we could write exception-throwing code fluently as follows:

        namespace fs = std::filesystem::nothrow;
        void truncate_if_large(const fs::path& p) noexcept
        {
            uintmax_t oldsize = fs::file_size(p).or_throw();
            if (oldsize > 1000) {
                fs::resize_file(p, 1000).or_throw();
            }
        }
    
Here we assume that the template class Expected<void> is marked with the standard [[nodiscard]] attribute, so that if the programmer accidentally leaves off the final or_throw() the compiler will emit a warning.

SG14 seems not to have a great answer for how to avoid "two-API" libraries such as <filesystem> going forward; but we believe that "two-API" libraries should be avoided. The Networking TS seems to be shaping up to be another "two-API" library. We believe this is unfortunate.

4.3. No wording sets aside the 0 enumerator

The current Standard strongly implies the best-practice mentioned above: that every error-code-enumerator and every error-condition-enumerator should set aside success = 0 as a special case.

        enum class TroublesomeCode { out_of_memory, out_of_files };
        struct TroublesomeCategory : public std::error_category {
            const char *name() const noexcept override { return ""; }
            std::string message(int e) const override {
                switch (e) {
                    case TroublesomeCode::out_of_memory: return "out of memory";
                    case TroublesomeCode::out_of_files: return "out of files";
                    default: __builtin_unreachable();
                }
            }
        };
        const std::error_category& troublesome_category() {
            static const TroublesomeCategory instance;
            return instance;
        }

        template<> struct std::is_error_code_enum<TroublesomeCode> : std::true_type {};
        std::error_code make_error_code(TroublesomeCode e) {
            return std::error_code((int)e, troublesome_category());
        }

        int main() {
            std::error_code ec = TroublesomeCode::out_of_memory;
            if (ec) {
                puts("This line will not be printed.");
            }
        }
    
If the current specification of std::error_code is to remain untouched, then SG14 would like to see some explicit acknowledgment in the Standard that error-code enumerators with value 0 are "special," i.e., they will not be treated as "errors" by any of the machinery in the Standard. Error codes with value 0 are effectively reserved for the "success" case, and programmers should not attempt to use them for any other purpose.

Vice versa, programmers should be aware that using a non-zero integer value to represent "success" will not work as expected. Consider an HTTP library that naively attempts to use ok = 200 as its "success" code, and then provides an ADL make_error_code like this:

        enum class HTTPStatusCode { ok = 200, created = 201, /* ... */ };

        template<> struct std::is_error_code_enum<HTTPStatusCode> : std::true_type {};
        std::error_code make_error_code(HTTPStatusCode e) {
            return std::error_code((e == ok) ? 0 : (int)e, http_status_category());
        }

        std::string HTTPStatusCategory::message(int e) const {
            switch (e) {
                case 0: return "200 OK";
                case 201: return "201 Created";
                // ...
            }
        }
    
The programmer may head far down this "garden path" under the assumption that his goal of a non-zero "ok" code is attainable; but we on the Committee know that it is not attainable. We should save the programmer some time and some headaches, by explicitly reserving error-code 0 in the standard.

However, there is an alternative and perhaps better solution, which is to replace certain constructors and member functions of std::error_code and std::error_condition as follows: Rather than

        error_code() noexcept : error_code(0, std::system_category()) {}
        error_condition() noexcept : error_code(0, std::generic_category()) {}
        explicit operator bool() const noexcept { return value_ != 0; }
we could propose
        error_code() noexcept : error_code(0, std::null_category()) {}
        error_condition() noexcept : error_code(0, std::null_category()) {}
        explicit operator bool() const noexcept { return cat_ != &std::null_category(); }
This would eliminate the current requirement that error-code 0 always be reserved. Providing a default "null category" could also help to clarify the expected semantics of a default-constructed error_condition, which at present is implementation-defined and varies by vendor.

4.4. Reliance on singletons

std::error_category implicitly relies on singletons. Even if the programmer can somehow get away with using non-singletons for his own categories, the standard library's own categories (e.g. std::generic_category) are singletons: their operator== is explicitly defined to compare instances for address-equality. std::error_category is not the only standard C++ feature to rely on singletons: we have prior art in the form of the std::type_info singletons which are used by dynamic_cast and also by catch.

Prior to the SG14 telecon, Niall Douglas raised the point that singletons do not play well with DLLs (a.k.a. shared objects, a.k.a. dylibs). On some platforms, there are common programming idioms which can cause a C++ library's "singletons" to become duplicated. These include at least:

Arthur O'Dwyer's (admittedly uninformed) opinion is that these sound like antipatterns that could reasonably be avoided. Niall Douglas's opinion is that these are patterns in use in big consumers, e.g. Python loads its C/C++ modules with RTLD_LOCAL, and indeed has no choice but to do so. Charley Bay has also seen these failure modes in practice.

The practical difficulty of using singletons in DLLs seems to be a continuing pain point for certain programmers. The use of singletons by RTTI seems to be a continuing reason that some programmers avoid dynamic_cast and exception handling. The use of singletons by std::error_category will cause the same kinds of problems in practice as the use of singletons by RTTI in exception-handling. If we cannot figure out how to eliminate these practical problems, then std::error_code cannot possibly be a suitable replacement for exception-handling because it will continue to have the same problems (namely, a reliance on singletons compared for address-equality).

In practice, some platforms (notably MSVC, and libc++ if built with a compile-time flag) work around the above problems for type_info singletons by implementing type_info::operator== as a string-equality (i.e. !strcmp(this->name(), rhs.name())) instead of an address-equality (i.e. this == &rhs). This workaround is blessed by the Standard; type_info::operator== is specified to return "true if the two values describe the same type" with no constraints on how this "sameness" is determined. In contrast, error_category::operator== is specified to return exactly "this == &rhs" with no wiggle room at all.

SG14 suggests that perhaps error_category::operator== is overspecified, and that it could be relaxed to allow for string-equality comparison.

SG14 identifies "singletons in DLLs" as a problem area. It is currently unclear how to handle "singletons in DLLs" in C++. Whoever knows the best practice in this area should speak up.

4.5. No error_category subclass can be a literal type

During the SG14 telecon, Odin Holmes and Charley Bay raised the issue that std::error_code is not usable in constexpr contexts; for example, bool operator==(const error_code& lhs, const error_code& rhs) noexcept is not constexpr. Even constructing a std::error_code instance cannot be done constexprly, because error_code(int val, const error_category& cat) noexcept is not constexpr. Even if it were constexpr, we still wouldn't be able to construct an error_code constexprly, because we couldn't get the appropriate const error_category&, because for example const error_category& generic_category() noexcept is not constexpr!

It is impossible to manipulate any error_category instance constexprly, because error_category is not a literal type: it has a virtual destructor, whereas literal types require trivial destructors. This is especially unfortunate because error_category does not need polymorphic destruction.

During the SG14 telecon, Odin Holmes explained that in his programs he often uses constexpr on functions that he does not expect to be evaluated at compile-time. The reason he uses constexpr is to demonstrate and enforce that the functions are pure; and their compile-time-evaluability is just an occasional bonus. This usage of constexpr would have been infeasible in C++11, where constexpr functions were constrained to single statements; but it is feasible in C++14 and later. In C++11, the programmer must learn two different programming styles: a convenient and fluid style for "run-time" functions, and a highly convoluted and obfuscated style for "compile-time" constexpr functions. In C++14, the programmer can generally use a single, convenient, fluid style for both "run-time" functions and "compile-time" constexpr functions; the difference between a non-constexpr function and a constexpr function in C++14 is usually just the addition or subtraction of the keyword constexpr.

However, if a function A's body unconditionally uses some non-constexpr function or constructs some non-literal type B, then adding the keyword constexpr to A will trigger a compiler error. Examples of non-constexpr functions and types include std::vector and std::regex... but also std::error_code!

    constexpr int f(int i) {
        std::error_code ec;
        return i;
    }

    error: variable of non-literal type 'std::error_code' cannot be defined in a constexpr function
        std::error_code ec;
                        ^

Even if we forget about the potentially tricky code to raise and handle error_codes, and just try to propagate an error code up from the lower level to the higher level, we find that we cannot do it in a constexpr way. (This error message is from GCC with libstdc++. libc++ unilaterally adds constexpr to error_code::operator bool(), which is a conforming extension.)

    constexpr int constexpr_4throot(int i, std::error_code& ec) {
        int j = constexpr_sqrt(i, ec);
        if (ec) return 0;
        return constexpr_sqrt(j, ec);
    }

    error: call to non-constexpr function 'std::error_code::operator bool() const'
    if (ec) return 0;
          ^

SG14 would like to develop some good idioms for error handing with std::error_code; but at present, these idioms (such as the if (ec) in the above code) cannot be used in constexpr functions. This sends C++ programmers back to the dark ages of C++11, where we need to learn two different styles of programming: a convenient, fluid style (using error_code) for "run-time" functions and a constrained, convoluted style (eschewing error_code) for functions we want to mark constexpr.

SG14 suggests that std::error_category's destructor should originally have been non-virtual. It is too late to change std::error_category at this point, though, because removing the virtual specifier would noisily break idiomatic C++11 code such as the following:

        struct my_category : public std::error_category {
           // ...
            ~my_category() override;  // OK iff ~error_category is virtual
        }
    
So it seems that std::error_category cannot be made constexpr-friendly. We have not investigated whether std::error_code itself can be made constexpr-friendly, but we would like to see some work in this area.

4.6. No guidance on attaching extra information to error_code

Niall Douglas has attempted to subclass std::error_code in order to create a kind of extended_error_code that contains not only a category pointer and an integer but also some kind of "payload", such as a string or variant holding the arguments of the operation that failed. This is very similar in intent to the C++17 standard library's std::filesystem::filesystem_error, which holds a std::error_code and two instances of std::filesystem::path holding the arguments of the operation that failed. However, std::filesystem::filesystem_error is a subclass of std::runtime_error, whereas Niall and Charley are trying to make something non-polymorphic to be used in the absence of exception-handling.

Niall's further experiments with "error_code plus payload" have resulted in a design he calls "status_code", described in a thread on the SG14 reflector.

Charley Bay has attempted a similar goal via different means. Charley created a subclass of std::error_category which maintained a lookup table of the "payload" for each error_code currently in flight. (This lookup table needs some mechanism for "garbage-collection," since error_code objects are trivially destructible and have no built-in hook by which they could be reference-counted.)

The standard library does not currently provide an extended_error_code type. For people who need (or think they need) something along these lines, it's hard to tell whether they should be inheriting from error_code (will that lead to slicing pitfalls? overload resolution gaffes?), or aggregating á là filesystem_error, or simply passing the extra data around manually via a second out-parameter or something like pair<error_code, string>. Anyway, C++ does not provide a standard wheel, so different programmers may end up inventing slightly different wheels here.

4.7. Reliance on a surprising overload of operator==

The library design relies on the ability to perform a "semantic match" operation between a std::error_code (a low-level error indicator) and a std::error_condition (a high-level condition). The "semantic match" operation is expressed in C++ source code as (ec == econd). During the SG14 telecon, Charley Bay remarked that this use of operator== is surprising because it represents an operation that is fundamentally unlike equality.

For example, "equality" is transitive, but "semantic match" is not necessarily transitive, in that we can have ec1 == econd1 && econd1 == ec2 && ec1 != ec2 or econd1 == ec1 && ec1 == econd2 && econd1 != econd2.

Also, "semantic match" is not necessarily (notionally) symmetric or reflexive, in that we can construct examples where make_error_code(e1) == make_error_condition(e2) && make_error_code(e2) != make_error_condition(e1) or even make_error_code(e1) != make_error_condition(e1). However, these examples require that the enumeration type of e1 define both make_error_code and make_error_condition, which contradicts Arthur's suggested best practices.

It might have been better for the library's API to use a meaningful identifier such as ec.matches(econd), rather than hijacking the == operator for this distinct semantic-match operation.

4.8. error_category should properly have been named error_domain

SG14 generally concluded that the best way to explain the intention of error_category is to say that each singleton derived from error_category represents a particular domain of errors. Each category object understands and manages error codes in a particular domain — for example, "POSIX errors" or "filesystem errors" or "rdkafka errors." It was generally concluded that the English word "domain" expresses this notion more clearly and appropriately than the English word "category."

It is obviously too late to change the name of std::error_category to std::error_domain; that ship has sailed. In the SG14 telecon it was remarked that maybe it's fortunate that the name std::error_domain is still available. If someone designs a "better error_category" (one without singletons, or without std::string, or with constexpr support), then the new thing can be called std::error_domain and we can stop teaching the "awkwardly named" std::error_category.

4.9. Standard error_code-yielding functions can throw exceptions anyway

In LWG discussion of LWG issue 3014, it was pointed out that [fs.err.report] makes several guarantees about the error-reporting behavior of functions in the std::filesystem namespace.

Functions not having an argument of type error_code& handle errors as follows, unless otherwise specified: Functions having an argument of type error_code& handle errors as follows, unless otherwise specified:

It is unclear whether functions using the "error_code" API (e.g. std::filesystem::copy_file) are meant to report "out of memory" errors according to (3.1), or whether the lack of any explicit instruction for "out of memory" errors means that vendors are free to report them by throwing an exception. In practice, it appears that vendors do feel free to report "out of memory" conditions via exception-handling.

This arguably sets a precedent that makes the "two-API solution" even harder to teach and use. A programmer might reasonably expect that if an API provides an error_code& out-parameter, then errors will be reported via that out-parameter. If, as in the case of std::filesystem, the presence of an out-parameter cannot be used to detect the error-reporting mechanism — if standard APIs are free to mix several different error-reporting mechanisms within the same function — writing reliable code becomes more challenging.

4.10. Underspecified error_code comparison semantics

Arthur observes that the following program's behavior differs between libc++ and libstdc++:

    #include <system_error>

    int main() {
        std::error_code ec;          // that is, {0, system_category}
        std::error_condition econd;  // that is, {0, generic_category}
        return ec == econd;          // libc++ true, libstdc++ false
    }

Since error_condition objects are meant only for "catching" (testing) error_code values, perhaps it is incorrect for the programmer to even consider default-constructing an error_condition variable. However, it might still be worthwhile for the Standard to provide guidance in this area.

Overload journal 141 (October 2017) includes an article by Ralph McArdell titled "C++11 (and beyond) Exception Handling," which presents a set of proposed "best practices" for using <system_error>. These best practices basically align with Arthur O'Dwyer's best practices in Section 3 above. A notable difference is that McArdell recommends constructing the "no error" error_code value differently:

    class ErrCategory : std::error_category { ... };
    enum ErrCode { success = 0, failure1 = 1, failure2 = 2 };
    template<> struct std::is_error_code_enum : std::true_type {};
    inline std::error_code make_error_code(ErrCode e) {
        static const ErrCategory c;
        return std::error_code((int)e, c);
    }

    std::error_code McArdellNoError() {
        return ErrCode{};  // that is, {0, ErrCategory}
    }

    std::error_code ODwyerNoError() {
        return std::error_code{};  // that is, {0, system_category}
    }

A programmer working with one of these libraries, who merely wants to know whether std::error_code ec represents an error or not, must have a deep knowledge of all the libraries in his program in order to know whether !ec is trustworthy (see point 4.3); and furthermore, he cannot necessarily trust ec == std::error_condition{} either (point 4.10); and furthermore, if one of his libraries might use McArdell's best practices, he cannot necessarily trust ec == std::error_code{} either!