• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1+++
2title = "Major differences"
3weight = 20
4+++
5
6The major design differences between `<system_error>` and proposed `<system_error2>` are
7as follows:
8
91. `experimental::status_code<DomainType>` can represent warnings
10and form-of-success codes as well as failure codes. `experimental::errored_status_code<DomainType>`
11is more similar to `std::error_code`, in that it can only represent failures
12(this is enforced by C++ 20 contract or runtime assertion check).
13
142. The code's domain implementation defines the payload type to be transported around by
15`experimental::status_code<DomainType>`, rather than it being
16hardcoded to `int` as in `std::error_code`. The payload type can be anything
17you like, including non-trivially-copyable, move-only, complex etc types.
18
19    This facility is extremely useful. Extra failure metadata such as stack
20backtraces can be returned, for example. You can absolutely vary the payload
21depending on whether `NDEBUG` or `_DEBUG` is defined, too.
22
233. If your domain defines a payload type which is trivially copyable or
24move relocating[^1], it gains an implicit convertibility to a move-only
25`experimental::status_code<erased<T>>` where `T` is another
26trivially copyable or move relocating type. This permits global headers
27to use a single, common, type erased, status code type which is highly
28desirable for code bases of any complexity. However, unlike `std::error_code`,
29which fulfils the exact same role in `<system_error>` based code, the type
30erased payload can be bigger than the hardcoded `int` in `std::error_code`.
31
32    This facility is also extremely useful, as extra failure metadata can be
33type erased, transported through code with no knowledge of such things,
34and the original type with failure metadata resurrected at the handling point.
35Indeed P1095 proposed `std::error` is a type alias to
36`experimental::status_code<erased<intptr_t>>`, and it can transport erased
37`std::exception_ptr` instances, POSIX error codes, and much more besides.
38
394. Equality comparisons between status code's with non-identical domains are
40always <b><em>semantic</em></b> i.e. are they semantically equivalent, rather than exactly
41equal? This mirrors when you compare `std::error_code` to a `std::error_condition`,
42but as there is no equivalent for the latter in `<system_error2>`, this means
43that when you see the code:
44
45    ```c++
46    if(code1 == code2) ...
47    ```
48
49    ... you can be highly confident that this is an inexact, semantic, match operation.
50The same code under `<system_error>` is highly ambiguous as to whether exact
51or inexact comparison is being performed (after all, all there is is "`code1 == code2`",
52so it depends on the types of `code1` and `code2` which usually is not obvious).
53
54    The ambiguity, and high cognitive load during auditing `<system_error>` code for correctness, has
55led to many surprising and unexpected failure handling bugs during the past
56decade in production C++.
57
585. `<system_error2>`, being a new design, has all-constexpr construction and
59destruction which avoids the need for static global variables, as `<system_error>`
60has. Each of those static global variables requires an atomic fence just in
61case it has not been initialised, thus every retrieval of an error category bloats
62code and inhibits optimisation, plus makes the use of custom error code categories
63in header-only libraries unreliable. Boost.System has replicated the all-constexpr
64construction and destruction from `<system_error2>`, and thus now has similar
65characteristics in this regard.
66
676. Finally, this is a small but important difference. Under `<system_error>`,
68this extremely common use case is ambiguous:
69
70    ```c++
71    if(ec) ...
72    ```
73
74    Does this code mean "if there was an error?", or "if the error code is set?",
75or "is the error code non-zero?". The correct answer according to the standard is the last choice, but
76a quick survey of open source `<system_error>` based code on github quickly
77demonstrates there is widespread confusion regarding correct usage.
78
79    `<system_error2>` solves this by removing the boolean test entirely. One
80now writes `if(sc.success()) ...`, `if(sc.failure()) ...`, `if(sc.empty()) ...`
81and so on, thus eliminating ambiguity.
82
83
84[^1]: [Move relocating is not in the standard, nor has been reviewed by WG21 yet](http://wg21.link/P1029). It is defined to be a type whose move constructor `memcpy()`'s the bits from source to destination, followed by `memcpy()` of the bits of a default constructed instance to source, and with a programmer-given guarantee that the destructor, when called on a default constructed instance, has no observable side effects. A surprising number of standard library types can meet this definition of move relocating, including `std::vector<T>`, `std::shared_ptr<T>`, and `std::exception_ptr`.