This is a plain wording fixing proposal for the table-based requirement set of class template fpos. A paper is provided instead of drafting wording in the referenced issues, to have more room available for presentation and explanation.The wording changes suggested by this paper should not have effects on any existing implementations, because it mainly intended to clearify several underspecified parts of the current specification. Nonetheless during the analysis of the paper there occurred some curiosities in the specification that caused to ask some questions to the committee which might lead to decisions to change existing implementations, but the author has tried to be as conservative as possible in regard to such suggestions.
The fpos type family is one of the rarely and sparely explicitly mentioned templates of the Standard Library specified in 188.8.131.52 [fpos.operations] including Table 112 — "Position type requirements" (all references within this proposal to the C++ working draft are based on 1). Specializations of that template are used in the IO Library to denote file positions and to allow some arithmetic operations and to store/retrieve a rather opaque "state"; furthermore it connects std::char_traits with the IO streams due to its pos_type member which is an fpos instantiation whose state type is equal to char_traits<CharT>::state_type. The latter property is not a strict requirement, since the specification leaves it implementation-defined what happens when traits::pos_type differs from fpos<char_traits<CharT>::state_type>. The current wording also says (30.9.2 [filebuf] p4) that the behaviour is undefined, if traits::pos_type of basic_filebuf is different from fpos<traits::state_type>. For the pre-defined char_traits, traits::state_type is equal to mbstate_t, just another sparely-specified type inherited from the <cwchar> header. The C specification tells us, that mbstate_t is a complete, non-array object type that can be default-initialized and value-initialized (freely translated into C++ language terms) and "can hold the conversion state information necessary to convert between sequences of multibyte characters and wide characters". In C++ (see 24.2.2 [char.traits.typedefs] p4) the requirements imposed on char_traits<>::state_type are CopyAssignable, CopyConstructible, and DefaultConstructible. Associated with both char_traits and fpos are also the types streamoff and streamsize which are both signed basic integral types (30.5.2 [stream.types] p1+2). An alias for std::fpos<std::mbstate_t> is std::streampos.
Summarizing the essence of 184.108.40.206 [fpos.members] and 220.127.116.11 [fpos.operations], we can conclude the following:
fpos<stateT> holds a single object of type stateT, that can be read from and written to via a state attribute.
fpos<stateT> objects are EqualityComparable (This is not explicitly said but it can be indirectly deduced from the definition of EqualityComparable in Table 20 and the imposed requirements of fpos in regard to == in Table 112).
fpos<stateT> objects can be direct-initialized or copy-initialized from int values.
fpos<stateT> objects p can be direct-initialized from streamoff objects o and streamoff objects can be direct-initialized from fpos<stateT> objects, such that the conversion sequence p1 → o → p2 ensures that p1 == p2.
streamoff values can be added to and subtracted from fpos and the difference of two fpos values returns a streamoff.
The following bullets list some concrete unclear points. To simplify the descriptions below, P denotes a class type that is an instance of the std::fpos template and O denotes the type std::streamoff.
What is the semantics of initializing an fpos from an int value? (see LWG 2832). Note that contrary to the construction from streamoff there exists no similar conversion sequence cycle guarantee for int values. In theory, an (insane) implementation could provide different constructors of fpos for int and for streamoff, both having different semantics. Or, the constructor taking an int could be non-explicit and the constructor taking an streamoff could be explicit. The author is convinced that this interpretation is not intended and recommends to specify the semantics of fpos initialization from int to be equivalent to the semantics of fpos initialization from streamoff, where the streamoff value has been initialized from int before (Note that due to the type nature of streamoff this conversion is statically possible, and it should always be dynamically be well-defined if the conversion of int to streamoff is well-defined). The indirect effect of doing that is that the specification would now relax the direct-initialization requirement from O to P to an implicit conversion from O to P, which is in sync with what existing implementations do anyway.
Related to the previous item, some entries in Table 112 are very likely debris of the last cleanup paper in this area, specifically by N2884. Before that paper, type streamoff was an implementation-defined type, not necessarily a built-in type. With the resolution of that paper streamoff became a guaranteed alias for a signed integral type. Since all integer types can be implicitly converted into each other, there is not much reason to provide the additional guarantee that fpos instances can be initialized by int values, because relying on the initialization by streamoff should be sufficient (This assumes the absence of implementations that used SFINAE-constrained constructors to accept exactly type streamoff for one constructor and exactly type int for another constructor). For similar reasons is the last row in Table 112 obsolete, which specifies the validity to convert between streamsize and streamoff, because streamsize is (and was) also a synonym for a signed integral type. This proposal therefore recommends to remove this last row as well as the explicit initialization specification of fpos by int values.
What requirements are imposed on the template type parameter stateT? (see LWG 2808). The current specification allows to conclude that stateT is required to be DefaultConstructible, because there exists no portable way to initialize fpos with stateT, assuming that an implementation is not required to in-place construct stateT. The semantics of the state attribute functions allows the assumption that stateT needs to be also CopyConstructible, CopyAssignable, and Destructible. This overlaps nicely with the requirements that come from 24.2.2 [char.traits.typedefs] p4, but adds a missing Destructible requirement.A related open question is that the current specification doesn't tell whether the specified fpos<stateT> constructors will default-initialize or value-initialize the exposition-only stateT member, because this has impact on conforming programs that attempt to read the state before calling the setter. After inspection of existing implementations it turns out that all do the same thing, they value-initialize the state member. The author therefore recommends to specify this existing practice. If the committee would consider this as an unwanted strong specification, the wording should at least emphasize that the stateT member may be default-initialized after construction and may thus have an indeterminate value, which implies that an initial call of the stateT state() const signature would be undefined behaviour.
Are fpos instances DefaultConstructible, CopyConstructible, CopyAssignable, and Destructible? The requirements for CopyConstructible and Destructible follow indirectly from the usage of fpos in return types and function arguments from a pre-C++11 era, where every move was a copy. There is currently clearly no DefaultConstructible requirement, but existing implementations such as from Visual Studio 2017/Dinkumware, clang/libc++, and gcc/libstd++ provide default-constructible fpos types. In all cases the default-constructed fpos is equivalent to an fpos initialized by streamoff(0). Therefore this proposal suggests to require that fpos instances are CopyAssignable and DefaultConstructible.
Existing implementations of fpos provide implicit conversion to streamoff, the specification requires only explicit conversions. The current implementation state can be explained by the pre-C++11 existence of that type, where no explicit conversion functions where available. Nonetheless the author believes that the currently stricter specification is reasonable for now and doesn't recommend any change. The author is open to be requested to weaken the specification, therefore points this characteristics out here, asking for feedback from the committee.
Are there any trivial type guarantees of fpos, conditionally depending on triviallity characteristics of stateT? Albeit current implementations provide a default-constructor for fpos, this one is never trivial. So regardless whether our recommendation is accepted to require the existence of an fpos default constructor, we can't give trivial type guarantees. But all implementations available to the author meet the following characteristics:
and thus are trivially copyable if stateT is also trivially copyable. The author of this proposal suggest to provide individual trivial special member guarantees depending on corresponding is_trivially_xxx<stateT> properties, similar to the special member specifications of std::istream_iterator.
The mixed additions of fpos and streamoff are not symmetric (fpos on the right side is not specified), is this intended?Let p denote a (possibly const) value of P and o denote a (possibly const) value of O, the existing requirements support the following expression
p + o;
but not the other way around:
o + p;
The observation that o + p is undefined by the letters of the Standard is unexpected for users and according to the author's opinion user code uses either form obviously unaware of that fact. The interesting point is that all implementations seem to behave the same way as follows:
P p(0); O o(0); using add1 = decltype(p + o); using add2 = decltype(o + p); static_assert(std::is_same_v<add1, P>); static_assert(std::is_same_v<add2, O>);
This program is well-formed for all tested implementations, which means that p + o and o + p do have different result types: The first one has type P, the second one has type O. The technical reason for this is that existing implementations only define a member signature fpos operator+(streamoff offs) const but also always a conversion function operator streamoff() const, the net result being that the expression o + p is well-formed, because it finds the conversion function and after the conversion of p to streamoff it just adds two streamoff values.The author of this proposal believes that the binary plus operation of P and O should be well-defined to reduce user surprises. Unfortunately existing implementations behave non-symmetric in regard to the semantics and result type. It is unclear how to proceed here:
Keep the current state and leave o + p undefined. User-code cannot call that combination of this binary operator at all in conforming code.
Change the specification to make o + p well-defined, allow one of P or O to be valid. User-code could thus either use auto for the result or rely on an implicit conversion to P in conforming code.
Change the specification to make o + p well-defined, and say the the return type is convertible to P. User-code could thus either use auto for the result or rely on an implicit conversion to P in conforming code.
Change the specification to make o + p well-defined, but specify what existing implementations do. User-code could thus either use O for the result or rely on implicit conversion to P in conforming code.
Change the specification to make o + p well-defined and require that all existing implementations change so that the return type is P instead of O. User-code could thus either use P for the result or rely on direct initialization to O in conforming code.
Some short observations to some of the suggested variants: Variant (2) would be a very unusual specification, because we typically don't say that the result is one of several choices. In diffuse cases we typically make an implicit conversion to some other type part of the spec (See for example the result of == or != in most requirement sets). Interestingly, all variants have in common that the result is convertible to P, which is appealing because it emphasizes the "right" type. Variant (3) uses this commonality as the specification directly.The author would like to gather feedback from the committee in regard to the preferred outcome, but suggests to use the non-breaking variant (3), which has the following advantages: First, its specification emphasizes the intuitively "correct" result type. Second, if the committee is in fond of that idea, existing implementations could decide in the future to add the missing P operator+(O, P). Then, in a later future, if a new survey of existing practice would be performed again and when now all existing implementations have added this member, we could now strengthen the specification to impose that the return value of o + p is P. The author would also express his recommendation not to standardize the behaviour of existing implementations (variant (4)). This seems an unusual recommendation at first, but in this particular case, the author suspects that the status quo of implementations doesn't exist because of an intentional design, but due to an unintended artifact of strictly implementing what Table 112 describes. The author would like to get feedback from existing implementations whether this assumption is correct or wrong.
According to Table 112 it looks as if the return types of assignment operator and compound assignment operator of fpos have type fpos and not fpos& (similar for streamoff). The presumable intend is to specify the return types of the pure binary arithmetic expressions and not of the denoted assignments (except for the compound assignments). This paper proposes to isolate the different expressions and fix the wrong compound assignment return types.
Some rows in Table 112 specify a return type fpos for expressions such as P(o). This is invalid, because fpos is a class template, but the return type for an expression P(o) needs to be P, which is the same specialization of fpos that has been used to form the actual expression. The recommended resolution is to replace such uses of fpos in Table 112 by P instead.
If the proposed resolution will be accepted, the following library issues will be resolved:
|Number||Description||C++17 NB comment|
|2808||Requirements for fpos and stateT||GB 60|
|2832||§[fpos.operations] strange requirement for P(i)||—|
At some places below, additional markup of the form
[Drafting notes: whatever — end drafting notes]
is provided, which is not part of the normative wording, but is solely shown to provide additional information to the reader about the rationale of the concrete wording.
The proposed wording changes refer in all cases to N4659.
Change 18.104.22.168 [fpos.operations] as indicated:
Operations specified in Table 112 are permitted.In that table,
(1.1) — P refers to an instance of fpos,
(1.2) — p and q refer to values of type P ,
(1.3) — O refers to type streamoff,
(1.4) — o refers to a value of type streamoff ,
(1.5) — sz refers to a value of type streamsize and
(1.6) — i refers to a value of type int.
[Drafting notes: It is recommend to strike the non-normative note in p2 completely. This seems to be very out-dated wording and what is says is already said in 22.214.171.124 [structure.requirements], in particular in p3+p4 — end drafting notes]
-3- Stream operations that return a value of type traits::pos_type return P(O(-1)) as an invalid value to signal an error. If this value is used as an argument to any istream, ostream, or streambuf member that accepts a value of type traits::pos_type then the behavior of that function is undefined.
-2- [Note: Every implementation is required to supply overloaded operators on fpos objects to satisfy the requirements of 126.96.36.199 [fpos.operations]. It is unspecified whether these operators are members of fpos, global operators, or provided in some other way. — end note]
Change Table 112 — "Position type requirements", as indicated:
[Drafting notes: We can strike the wording about the destructor and ==, because we have now the EqualityComparable and Destructible requirements specified in 188.8.131.52 [fpos.operations] — end drafting notes]
Table 112 — Position type requirements Expression Return type Operational semantics Assertion/note
P(i) p == P(i)
note: a destructor is assumed.
P p = i;
Postconditions: p == P(i). P(o) fpos converts from offset
O(p) streamoff converts to offset P(O(p)) == p p == q convertible to bool == is an equivalence relation p != q convertible to bool !(p == q) q =p + o
p += o fpos + offset q - o == p q =p - o
p -= o fpos - offset q + o == p o =p - q streamoff distance
q + o == p
streamsize(O(sz)) == sz
streamsize(O(sz)) == sz
N4659 Richard Smith: "Working Draft, Standard for Programming Language C++"
Thanks to Jonathan Wakely and Billy Robert O'Neal III for a review of this proposal and for encouragement to make a little more changes as I originally believed I should make.