+++++++++++++++++++++++++++++++++++++ Iterator concept and adapter issues +++++++++++++++++++++++++++++++++++++ :date: $Date$ :copyright: Copyright David Abrahams, Jeremy Siek, and Thomas Witt 2003. =================================== Issues from Matt's TR issues list =================================== .. contents:: Index 9.1 iterator_access overspecified? ================================== :Submitter: Pete Becker :Status: New The proposal includes:: enum iterator_access { readable_iterator = 1, writable_iterator = 2, swappable_iterator = 4, lvalue_iterator = 8 }; In general, the standard specifies thing like this as a bitmask type with a list of defined names, and specifies neither the exact type nor the specific values. Is there a reason for iterator_access to be more specific? :Proposed resolution: The ``iterator_access`` enum will be removed, so this is no longer an issue. See the resolution to 9.15. 9.2 operators of iterator_facade overspecified ============================================== :Submitter: Pete Becker :Status: New In general, we've provided operational semantics for things like operator++. That is, we've said that ++iter must work, without requiring either a member function or a non-member function. iterator_facade specifies most operators as member functions. There's no inherent reason for these to be members, so we should remove this requirement. Similarly, some operations are specified as non-member functions but could be implemented as members. Again, the standard doesn't make either of these choices, and TR1 shouldn't, either. So: ``operator*()``, ``operator++()``, ``operator++(int)``, ``operator--()``, ``operator--(int)``, ``operator+=``, ``operator-=``, ``operator-(difference_type)``, ``operator-(iterator_facade instance)``, and ``operator+`` should be specified with operational semantics and not explicitly required to be members or non-members. :Proposed resolution: Not a defect. :Rationale: The standard uses valid expressions such as ``++iter`` in requirements tables, such as for input iterator. However, for classes, such as ``reverse_iterator``, the standard uses function prototypes, as we have done here for ``iterator_facade``. Further, the prototype specification does not prevent the implementor from using members or non-members. 9.3 enable_if_interoperable needs standardese ============================================= :Submitter: Pete Becker :Status: New The only discussion of what this means is in a note, so is non-normative. Further, the note seems to be incorrect. It says that enable_if_interoperable only works for types that "are interoperable, by which we mean they are convertible to each other." This requirement is too strong: it should be that one of the types is convertible to the other. N1541 48 :Proposed resolution: Pete proposed: Remove the enable_if_interoperable stuff, and just write all the comparisons to return bool. Then add a blanket statement that the behavior of these functions is undefined if the two types aren't interoperable. **Needs diffs** (fixed in text). 9.4 enable_if_convertible unspecified, conflicts with requires ============================================================== :Submitter: Pete Becker :Status: New In every place where enable_if_convertible is used it's used like this (simplified):: template struct C { template C(T1, enable_if_convertible::type* = 0); }; The idea being that this constructor won't compile if T1 isn't convertible to T. As a result, the constructor won't be considered as a possible overload when constructing from an object x where the type of x isn't convertible to T. In addition, however, each of these constructors has a requires clause that requires convertibility, so the behavior of a program that attempts such a construction is undefined. Seems like the enable_if_convertible part is irrelevant, and should be removed. There are two problems. First, enable_if_convertible is never specified, so we don’t know what this is supposed to do. Second: we could reasonably say that this overload should be disabled in certain cases or we could reasonably say that behavior is undefined, but we can’t say both. Thomas Witt writes that the goal of putting in enable_if_convertible here is to make sure that a specific overload doesn’t interfere with the generic case except when that overload makes sense. He agrees that what we currently have is deficient. Dave Abrahams writes that there is no conflict with the requires cause because the requires clause only takes effect when the function is actually called. The presence of the constructor signature can/will be detected by is_convertible without violating the requires clause, and thus it makes a difference to disable those constructor instantiations that would be disabled by enable_if_convertible even if calling them invokes undefined behavior. There was more discussion on the reflector: c++std-lib-12312, c++std-lib-12325, c++std-lib- 12330, c++std-lib-12334, c++std-lib-12335, c++std-lib-12336, c++std-lib-12338, c++std-lib- 12362. :Proposed resolution: Specify enable_if_convertible to be as-if :: template enable_if_convertible_impl {}; template <> enable_if_convertible_impl { struct type; }; template struct enable_if_convertible : enable_if_convertible_impl< is_convertible::value> {}; 9.5 ``iterator_adaptor`` has an extraneous '``bool``' at the start of the template definition ============================================================================================= :Submitter: Pete Becker :Status: New The title says it all; this is probably just a typo. :Proposed resolution: Remove the 'bool'. 9.6 Name of private member shouldn't be normative ================================================= :Submitter: Pete Becker :Status: New ``iterator_adaptor`` has a private member named ``m_iterator``. Presumably this is for exposition only, since it's an implementation detail. It needs to be marked as such. :Proposed resolution: Mark the member ``m_iterator`` as exposition only. Note/DWA: I think this is NAD because the user can't detect it, though I'm happy to mark it exposition only. 9.7 ``iterator_adaptor`` operations specifications are a bit inconsistent ========================================================================= :Submitter: Pete Becker :Status: New ``iterator_adpator()`` has a Requires clause, that ``Base`` must be default constructible. ``iterator_adaptor(Base)`` has no Requires clause, although the Returns clause says that the Base member is copy construced from the argument (this may actually be an oversight in N1550, which doesn't require iterators to be copy constructible or assignable). :Proposed resolution: Add a requirements section for the template parameters of ``iterator_adaptor``, and state that ``Base`` must be Copy Constructible and Assignable. N1550 does in fact include requirements for copy constructible and assignable in the requirements tables. To clarify, we've also added the requirements to the text. 9.8 Specialized adaptors text should be normative ================================================= :Submitter: Pete Becker :Status: New similar to 9.3, "Specialized Adaptors" has a note describing ``enable_if_convertible``. This should be normative text. :Proposed resolution: Changed it to normative text. See the resolution of 9.4 9.9 Reverse_iterator text is too informal ========================================= :Submitter: Pete Becker :Status: New reverse iterator "flips the direction of the base iterator's motion". This needs to be more formal, as in the current standard. Something like: "iterates through the controlled sequence in the opposite direction" :Proposed resolution: Change the introduction to: The reverse iterator adaptor iterates through the adapted iterator range in the opposite direction. 9.10 'prior' is undefined ========================= :Submitter: Pete Becker :Status: New reverse_iterator::dereference is specified as calling a function named 'prior' which has no specification. :Proposed resolution: Change the specification to avoid using ``prior`` as follows. :Effects: :: Iterator tmp = m_iterator; return *--tmp; 9.11 "In other words" is bad wording ==================================== :Submitter: Pete Becker :Status: New Transform iterator has a two-part specification: it does this, in other words, it does that. "In other words" always means "I didn't say it right, so I'll try again." We need to say it once. :Proposed resolution: Change: The transform iterator adapts an iterator by applying some function object to the result of dereferencing the iterator. In other words, the ``operator*`` of the transform iterator first dereferences the base iterator, passes the result of this to the function object, and then returns the result. to: The transform iterator adapts an iterator by modifying the ``operator*`` to apply a function object to the result of dereferencing the iterator and returning the result. 9.12 Transform_iterator shouldn't mandate private member ======================================================== :Submitter: Pete Becker :Status: New transform_iterator has a private member named 'm_f' which should be marked "exposition only." :Proposed resolution: Mark the member ``m_f`` as exposition only. Note/DWA: I think this is NAD because the user can't detect it, though I'm happy to mark it exposition only. Change:: UnaryFunction m_f; to:: UnaryFunction m_f; // exposition only 9.13 Unclear description of counting iterator ============================================= :Submitter: Pete Becker :Status: New The description of Counting iterator is unclear. "The counting iterator adaptor implements dereference by returning a reference to the base object. The other operations are implemented by the base m_iterator, as per the inheritance from ``iterator_adaptor``." :Proposed resolution: Change the introduction to: ``counting_iterator`` adapts an object by adding an ``operator*`` that returns the current value of the object. All other iterator operations are forwarded to the adapted object. 9.14 Counting_iterator's difference type ======================================== :Submitter: Pete Becker :Status: New Counting iterator has the following note: [Note: implementers are encouraged to provide an implementation of distance_to and a difference_type that avoids overflows in the cases when the Incrementable type is a numeric type.] I'm not sure what this means. The user provides a template argument named Difference, but there's no difference_type. I assume this is just a glitch in the wording. But if implementors are encouraged to ignore this argument if it won't work right, why is it there? :Proposed resolution: The ``difference_type`` was inherited from ``iterator_adaptor``. However, we've removed the explicit inheritance, so explicit typedefs have been added. 9.15 How to detect lvalueness? ============================== :Submitter: Dave Abrahams :Status: New Shortly after N1550 was accepted, we discovered that an iterator's lvalueness can be determined knowing only its value_type. This predicate can be calculated even for old-style iterators (on whose reference type the standard places few requirements). A trait in the Boost iterator library does it by relying on the compiler's unwillingness to bind an rvalue to a T& function template parameter. Similarly, it is possible to detect an iterator's readability knowing only its value_type. Thus, any interface which asks the user to explicitly describe an iterator's lvalue-ness or readability seems to introduce needless complexity. :Proposed resolution: 1. Remove the ``is_writable`` and ``is_swappable`` traits, and remove the requirements in the Writable Iterator and Swappable Iterator concepts that require their models to support these traits. 2. Change the ``is_readable`` specification to be: ``is_readable::type`` is ``true_type`` if the result type of ``X::operator*`` is convertible to ``iterator_traits::value_type`` and is ``false_type`` otherwise. Also, ``is_readable`` is required to satisfy the requirements for the UnaryTypeTrait concept (defined in the type traits proposal). Remove the requirement for support of the ``is_readable`` trait from the Readable Iterator concept. 3. Remove the ``iterator_tag`` class. 4. Change the specification of ``traversal_category`` to:: traversal-category(Iterator) = let cat = iterator_traits::iterator_category if (cat is convertible to incrementable_iterator_tag) return cat; // Iterator is a new iterator else if (cat is convertible to random_access_iterator_tag) return random_access_traversal_tag; else if (cat is convertible to bidirectional_iterator_tag) return bidirectional_traversal_tag; else if (cat is convertible to forward_iterator_tag) return forward_traversal_tag; else if (cat is convertible to input_iterator_tag) return single_pass_iterator_tag; else if (cat is convertible to output_iterator_tag) return incrementable_iterator_tag; else return null_category_tag; In [lib.iterator.special.adaptors]: Change:: template < class Iterator , class Value = use_default , unsigned Access = use_default_access , class Traversal = use_default , class Reference = use_default , class Difference = use_default > class indirect_iterator to:: template < class Iterator , class Value = use_default , class CategoryOrTraversal = use_default , class Reference = use_default , class Difference = use_default > class indirect_iterator Change:: template < class Iterator2, class Value2, unsigned Access2, class Traversal2 , class Reference2, class Difference2 > indirect_iterator( to:: template < class Iterator2, class Value2, class Category2 , class Reference2, class Difference2 > indirect_iterator( :Rationale: 1. There are two reasons for removing ``is_writable`` and ``is_swappable``. The first is that we do not know of a way to fix the specification so that it gives the correct answer for all iterators. Second, there was only a weak motivation for having ``is_writable`` and ``is_swappable`` there in the first place. The main motivation was simply uniformity: we have tags for the old iterator categories so we should have tags for the new iterator categories. While having tags and the capability to dispatch based on the traversal categories is often used, we see less of a need for dispatching based on writability and swappability, since typically algorithms that need these capabilities have no alternative if they are not provided. 2. We discovered that the ``is_readable`` trait can be implemented using only the iterator type itself and its ``value_type``. Therefore we remove the requirement for ``is_readable`` from the Readable Iterator concept, and change the definition of ``is_readable`` so that it works for any iterator type. 3. The purpose of the ``iterator_tag`` class was to bundle the traversal and access category tags into the ``iterator_category`` typedef. With ``is_writable`` and ``is_swappable`` gone, and ``is_readable`` no longer in need of special hints, there is no reason for iterators to provide information about the access capabilities of an iterator. Thus there is no need for the ``iterator_tag``. The traversal tag can be directly used for the ``iterator_category``. If a new iterator is intended to be backward compatible with old iterator concepts, a tag type that is convertible to both one of the new traversal tags and also to an old iterator tag can be created and use for the ``iterator_category``. 4. The changes to the specification of ``traversal_category`` are a direct result of the removal of ``iterator_tag``. 9.16 is_writable_iterator returns false positives ================================================= :Submitter: Dave Abrahams :Status: New is_writable_iterator returns false positives for forward iterators whose value_type has a private assignment operator, or whose reference type is not a reference (currently legal). :Proposed Resolution: See the resolution to 9.15. 9.17 is_swappable_iterator returns false positives ================================================== :Submitter: Dave Abrahams :Status: New is_swappable_iterator has the same problems as is_writable_iterator. In addition, if we allow users to write their own iter_swap functions it's easy to imagine old-style iterators for which is_swappable returns false negatives. :Proposed Resolution: See the resolution to 9.15. 9.18 Are is_readable, is_writable, and is_swappable useful? =========================================================== :Submitter: Dave Abrahams :Status: New I am concerned that there is little use for any of is_readable, is_writable, or is_swappable, and that not only do they unduly constrain iterator implementors but they add overhead to ``iterator_facade`` and ``iterator_adaptor`` in the form of a template parameter which would otherwise be unneeded. Since we can't implement two of them accurately for old-style iterators, I am having a hard time justifying their impact on the rest of the proposal(s). :Proposed Resolution: See the resolution to 9.15. 9.19 Non-Uniformity of the "lvalue_iterator Bit" ================================================ :Submitter: Dave Abrahams :Status: New The proposed iterator_tag class template accepts an "access bits" parameter which includes a bit to indicate the iterator's lvalueness (whether its dereference operator returns a reference to its value_type. The relevant part of N1550 says: The purpose of the lvalue_iterator part of the iterator_access enum is to communicate to iterator_tagwhether the reference type is an lvalue so that the appropriate old category can be chosen for the base class. The lvalue_iterator bit is not recorded in the iterator_tag::access data member. The lvalue_iterator bit is not recorded because N1550 aims to improve orthogonality of the iterator concepts, and a new-style iterator's lvalueness is detectable by examining its reference type. This inside/outside difference is awkward and confusing. :Proposed Resolution: The iterator_tag class will be removed, so this is no longer an issue. See the resolution to 9.15. 9.20 Traversal Concepts and Tags ================================ :Submitter: Dave Abrahams :Status: New Howard Hinnant pointed out some inconsistencies with the naming of these tag types:: incrementable_iterator_tag // ++r, r++ single_pass_iterator_tag // adds a == b, a != b forward_traversal_iterator_tag // adds multi-pass bidirectional_traversal_iterator_tag // adds --r, r-- random_access_traversal_iterator_tag // adds r+n,n+r,etc. Howard thought that it might be better if all tag names contained the word "traversal". It's not clear that would result in the best possible names, though. For example, incrementable iterators can only make a single pass over their input. What really distinguishes single pass iterators from incrementable iterators is not that they can make a single pass, but that they are equality comparable. Forward traversal iterators really distinguish themselves by introducing multi-pass capability. Without entering a "Parkinson's Bicycle Shed" type of discussion, it might be worth giving the names of these tags (and the associated concepts) some extra attention. :Proposed resolution: Change the names of the traversal tags to the following names:: incrementable_traversal_tag single_pass_traversal_tag forward_traversal_tag bidirectional_traversal_tag random_access_traversal_tag ** Needs work ** (thw) I still believe that implicit_traversal_tag is more appropriate than incrementable_traversal_tag 9.21 iterator_facade Derived template argument underspecified ============================================================= :Submitter: Pete Becker :Status: New The first template argument to iterator_facade is named Derived, and the proposal says: The Derived template parameter must be a class derived from iterator_facade. First, iterator_facade is a template, so cannot be derived from. Rather, the class must be derived from a specialization of iterator_facade. More important, isn't Derived required to be the class that is being defined? That is, if I understand it right, the definition of D here this is not valid:: class C : public iterator_facade { ... }; class D : public iterator_facade { ... }; In the definition of D, the Derived argument to iterator_facade is a class derived from a specialization of iterator_facade, so the requirement is met. Shouldn't the requirement be more like "when using iterator_facade to define an iterator class Iter, the class Iter must be derived from a specialization of iterator_facade whose first template argument is Iter." That's a bit awkward, but at the moment I don't see a better way of phrasing it. :Proposed resolution: Add: In addition, ``static_cast(iterator_facade*)`` shall be well-formed. 9.22 return type of Iterator difference for iterator facade =========================================================== :Submitter: Pete Becker :Status: New The proposal says:: template typename enable_if_interoperable::type operator -(iterator_facade const& lhs, iterator_facade const& rhs); Shouldn't the return type be one of the two iterator types? Which one? The idea is that if one of the iterator types can be converted to the other type, then the subtraction is okay. Seems like the return type should then be the type that was converted to. Is that right? :Proposed resolution: **Needs diffs** (fixed in text) 9.23 Iterator_facade: minor wording Issue ========================================= :Submitter: Pete Becker :Status: New In the table that lists the required (sort of) member functions of iterator types that are based on iterator_facade, the entry for c.equal(y) says: true iff c and y refer to the same position. Implements c == y and c != y. The second sentence is inside out. c.equal(y) does not implement either of these operations. It is used to implement them. Same thing in the description of c.distance_to(z). :Proposed resolution: remove "implements" descriptions from table. See resolution to 9.34 9.24 Use of undefined name in iterator_facade table =================================================== :Submitter: Pete Becker :Status: New Several of the descriptions use the name X without defining it. This seems to be a carryover from the table immediately above this section, but the text preceding that table says "In the table below, X is the derived iterator type." Looks like the X:: qualifiers aren't really needed; X::reference can simply be reference, since that's defined by the iterator_facade specialization itself. :Proposed resolution: Removed the use of X. 9.25 Iterator_facade: wrong return type ======================================= :Submitter: Pete Becker :Status: New Several of the member functions return a Derived object or a Derived&. Their Effects clauses end with:: return *this; This should be :: return *static_cast(this); :Proposed resolution: Change the returns clause to:: return *static_cast(this); Note/DWA: I think this needs to be a more detailed change list, so I'm marking it **Needs language**. 9.26 Iterator_facade: unclear returns clause for operator[] =========================================================== :Submitter: Pete Becker :Status: New The returns clause for ``operator[](difference_type n)`` const says: Returns: an object convertible to X::reference and holding a copy p of a+n such that, for a constant object v of type X::value_type, X::reference(a[n] = v) is equivalent to p = v. This needs to define 'a', but assuming it's supposed to be ``*this`` (or maybe ``*(Derived*)this``), it still isn't clear what this says. Presumably, the idea is that you can index off of an iterator and assign to the result. But why the requirement that it hold a copy of a+n? Granted, that's probably how it's implemented, but it seems over-constrained. And the last phrase seems wrong. p is an iterator; there's no requirement that you can assign a value_type object to it. Should that be ``*p = v``? But why the cast in reference(a[n] = v)? :Proposed resolution: Reworded. 9.27 Iterator_facade: redundant clause ====================================== :Submitter: Pete Becker :Status: New ``operator-`` has both an effects clause and a returns clause. Looks like the returns clause should be removed. :Proposed resolution: Remove the returns clause. 9.28 indirect_iterator: incorrect specification of default constructor ====================================================================== :Submitter: Pete Becker :Status: New The default constructor returns "An instance of indirect_iterator with a default constructed base object", but the constructor that takes an Iterator object returns "An instance of indirect_iterator with the ``iterator_adaptor`` subobject copy constructed from ``x``." The latter is the correct form, since it does not reach inside the base class for its semantics. So the default constructor should return "An instance of ``indirect_iterator`` with a default-constructed ``iterator_adaptor`` subobject." :Proposed resolution: Change: :Returns: An instance of ``indirect_iterator`` with a default constructed base object. to: :Returns: An instance of ``indirect_iterator`` with a default-constructed ``m_iterator``. :Rationale: Inheritance from ``iterator_adaptor`` has been removed, so we instead give the semantics in terms of the (exposition only) member ``m_iterator``. 9.29 indirect_iterator: unclear specification of template constructor ===================================================================== :Submitter: Pete Becker :Status: New The templated constructor that takes an indirect_iterator with a different set of template arguments says that it returns "An instance of indirect_iterator that is a copy of [the argument]". But the type of the argument is different from the type of the object being constructed, and there is no description of what a "copy" means. The Iterator template parameter for the argument must be convertible to the Iterator template parameter for the type being constructed, which suggests that the argument's contained Iterator object should be converted to the target type's Iterator type. Is that what's meant here? (Pete later writes: In fact, this problem is present in all of the specialized adaptors that have a constructor like this: the constructor returns "a copy" of the argument without saying what a copy is.) :Proposed resolution: Change: :Returns: An instance of ``indirect_iterator`` that is a copy of ``y``. to: :Returns: An instance of ``indirect_iterator`` whose ``m_iterator`` subobject is constructed from ``y.base()``. :Rationale: Inheritance from ``iterator_adaptor`` has been removed, so we instead give the semantics in terms of the member ``m_iterator``. 9.30 transform_iterator argument irregularity ============================================= :Submitter: Pete Becker :Status: New The specialized adaptors that take both a Value and a Reference template argument all take them in that order, i.e. Value precedes Reference in the template argument list, with the exception of transform_iterator, where Reference precedes Value. This seems like a possible source of confusion. Is there a reason why this order is preferable? :Proposed resolution: NAD :Rationale: defaults for Value depend on Reference. A sensible Value can almost always be computed from Reference. The first parameter is UnaryFunction, so the argument order is already different from the other adapters. 9.31 function_output_iterator overconstrained ============================================= :Submitter: Pete Becker :Status: New function_output_iterator requirements says: "The UnaryFunction must be Assignable, Copy Constructible, and the expression f(x) must be valid, where f is an object of type UnaryFunction and x is an object of a type accepted by f." Everything starting with "and," somewhat reworded, is actually a constraint on output_proxy::operator=. All that's needed to create a function_output_iterator object is that the UnaryFunction type be Assignable and CopyConstructible. That's also sufficient to dereference and to increment such an object. It's only when you try to assign through a dereferenced iterator that f(x) has to work, and then only for the particular function object that the iterator holds and for the particular value that is being assigned. :Proposed resolution: Remove the part of the sentence after "and". Remove the use of ``output_proxy`` and instead specify ``operator*`` in the following way. ``/* implementation defined */ operator*();`` :Returns: An object ``r`` of implementation defined type such that if ``f(t)`` is a valid expression for some object ``t`` then ``r = t`` is a valid expression. ``r = t`` will have the same effect as ``f(t)``. 9.32 Should output_proxy really be a named type? ================================================ :Submitter: Pete Becker :Status: New This means someone can store an output_proxy object for later use, whatever that means. It also constrains output_proxy to hold a copy of the function object, rather than a pointer to the iterator object. Is all this mechanism really necessary? :Proposed resolution: See issue 9.31. 9.33 istreambuf_iterator isn't a Readable Iterator ================================================== :Submitter: Pete Becker :Status: New c++std-lib-12333: N1550 requires that for a Readable Iterator a of type X, ``*a`` returns an object of type ``iterator_traits::reference``. ``istreambuf_iterator::operator*`` returns ``charT``, but ``istreambuf_iterator::reference`` is ``charT&``. So am I overlooking something, or is ``istreambuf_iterator`` not Readable. :Proposed resolution: Remove all constraints on ``iterator_traits::reference`` in Readable Iterator and Lvalue Iterator. Change Lvalue Iterator to refer to ``T&`` instead of ``iterator_traits::reference``. :Rationale: Ideally there should be requirements on the reference type, however, since Readable Iterator is suppose to correspond to the current standard iterator requirements (which do not place requirements on the reference type) we will leave them off for now. There is a DR in process with respect to the reference type in the stadard iterator requirements. Once that is resolved we will revisit this issue for Readable Iterator and Lvalue Iterator. 9.34 iterator_facade free functions unspecified =============================================== :Submitter: Pete Becker :Status: New c++std-lib-12562: The template functions ``operator==``, ``operator!=``, ``operator<``, ``operator<=``, ``operator>``, ``operator>=``, and ``operator-`` that take two arguments that are specializations of iterator_facade have no specification. The template function operator+ that takes an argument that is a specialization of iterator_facade and an argument of type difference_type has no specification. :Proposed resolution: **Needs diffs** Added the missing specifications. 9.35 iterator_facade: too many equals? ====================================== :Submitter: Pete Becker :Status: New c++std-lib-12563: The table listing the functions required for types derived from iterator_facade has two functions named equal and two named distance_to:: c.equal(b) c.equal(y) c.distance_to(b) c.distance_to(z) where b and c are const objects of the derived type, y and z are constant objects of certain iterator types that are interoperable with the derived type. Seems like the 'b' versions are redundant: in both cases, the other version will take a 'b'. In fact, ``iterator_adaptor`` is specified to use iterator_facade, but does not provide the 'b' versions of these functions. Are the 'b' versions needed? :Proposed resolution: Remove the 'b' versions. 9.36 iterator_facade function requirements ========================================== :Submitter: Pete Becker :Status: New c++std-lib-12636: The table that lists required functions for the derived type X passed to iterator_facade lists, among others: for a single pass iterator:: c.equal(b) c.equal(y) where b and c are const X objects, and y is a const object of a single pass iterator that is interoperable with X. Since X is interoperable with itself, c.equal(b) is redundant. There is a difference in their descriptions, but its meaning isn't clear. The first is "true iff b and c are equivalent", and the second is "true iff c and y refer to the same position." Is there a difference between the undefined term "equivalent" and "refer to the same position"? Similarly, for a random access traversal iterator:: c.distance_to(b) c.distance_to(z) where z is a constant object of a random access traversal iterator that is interoperable with X. Again, X is interoperable with itself, so c.distance_to(b) is redundant. Also, the specification for c.distance_to(z) isn't valid. It's written as "equivalent to distance(c, z)". The template function distance takes two arguments of the same type, so distance(c, z) isn't valid if c and z are different types. Should it be distance(c, (X)z)? :Proposed resolution: Removed the 'b' versions, added the cast. ==================================== More Issues (not from Matt's list) ==================================== 9.37x Inheritance in ``iterator_adaptor`` and other adaptors is an overspecification ===================================================================================== :Submitter: Pete Becker :Status: New c++std-lib-12696: The paper requires that ``iterator_adaptor`` be derived from an appropriate instance of ``iterator_facade``, and that most of the specific forms of adaptors be derived from appropriate instances of ``iterator_adaptor``. That seems like overspecification, and we ought to look at specifying these things in terms of what the various templates provide rather than how they're implemented. :Proposed resolution: Remove the specfication of inheritance, and add explicit specification of all the functionality that was inherited from the specialized iterators. In ``iterator_adaptor``, inheritance is retained, sorry NAD. Also, the Interoperable Iterators concept is added to the new iterator concepts, and this concept is used in the specification of the iterator adaptors. In [lib.iterator.special.adaptors] Change:: class indirect_iterator : public iterator_adaptor { friend class iterator_core_access; to:: class indirect_iterator { public: typedef /* see below */ value_type; typedef /* see below */ reference; typedef /* see below */ pointer; typedef /* see below */ difference_type; typedef /* see below */ iterator_category; Change:: private: // as-if specification typename indirect_iterator::reference dereference() const { return **this->base(); } to:: Iterator const& base() const; reference operator*() const; indirect_iterator& operator++(); indirect_iterator& operator--(); private: Iterator m_iterator; // exposition After the synopsis add: The member types of ``indirect_iterator`` are defined according to the following pseudo-code, where ``V`` is ``iterator_traits::value_type`` .. parsed-literal:: if (Value is use_default) then typedef remove_const::type>::type value_type; else typedef remove_const::type value_type; if (Reference is use_default) then if (Value is use_default) then typedef indirect_reference::type reference; else typedef Value& reference; else typedef Reference reference; if (Value is use_default) then typedef pointee::type\* pointer; else typedef Value\* pointer; if (Difference is use_default) typedef iterator_traits::difference_type difference_type; else typedef Difference difference_type; if (CategoryOrTraversal is use_default) typedef |iterator-category|\ ( iterator_traversal::type,``reference``,``value_type`` ) iterator_category; else typedef |iterator-category|\ ( CategoryOrTraversal,``reference``,``value_type`` ) iterator_category; Add a models section after the requirements section with the following contents: In addition to the concepts indicated by ``iterator_category`` and by ``iterator_traversal::type``, a specialization of ``indirect_iterator`` models the following concepts, Where ``v`` is an object of ``iterator_traits::value_type``: * Readable Iterator if ``reference(*v)`` is convertible to ``value_type``. * Writable Iterator if ``reference(*v) = t`` is a valid expression (where ``t`` is an object of type ``indirect_iterator::value_type``) * Lvalue Iterator if ``reference`` is a reference type. ``indirect_iterator`` is interoperable with ``indirect_iterator`` if and only if ``X`` is interoperable with ``Y``. Before ``indirect_iterator();`` add: In addition to the operations required by the concepts described above, specializations of ``indirect_iterator`` provide the following operations. Change: :Returns: An instance of ``indirect_iterator`` with the ``iterator_adaptor`` subobject copy constructed from ``x``. to: :Returns: An instance of ``indirect_iterator`` with ``m_iterator`` copy constructed from ``x``. At the end of the indirect_iterator operations add: ``Iterator const& base() const;`` :Returns: ``m_iterator`` ``reference operator*() const;`` :Returns: ``**m_iterator`` ``indirect_iterator& operator++();`` :Effects: ``++m_iterator`` :Returns: ``*this`` ``indirect_iterator& operator--();`` :Effects: ``--m_iterator`` :Returns: ``*this`` Change:: class transform_iterator : public iterator_adaptor { friend class iterator_core_access; to:: class transform_iterator { public: typedef /* see below */ value_type; typedef /* see below */ reference; typedef /* see below */ pointer; typedef iterator_traits::difference_type difference_type; typedef /* see below */ iterator_category; After ``UnaryFunction functor() const;`` add:: Iterator const& base() const; reference operator*() const; transform_iterator& operator++(); transform_iterator& operator--(); Change:: private: typename transform_iterator::value_type dereference() const; UnaryFunction m_f; }; to:: private: Iterator m_iterator; // exposition only UnaryFunction m_f; // exposition only }; After the synopsis, add: If ``Iterator`` models Readable Lvalue Iterator and if ``Iterator`` models Random Access Traversal Iterator, then ``iterator_category`` is convertible to ``random_access_iterator_tag``. Otherwise, if ``Iterator`` models Bidirectional Traversal Iterator, then ``iterator_category`` is convertible to ``bidirectional_iterator_tag``. Otherwise ``iterator_category`` is convertible to ``forward_iterator_tag``. If ``Iterator`` does not model Readable Lvalue Iterator then ``iterator_category`` is convertible to ``input_iterator_tag``. In the requirements section, change: The type ``Iterator`` must at least model Readable Iterator. The resulting ``transform_iterator`` models the most refined of the following that is also modeled by ``Iterator``. * Writable Lvalue Iterator if ``result_of::reference)>::type`` is a non-const reference. * Readable Lvalue Iterator if ``result_of::reference)>::type`` is a const reference. * Readable Iterator otherwise. The ``transform_iterator`` models the most refined standard traversal concept that is modeled by ``Iterator``. The ``reference`` type of ``transform_iterator`` is ``result_of::reference)>::type``. The ``value_type`` is ``remove_cv >::type``. to: The argument ``Iterator`` shall model Readable Iterator. Add a models section after the requirements section with the following contents: The resulting ``transform_iterator`` models the most refined of the following options that is also modeled by ``Iterator``. * Writable Lvalue Iterator if ``transform_iterator::reference`` is a non-const reference. * Readable Lvalue Iterator if ``transform_iterator::reference`` is a const reference. * Readable Iterator otherwise. The ``transform_iterator`` models the most refined standard traversal concept that is modeled by the ``Iterator`` argument. If ``transform_iterator`` is a model of Readable Lvalue Iterator then it models the following original iterator concepts depending on what the ``Iterator`` argument models. +-----------------------------------+---------------------------------+ | If ``Iterator`` models | then ``filter_iterator`` models | +===================================+=================================+ | Single Pass Iterator | Input Iterator | +-----------------------------------+---------------------------------+ | Forward Traversal Iterator | Forward Iterator | +-----------------------------------+---------------------------------+ | Bidirectional Traversal Iterator | Bidirectional Iterator | +-----------------------------------+---------------------------------+ | Random Access Traversal Iterator | Random Access Iterator | +-----------------------------------+---------------------------------+ If ``transform_iterator`` models Writable Lvalue Iterator then it is a mutable iterator (as defined in the old iterator requirements). ``transform_iterator`` is interoperable with ``transform_iterator`` if and only if ``X`` is interoperable with ``Y``. Remove the private operations section heading and remove:: ``typename transform_iterator::value_type dereference() const;`` :Returns: ``m_f(transform_iterator::dereference());`` After the entry for ``functor()``, add:: ``Iterator const& base() const;`` :Returns: ``m_iterator`` ``reference operator*() const;`` :Returns: ``m_f(*m_iterator)`` ``transform_iterator& operator++();`` :Effects: ``++m_iterator`` :Returns: ``*this`` ``transform_iterator& operator--();`` :Effects: ``--m_iterator`` :Returns: ``*this`` .. |iterator-category| replace:: *iterator-category* 9.38x Problem with specification of a->m in Readable Iterator ============================================================= :Submitter: Howard Hinnant :Status: New c++std-lib-12585: Readable Iterator Requirements says:: a->m U& pre: (*a).m is well-defined. Equivalent to (*a).m Do we mean to outlaw iterators with proxy references from meeting the readable requirements? Would it be better for the requirements to read ``static_cast(*a).m`` instead of ``(*a).m`` ? :Proposed resolution: NAD. :Rationale: We think you're misreading "pre:". If ``(*a).m`` is not well defined, then the iterator is not required to provide ``a->m``. So a proxy iterator is not required to provide ``a->m``. As an aside, it is possible for proxy iterators to support ``->``, so changing the requirements to read ``static_cast(*a).m`` is interesting. However, such a change to Readable Iterator would mean that it no longer corresponds to the input iterator requirements. So old iterators would not necessarily conform to new iterator requirements. 9.39x counting_iterator Traversal argument unspecified ====================================================== :Submitter: Pete Becker c++std-lib-12635: counting_iterator takes an argument for its Traversal type, with a default value of use_default. It is derived from an instance of ``iterator_adaptor``, where the argument passed for the Traversal type is described as "\ ``/* see details for traversal category */``". The details for counting_iterator describe constraints on the Incrementable type imposed by various traversal categories. There is no description of what the argument to ``iterator_adaptor`` should be. :Proposed resolution: We no longer inherit from ``iterator_adaptor``. So instead, we specify the ``iterator_category`` in terms of the Traversal type (which is now called CategoryOrTraversal). Also the requirements and models section was reorganized to match these changes and to make more sense. 9.40x indirect_iterator requirements muddled ============================================ :Submitter: Pete Becker c++std-lib-12640: The value_type of the Iterator template parameter should itself be dereferenceable. The return type of the ``operator*`` for the value_type must be the same type as the Reference template parameter. I'd say this a bit differently, to emphasize what's required: iterator_traits::value_type must be dereferenceable. The Reference template parameter must be the same type as ``*iterator_traits::value_type()``. The Value template parameter will be the value_type for the indirect_iterator, unless Value is const. If Value is const X, then value_type will be non- const X. Also non-volatile, right? In other words, if Value isn't use_default, it just gets passed as the Value argument for ``iterator_adaptor``. The default for Value is:: iterator_traits< iterator_traits::value_type >::value_type If the default is used for Value, then there must be a valid specialization of iterator_traits for the value type of the base iterator. The earlier requirement is that ``iterator_traits::value_type`` must be dereferenceable. Now it's being treated as an iterator. Is this just a pun, or is ``iterator_traits::value_type`` required to be some form of iterator? If it's the former we need to find a different way to say it. If it's the latter we need to say so. :Proposed resolution: Change: The ``value_type`` of the ``Iterator`` template parameter should itself be dereferenceable. The return type of the ``operator*`` for the ``value_type`` must be the same type as the ``Reference`` template parameter. The ``Value`` template parameter will be the ``value_type`` for the ``indirect_iterator``, unless ``Value`` is const. If ``Value`` is ``const X``, then ``value_type`` will be *non-* ``const X``. The default for ``Value`` is:: iterator_traits< iterator_traits::value_type >::value_type If the default is used for ``Value``, then there must be a valid specialization of ``iterator_traits`` for the value type of the base iterator. The ``Reference`` parameter will be the ``reference`` type of the ``indirect_iterator``. The default is ``Value&``. The ``Access`` and ``Traversal`` parameters are passed unchanged to the corresponding parameters of the ``iterator_adaptor`` base class, and the ``Iterator`` parameter is passed unchanged as the ``Base`` parameter to the ``iterator_adaptor`` base class. to: The expression ``*v``, where ``v`` is an object of ``iterator_traits::value_type``, shall be valid expression and convertible to ``reference``. ``Iterator`` shall model the traversal concept indicated by ``iterator_category``. ``Value``, ``Reference``, and ``Difference`` shall be chosen so that ``value_type``, ``reference``, and ``difference_type`` meet the requirements indicated by ``iterator_category``. [Note: there are further requirements on the ``iterator_traits::value_type`` if the ``Value`` parameter is not ``use_default``, as implied by the algorithm for deducing the default for the ``value_type`` member.] :Rationale: Not included above is the specification of the ``value_type``, ``reference``, etc., members, which is handled by the changes in 9.37x. 9.41x Problem with transform_iterator requirements ================================================== :Submitter: Pete Becker c++std-lib-12641: The reference type of transform_iterator is ``result_of::reference)>::type``. The ``value_type`` is ``remove_cv >::type``. These are the defaults, right? If the user supplies their own types that's what gets passed to ``iterator_adaptor``. And again, the specification should be in terms of the specialization of ``iterator_adaptor``, and not in terms of the result: Reference argument to iterator_adaptor:: if (Reference != use_default) Reference else result_of::reference)>::type Value argument to iterator_adaptor:: if (Value != use_default) Value else if (Reference != use_default) remove_reference::type else remove_reference::reference)>::type>::type There's probably a better way to specify that last alternative, but I've been at this too long, and it's all turning into a maze of twisty passages, all alike. :Proposed resolution: Replace: The reference type of transform_iterator is ``result_of::reference)>::type``. The ``value_type`` is ``remove_cv >::type``. with: If ``Reference`` is ``use_default`` then the ``reference`` member of ``transform_iterator`` is ``result_of::reference)>::type``. Otherwise, ``reference`` is ``Reference``. If ``Value`` is ``use_default`` then the ``value_type`` member is ``remove_cv >::type``. Otherwise, ``value_type`` is ``Value``. 9.42x filter_iterator details unspecified ========================================= :Submitter: Pete Becker c++std-lib-12642: The paper says:: template class filter_iterator : public iterator_adaptor< filter_iterator, Iterator, use_default, /* see details */ > That comment covers the Access, Traversal, Reference, and Difference arguments. The only specification for any of these in the details is: The access category of the filter_iterator will be the same as the access category of Iterator. Needs more. :Proposed resolution: Add to the synopsis:: typedef iterator_traits::value_type value_type; typedef iterator_traits::reference reference; typedef iterator_traits::pointer pointer; typedef iterator_traits::difference_type difference_type; typedef /* see below */ iterator_category; and add just after the synopsis: If ``Iterator`` models Readable Lvalue Iterator and Forward Traversal Iterator then ``iterator_category`` is convertible to ``std::forward_iterator_tag``. Otherwise ``iterator_category`` is convertible to ``std::input_iterator_tag``. 9.43x transform_iterator interoperability too restrictive ========================================================= We do not need to require that the function objects have the same type, just that they be convertible. :Proposed resolution: Change:: template transform_iterator( transform_iterator const& t , typename enable_if_convertible::type* = 0 // exposition ); to:: template transform_iterator( transform_iterator const& t , typename enable_if_convertible::type* = 0 // exposition only , typename enable_if_convertible::type* = 0 // exposition only ); 9.44y ``indirect_iterator`` and smart pointers ============================================== :Submitter: Dave Abrahams ``indirect_iterator`` should be able to iterate over containers of smart pointers, but the mechanism that allows it was left out of the specification, even though it's present in the Boost specification :Proposed resolution: Add ``pointee`` and ``indirect_reference`` 9.45y N1530: Typos in proposal text (not standardese) ===================================================== :Submitter: Dave Abrahams 1. "because specification helps to highlight that the ``Reference`` template parameter may not always be identical to the iterator's ``reference`` type, and will keep users making mistakes based on that assumption." :Proposed resolution: add "from" before "making" 9.46y N1530: ``base()`` return-by-value is costly ================================================= :Submitter: Dave Abrahams We've had some real-life reports that iterators that use ``iterator_adaptor``\ 's ``base()`` function can be inefficient when the ``Base`` iterator is expensive to copy. Iterators, of all things, should be efficient. :Proposed resolution: Change:: Base base() const; to:: Base const const& base() const;