P2408R4
Ranges iterators as inputs to non-Ranges algorithms

Published Proposal,

This version:
https://wg21.link/p2408r4
Issue Tracking:
Inline In Spec
Author:
(NVIDIA)
Audience:
SG9, LEWG
Toggle Diffs:
Project:
ISO/IEC JTC1/SC22/WG21 14882: Programming Language — C++

1. Abstract

Change the iterator requirements for non-Ranges algorithms. For forward iterators and above that are constant iterators, instead of requiring that iterators meet certain Cpp17...Iterator requirements, require that the iterators model certain iterator concepts. This makes iterators from several standard views usable with non-Ranges algorithms that require forward iterators or above, such as the parallel overloads of most algorithms.

2. Revision history

2.1. R0

This paper arose out of a private e-mail discussion where Bryce Adelstein Lelbach questioned why code that is similar to the first example in § 3 Motivation didn’t compile with MSVC.

2.2. R1

Add the section § 8 Sentinels based on discussion in SG9 (Ranges) on 9 Aug 2021. This is just background information; there is no change to what is being proposed.

2.3. R2

Change the title from "Ranges views as inputs to non-Ranges algorithms" to "Ranges iterators as inputs to non-Ranges algorithms".

Based on discussion in SG9 (Ranges) on 13 Sep 2021, withdraw the portion of the paper that changed input iterators and output iterators to use iterator concepts. The requirements for input and output iterators remain unchanged from the current standard, using Cpp17Iterator requirements.

Add the section Mutable Iterators. This is just background information; there is no change to what is being proposed.

In § 4 Analysis, discuss the iterators for zip_view, et al., that are being proposed in [P2321].

2.4. R3

Based on discussion in SG9 (Ranges) on 11 Oct 2021, change the proposal to handle proxy iterators correctly. This turned out to be a significant change, both to the wording and to the design section. In the wording, the change from Cpp17...Iterator requirements to iterator concepts now only happens for constant iterators, not all iterators. In the design section, § 5.3 Algorithms require concepts instead of categories and § 6 Impact and Details had significant changes and § 7 Mutable Iterators and Proxy Iterators was completely rewritten.

2.5. R4

Gather implementation experience, described in § 9 Implementation Experience. Update the § 6.4 Implementation impact section based on the implementation experience.

3. Motivation

These two snippets of code should be well-formed:

std::vector<int> data = ...;
auto v = data | std::views::transform([](int x){ return x * x; });
int sum_of_squares = std::reduce(std::execution::par, begin(v), end(v));
auto idxs = std::views::iota(0, N);
std::transform(std::execution::par, begin(idxs), end(idxs), begin(sqrts),
               [](int x) { return std::sqrt(float(x)); });

It should be possible, in most cases, to use the iterators from ranges and views as the inputs to C++17 parallel algorithms and to other algorithms that require forward iterators or greater.

4. Analysis

std::reduce requires that its iterator parameters satisfy the requirements of Cpp17ForwardIterator. One of those requirements is that std::iterator_traits<I>::reference be a reference type, either T& or const T&. ([forward.iterators]/p1.3) The iterator category of transform_view::iterator matches the iterator category of the underlying range’s iterator only if the transformation function returns an lvalue reference. If the return type of the transformation function is a value rather than a reference, the iterator category of transform_view::iterator is input_iterator_tag, because the iterator’s reference type alias can’t be a reference type. ([range.transform.iterator]/p2)

In this example, the lambda transformation function returns a value, so the iterators passed to std::reduce are only input iterators in the classic iterator taxonomy used by the non-Ranges algorithms, even though the iterators satisfy the random_access_iterator concept in the Ranges iterator taxonomy.

Several other views have the same issue, where the iterator category of the view’s iterator is input_iterator_tag even when the iterator concept is something else.

4.1. Current situation

MSVC checks that iterator arguments to parallel algorithms have the correct iterator category. If either of the code snippets in § 3 Motivation are compiled, the compilation will fail with "error C2338: Parallel algorithms require forward iterators or stronger."

libstd++ does not do any up-front checking of the iterator category of algorithm arguments. Its implementation of algorithms will accept iterators of any category as long as the iterators support the operations that are actually used in the implementation. The code snippets in § 3 Motivation can be successfully compiled with GCC 11.1.

5. Possible Solutions

5.1. Do nothing

We could simply wait for parallel versions of Range-based algorithms to make their way into the standard. Until then, users who want to use certain view iterators as inputs to parallel algorithms are out of luck.

This solution might be worth considering if parallel Range-based algorithms were on track for C++23. But [P2214] "A Plan for C++23 Ranges" puts parallel Range-based algorithms in Tier 2 of Ranges work, with the caveat that the work needs to be coordinated with Executors to make sure the interface is correct. That pushes the work well past C++23.

Even if parallel Range-based algorithms were already in the draft IS, it would still be worth investigating how to get Ranges iterators to work better with non-Ranges algorithms, in the interest of having the various parts of the Standard work well together.

5.2. Relax the Cpp17 iterator requirements

We could change the definition of Cpp17ForwardIterator ([forward.iterators]), removing the requirement that reference be defined for constant iterators. reference would still need to be defined, and be T&, for mutable iterators.

The authors do not consider this option to be feasible. This change would break existing code that assumes that iterator_traits<I>::reference exists and is a reference type if the iterator category is forward_iterator_tag or greater. There are other possible solutions that don’t break existing code.

5.3. Algorithms require concepts instead of categories

We could change the wording in [algorithms.requirements] to say that the template arguments of the algorithms shall model an iterator concept rather than meet a set of Cpp17 requirements when the iterator is used as a constant iterator. For example:

If an algorithm’s template parameter is named ForwardIterator, ForwardIterator1, or ForwardIterator2, the template argument shall meet the Cpp17ForwardIterator requirements ([forward.iterators]) if it is required to be a mutable iterator, or model forward_iterator ([iterator.concept.forward]) otherwise .

This change relaxes the requirements on some of the template arguments for non-Ranges algorithms. Implementations may need to change to no longer depend on requirements that are no longer required. It is not expected that this will be a big burden, since a well-written algorithm shouldn’t be depending on iterator_traits<I>::reference being a reference type for an iterator that is only used for input.

This change will not break any existing code because the iterator concept imposes fewer requirements on the type than the corresponding Cpp17...Iterator requirements. Any well-formed program will continue to be well-formed and have the same behavior. Some programs that were not well-formed may become well-formed, such as the two motivating examples in this paper; these will have reasonable, non-surprising behavior.

This is the solution proposed by this paper. See immediately below for more details and analysis of this solution.

Making this change only for constant iterators, not mutable iterators, is important. See § 7 Mutable Iterators and Proxy Iterators for details.

For reasons explained below, this change is proposed only for forward, bidirectional, and random access iterators. The requirements for input and output iterators remain unchanged.

6. Impact and Details

6.1. Changes

In the bullets of [algorithms.requirements]/p4 for forward, bidirectional, and random access iterators leave the requirements unchanged for mutable iterators and change the requirements on constant iterators from meeting a set of Cpp17...Iterator requirements to modeling the corresponding iterator concept.

6.2. Relaxed requirements

The Cpp17...Iterator requirements and the iterator concepts are worded very differently, making it challenging to compare them directly.

The only thing that I can find that the iterator concepts require that the Cpp17...Iterator requirements do not is that iterator destructors must be declared noexcept. (Cpp17Iterator requires Cpp17Destructible, and Cpp17Destructible requires that the destructor not throw any exceptions. But Cpp17Destructible allows the destructor to be declared noexcept(false), while the input_or_output_iterator concept requires that the destructor be declared noexcept(true).) Since destructors are implicitly declared noexcept, and no standard iterator has a noexcept(false) destructor, this difference is not meaningful and is not expected to cause any problems.

The one thing that Cpp17ForwardIterator requires that the forward_iterator concept does not is that iterator_traits<I>::reference be a reference type, either T& or const T&. forward_iterator requires that iter_reference_t<I> exist, but doesn’t require that it be a reference type. This same difference also applies to bidirectional iterators and random access iterators. This difference is the root cause of the problems, and is the reason for this paper.

6.3. Other uses

The Cpp17...Iterator requirements are used in several other places in the standard besides [algorithms.requirements]. It is proposed that in places where the requirements are on forward or above non-mutable iterator types that the program passes to standard algorithms, the wording is changed from meeting Cpp17...Iterator requirements to modeling an iterator concept, just like is being done for [algorithms.requirements]. See § 11 Wording for the five places where this happens. Sections where the only requirements on program-supplied iterators are Cpp17InputIterator or Cpp17OutputIterator are not changed. These are [rand.dist.samp.discrete], [rand.dist.samp.pconst], [rand.dist.samp.plinear], and [re.results.form]. Sections where the requirements on forward or above iterators are unchanged because they are required to be mutable are [rand.util.seedseq], [alg.shift], [alg.partitions].

Other uses of Cpp17...Iterator requirements are not changed by this proposal. In some of those cases the requirement is on a standard iterator, not a program iterator, so relaxing the requirements could break existing code. The other uses are in [sequence.reqmts], [associative.reqmts.general], [move.iter.requirements], [locale.category], [fs.req], [fs.path.req], [fs.class.directory.iterator.general], [time.zone.db.list], [reverse.iter.requirements], [allocator.requirements.general], [stacktrace.basic.obs], [string.view.iterators], [container.requirements.general], [span.iterators], [iterator.operations], [alg.equal], and [valarray.range].

6.4. Implementation impact

If any standard algorithm implementations rely on the return type of dereferencing a forward (or higher), non-mutable iterator being a reference type (see § 6.2 Relaxed requirements) or the iterator’s value_type (see § 7 Mutable Iterators and Proxy Iterators), those implementations will have to change to not rely on that. Based on testing (see § 9 Implementation Experience), the number of changes of this sort will be small.

Any implementation, such as MSVC, that checks the iterator category of forward iterators passed to algorithms will have to change the way those checks happen. That change, while a little bit tedious, is not at all hard.

If an algorithm checks the iterator category to choose the most efficient implementation, in some cases that code should be changed to check the iterator concept instead. Testing turned up two cases of this in libstdc++ (see § 9 Implementation Experience), but testing probably won’t find all cases like this. Visual code inspection might be necessary.

I expect that the biggest impact on implementations will be expanding their test suites to cover the code patterns that will become well-formed once this proposal is adopted.

6.5. User impact

No well-formed programs will become ill-formed (unless there are user-defined iterators with noexcept(false) destructors) or have different behavior as a result of this change. Some ill-formed programs will become well-formed with the behaviors users would expect. This change will reduce the number of surprises by making code that users reasonably expect to be well-formed actually well-formed. It is not expected that any users would be negatively impacted by this change.

7. Mutable Iterators and Proxy Iterators

The change from Cpp17...Iterator requirements to iterator concepts is only being done for constant iterators. If the requirements for mutable iterators were changed in the same way, that would allow the passing of proxy iterators to algorithms that wouldn’t handle them correctly.

C++ has always had at least one standard proxy iterator: vector<bool>::iterator. C++23 will get several additional proxy iterators when [P2321] adds zip and related views to the standard library. Proxy iterators often satisfy the requirements of mutable iterators, because they have been designed so that *it = v; does what users expect. But things may go wrong, either compilation errors or unexpected runtime behavior, when other mutating operations are done on proxy iterators, such as v = std::move(*it); or swap(*it1, *it2);. For mutable non-proxy iterators, decltype(*it) is usually T&, but for mutable proxy iterators, decltype(*it) is something different. (For zip_view::iterator, I believe decltype(*it) is a pair or tuple of reference types, not a reference to a pair or tuple.)

Ranges algorithms are designed to correctly handle proxy iterators. For example, their specifications say that they call iter_move or iter_swap, which proxy iterator types can customize, when elements need to be moved or swapped. Non-Ranges algorithms are usually not prepared to handle proxy iterators, with their specifications calling std::move or swap directly on the dereferenced iterators. See the algorithms move and swap_ranges for examples of how std:: algorithms and std::ranges:: algorithms are specified differently.

Proxy iterators can be used as constant iterators without problems. It is only their use as mutable iterators in algorithms that aren’t designed for them that causes problems.

Keeping the Cpp17...Iterator requirements for mutable iterators passed to non-Ranges algorithms prevents using proxy iterators in those situations from being well-formed, because proxy iterators do not satisfy the Cpp17ForwardIterator requirement that iterator_traits<I>::reference be T&.

8. Sentinels

Many ranges use iterator/sentinel pairs, rather than a pair of iterators, as the bounds of the range. Non-Ranges algorithms always require a pair of iterators that have the same type, and will not work with iterator/sentinel pairs. If this proposal is adopted, more ranges than before will be usable with classic algorithms, but the sentinel issue will still prevent some ranges from being useful as inputs to classic algorithms.

The two examples in § 3 Motivation use iterator pairs and don’t suffer from the sentinel issue. transform_view::end() returns an iterator rather than a sentinel if the base range is a common range. In the example, the base range is std::vector, which is a common range. iota_view::end() returns an iterator rather than a sentinel when the range has an upper bound of the same type as the lower bound. It is primarily unbounded iota_views that use a sentinel. Passing an unbounded iota_view to any classic algorithm function as a first/last pair isn’t going to work well, even if first and last are of the same type. I don’t think iota_view's use of sentinels will be a problem in practice in this area.

If a range with iterator/sentinel pairs needs to be used with a classic algorithm, there is a good chance that the user can simply wrap the range in a common_view, or wrap the range’s iterator and sentinel in a common_iterator. The common_view's iterators will be random access if the base range is random access and sized, and will be forward iterators if the base range’s iterator is at least forward. That will allow the common_view to be used in most places where the base range would have been usable if it had used an iterator rather than a sentinel.

I am in favor of changing the non-Ranges algorithms to use iterator/sentinel pairs. But that change can be done independently, not as part of this paper. Both changes are useful on their own, independent of the other, and wouldn’t benefit from being tied together.

9. Implementation Experience

I took a smallish test suite of the C++ parallel algorithms that I wrote a few years ago and adapted it to test this paper. I made two copies of the test suite. In one copy, I used transform_view's iterators, or occasionally iota_view's iterators, wherever possible. In the other copy, I used a hand-written proxy iterator adapter wherever possible. The proxy iterator was designed to have minimal functionality and to catch cases where code assumes that dereferencing an iterator gives you a reference type or a prvalue of the iterator’s value_type.

I ran this test suite with GCC 11 in C++20 mode using its libstdc++. That turned up three places that will need to be changed if this proposal is adopted. In two algorithms, find_end and search_n, the code checked the iterator category and failed to compile if the category was not at least forward_iterator_tag. With this proposal, the code will have to be changed to check the iterator concept instead of iterator category. The other failure was in the overload of inclusive_scan with no initial value. That failed when passed a proxy iterator because it got the type of the intermediate value wrong.

I also ran the test suite with Visual Studio 2022 with /std:c++latest, with the standard library cloned from the STL project on GitHub. I edited the parallel algorithms in the library (see § 4.1 Current situation), changing the requirements checks on the iterators to match this proposal, which meant changing from checking that the iterator’s category was at least forward_iterator_tag to checking that the iterator satisfied the forward_iterator concept. Once I did that, all the tests passed.

The parallel algorithms that were covered by my test suite are adjacent_difference, adjacent_find, all_of, any_of, count, count_if, exclusive_scan, find, find_end, find_first_of, find_if, find_if_not, fill, for_each, inclusive_scan, mismatch, none_of, reduce, search, search_n, transform, transform_exclusive_scan, transform_inclusive_scan, and transform_reduce. This was not meant to be an exhaustive test suite. Its purpose was to have enough coverage to get a sense for how much work it would be to implement this proposal.

Based on these test results, this proposal is implementable with reasonable effort. Some changes will have to be made to the compiler’s standard library, but those changes should not be difficult to make. I expect that most of the effort will be spent writing tests, or adapting existing tests, to cover this area.

10. Feature Test Macro

This change affects the well-formed-ness of certain code. Some users might want to write their code two different ways based on whether or not a standard library has implemented this change. Therefore, I believe that the benefit to users of a feature test macro is greater than the cost to implementers of defining it. This proposal adds the new macro __cpp_lib_algorithm_iterator_requirements.

11. Wording

Changes are relative to N4888 from June 2021.

Change 25.2 "Algorithms requirements" [algorithms.requirements] paragraphs 4 and 5 as follows, combining them into a single paragraph:

Throughout this Clause, where the template parameters are not constrained, the names of template parameters are used to express type requirements. If an algorithm’s Effects: element specifies that a value pointed to by any iterator passed as an argument is modified, then that algorithm has an additional type requirement: The type of that argument shall meet the requirements of a mutable iterator ([iterator.requirements]).

[Note 1: This requirement does These requirements do not affect iterator arguments that are named OutputIterator, OutputIterator1, or OutputIterator2, because output iterators must always be mutable, nor does it affect arguments that are constrained, for which iterator category and mutability requirements are expressed explicitly. — end note]

As currently worded, before these changes, the implementation is required to issue a diagnostic if the iterator type doesn’t meet the given requirements. The requirements contain some semantic elements, which the implementation cannot reliably check at compile time, making it difficult to always issue the required diagnostic. The proposed wording change does not fix this issue, since the concepts also have semantic requirements that can’t be reliably checked at compile time. Should the wording be changed to make the failure to meet the iterator requirements IFNDR or undefined behavior?

Change 25.7.12 "Sample" [alg.random.sample] paragraphs 2 and 5 as follows:

Paragraph 2:

Preconditions: out is not in the range [first, last). For the overload in namespace std:

Paragraph 5:

Remarks:

Change 25.7.9 "Unique" [alg.unique] paragraph 8.2.2 as follows:

Change 30.10.2 "regex_match" [re.alg.match] paragraph 1 as follows:

Preconditions: BidirectionalIterator meets the Cpp17BidirectionalIterator requirements ([bidirectional.iterators]) models bidirectional_iterator ([iterator.concept.bidir]) .

Change 30.10.3 "regex_search" [re.alg.search] paragraph 1 as follows:

Preconditions: BidirectionalIterator meets the Cpp17BidirectionalIterator requirements ([bidirectional.iterators]) models bidirectional_iterator ([iterator.concept.bidir]) .

Add the following feature test macro to [version.syn]:

#define __cpp_lib_algorithm_iterator_requirements date // also in <algorithm>, <numeric>, <memory>

References

Informative References

[P2214]
Barry Rezvin; Conor Hoekstra; Tim Song. A Plan for C++23 Ranges. URL: http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2020/p2214r0.html
[P2321]
Tim Song. zip. URL: http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2021/p2321r2.html

Issues Index

As currently worded, before these changes, the implementation is required to issue a diagnostic if the iterator type doesn’t meet the given requirements. The requirements contain some semantic elements, which the implementation cannot reliably check at compile time, making it difficult to always issue the required diagnostic. The proposed wording change does not fix this issue, since the concepts also have semantic requirements that can’t be reliably checked at compile time. Should the wording be changed to make the failure to meet the iterator requirements IFNDR or undefined behavior?