Skip to content

[libc++] Optimize {std,ranges}::distance for segmented iterators #133612

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
wants to merge 3 commits into
base: main
Choose a base branch
from
+277 −103
Open

[libc++] Optimize {std,ranges}::distance for segmented iterators #133612

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
ec7986a
Optimize {std,ranges}::distance for segmented iterators
winner245 Mar 30, 2025
4ef096d
Address reviewer comments
winner245 Jun 26, 2025
78003ca
Merge branch 'main' into segmented-distance
winner245 Jun 28, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
66 changes: 52 additions & 14 deletions libcxx/include/__iterator/distance.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -10,41 +10,82 @@
#ifndef _LIBCPP___ITERATOR_DISTANCE_H
#define _LIBCPP___ITERATOR_DISTANCE_H

#include <__algorithm/for_each_segment.h>
#include <__config>
#include <__iterator/concepts.h>
#include <__iterator/incrementable_traits.h>
#include <__iterator/iterator_traits.h>
#include <__iterator/segmented_iterator.h>
#include <__ranges/access.h>
#include <__ranges/concepts.h>
#include <__ranges/size.h>
#include <__type_traits/decay.h>
#include <__type_traits/enable_if.h>
#include <__type_traits/remove_cvref.h>
#include <__utility/move.h>

#if !defined(_LIBCPP_HAS_NO_PRAGMA_SYSTEM_HEADER)
# pragma GCC system_header
#endif

_LIBCPP_PUSH_MACROS
#include <__undef_macros>

_LIBCPP_BEGIN_NAMESPACE_STD

template <class _InputIter>
inline _LIBCPP_HIDE_FROM_ABI _LIBCPP_CONSTEXPR_SINCE_CXX17 typename iterator_traits<_InputIter>::difference_type
__distance(_InputIter __first, _InputIter __last, input_iterator_tag) {
typename iterator_traits<_InputIter>::difference_type __r(0);
#if _LIBCPP_STD_VER >= 20
template <class _Iter>
using __iter_distance_t _LIBCPP_NODEBUG = std::iter_difference_t<_Iter>;
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

We already have __iter_diff_t in iterator_traits.h. It doesn't do exactly what you want, but I think we should make it do what you need instead of introducing a new thing that's essentially identical.

Basically, we should try re-defining __iter_diff_t in iterator_traits.h as:

#if _LIBCPP_STD_VER >= 20
template <class _Iter>
using __iter_diff_t _LIBCPP_NODEBUG = std::iter_difference_t<_Iter>;
#else
template <class _Iter>
using __iter_diff_t _LIBCPP_NODEBUG = typename iterator_traits<_Iter>::difference_type;
#endif

And if that's not possible for some reason, it would be interesting to understand why.

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

It seems that the standard hasn't requires iterator_traits<I>::difference_type to be same as iter_difference_t<I> yet when I is a non-random-access iterator type. One is allowed to make them different by specializing incrementable_traits. This is perhaps a defect, for which I've submitted LWG4080.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Given the potential mismatch between iterator_traits<I>::difference_type and iter_difference_t<I> (as pointed out by @frederick-vs-ja), I am now consistently using iterator_traits<I>::difference_type as the return type for the non-range std::__distance algorithm, and iter_difference_t<I> for the range algorithm.

To enable the range algorithm to reuse the same optimization introduced for the non-range std::__distance, I am now adding a static_cast to explicitly cast the result of std::__distance to the desired type iter_difference_t<I>.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I am now consistently using iterator_traits<I>::difference_type as the return type for the non-range std::__distance algorithm.

However, this approach fails when the range algorithm std::ranges::distance forwards to std::__distance with I = cpp20_input_iterator, resulting in a hard substitution error: no type named 'difference_type' in 'std::iterator_traits<cpp20_input_iterator<int *>>', as captured by the tests range.pass.cpp and iterator_sentinel.pass.cpp (see the full log: stage1 (generic-modules, clang-21, clang++-21)).

The root cause for this error is that cpp20_input_iterator does not have the member type iterator_category (although it provides the iterator_concept member type), leading to a hard substitution failure in std::iterator_traits<cpp20_input_iterator<Iter>>::difference_type (See A reproducer).

To resolve this, I still need the above internal definition of __iter_distance_t, which is used as the return type of the internal function std::__distance. This allows both the non-range std::distance and range version std::ranges::distance to share the same implementation and optimization. However, to handle the potential mismatch between iterator_traits<I>::difference_type and iter_difference_t<I>, I explicitly specified iterator_traits<I>::difference_type as the return type of std::distance and iter_difference_t<I> as the return type of std::ranges::distance. This way the underlying implicit conversion would automatically take place when there is a type mismatch.

Basically, we should try re-defining __iter_diff_t in iterator_traits.h as:

#if _LIBCPP_STD_VER >= 20
template <class _Iter>
using __iter_diff_t _LIBCPP_NODEBUG = std::iter_difference_t<_Iter>;
#else
template <class _Iter>
using __iter_diff_t _LIBCPP_NODEBUG = typename iterator_traits<_Iter>::difference_type;
#endif

And if that's not possible for some reason, it would be interesting to understand why.

I am a bit hesitant to make this change because __iter_diff_t is currently unconditionally defined as iterator_traits<_Iter>::difference_type, and it is used by many algorithms. Making this change would introduce the potential type mismatch to many algorithms. My current approach with the introduction of __iter_distance_t ensures that the potential type mismatch is confined and handled within distance.h.

#else
template <class _Iter>
using __iter_distance_t _LIBCPP_NODEBUG = typename iterator_traits<_Iter>::difference_type;
#endif

template <class _InputIter, class _Sent>
inline _LIBCPP_HIDE_FROM_ABI _LIBCPP_CONSTEXPR_SINCE_CXX17 __iter_distance_t<_InputIter>
__distance(_InputIter __first, _Sent __last) {
__iter_distance_t<_InputIter> __r(0);
for (; __first != __last; ++__first)
++__r;
return __r;
}

template <class _RandIter>
inline _LIBCPP_HIDE_FROM_ABI _LIBCPP_CONSTEXPR_SINCE_CXX17 typename iterator_traits<_RandIter>::difference_type
__distance(_RandIter __first, _RandIter __last, random_access_iterator_tag) {
template <class _RandIter, __enable_if_t<__has_random_access_iterator_category<_RandIter>::value, int> = 0>
inline _LIBCPP_HIDE_FROM_ABI _LIBCPP_CONSTEXPR_SINCE_CXX17 __iter_distance_t<_RandIter>
__distance(_RandIter __first, _RandIter __last) {
return __last - __first;
}

template <class _SegmentedIter, class _Difference>
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Is there a reason to apply this optimization before C++20? AFAICT this only applies to join_view and guarding this for C++20 would make the whole thing a lot simpler.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Thanks! I've now made this optimization only for >= C++20.

struct __segment_distance {
using _Traits _LIBCPP_NODEBUG = __segmented_iterator_traits<_SegmentedIter>;

_Difference& __r_;

_LIBCPP_HIDE_FROM_ABI _LIBCPP_CONSTEXPR_SINCE_CXX14 explicit __segment_distance(_Difference& __r) : __r_(__r) {}

_LIBCPP_HIDE_FROM_ABI _LIBCPP_CONSTEXPR_SINCE_CXX14 void
operator()(typename _Traits::__local_iterator __lfirst, typename _Traits::__local_iterator __llast) {
__r_ += std::__distance(__lfirst, __llast);
}
};

template <class _SegmentedIter,
__enable_if_t<!__has_random_access_iterator_category<_SegmentedIter>::value &&
__is_segmented_iterator<_SegmentedIter>::value,
int> = 0>
inline _LIBCPP_HIDE_FROM_ABI _LIBCPP_CONSTEXPR_SINCE_CXX17 __iter_distance_t<_SegmentedIter>
__distance(_SegmentedIter __first, _SegmentedIter __last) {
using __difference_type = __iter_distance_t<_SegmentedIter>;
__difference_type __r(0);
std::__for_each_segment(__first, __last, std::__segment_distance<_SegmentedIter, __difference_type>(__r));
return __r;
}

template <class _InputIter>
inline _LIBCPP_HIDE_FROM_ABI _LIBCPP_CONSTEXPR_SINCE_CXX17 typename iterator_traits<_InputIter>::difference_type
distance(_InputIter __first, _InputIter __last) {
return std::__distance(__first, __last, typename iterator_traits<_InputIter>::iterator_category());
return std::__distance(__first, __last);
}

#if _LIBCPP_STD_VER >= 20
Expand All @@ -56,12 +97,7 @@ struct __distance {
template <class _Ip, sentinel_for<_Ip> _Sp>
requires(!sized_sentinel_for<_Sp, _Ip>)
_LIBCPP_HIDE_FROM_ABI constexpr iter_difference_t<_Ip> operator()(_Ip __first, _Sp __last) const {
iter_difference_t<_Ip> __n = 0;
while (__first != __last) {
++__first;
++__n;
}
return __n;
return std::__distance(std::move(__first), std::move(__last));
}

template <class _Ip, sized_sentinel_for<decay_t<_Ip>> _Sp>
Expand Down Expand Up @@ -92,4 +128,6 @@ inline constexpr auto distance = __distance{};

_LIBCPP_END_NAMESPACE_STD

_LIBCPP_POP_MACROS

#endif // _LIBCPP___ITERATOR_DISTANCE_H
73 changes: 73 additions & 0 deletions libcxx/test/benchmarks/iterators/distance.bench.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,73 @@
//===----------------------------------------------------------------------===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//

// UNSUPPORTED: c++03, c++11, c++14, c++17

#include <cstddef>
#include <deque>
#include <iterator>
#include <ranges>
#include <vector>

#include <benchmark/benchmark.h>

int main(int argc, char** argv) {
auto std_distance = [](auto first, auto last) { return std::distance(first, last); };

// {std,ranges}::distance
{
auto bm = []<class Container>(std::string name, auto distance, std::size_t seg_size) {
benchmark::RegisterBenchmark(
name,
[distance, seg_size](auto& st) {
std::size_t const size = st.range(0);
std::size_t const segments = (size + seg_size - 1) / seg_size;
Container c(segments);
for (std::size_t i = 0, n = size; i < segments; ++i, n -= seg_size) {
c[i].resize(std::min(seg_size, n));
}

auto view = c | std::views::join;
auto first = view.begin();
auto last = view.end();

for ([[maybe_unused]] auto _ : st) {
benchmark::DoNotOptimize(c);
auto result = distance(first, last);
benchmark::DoNotOptimize(result);
}
})
->Arg(50) // non power-of-two
->Arg(1024)
->Arg(4096)
->Arg(8192);
// ->Arg(1 << 14)
// ->Arg(1 << 16)
// ->Arg(1 << 18)
// ->Arg(1 << 20);
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Did you mean to leave those commented? Let's either keep them uncommented or drop them.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I've dropped these large benchmark sizes, which were only used when running the benchmarks for this patch.

};
bm.operator()<std::vector<std::vector<int>>>("std::distance(join_view(vector<vector<int>>))", std_distance, 256);
bm.operator()<std::deque<std::deque<int>>>("std::distance(join_view(deque<deque<int>>))", std_distance, 256);
bm.operator()<std::vector<std::vector<int>>>(
"rng::distance(join_view(vector<vector<int>>)", std::ranges::distance, 256);
bm.operator()<std::deque<std::deque<int>>>(
"rng::distance(join_view(deque<deque<int>>)", std::ranges::distance, 256);

bm.operator()<std::vector<std::vector<int>>>("std::distance(join_view(vector<vector<int>>))", std_distance, 1024);
bm.operator()<std::deque<std::deque<int>>>("std::distance(join_view(deque<deque<int>>))", std_distance, 1024);
bm.operator()<std::vector<std::vector<int>>>(
"rng::distance(join_view(vector<vector<int>>)", std::ranges::distance, 1024);
bm.operator()<std::deque<std::deque<int>>>(
"rng::distance(join_view(deque<deque<int>>)", std::ranges::distance, 1024);
}

benchmark::Initialize(&argc, argv);
benchmark::RunSpecifiedBenchmarks();
benchmark::Shutdown();
return 0;
}
77 changes: 56 additions & 21 deletions libcxx/test/std/iterators/iterator.primitives/iterator.operations/distance.pass.cpp
View file
Open in desktop
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Not attached to this file: let's add a release note!

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Added.

Original file line number Diff line number Diff line change
Expand Up @@ -16,38 +16,73 @@
// Iter::difference_type
// distance(Iter first, Iter last); // constexpr in C++17

#include <iterator>
#include <array>
#include <cassert>
#include <deque>
#include <iterator>
#include <vector>
#include <type_traits>

#include "test_macros.h"
#include "test_iterators.h"

template <class It>
TEST_CONSTEXPR_CXX17
void check_distance(It first, It last, typename std::iterator_traits<It>::difference_type dist)
{
typedef typename std::iterator_traits<It>::difference_type Difference;
static_assert(std::is_same<decltype(std::distance(first, last)), Difference>::value, "");
assert(std::distance(first, last) == dist);
TEST_CONSTEXPR_CXX17 void check_distance(It first, It last, typename std::iterator_traits<It>::difference_type dist) {
typedef typename std::iterator_traits<It>::difference_type Difference;
static_assert(std::is_same<decltype(std::distance(first, last)), Difference>::value, "");
assert(std::distance(first, last) == dist);
}

TEST_CONSTEXPR_CXX17 bool tests()
{
const char* s = "1234567890";
check_distance(cpp17_input_iterator<const char*>(s), cpp17_input_iterator<const char*>(s+10), 10);
check_distance(forward_iterator<const char*>(s), forward_iterator<const char*>(s+10), 10);
check_distance(bidirectional_iterator<const char*>(s), bidirectional_iterator<const char*>(s+10), 10);
check_distance(random_access_iterator<const char*>(s), random_access_iterator<const char*>(s+10), 10);
check_distance(s, s+10, 10);
return true;
#if TEST_STD_VER >= 20
/*TEST_CONSTEXPR_CXX26*/ void test_deque() { // TODO: Mark as TEST_CONSTEXPR_CXX26 once std::deque is constexpr
using Container = std::deque<std::deque<double>>;
Container c;
auto view = c | std::views::join;
Container::difference_type n = 0;
for (std::size_t i = 0; i < 10; ++i) {
n += i;
c.push_back(Container::value_type(i));
}
assert(std::distance(view.begin(), view.end()) == n);
}
#endif

TEST_CONSTEXPR_CXX17 bool tests() {
const char* s = "1234567890";
check_distance(cpp17_input_iterator<const char*>(s), cpp17_input_iterator<const char*>(s + 10), 10);
check_distance(forward_iterator<const char*>(s), forward_iterator<const char*>(s + 10), 10);
check_distance(bidirectional_iterator<const char*>(s), bidirectional_iterator<const char*>(s + 10), 10);
check_distance(random_access_iterator<const char*>(s), random_access_iterator<const char*>(s + 10), 10);
check_distance(s, s + 10, 10);

#if TEST_STD_VER >= 20
{
using Container = std::vector<std::vector<int>>;
Container c;
auto view = c | std::views::join;
Container::difference_type n = 0;
for (std::size_t i = 0; i < 10; ++i) {
n += i;
c.push_back(Container::value_type(i));
}
assert(std::distance(view.begin(), view.end()) == n);
}
{
using Container = std::array<std::array<char, 3>, 10>;
Container c;
auto view = c | std::views::join;
assert(std::distance(view.begin(), view.end()) == 30);
}
if (!TEST_IS_CONSTANT_EVALUATED) // TODO: Use TEST_STD_AT_LEAST_26_OR_RUNTIME_EVALUATED when std::deque is made constexpr
test_deque();
#endif
return true;
}

int main(int, char**)
{
tests();
int main(int, char**) {
tests();
#if TEST_STD_VER >= 17
static_assert(tests(), "");
static_assert(tests(), "");
#endif
return 0;
return 0;
}
Loading
Loading