Automerge: [libc++] Optimize ranges::{for_each, for_each_n} for segmented iterators (#132896)

Previously, the segmented iterator optimization was limited to `std::{for_each, for_each_n}`. This patch
extends the optimization to `std::ranges::for_each` and `std::ranges::for_each_n`, ensuring consistent
optimizations across these algorithms. This patch first generalizes the `std` algorithms by introducing
a `Projection` parameter, which is set to `__identity` for the `std` algorithms. Then we let the `ranges`
algorithms to directly call their `std` counterparts with a general `__proj` argument. Benchmarks
demonstrate performance improvements of up to 21.4x for ``std::deque::iterator`` and 22.3x for
``join_view`` of ``vector<vector<char>>``.

Addresses a subtask of #102817.

Author: winner245

libcxx/docs/ReleaseNotes/21.rst

@@ -70,8 +70,9 @@ Improvements and New Features
 - The segmented iterator optimization for ``std::for_each`` has been backported to C++11. Previously it was only available
   in C++23 and later.
 
-- The ``std::for_each_n`` algorithm has been optimized for segmented iterators, resulting in a performance improvement of
-  up to 17.7x for ``std::deque<short>`` iterators, and up to 13.9x for ``std::join_view<vector<vector<short>>>`` iterators.
+- The ``std::for_each_n``, ``std::ranges::for_each`` and ``std::ranges::for_each_n`` algorithms have been optimized for
+  segmented iterators, resulting in a performance improvement of up to 17.7x for ``std::deque<short>`` iterators, and up
+  to 13.9x for ``std::join_view<vector<vector<short>>>`` iterators.
 
 - The ``bitset::to_string`` function has been optimized, resulting in a performance improvement of up to 8.3x for bitsets
   with uniformly distributed zeros and ones, and up to 13.5x and 16.1x for sparse and dense bitsets, respectively.

libcxx/include/__algorithm/for_each.h

@@ -12,41 +12,54 @@
 
 #include <__algorithm/for_each_segment.h>
 #include <__config>
+#include <__functional/identity.h>
 #include <__iterator/segmented_iterator.h>
 #include <__type_traits/enable_if.h>
+#include <__type_traits/invoke.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 _InputIterator, class _Sent, class _Func>
-_LIBCPP_HIDE_FROM_ABI _LIBCPP_CONSTEXPR_SINCE_CXX20 void __for_each(_InputIterator __first, _Sent __last, _Func& __f) {
+template <class _InputIterator, class _Sent, class _Func, class _Proj>
+_LIBCPP_HIDE_FROM_ABI _LIBCPP_CONSTEXPR_SINCE_CXX20 _InputIterator
+__for_each(_InputIterator __first, _Sent __last, _Func& __f, _Proj& __proj) {
   for (; __first != __last; ++__first)
-    __f(*__first);
+    std::__invoke(__f, std::__invoke(__proj, *__first));
+  return __first;
 }
 
 #ifndef _LIBCPP_CXX03_LANG
 template <class _SegmentedIterator,
-          class _Function,
+          class _Func,
+          class _Proj,
           __enable_if_t<__is_segmented_iterator<_SegmentedIterator>::value, int> = 0>
-_LIBCPP_HIDE_FROM_ABI _LIBCPP_CONSTEXPR_SINCE_CXX20 void
-__for_each(_SegmentedIterator __first, _SegmentedIterator __last, _Function& __func) {
+_LIBCPP_HIDE_FROM_ABI _LIBCPP_CONSTEXPR_SINCE_CXX20 _SegmentedIterator
+__for_each(_SegmentedIterator __first, _SegmentedIterator __last, _Func& __func, _Proj& __proj) {
   using __local_iterator_t = typename __segmented_iterator_traits<_SegmentedIterator>::__local_iterator;
   std::__for_each_segment(__first, __last, [&](__local_iterator_t __lfirst, __local_iterator_t __llast) {
-    std::__for_each(__lfirst, __llast, __func);
+    std::__for_each(__lfirst, __llast, __func, __proj);
   });
+  return __last;
 }
 #endif // !_LIBCPP_CXX03_LANG
 
-template <class _InputIterator, class _Function>
-_LIBCPP_HIDE_FROM_ABI _LIBCPP_CONSTEXPR_SINCE_CXX20 _Function
-for_each(_InputIterator __first, _InputIterator __last, _Function __f) {
-  std::__for_each(__first, __last, __f);
+template <class _InputIterator, class _Func>
+_LIBCPP_HIDE_FROM_ABI _LIBCPP_CONSTEXPR_SINCE_CXX20 _Func
+for_each(_InputIterator __first, _InputIterator __last, _Func __f) {
+  __identity __proj;
+  std::__for_each(__first, __last, __f, __proj);
   return __f;
 }
 
 _LIBCPP_END_NAMESPACE_STD
 
+_LIBCPP_POP_MACROS
+
 #endif // _LIBCPP___ALGORITHM_FOR_EACH_H

libcxx/include/__algorithm/for_each_n.h

@@ -13,10 +13,12 @@
 #include <__algorithm/for_each.h>
 #include <__algorithm/for_each_n_segment.h>
 #include <__config>
+#include <__functional/identity.h>
 #include <__iterator/iterator_traits.h>
 #include <__iterator/segmented_iterator.h>
 #include <__type_traits/disjunction.h>
 #include <__type_traits/enable_if.h>
+#include <__type_traits/invoke.h>
 #include <__type_traits/negation.h>
 #include <__utility/convert_to_integral.h>
 #include <__utility/move.h>
@@ -33,16 +35,17 @@ _LIBCPP_BEGIN_NAMESPACE_STD
 template <class _InputIterator,
           class _Size,
           class _Func,
+          class _Proj,
           __enable_if_t<!__has_random_access_iterator_category<_InputIterator>::value &&
                             _Or< _Not<__is_segmented_iterator<_InputIterator> >,
                                  _Not<__has_random_access_local_iterator<_InputIterator> > >::value,
                         int> = 0>
 _LIBCPP_HIDE_FROM_ABI _LIBCPP_CONSTEXPR_SINCE_CXX20 _InputIterator
-__for_each_n(_InputIterator __first, _Size __orig_n, _Func& __f) {
+__for_each_n(_InputIterator __first, _Size __orig_n, _Func& __f, _Proj& __proj) {
   typedef decltype(std::__convert_to_integral(__orig_n)) _IntegralSize;
   _IntegralSize __n = __orig_n;
   while (__n > 0) {
-    __f(*__first);
+    std::__invoke(__f, std::__invoke(__proj, *__first));
     ++__first;
     --__n;
   }
@@ -52,39 +55,42 @@ __for_each_n(_InputIterator __first, _Size __orig_n, _Func& __f) {
 template <class _RandIter,
           class _Size,
           class _Func,
+          class _Proj,
           __enable_if_t<__has_random_access_iterator_category<_RandIter>::value, int> = 0>
 _LIBCPP_HIDE_FROM_ABI _LIBCPP_CONSTEXPR_SINCE_CXX20 _RandIter
-__for_each_n(_RandIter __first, _Size __orig_n, _Func& __f) {
+__for_each_n(_RandIter __first, _Size __orig_n, _Func& __f, _Proj& __proj) {
   typename std::iterator_traits<_RandIter>::difference_type __n = __orig_n;
   auto __last                                                   = __first + __n;
-  std::__for_each(__first, __last, __f);
-  return std::move(__last);
+  std::__for_each(__first, __last, __f, __proj);
+  return __last;
 }
 
 #ifndef _LIBCPP_CXX03_LANG
 template <class _SegmentedIterator,
           class _Size,
           class _Func,
+          class _Proj,
           __enable_if_t<!__has_random_access_iterator_category<_SegmentedIterator>::value &&
                             __is_segmented_iterator<_SegmentedIterator>::value &&
                             __has_random_access_iterator_category<
                                 typename __segmented_iterator_traits<_SegmentedIterator>::__local_iterator>::value,
                         int> = 0>
 _LIBCPP_HIDE_FROM_ABI _LIBCPP_CONSTEXPR_SINCE_CXX20 _SegmentedIterator
-__for_each_n(_SegmentedIterator __first, _Size __orig_n, _Func& __f) {
+__for_each_n(_SegmentedIterator __first, _Size __orig_n, _Func& __f, _Proj& __proj) {
   using __local_iterator_t = typename __segmented_iterator_traits<_SegmentedIterator>::__local_iterator;
   return std::__for_each_n_segment(__first, __orig_n, [&](__local_iterator_t __lfirst, __local_iterator_t __llast) {
-    std::__for_each(__lfirst, __llast, __f);
+    std::__for_each(__lfirst, __llast, __f, __proj);
   });
 }
 #endif // !_LIBCPP_CXX03_LANG
 
 #if _LIBCPP_STD_VER >= 17
 
-template <class _InputIterator, class _Size, class _Function>
+template <class _InputIterator, class _Size, class _Func>
 inline _LIBCPP_HIDE_FROM_ABI _LIBCPP_CONSTEXPR_SINCE_CXX20 _InputIterator
-for_each_n(_InputIterator __first, _Size __orig_n, _Function __f) {
-  return std::__for_each_n(__first, __orig_n, __f);
+for_each_n(_InputIterator __first, _Size __orig_n, _Func __f) {
+  __identity __proj;
+  return std::__for_each_n(__first, __orig_n, __f, __proj);
 }
 
 #endif // _LIBCPP_STD_VER >= 17

libcxx/include/__algorithm/ranges_for_each.h

@@ -9,10 +9,12 @@
 #ifndef _LIBCPP___ALGORITHM_RANGES_FOR_EACH_H
 #define _LIBCPP___ALGORITHM_RANGES_FOR_EACH_H
 
+#include <__algorithm/for_each.h>
+#include <__algorithm/for_each_n.h>
 #include <__algorithm/in_fun_result.h>
+#include <__concepts/assignable.h>
 #include <__config>
 #include <__functional/identity.h>
-#include <__functional/invoke.h>
 #include <__iterator/concepts.h>
 #include <__iterator/projected.h>
 #include <__ranges/access.h>
@@ -41,9 +43,17 @@ struct __for_each {
   template <class _Iter, class _Sent, class _Proj, class _Func>
   _LIBCPP_HIDE_FROM_ABI constexpr static for_each_result<_Iter, _Func>
   __for_each_impl(_Iter __first, _Sent __last, _Func& __func, _Proj& __proj) {
-    for (; __first != __last; ++__first)
-      std::invoke(__func, std::invoke(__proj, *__first));
-    return {std::move(__first), std::move(__func)};
+    // In the case where we have different iterator and sentinel types, the segmented iterator optimization
+    // in std::for_each will not kick in. Therefore, we prefer std::for_each_n in that case (whenever we can
+    // obtain the `n`).
+    if constexpr (!std::assignable_from<_Iter&, _Sent> && std::sized_sentinel_for<_Sent, _Iter>) {
+      auto __n   = __last - __first;
+      auto __end = std::__for_each_n(std::move(__first), __n, __func, __proj);
+      return {std::move(__end), std::move(__func)};
+    } else {
+      auto __end = std::__for_each(std::move(__first), std::move(__last), __func, __proj);
+      return {std::move(__end), std::move(__func)};
+    }
   }
 
 public:

libcxx/include/__algorithm/ranges_for_each_n.h

@@ -9,10 +9,10 @@
 #ifndef _LIBCPP___ALGORITHM_RANGES_FOR_EACH_N_H
 #define _LIBCPP___ALGORITHM_RANGES_FOR_EACH_N_H
 
+#include <__algorithm/for_each_n.h>
 #include <__algorithm/in_fun_result.h>
 #include <__config>
 #include <__functional/identity.h>
-#include <__functional/invoke.h>
 #include <__iterator/concepts.h>
 #include <__iterator/incrementable_traits.h>
 #include <__iterator/iterator_traits.h>
@@ -40,11 +40,8 @@ struct __for_each_n {
   template <input_iterator _Iter, class _Proj = identity, indirectly_unary_invocable<projected<_Iter, _Proj>> _Func>
   _LIBCPP_HIDE_FROM_ABI constexpr for_each_n_result<_Iter, _Func>
   operator()(_Iter __first, iter_difference_t<_Iter> __count, _Func __func, _Proj __proj = {}) const {
-    while (__count-- > 0) {
-      std::invoke(__func, std::invoke(__proj, *__first));
-      ++__first;
-    }
-    return {std::move(__first), std::move(__func)};
+    auto __last = std::__for_each_n(std::move(__first), __count, __func, __proj);
+    return {std::move(__last), std::move(__func)};
   }
 };
 

libcxx/include/experimental/iterator

@@ -127,6 +127,7 @@ _LIBCPP_POP_MACROS
 #  if !defined(_LIBCPP_REMOVE_TRANSITIVE_INCLUDES) && _LIBCPP_STD_VER <= 20
 #    include <cstddef>
 #    include <iosfwd>
+#    include <optional>
 #    include <type_traits>
 #  endif
 #endif // __cplusplus < 201103L && defined(_LIBCPP_USE_FROZEN_CXX03_HEADERS)

libcxx/include/mutex

@@ -504,6 +504,7 @@ _LIBCPP_POP_MACROS
 #    include <initializer_list>
 #    include <iosfwd>
 #    include <new>
+#    include <optional>
 #    include <stdexcept>
 #    include <system_error>
 #    include <type_traits>

libcxx/include/shared_mutex

@@ -457,6 +457,7 @@ _LIBCPP_POP_MACROS
 #  endif // _LIBCPP_HAS_THREADS
 
 #  if !defined(_LIBCPP_REMOVE_TRANSITIVE_INCLUDES) && _LIBCPP_STD_VER <= 20
+#    include <optional>
 #    include <system_error>
 #  endif
 #endif // __cplusplus < 201103L && defined(_LIBCPP_USE_FROZEN_CXX03_HEADERS)

libcxx/test/benchmarks/algorithms/nonmodifying/for_each.bench.cpp

@@ -12,6 +12,7 @@
 #include <cstddef>
 #include <deque>
 #include <list>
+#include <ranges>
 #include <string>
 #include <vector>
 
@@ -23,6 +24,7 @@ int main(int argc, char** argv) {
   // {std,ranges}::for_each
   {
     auto bm = []<class Container>(std::string name, auto for_each) {
+      using ElemType = typename Container::value_type;
       benchmark::RegisterBenchmark(
           name,
           [for_each](auto& st) {
@@ -33,15 +35,14 @@ int main(int argc, char** argv) {
 
             for ([[maybe_unused]] auto _ : st) {
               benchmark::DoNotOptimize(c);
-              auto result = for_each(first, last, [](int& x) { x = std::clamp(x, 10, 100); });
+              auto result = for_each(first, last, [](ElemType& x) { x = std::clamp<ElemType>(x, 10, 100); });
               benchmark::DoNotOptimize(result);
             }
           })
           ->Arg(8)
           ->Arg(32)
           ->Arg(50) // non power-of-two
-          ->Arg(8192)
-          ->Arg(1 << 20);
+          ->Arg(8192);
     };
     bm.operator()<std::vector<int>>("std::for_each(vector<int>)", std_for_each);
     bm.operator()<std::deque<int>>("std::for_each(deque<int>)", std_for_each);
@@ -51,6 +52,42 @@ int main(int argc, char** argv) {
     bm.operator()<std::list<int>>("rng::for_each(list<int>)", std::ranges::for_each);
   }
 
+  // {std,ranges}::for_each for join_view
+  {
+    auto bm = []<class Container>(std::string name, auto for_each) {
+      using C1       = typename Container::value_type;
+      using ElemType = typename C1::value_type;
+
+      benchmark::RegisterBenchmark(
+          name,
+          [for_each](auto& st) {
+            std::size_t const size     = st.range(0);
+            std::size_t const seg_size = 256;
+            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), ElemType(1));
+            }
+
+            auto view  = c | std::views::join;
+            auto first = view.begin();
+            auto last  = view.end();
+
+            for ([[maybe_unused]] auto _ : st) {
+              benchmark::DoNotOptimize(c);
+              auto result = for_each(first, last, [](ElemType& x) { x = std::clamp<ElemType>(x, 10, 100); });
+              benchmark::DoNotOptimize(result);
+            }
+          })
+          ->Arg(8)
+          ->Arg(32)
+          ->Arg(50) // non power-of-two
+          ->Arg(8192);
+    };
+    bm.operator()<std::vector<std::vector<int>>>("std::for_each(join_view(vector<vector<int>>))", std_for_each);
+    bm.operator()<std::vector<std::vector<int>>>("rng::for_each(join_view(vector<vector<int>>)", std::ranges::for_each);
+  }
+
   benchmark::Initialize(&argc, argv);
   benchmark::RunSpecifiedBenchmarks();
   benchmark::Shutdown();

libcxx/test/benchmarks/algorithms/nonmodifying/for_each_n.bench.cpp

@@ -21,7 +21,7 @@
 int main(int argc, char** argv) {
   auto std_for_each_n = [](auto first, auto n, auto f) { return std::for_each_n(first, n, f); };
 
-  // std::for_each_n
+  // {std,ranges}::for_each_n
   {
     auto bm = []<class Container>(std::string name, auto for_each_n) {
       using ElemType = typename Container::value_type;
@@ -41,19 +41,17 @@ int main(int argc, char** argv) {
           ->Arg(8)
           ->Arg(32)
           ->Arg(50) // non power-of-two
-          ->Arg(1024)
-          ->Arg(4096)
-          ->Arg(8192)
-          ->Arg(1 << 14)
-          ->Arg(1 << 16)
-          ->Arg(1 << 18);
+          ->Arg(8192);
     };
     bm.operator()<std::vector<int>>("std::for_each_n(vector<int>)", std_for_each_n);
     bm.operator()<std::deque<int>>("std::for_each_n(deque<int>)", std_for_each_n);
     bm.operator()<std::list<int>>("std::for_each_n(list<int>)", std_for_each_n);
+    bm.operator()<std::vector<int>>("rng::for_each_n(vector<int>)", std::ranges::for_each_n);
+    bm.operator()<std::deque<int>>("rng::for_each_n(deque<int>)", std::ranges::for_each_n);
+    bm.operator()<std::list<int>>("rng::for_each_n(list<int>)", std::ranges::for_each_n);
   }
 
-  // std::for_each_n for join_view
+  // {std,ranges}::for_each_n for join_view
   {
     auto bm = []<class Container>(std::string name, auto for_each_n) {
       using C1       = typename Container::value_type;
@@ -81,14 +79,11 @@ int main(int argc, char** argv) {
           ->Arg(8)
           ->Arg(32)
           ->Arg(50) // non power-of-two
-          ->Arg(1024)
-          ->Arg(4096)
-          ->Arg(8192)
-          ->Arg(1 << 14)
-          ->Arg(1 << 16)
-          ->Arg(1 << 18);
+          ->Arg(8192);
     };
     bm.operator()<std::vector<std::vector<int>>>("std::for_each_n(join_view(vector<vector<int>>))", std_for_each_n);
+    bm.operator()<std::vector<std::vector<int>>>(
+        "rng::for_each_n(join_view(vector<vector<int>>)", std::ranges::for_each_n);
   }
 
   benchmark::Initialize(&argc, argv);