[libc++] Optimize ranges::copy_backward for vector<bool>::iterator (#121026)

winner245 · web-flow · commit edc3dc6abd9d · 2025-01-30T14:55:05.000-05:00
As a follow-up to #121013 (which focused on `std::ranges::copy`), this PR optimizes the performance of `std::ranges::copy_backward` for `vector<bool>::iterator`, addressing a subtask outlined in issue #64038. The optimizations yield performance improvements of up to 2000x for aligned copies and 60x for unaligned copies.
diff --git a/libcxx/docs/ReleaseNotes/21.rst b/libcxx/docs/ReleaseNotes/21.rst
@@ -44,7 +44,7 @@ Implemented Papers
 Improvements and New Features
 -----------------------------
 
-- The ``std::ranges::copy`` and ``std::ranges::copy_n`` algorithms have been optimized for ``std::vector<bool>::iterator``\s,
+- The ``std::ranges::{copy, copy_n, copy_backward}`` algorithms have been optimized for ``std::vector<bool>::iterator``\s,
   resulting in a performance improvement of up to 2000x.
 
 
diff --git a/libcxx/include/__algorithm/copy_backward.h b/libcxx/include/__algorithm/copy_backward.h
@@ -10,11 +10,14 @@
 #define _LIBCPP___ALGORITHM_COPY_BACKWARD_H
 
 #include <__algorithm/copy_move_common.h>
+#include <__algorithm/copy_n.h>
 #include <__algorithm/iterator_operations.h>
 #include <__algorithm/min.h>
 #include <__config>
+#include <__fwd/bit_reference.h>
 #include <__iterator/iterator_traits.h>
 #include <__iterator/segmented_iterator.h>
+#include <__memory/pointer_traits.h>
 #include <__type_traits/common_type.h>
 #include <__type_traits/enable_if.h>
 #include <__type_traits/is_constructible.h>
@@ -34,6 +37,124 @@ template <class _AlgPolicy, class _InIter, class _Sent, class _OutIter>
 _LIBCPP_HIDE_FROM_ABI _LIBCPP_CONSTEXPR_SINCE_CXX20 pair<_InIter, _OutIter>
 __copy_backward(_InIter __first, _Sent __last, _OutIter __result);
 
+template <class _Cp, bool _IsConst>
+_LIBCPP_CONSTEXPR_SINCE_CXX20 _LIBCPP_HIDE_FROM_ABI __bit_iterator<_Cp, false> __copy_backward_aligned(
+    __bit_iterator<_Cp, _IsConst> __first, __bit_iterator<_Cp, _IsConst> __last, __bit_iterator<_Cp, false> __result) {
+  using _In             = __bit_iterator<_Cp, _IsConst>;
+  using difference_type = typename _In::difference_type;
+  using __storage_type  = typename _In::__storage_type;
+
+  const int __bits_per_word = _In::__bits_per_word;
+  difference_type __n       = __last - __first;
+  if (__n > 0) {
+    // do first word
+    if (__last.__ctz_ != 0) {
+      difference_type __dn = std::min(static_cast<difference_type>(__last.__ctz_), __n);
+      __n -= __dn;
+      unsigned __clz     = __bits_per_word - __last.__ctz_;
+      __storage_type __m = (~__storage_type(0) << (__last.__ctz_ - __dn)) & (~__storage_type(0) >> __clz);
+      __storage_type __b = *__last.__seg_ & __m;
+      *__result.__seg_ &= ~__m;
+      *__result.__seg_ |= __b;
+      __result.__ctz_ = static_cast<unsigned>(((-__dn & (__bits_per_word - 1)) + __result.__ctz_) % __bits_per_word);
+      // __last.__ctz_ = 0
+    }
+    // __last.__ctz_ == 0 || __n == 0
+    // __result.__ctz_ == 0 || __n == 0
+    // do middle words
+    __storage_type __nw = __n / __bits_per_word;
+    __result.__seg_ -= __nw;
+    __last.__seg_ -= __nw;
+    std::copy_n(std::__to_address(__last.__seg_), __nw, std::__to_address(__result.__seg_));
+    __n -= __nw * __bits_per_word;
+    // do last word
+    if (__n > 0) {
+      __storage_type __m = ~__storage_type(0) << (__bits_per_word - __n);
+      __storage_type __b = *--__last.__seg_ & __m;
+      *--__result.__seg_ &= ~__m;
+      *__result.__seg_ |= __b;
+      __result.__ctz_ = static_cast<unsigned>(-__n & (__bits_per_word - 1));
+    }
+  }
+  return __result;
+}
+
+template <class _Cp, bool _IsConst>
+_LIBCPP_CONSTEXPR_SINCE_CXX20 _LIBCPP_HIDE_FROM_ABI __bit_iterator<_Cp, false> __copy_backward_unaligned(
+    __bit_iterator<_Cp, _IsConst> __first, __bit_iterator<_Cp, _IsConst> __last, __bit_iterator<_Cp, false> __result) {
+  using _In             = __bit_iterator<_Cp, _IsConst>;
+  using difference_type = typename _In::difference_type;
+  using __storage_type  = typename _In::__storage_type;
+
+  const int __bits_per_word = _In::__bits_per_word;
+  difference_type __n       = __last - __first;
+  if (__n > 0) {
+    // do first word
+    if (__last.__ctz_ != 0) {
+      difference_type __dn = std::min(static_cast<difference_type>(__last.__ctz_), __n);
+      __n -= __dn;
+      unsigned __clz_l     = __bits_per_word - __last.__ctz_;
+      __storage_type __m   = (~__storage_type(0) << (__last.__ctz_ - __dn)) & (~__storage_type(0) >> __clz_l);
+      __storage_type __b   = *__last.__seg_ & __m;
+      unsigned __clz_r     = __bits_per_word - __result.__ctz_;
+      __storage_type __ddn = std::min(__dn, static_cast<difference_type>(__result.__ctz_));
+      if (__ddn > 0) {
+        __m = (~__storage_type(0) << (__result.__ctz_ - __ddn)) & (~__storage_type(0) >> __clz_r);
+        *__result.__seg_ &= ~__m;
+        if (__result.__ctz_ > __last.__ctz_)
+          *__result.__seg_ |= __b << (__result.__ctz_ - __last.__ctz_);
+        else
+          *__result.__seg_ |= __b >> (__last.__ctz_ - __result.__ctz_);
+        __result.__ctz_ = static_cast<unsigned>(((-__ddn & (__bits_per_word - 1)) + __result.__ctz_) % __bits_per_word);
+        __dn -= __ddn;
+      }
+      if (__dn > 0) {
+        // __result.__ctz_ == 0
+        --__result.__seg_;
+        __result.__ctz_ = static_cast<unsigned>(-__dn & (__bits_per_word - 1));
+        __m             = ~__storage_type(0) << __result.__ctz_;
+        *__result.__seg_ &= ~__m;
+        __last.__ctz_ -= __dn + __ddn;
+        *__result.__seg_ |= __b << (__result.__ctz_ - __last.__ctz_);
+      }
+      // __last.__ctz_ = 0
+    }
+    // __last.__ctz_ == 0 || __n == 0
+    // __result.__ctz_ != 0 || __n == 0
+    // do middle words
+    unsigned __clz_r   = __bits_per_word - __result.__ctz_;
+    __storage_type __m = ~__storage_type(0) >> __clz_r;
+    for (; __n >= __bits_per_word; __n -= __bits_per_word) {
+      __storage_type __b = *--__last.__seg_;
+      *__result.__seg_ &= ~__m;
+      *__result.__seg_ |= __b >> __clz_r;
+      *--__result.__seg_ &= __m;
+      *__result.__seg_ |= __b << __result.__ctz_;
+    }
+    // do last word
+    if (__n > 0) {
+      __m                 = ~__storage_type(0) << (__bits_per_word - __n);
+      __storage_type __b  = *--__last.__seg_ & __m;
+      __clz_r             = __bits_per_word - __result.__ctz_;
+      __storage_type __dn = std::min(__n, static_cast<difference_type>(__result.__ctz_));
+      __m                 = (~__storage_type(0) << (__result.__ctz_ - __dn)) & (~__storage_type(0) >> __clz_r);
+      *__result.__seg_ &= ~__m;
+      *__result.__seg_ |= __b >> (__bits_per_word - __result.__ctz_);
+      __result.__ctz_ = static_cast<unsigned>(((-__dn & (__bits_per_word - 1)) + __result.__ctz_) % __bits_per_word);
+      __n -= __dn;
+      if (__n > 0) {
+        // __result.__ctz_ == 0
+        --__result.__seg_;
+        __result.__ctz_ = static_cast<unsigned>(-__n & (__bits_per_word - 1));
+        __m             = ~__storage_type(0) << __result.__ctz_;
+        *__result.__seg_ &= ~__m;
+        *__result.__seg_ |= __b << (__result.__ctz_ - (__bits_per_word - __n - __dn));
+      }
+    }
+  }
+  return __result;
+}
+
 template <class _AlgPolicy>
 struct __copy_backward_impl {
   template <class _InIter, class _Sent, class _OutIter>
@@ -107,6 +228,16 @@ struct __copy_backward_impl {
     }
   }
 
+  template <class _Cp, bool _IsConst>
+  _LIBCPP_HIDE_FROM_ABI _LIBCPP_CONSTEXPR_SINCE_CXX20 pair<__bit_iterator<_Cp, _IsConst>, __bit_iterator<_Cp, false> >
+  operator()(__bit_iterator<_Cp, _IsConst> __first,
+             __bit_iterator<_Cp, _IsConst> __last,
+             __bit_iterator<_Cp, false> __result) {
+    if (__last.__ctz_ == __result.__ctz_)
+      return std::make_pair(__last, std::__copy_backward_aligned(__first, __last, __result));
+    return std::make_pair(__last, std::__copy_backward_unaligned(__first, __last, __result));
+  }
+
   // At this point, the iterators have been unwrapped so any `contiguous_iterator` has been unwrapped to a pointer.
   template <class _In, class _Out, __enable_if_t<__can_lower_copy_assignment_to_memmove<_In, _Out>::value, int> = 0>
   _LIBCPP_HIDE_FROM_ABI _LIBCPP_CONSTEXPR_SINCE_CXX14 pair<_In*, _Out*>
diff --git a/libcxx/include/__bit_reference b/libcxx/include/__bit_reference
@@ -11,6 +11,7 @@
 #define _LIBCPP___BIT_REFERENCE
 
 #include <__algorithm/copy.h>
+#include <__algorithm/copy_backward.h>
 #include <__algorithm/copy_n.h>
 #include <__algorithm/min.h>
 #include <__bit/countr.h>
@@ -185,134 +186,6 @@ private:
         __mask_(__m) {}
 };
 
-// copy_backward
-
-template <class _Cp, bool _IsConst>
-_LIBCPP_CONSTEXPR_SINCE_CXX20 _LIBCPP_HIDE_FROM_ABI __bit_iterator<_Cp, false> __copy_backward_aligned(
-    __bit_iterator<_Cp, _IsConst> __first, __bit_iterator<_Cp, _IsConst> __last, __bit_iterator<_Cp, false> __result) {
-  using _In             = __bit_iterator<_Cp, _IsConst>;
-  using difference_type = typename _In::difference_type;
-  using __storage_type  = typename _In::__storage_type;
-
-  const int __bits_per_word = _In::__bits_per_word;
-  difference_type __n       = __last - __first;
-  if (__n > 0) {
-    // do first word
-    if (__last.__ctz_ != 0) {
-      difference_type __dn = std::min(static_cast<difference_type>(__last.__ctz_), __n);
-      __n -= __dn;
-      unsigned __clz     = __bits_per_word - __last.__ctz_;
-      __storage_type __m = (~__storage_type(0) << (__last.__ctz_ - __dn)) & (~__storage_type(0) >> __clz);
-      __storage_type __b = *__last.__seg_ & __m;
-      *__result.__seg_ &= ~__m;
-      *__result.__seg_ |= __b;
-      __result.__ctz_ = static_cast<unsigned>(((-__dn & (__bits_per_word - 1)) + __result.__ctz_) % __bits_per_word);
-      // __last.__ctz_ = 0
-    }
-    // __last.__ctz_ == 0 || __n == 0
-    // __result.__ctz_ == 0 || __n == 0
-    // do middle words
-    __storage_type __nw = __n / __bits_per_word;
-    __result.__seg_ -= __nw;
-    __last.__seg_ -= __nw;
-    std::copy_n(std::__to_address(__last.__seg_), __nw, std::__to_address(__result.__seg_));
-    __n -= __nw * __bits_per_word;
-    // do last word
-    if (__n > 0) {
-      __storage_type __m = ~__storage_type(0) << (__bits_per_word - __n);
-      __storage_type __b = *--__last.__seg_ & __m;
-      *--__result.__seg_ &= ~__m;
-      *__result.__seg_ |= __b;
-      __result.__ctz_ = static_cast<unsigned>(-__n & (__bits_per_word - 1));
-    }
-  }
-  return __result;
-}
-
-template <class _Cp, bool _IsConst>
-_LIBCPP_CONSTEXPR_SINCE_CXX20 _LIBCPP_HIDE_FROM_ABI __bit_iterator<_Cp, false> __copy_backward_unaligned(
-    __bit_iterator<_Cp, _IsConst> __first, __bit_iterator<_Cp, _IsConst> __last, __bit_iterator<_Cp, false> __result) {
-  using _In             = __bit_iterator<_Cp, _IsConst>;
-  using difference_type = typename _In::difference_type;
-  using __storage_type  = typename _In::__storage_type;
-
-  const int __bits_per_word = _In::__bits_per_word;
-  difference_type __n       = __last - __first;
-  if (__n > 0) {
-    // do first word
-    if (__last.__ctz_ != 0) {
-      difference_type __dn = std::min(static_cast<difference_type>(__last.__ctz_), __n);
-      __n -= __dn;
-      unsigned __clz_l     = __bits_per_word - __last.__ctz_;
-      __storage_type __m   = (~__storage_type(0) << (__last.__ctz_ - __dn)) & (~__storage_type(0) >> __clz_l);
-      __storage_type __b   = *__last.__seg_ & __m;
-      unsigned __clz_r     = __bits_per_word - __result.__ctz_;
-      __storage_type __ddn = std::min(__dn, static_cast<difference_type>(__result.__ctz_));
-      if (__ddn > 0) {
-        __m = (~__storage_type(0) << (__result.__ctz_ - __ddn)) & (~__storage_type(0) >> __clz_r);
-        *__result.__seg_ &= ~__m;
-        if (__result.__ctz_ > __last.__ctz_)
-          *__result.__seg_ |= __b << (__result.__ctz_ - __last.__ctz_);
-        else
-          *__result.__seg_ |= __b >> (__last.__ctz_ - __result.__ctz_);
-        __result.__ctz_ = static_cast<unsigned>(((-__ddn & (__bits_per_word - 1)) + __result.__ctz_) % __bits_per_word);
-        __dn -= __ddn;
-      }
-      if (__dn > 0) {
-        // __result.__ctz_ == 0
-        --__result.__seg_;
-        __result.__ctz_ = static_cast<unsigned>(-__dn & (__bits_per_word - 1));
-        __m             = ~__storage_type(0) << __result.__ctz_;
-        *__result.__seg_ &= ~__m;
-        __last.__ctz_ -= __dn + __ddn;
-        *__result.__seg_ |= __b << (__result.__ctz_ - __last.__ctz_);
-      }
-      // __last.__ctz_ = 0
-    }
-    // __last.__ctz_ == 0 || __n == 0
-    // __result.__ctz_ != 0 || __n == 0
-    // do middle words
-    unsigned __clz_r   = __bits_per_word - __result.__ctz_;
-    __storage_type __m = ~__storage_type(0) >> __clz_r;
-    for (; __n >= __bits_per_word; __n -= __bits_per_word) {
-      __storage_type __b = *--__last.__seg_;
-      *__result.__seg_ &= ~__m;
-      *__result.__seg_ |= __b >> __clz_r;
-      *--__result.__seg_ &= __m;
-      *__result.__seg_ |= __b << __result.__ctz_;
-    }
-    // do last word
-    if (__n > 0) {
-      __m                 = ~__storage_type(0) << (__bits_per_word - __n);
-      __storage_type __b  = *--__last.__seg_ & __m;
-      __clz_r             = __bits_per_word - __result.__ctz_;
-      __storage_type __dn = std::min(__n, static_cast<difference_type>(__result.__ctz_));
-      __m                 = (~__storage_type(0) << (__result.__ctz_ - __dn)) & (~__storage_type(0) >> __clz_r);
-      *__result.__seg_ &= ~__m;
-      *__result.__seg_ |= __b >> (__bits_per_word - __result.__ctz_);
-      __result.__ctz_ = static_cast<unsigned>(((-__dn & (__bits_per_word - 1)) + __result.__ctz_) % __bits_per_word);
-      __n -= __dn;
-      if (__n > 0) {
-        // __result.__ctz_ == 0
-        --__result.__seg_;
-        __result.__ctz_ = static_cast<unsigned>(-__n & (__bits_per_word - 1));
-        __m             = ~__storage_type(0) << __result.__ctz_;
-        *__result.__seg_ &= ~__m;
-        *__result.__seg_ |= __b << (__result.__ctz_ - (__bits_per_word - __n - __dn));
-      }
-    }
-  }
-  return __result;
-}
-
-template <class _Cp, bool _IsConst>
-inline _LIBCPP_HIDE_FROM_ABI _LIBCPP_CONSTEXPR_SINCE_CXX20 __bit_iterator<_Cp, false> copy_backward(
-    __bit_iterator<_Cp, _IsConst> __first, __bit_iterator<_Cp, _IsConst> __last, __bit_iterator<_Cp, false> __result) {
-  if (__last.__ctz_ == __result.__ctz_)
-    return std::__copy_backward_aligned(__first, __last, __result);
-  return std::__copy_backward_unaligned(__first, __last, __result);
-}
-
 // move
 
 template <class _Cp, bool _IsConst>
@@ -876,9 +749,8 @@ private:
   template <class _Dp, bool _IC>
   _LIBCPP_CONSTEXPR_SINCE_CXX20 friend __bit_iterator<_Dp, false> __copy_backward_unaligned(
       __bit_iterator<_Dp, _IC> __first, __bit_iterator<_Dp, _IC> __last, __bit_iterator<_Dp, false> __result);
-  template <class _Dp, bool _IC>
-  _LIBCPP_CONSTEXPR_SINCE_CXX20 friend __bit_iterator<_Dp, false>
-  copy_backward(__bit_iterator<_Dp, _IC> __first, __bit_iterator<_Dp, _IC> __last, __bit_iterator<_Dp, false> __result);
+  template <class _AlgPolicy>
+  friend struct __copy_backward_impl;
   template <class _Cl, class _Cr>
   friend __bit_iterator<_Cr, false>
       __swap_ranges_aligned(__bit_iterator<_Cl, false>, __bit_iterator<_Cl, false>, __bit_iterator<_Cr, false>);
diff --git a/libcxx/include/__vector/vector_bool.h b/libcxx/include/__vector/vector_bool.h
@@ -10,6 +10,7 @@
 #define _LIBCPP___VECTOR_VECTOR_BOOL_H
 
 #include <__algorithm/copy.h>
+#include <__algorithm/copy_backward.h>
 #include <__algorithm/fill_n.h>
 #include <__algorithm/iterator_operations.h>
 #include <__algorithm/max.h>
diff --git a/libcxx/include/bitset b/libcxx/include/bitset
@@ -130,6 +130,7 @@ template <size_t N> struct hash<std::bitset<N>>;
 #  include <__cxx03/bitset>
 #else
 #  include <__algorithm/copy.h>
+#  include <__algorithm/copy_backward.h>
 #  include <__algorithm/count.h>
 #  include <__algorithm/fill.h>
 #  include <__algorithm/fill_n.h>
diff --git a/libcxx/test/benchmarks/algorithms/copy_backward.bench.cpp b/libcxx/test/benchmarks/algorithms/copy_backward.bench.cpp
@@ -0,0 +1,55 @@
+//===----------------------------------------------------------------------===//
+//
+// 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, c++20
+
+#include <algorithm>
+#include <benchmark/benchmark.h>
+#include <vector>
+
+static void bm_ranges_copy_backward_vb(benchmark::State& state, bool aligned) {
+  auto n = state.range();
+  std::vector<bool> in(n, true);
+  std::vector<bool> out(aligned ? n : n + 8);
+  benchmark::DoNotOptimize(&in);
+  auto dst = aligned ? out.end() : out.end() - 4;
+  for (auto _ : state) {
+    benchmark::DoNotOptimize(std::ranges::copy_backward(in, dst));
+    benchmark::DoNotOptimize(&out);
+  }
+}
+
+static void bm_copy_backward_vb(benchmark::State& state, bool aligned) {
+  auto n = state.range();
+  std::vector<bool> in(n, true);
+  std::vector<bool> out(aligned ? n : n + 8);
+  benchmark::DoNotOptimize(&in);
+  auto beg = in.begin();
+  auto end = in.end();
+  auto dst = aligned ? out.end() : out.end() - 4;
+  for (auto _ : state) {
+    benchmark::DoNotOptimize(std::copy_backward(beg, end, dst));
+    benchmark::DoNotOptimize(&out);
+  }
+}
+
+static void bm_ranges_copy_backward_vb_aligned(benchmark::State& state) { bm_ranges_copy_backward_vb(state, true); }
+static void bm_ranges_copy_backward_vb_unaligned(benchmark::State& state) { bm_ranges_copy_backward_vb(state, false); }
+
+static void bm_copy_backward_vb_aligned(benchmark::State& state) { bm_copy_backward_vb(state, true); }
+static void bm_copy_backward_vb_unaligned(benchmark::State& state) { bm_copy_backward_vb(state, false); }
+
+// Test std::ranges::copy_backward for vector<bool>::iterator
+BENCHMARK(bm_ranges_copy_backward_vb_aligned)->Range(8, 1 << 16)->DenseRange(102400, 204800, 4096);
+BENCHMARK(bm_ranges_copy_backward_vb_unaligned)->Range(8, 1 << 20);
+
+// Test std::copy_backward for vector<bool>::iterator
+BENCHMARK(bm_copy_backward_vb_aligned)->Range(8, 1 << 20);
+BENCHMARK(bm_copy_backward_vb_unaligned)->Range(8, 1 << 20);
+
+BENCHMARK_MAIN();
diff --git a/libcxx/test/std/algorithms/alg.modifying.operations/alg.copy/copy_backward.pass.cpp b/libcxx/test/std/algorithms/alg.modifying.operations/alg.copy/copy_backward.pass.cpp
diff --git a/libcxx/test/std/algorithms/alg.modifying.operations/alg.copy/ranges.copy_backward.pass.cpp b/libcxx/test/std/algorithms/alg.modifying.operations/alg.copy/ranges.copy_backward.pass.cpp
