microsoft · StephanTLavavej · Aug 5, 2020 · Jul 27, 2020 · Jul 28, 2020 · Jul 29, 2020
diff --git a/stl/inc/algorithm b/stl/inc/algorithm
@@ -4927,7 +4927,7 @@ _CONSTEXPR20 _OutIt reverse_copy(_BidIt _First, _BidIt _Last, _OutIt _Dest) {
 #if _HAS_IF_CONSTEXPR && _USE_STD_VECTOR_ALGORITHMS
     using _Elem                         = remove_pointer_t<decltype(_UFirst)>;
     using _DestElem                     = remove_pointer_t<decltype(_UDest)>;
-    constexpr bool _Allow_vectorization = conjunction_v<is_same<remove_const_t<_Elem>, remove_const_t<_DestElem>>,
+    constexpr bool _Allow_vectorization = conjunction_v<is_same<remove_const_t<_Elem>, _DestElem>,
         is_pointer<decltype(_UFirst)>, is_trivially_copyable<_Elem>, negation<is_volatile<_Elem>>>;
     constexpr size_t _Nx                = sizeof(_Elem);
 
@@ -4970,6 +4970,95 @@ _FwdIt reverse_copy(_ExPo&&, _BidIt _First, _BidIt _Last, _FwdIt _Dest) noexcept
     _REQUIRE_PARALLEL_ITERATOR(_FwdIt);
     return _STD reverse_copy(_First, _Last, _Dest);
 }
+
+#ifdef __cpp_lib_concepts
+namespace ranges {
+    // ALIAS TEMPLATE reverse_copy_result
+    template <class _In, class _Out>
+    using reverse_copy_result = in_out_result<_In, _Out>;
+
+    // VARIABLE ranges::reverse_copy
+    class _Reverse_copy_fn : private _Not_quite_object {
+    public:
+        using _Not_quite_object::_Not_quite_object;
+
+        // clang-format off
+        template <bidirectional_iterator _It, sentinel_for<_It> _Se, weakly_incrementable _Out>
+            requires indirectly_copyable<_It, _Out>
+        constexpr reverse_copy_result<_It, _Out> operator()(_It _First, _Se _Last, _Out _Result) const {
+            // clang-format on
+            _Adl_verify_range(_First, _Last);
+            auto _UFirst = _Get_unwrapped(_STD move(_First));
+            auto _ULast  = _Get_final_iterator_unwrapped<_It>(_UFirst, _STD move(_Last));
+            _Seek_wrapped(_First, _ULast);
+            _Result = _Reverse_copy_common(_STD move(_UFirst), _STD move(_ULast), _STD move(_Result));
+            return {_STD move(_First), _STD move(_Result)};
+        }
+
+        // clang-format off
+        template <bidirectional_range _Rng, weakly_incrementable _Out>
+            requires indirectly_copyable<iterator_t<_Rng>, _Out>
+        constexpr reverse_copy_result<borrowed_iterator_t<_Rng>, _Out> operator()(_Rng&& _Range, _Out _Result) const {
+            // clang-format on
+            if constexpr (common_range<_Rng>) {
+                _Result = _Reverse_copy_common(_Ubegin(_Range), _Uend(_Range), _STD move(_Result));
+                return {_RANGES end(_Range), _STD move(_Result)};
+            } else {
+                auto _ULast = _Get_final_iterator_unwrapped(_Range);
+                _Result     = _Reverse_copy_common(_Ubegin(_Range), _ULast, _STD move(_Result));
+                return {_Rewrap_iterator(_Range, _STD move(_ULast)), _STD move(_Result)};
+            }
+        }
+
+    private:
+        template <class _It, class _Out>
+        _NODISCARD static constexpr _Out _Reverse_copy_common(const _It _First, _It _Last, _Out _Result) {
+            _STL_INTERNAL_STATIC_ASSERT(bidirectional_iterator<_It>);
+            _STL_INTERNAL_STATIC_ASSERT(weakly_incrementable<_Out>);
+            _STL_INTERNAL_STATIC_ASSERT(indirectly_copyable<_It, _Out>);
+
+#if _USE_STD_VECTOR_ALGORITHMS
+            if constexpr (contiguous_iterator<_It> && contiguous_iterator<_Out>) {
+                using _Elem                         = remove_reference_t<iter_reference_t<_It>>;
+                using _DestElem                     = remove_reference_t<iter_reference_t<_Out>>;
+                constexpr bool _Allow_vectorization = conjunction_v<is_same<remove_const_t<_Elem>, _DestElem>,
+                    is_trivially_copyable<_Elem>, negation<is_volatile<_Elem>>>;
+                constexpr size_t _Nx                = sizeof(_Elem);
+
+#pragma warning(suppress : 6326) // Potential comparison of a constant with another constant
+                if constexpr (_Allow_vectorization && _Nx <= 8 && (_Nx & (_Nx - 1)) == 0) {
+                    if (!_STD is_constant_evaluated()) {
+                        _Elem* const _First_addr      = _STD to_address(_First);
+                        _Elem* const _Last_addr       = _STD to_address(_Last);
+                        _DestElem* const _Result_addr = _STD to_address(_Result);
+                        if constexpr (_Nx == 1) {
+                            __std_reverse_copy_trivially_copyable_1(_First_addr, _Last_addr, _Result_addr);
+                        } else if constexpr (_Nx == 2) {
+                            __std_reverse_copy_trivially_copyable_2(_First_addr, _Last_addr, _Result_addr);
+                        } else if constexpr (_Nx == 4) {
+                            __std_reverse_copy_trivially_copyable_4(_First_addr, _Last_addr, _Result_addr);
+                        } else {
+                            __std_reverse_copy_trivially_copyable_8(_First_addr, _Last_addr, _Result_addr);
+                        }
+
+                        _Result += _Last - _First;
+                        return _Result;
+                    }
+                }
+            }
+#endif // _USE_STD_VECTOR_ALGORITHMS
+
+            for (; _First != _Last; ++_Result) {
+                *_Result = *--_Last;
+            }
+
+            return _Result;
+        }
+    };
+
+    inline constexpr _Reverse_copy_fn reverse_copy{_Not_quite_object::_Construct_tag{}};
+} // namespace ranges
+#endif // __cpp_lib_concepts
 #endif // _HAS_CXX17
 
 #ifdef __cpp_lib_concepts

diff --git a/tests/std/include/range_algorithm_support.hpp b/tests/std/include/range_algorithm_support.hpp
@@ -1063,6 +1063,16 @@ constexpr void test_fwd_write() {
     with_forward_ranges<with_writable_iterators<Instantiator, Element2>, Element1>::call();
 }
 
+template <class Instantiator, class Element1, class Element2>
+constexpr void test_bidi_write() {
+    with_bidirectional_ranges<with_writable_iterators<Instantiator, Element2>, Element1>::call();
+}
+
+template <class Instantiator, class Element1, class Element2>
+constexpr void test_contiguous_write() {
+    with_contiguous_ranges<with_writable_iterators<Instantiator, Element2>, Element1>::call();
+}
+
 template <class Instantiator, class Element>
 constexpr void test_read() {
     with_input_iterators<Instantiator, Element>::call();

diff --git a/tests/std/test.lst b/tests/std/test.lst
@@ -238,7 +238,6 @@ tests\P0769R2_shift_left_shift_right
 tests\P0784R7_library_support_for_more_constexpr_containers
 tests\P0811R3_midpoint_lerp
 tests\P0896R4_P1614R2_comparisons
-tests\P0896R4_ranges_algorithm_machinery
 tests\P0896R4_ranges_alg_adjacent_find
 tests\P0896R4_ranges_alg_all_of
 tests\P0896R4_ranges_alg_any_of
@@ -282,6 +281,7 @@ tests\P0896R4_ranges_alg_replace_copy
 tests\P0896R4_ranges_alg_replace_copy_if
 tests\P0896R4_ranges_alg_replace_if
 tests\P0896R4_ranges_alg_reverse
+tests\P0896R4_ranges_alg_reverse_copy
 tests\P0896R4_ranges_alg_rotate
 tests\P0896R4_ranges_alg_rotate_copy
 tests\P0896R4_ranges_alg_sample
@@ -297,6 +297,7 @@ tests\P0896R4_ranges_alg_transform_binary
 tests\P0896R4_ranges_alg_transform_unary
 tests\P0896R4_ranges_alg_unique
 tests\P0896R4_ranges_alg_unique_copy
+tests\P0896R4_ranges_algorithm_machinery
 tests\P0896R4_ranges_iterator_machinery
 tests\P0896R4_ranges_range_machinery
 tests\P0896R4_ranges_subrange

diff --git a/tests/std/tests/P0896R4_ranges_alg_reverse_copy/env.lst b/tests/std/tests/P0896R4_ranges_alg_reverse_copy/env.lst
@@ -0,0 +1,4 @@
+# Copyright (c) Microsoft Corporation.
+# SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+
+RUNALL_INCLUDE ..\concepts_matrix.lst
diff --git a/tests/std/tests/P0896R4_ranges_alg_reverse_copy/test.cpp b/tests/std/tests/P0896R4_ranges_alg_reverse_copy/test.cpp
@@ -0,0 +1,163 @@
+// Copyright (c) Microsoft Corporation.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+
+#include <algorithm>
+#include <cassert>
+#include <compare>
+#include <concepts>
+#include <numeric>
+#include <ranges>
+#include <utility>
+
+#include <range_algorithm_support.hpp>
+
+using namespace std;
+
+// Validate that reverse_copy_result aliases in_out_result
+STATIC_ASSERT(same_as<ranges::reverse_copy_result<int, double>, ranges::in_out_result<int, double>>);
+
+// Validate dangling story
+STATIC_ASSERT(same_as<decltype(ranges::reverse_copy(borrowed<false>{}, nullptr_to<int>)),
+    ranges::reverse_copy_result<ranges::dangling, int*>>);
+STATIC_ASSERT(same_as<decltype(ranges::reverse_copy(borrowed<true>{}, nullptr_to<int>)),
+    ranges::reverse_copy_result<int*, int*>>);
+
+struct nontrivial_int {
+    int val;
+
+    constexpr nontrivial_int(int i = 0) noexcept : val{i} {}
+    constexpr nontrivial_int(const nontrivial_int& that) noexcept : val{that.val} {}
+    constexpr nontrivial_int& operator=(const nontrivial_int& that) noexcept {
+        val = that.val;
+        return *this;
+    }
+
+    auto operator<=>(const nontrivial_int&) const = default;
+};
+
+struct instantiator {
+    static constexpr nontrivial_int input[]    = {13, 42, 1367};
+    static constexpr nontrivial_int expected[] = {1367, 42, 13};
+
+    template <ranges::bidirectional_range In, indirectly_writable<ranges::range_reference_t<In>> Out>
+    static constexpr void call() {
+        if constexpr (!ranges::contiguous_range<In>) { // the vectorized tests below have plenty of contiguous coverage
+            using ranges::reverse_copy, ranges::reverse_copy_result, ranges::begin, ranges::end, ranges::equal,
+                ranges::iterator_t;
+
+            { // Validate iterator overload
+                nontrivial_int output[3];
+                In wrapped_input{input};
+                const same_as<reverse_copy_result<iterator_t<In>, Out>> auto result =
+                    reverse_copy(wrapped_input.begin(), wrapped_input.end(), Out{output});
+                assert(result.in == wrapped_input.end());
+                assert(result.out.peek() == end(output));
+                assert(equal(output, expected));
+            }
+            { // Validate range overload
+                nontrivial_int output[3];
+                In wrapped_input{input};
+                const same_as<reverse_copy_result<iterator_t<In>, Out>> auto result =
+                    reverse_copy(wrapped_input, Out{output});
+                assert(result.in == wrapped_input.end());
+                assert(result.out.peek() == end(output));
+                assert(equal(output, expected));
+            }
+
+            { // Validate iterator overload, empty range
+                nontrivial_int output[3];
+                In wrapped_input{};
+                const same_as<reverse_copy_result<iterator_t<In>, Out>> auto result =
+                    reverse_copy(wrapped_input.begin(), wrapped_input.end(), Out{output});
+                assert(result.in == wrapped_input.end());
+                assert(result.out.peek() == begin(output));
+            }
+            { // Validate range overload, empty range
+                nontrivial_int output[3];
+                In wrapped_input{};
+                const same_as<reverse_copy_result<iterator_t<In>, Out>> auto result =
+                    reverse_copy(wrapped_input, Out{output});
+                assert(result.in == wrapped_input.end());
+                assert(result.out.peek() == begin(output));
+            }
+        }
+    }
+};
+
+template <size_t N>
+struct bytes {
+    unsigned char storage[N];
+
+    constexpr bytes() {
+        ranges::fill(storage, static_cast<unsigned char>(-1));
+    }
+
+    constexpr bytes(unsigned char base) {
+        iota(storage, storage + N, base);
+    }
+
+    bool operator==(const bytes&) const = default;
+};
+
+struct test_vector {
+    template <ranges::contiguous_range In, indirectly_writable<ranges::range_reference_t<In>> Out>
+    static constexpr void call() {
+        using ranges::reverse_copy, ranges::reverse_copy_result, ranges::begin, ranges::end, ranges::equal,
+            ranges::iterator_t, ranges::range_value_t;
+
+        const range_value_t<In> input[3]    = {0x10, 0x20, 0x30};
+        const range_value_t<In> expected[3] = {0x30, 0x20, 0x10};
+
+        { // Validate iterator overload, vectorizable
+            range_value_t<In> output[3];
+            In wrapped_input{input};
+            const same_as<reverse_copy_result<iterator_t<In>, Out>> auto result =
+                reverse_copy(wrapped_input.begin(), wrapped_input.end(), Out{output});
+            assert(result.in == wrapped_input.end());
+            assert(result.out.peek() == end(output));
+            assert(equal(output, expected));
+        }
+        { // Validate range overload, vectorizable
+            range_value_t<In> output[3];
+            In wrapped_input{input};
+            const same_as<reverse_copy_result<iterator_t<In>, Out>> auto result =
+                reverse_copy(wrapped_input, Out{output});
+            assert(result.in == wrapped_input.end());
+            assert(result.out.peek() == end(output));
+            assert(equal(output, expected));
+        }
+
+        { // Validate iterator overload, vectorizable empty
+            range_value_t<In> output[3];
+            In wrapped_input{};
+            const same_as<reverse_copy_result<iterator_t<In>, Out>> auto result =
+                reverse_copy(wrapped_input.begin(), wrapped_input.end(), Out{output});
+            assert(result.in == wrapped_input.end());
+            assert(result.out.peek() == begin(output));
+        }
+        { // Validate range overload, vectorizable empty
+            range_value_t<In> output[3];
+            In wrapped_input{};
+            const same_as<reverse_copy_result<iterator_t<In>, Out>> auto result =
+                reverse_copy(wrapped_input, Out{output});
+            assert(result.in == wrapped_input.end());
+            assert(result.out.peek() == begin(output));
+        }
+    }
+};
+
+int main() {
+    STATIC_ASSERT((test_bidi_write<instantiator, const nontrivial_int, nontrivial_int>(), true));
+    test_bidi_write<instantiator, const nontrivial_int, nontrivial_int>();
+
+    STATIC_ASSERT((test_contiguous_write<test_vector, const bytes<1>, bytes<1>>(), true));
+    STATIC_ASSERT((test_contiguous_write<test_vector, const bytes<2>, bytes<2>>(), true));
+    STATIC_ASSERT((test_contiguous_write<test_vector, const bytes<4>, bytes<4>>(), true));
+    STATIC_ASSERT((test_contiguous_write<test_vector, const bytes<8>, bytes<8>>(), true));
+    STATIC_ASSERT((test_contiguous_write<test_vector, const bytes<3>, bytes<3>>(), true));
+    test_contiguous_write<test_vector, const bytes<1>, bytes<1>>();
+    test_contiguous_write<test_vector, const bytes<2>, bytes<2>>();
+    test_contiguous_write<test_vector, const bytes<4>, bytes<4>>();
+    test_contiguous_write<test_vector, const bytes<8>, bytes<8>>();
+    test_contiguous_write<test_vector, const bytes<3>, bytes<3>>();
+}