Fix uninitialized_meow memcpy optimization

stl/inc/memory

@@ -149,7 +149,9 @@ namespace ranges {
             auto _OFirst      = _Get_unwrapped(_STD move(_First2));
             const auto _OLast = _Get_unwrapped(_STD move(_Last2));
             if constexpr (_Ptr_copy_cat<_It, _Out>::_Really_trivial) {
-                _OFirst = _Copy_memcpy_common(_IFirst, _IFirst + _Count, _OFirst, _OLast);
+                auto _UResult = _Copy_memcpy_common(_IFirst, _IFirst + _Count, _OFirst, _OLast);
+                _IFirst       = _UResult.in;
+                _OFirst       = _UResult.out;
             } else {
                 _Uninitialized_backout _Backout{_STD move(_OFirst)};
 
@@ -283,7 +285,9 @@ namespace ranges {
             auto _OFirst      = _Get_unwrapped(_STD move(_First2));
             const auto _OLast = _Get_unwrapped(_STD move(_Last2));
             if constexpr (_Ptr_move_cat<_It, _Out>::_Really_trivial) {
-                _OFirst = _Copy_memcpy_common(_IFirst, _IFirst + _Count, _OFirst, _OLast);
+                auto _UResult = _Copy_memcpy_common(_IFirst, _IFirst + _Count, _OFirst, _OLast);
+                _IFirst       = _UResult.in;
+                _OFirst       = _UResult.out;
             } else {
                 _Uninitialized_backout _Backout{_STD move(_OFirst)};
 

stl/inc/xmemory

@@ -1551,6 +1551,18 @@ namespace ranges {
         && _No_throw_forward_iterator<iterator_t<_Rng>>;
     // clang-format on
 
+    template <class _InIt, class _OutIt>
+    in_out_result<_InIt, _OutIt> _Copy_memcpy_common(
+        _InIt _IFirst, _InIt _ILast, _OutIt _OFirst, _OutIt _OLast) noexcept {
+        const auto _IFirst_ch = const_cast<char*>(reinterpret_cast<const volatile char*>(_IFirst));
+        const auto _ILast_ch  = const_cast<const char*>(reinterpret_cast<const volatile char*>(_ILast));
+        const auto _OFirst_ch = const_cast<char*>(reinterpret_cast<volatile char*>(_OFirst));
+        const auto _OLast_ch  = const_cast<const char*>(reinterpret_cast<const volatile char*>(_OLast));
+        const auto _Count     = static_cast<size_t>((_STD min)(_ILast_ch - _IFirst_ch, _OLast_ch - _OFirst_ch));
+        _CSTD memcpy(_OFirst_ch, _IFirst_ch, _Count);
+        return {reinterpret_cast<_InIt>(_IFirst_ch + _Count), reinterpret_cast<_OutIt>(_OFirst_ch + _Count)};
+    }
+
     // ALIAS TEMPLATE uninitialized_move_result
     template <class _In, class _Out>
     using uninitialized_move_result = in_out_result<_In, _Out>;

stl/inc/xutility

@@ -4102,17 +4102,6 @@ _OutIt _Copy_memmove(move_iterator<_InIt> _First, move_iterator<_InIt> _Last, _O
     return _Copy_memmove(_First.base(), _Last.base(), _Dest);
 }
 
-template <class _InIt, class _OutIt>
-_OutIt _Copy_memcpy_common(_InIt _IFirst, _InIt _ILast, _OutIt _OFirst, _OutIt _OLast) noexcept {
-    const auto _IFirst_ch = const_cast<const char*>(reinterpret_cast<const volatile char*>(_IFirst));
-    const auto _ILast_ch  = const_cast<const char*>(reinterpret_cast<const volatile char*>(_ILast));
-    const auto _OFirst_ch = const_cast<char*>(reinterpret_cast<volatile char*>(_OFirst));
-    const auto _OLast_ch  = const_cast<const char*>(reinterpret_cast<const volatile char*>(_OLast));
-    const auto _Count     = static_cast<size_t>((_STD min)(_ILast_ch - _IFirst_ch, _OLast_ch - _OFirst_ch));
-    _CSTD memcpy(_OFirst_ch, _IFirst_ch, _Count);
-    return reinterpret_cast<_OutIt>(_OFirst_ch + _Count);
-}
-
 // VARIABLE TEMPLATE _Is_vb_iterator
 template <class _It, bool _RequiresMutable = false>
 _INLINE_VAR constexpr bool _Is_vb_iterator = false;

tests/std/tests/P0896R4_ranges_alg_uninitialized_copy/test.cpp

@@ -76,21 +76,16 @@ struct holder {
     }
 };
 
-template <class R>
-void not_ranges_destroy(R&& r) { // TRANSITION, ranges::destroy
-    for (auto& e : r) {
-        destroy_at(&e);
-    }
-}
-
 struct instantiator {
-    static constexpr int expected_output[] = {13, 55, 12345};
-    static constexpr int expected_input[]  = {13, 55, 12345};
+    static constexpr int expected_output[]      = {13, 55, 12345};
+    static constexpr int expected_output_long[] = {13, 55, 12345, -1};
+    static constexpr int expected_input[]       = {13, 55, 12345};
+    static constexpr int expected_input_long[]  = {13, 55, 12345, 42};
 
     template <ranges::input_range R, ranges::forward_range W>
     static void call() {
-        using ranges::uninitialized_copy, ranges::uninitialized_copy_result, ranges::equal, ranges::equal_to,
-            ranges::iterator_t;
+        using ranges::destroy, ranges::uninitialized_copy, ranges::uninitialized_copy_result, ranges::equal,
+            ranges::equal_to, ranges::iterator_t;
 
         { // Validate range overload
             int_wrapper input[3] = {13, 55, 12345};
@@ -107,7 +102,7 @@ struct instantiator {
             assert(result.out == wrapped_output.end());
             assert(equal(wrapped_output, expected_output, equal_to{}, &int_wrapper::val));
             assert(equal(input, expected_input, equal_to{}, &int_wrapper::val));
-            not_ranges_destroy(wrapped_output);
+            destroy(wrapped_output);
             assert(int_wrapper::constructions == 3);
             assert(int_wrapper::destructions == 3);
         }
@@ -127,10 +122,51 @@ struct instantiator {
             assert(result.out == wrapped_output.end());
             assert(equal(wrapped_output, expected_output, equal_to{}, &int_wrapper::val));
             assert(equal(input, expected_input, equal_to{}, &int_wrapper::val));
-            not_ranges_destroy(wrapped_output);
+            destroy(wrapped_output);
             assert(int_wrapper::constructions == 3);
             assert(int_wrapper::destructions == 3);
         }
+
+        { // Validate range overload shorter output
+            int_wrapper input[4] = {13, 55, 12345, 42};
+            R wrapped_input{input};
+            holder<int_wrapper, 3> mem;
+            W wrapped_output{mem.as_span()};
+
+            int_wrapper::clear_counts();
+            same_as<uninitialized_copy_result<iterator_t<R>, iterator_t<W>>> auto result =
+                uninitialized_copy(wrapped_input, wrapped_output);
+            assert(int_wrapper::constructions == 3);
+            assert(int_wrapper::destructions == 0);
+            assert(++result.in == wrapped_input.end());
+            assert(result.out == wrapped_output.end());
+            assert(equal(wrapped_output, expected_output, equal_to{}, &int_wrapper::val));
+            assert(equal(input, expected_input_long, equal_to{}, &int_wrapper::val));
+            destroy(wrapped_output);
+            assert(int_wrapper::constructions == 3);
+            assert(int_wrapper::destructions == 3);
+        }
+
+        { // Validate range overload shorter input
+            int_wrapper input[3] = {13, 55, 12345};
+            R wrapped_input{input};
+            holder<int_wrapper, 4> mem;
+            W wrapped_output{mem.as_span()};
+
+            int_wrapper::clear_counts();
+            same_as<uninitialized_copy_result<iterator_t<R>, iterator_t<W>>> auto result =
+                uninitialized_copy(wrapped_input, wrapped_output);
+            assert(int_wrapper::constructions == 3);
+            assert(int_wrapper::destructions == 0);
+            assert(result.in == wrapped_input.end());
+            construct_at(addressof(*result.out), -1); // Need to construct non written element for comparison
+            assert(++result.out == wrapped_output.end());
+            assert(equal(wrapped_output, expected_output_long, equal_to{}, &int_wrapper::val));
+            assert(equal(input, expected_input, equal_to{}, &int_wrapper::val));
+            destroy(wrapped_output);
+            assert(int_wrapper::constructions == 4);
+            assert(int_wrapper::destructions == 4);
+        }
     }
 };
 
@@ -160,6 +196,48 @@ struct throwing_test {
     }
 };
 
+struct memcpy_test {
+    static constexpr int expected_output[]      = {13, 55, 12345};
+    static constexpr int expected_output_long[] = {13, 55, 12345, -1};
+    static constexpr int expected_input[]       = {13, 55, 12345};
+    static constexpr int expected_input_long[]  = {13, 55, 12345, 42};
+
+    static void call() {
+        { // Validate only range overload
+            int input[]  = {13, 55, 12345};
+            int output[] = {-1, -1, -1};
+
+            const auto result = ranges::uninitialized_copy(input, output);
+            assert(result.in == end(input));
+            assert(result.out == end(output));
+            assert(ranges::equal(input, expected_input));
+            assert(ranges::equal(output, expected_output));
+        }
+
+        { // Validate input shorter
+            int input[]  = {13, 55, 12345};
+            int output[] = {-1, -1, -1, -1};
+
+            auto result = ranges::uninitialized_copy(input, output);
+            assert(result.in == end(input));
+            assert(++result.out == end(output));
+            assert(ranges::equal(input, expected_input));
+            assert(ranges::equal(output, expected_output_long));
+        }
+
+        { // Validate output shorter
+            int input[]  = {13, 55, 12345, 42};
+            int output[] = {-1, -1, -1};
+
+            auto result = ranges::uninitialized_copy(input, output);
+            assert(++result.in == end(input));
+            assert(result.out == end(output));
+            assert(ranges::equal(input, expected_input_long));
+            assert(ranges::equal(output, expected_output));
+        }
+    }
+};
+
 template <test::ProxyRef IsProxy>
 using test_input  = test::range<test::input, int_wrapper, test::Sized::no, test::CanDifference::no, test::Common::no,
     test::CanCompare::yes, IsProxy>;
@@ -174,4 +252,5 @@ int main() {
     instantiator::call<test_input<test::ProxyRef::yes>, test_output>();
     throwing_test::call<test_input<test::ProxyRef::no>, test_output>();
     throwing_test::call<test_input<test::ProxyRef::yes>, test_output>();
+    memcpy_test::call();
 }

tests/std/tests/P0896R4_ranges_alg_uninitialized_copy_n/test.cpp

@@ -9,6 +9,7 @@
 #include <ranges>
 #include <span>
 #include <utility>
+#include <vector>
 
 #include <range_algorithm_support.hpp>
 
@@ -66,39 +67,77 @@ struct holder {
     }
 };
 
-template <class R>
-void not_ranges_destroy(R&& r) { // TRANSITION, ranges::destroy
-    for (auto& e : r) {
-        destroy_at(&e);
-    }
-}
-
 struct instantiator {
-    static constexpr int expected_output[] = {13, 55, 12345};
-    static constexpr int expected_input[]  = {13, 55, 12345};
+    static constexpr int expected_output[]      = {13, 55, 12345};
+    static constexpr int expected_output_long[] = {13, 55, 12345, -1};
+    static constexpr int expected_input[]       = {13, 55, 12345};
+    static constexpr int expected_input_long[]  = {13, 55, 12345, 42};
 
     template <ranges::input_range Read, ranges::forward_range Write>
     static void call() {
-        using ranges::uninitialized_copy_n, ranges::uninitialized_copy_n_result, ranges::equal, ranges::equal_to,
-            ranges::iterator_t;
+        using ranges::destroy, ranges::uninitialized_copy_n, ranges::uninitialized_copy_n_result, ranges::equal,
+            ranges::equal_to, ranges::iterator_t;
+
+        { // Validate equal ranges
+            int_wrapper input[3] = {13, 55, 12345};
+            Read wrapped_input{input};
+            holder<int_wrapper, 3> mem;
+            Write wrapped_output{mem.as_span()};
+
+            int_wrapper::clear_counts();
+            const same_as<uninitialized_copy_n_result<iterator_t<Read>, iterator_t<Write>>> auto result =
+                uninitialized_copy_n(wrapped_input.begin(), 3, wrapped_output.begin(), wrapped_output.end());
+            assert(int_wrapper::constructions == 3);
+            assert(int_wrapper::destructions == 0);
+            assert(result.in == wrapped_input.end());
+            assert(result.out == wrapped_output.end());
+            assert(equal(wrapped_output, expected_output, equal_to{}, &int_wrapper::val));
+            assert(equal(input, expected_input, equal_to{}, &int_wrapper::val));
+            destroy(wrapped_output);
+            assert(int_wrapper::constructions == 3);
+            assert(int_wrapper::destructions == 3);
+        }
 
-        int_wrapper input[3] = {13, 55, 12345};
-        Read wrapped_input{input};
-        holder<int_wrapper, 3> mem;
-        Write wrapped_output{mem.as_span()};
+        { // Validate shorter output
+            int_wrapper input[4] = {13, 55, 12345, 42};
+            Read wrapped_input{input};
+            holder<int_wrapper, 3> mem;
+            Write wrapped_output{mem.as_span()};
+
+            int_wrapper::clear_counts();
+            same_as<uninitialized_copy_n_result<iterator_t<Read>, iterator_t<Write>>> auto result =
+                uninitialized_copy_n(wrapped_input.begin(), 3, wrapped_output.begin(), wrapped_output.end());
+            assert(int_wrapper::constructions == 3);
+            assert(int_wrapper::destructions == 0);
+            assert(++result.in == wrapped_input.end());
+            assert(result.out == wrapped_output.end());
+            assert(equal(wrapped_output, expected_output, equal_to{}, &int_wrapper::val));
+            assert(equal(input, expected_input_long, equal_to{}, &int_wrapper::val));
+            destroy(wrapped_output);
+            assert(int_wrapper::constructions == 3);
+            assert(int_wrapper::destructions == 3);
+        }
 
-        int_wrapper::clear_counts();
-        const same_as<uninitialized_copy_n_result<iterator_t<Read>, iterator_t<Write>>> auto result =
-            uninitialized_copy_n(wrapped_input.begin(), 3, wrapped_output.begin(), wrapped_output.end());
-        assert(int_wrapper::constructions == 3);
-        assert(int_wrapper::destructions == 0);
-        assert(result.in == wrapped_input.end());
-        assert(result.out == wrapped_output.end());
-        assert(equal(wrapped_output, expected_output, equal_to{}, &int_wrapper::val));
-        assert(equal(input, expected_input, equal_to{}, &int_wrapper::val));
-        not_ranges_destroy(wrapped_output);
-        assert(int_wrapper::constructions == 3);
-        assert(int_wrapper::destructions == 3);
+        { // Validate shorter input
+            int_wrapper input[3] = {13, 55, 12345};
+            Read wrapped_input{input};
+            holder<int_wrapper, 4> mem;
+            Write wrapped_output{mem.as_span()};
+
+            int_wrapper::clear_counts();
+            same_as<uninitialized_copy_n_result<iterator_t<Read>, iterator_t<Write>>> auto result =
+                uninitialized_copy_n(wrapped_input.begin(), 3, wrapped_output.begin(), wrapped_output.end());
+            assert(int_wrapper::constructions == 3);
+            assert(int_wrapper::destructions == 0);
+            assert(result.in == wrapped_input.end());
+            construct_at(addressof(*result.out), -1); // Need to construct non written element for comparison
+            assert(++result.out == wrapped_output.end());
+            assert(equal(wrapped_output, expected_output_long, equal_to{}, &int_wrapper::val));
+            assert(equal(input, expected_input, equal_to{}, &int_wrapper::val));
+            destroy(wrapped_output);
+            assert(int_wrapper::constructions == 4);
+            assert(int_wrapper::destructions == 4);
+        }
     }
 };
 
@@ -126,6 +165,53 @@ struct throwing_test {
     }
 };
 
+struct memcpy_test {
+    static constexpr int expected_output[]      = {13, 55, 12345, -1};
+    static constexpr int expected_output_long[] = {13, 55, -1, -1};
+    static constexpr int expected_input[]       = {13, 55, 12345, 42};
+    static constexpr int expected_input_short[] = {13, 55};
+    static constexpr int expected_input_long[]  = {13, 55, 12345, 42};
+
+    static void call() {
+        using ranges::uninitialized_copy_n, ranges::uninitialized_copy_n_result, ranges::equal, ranges::iterator_t;
+        { // Validate range overload
+            vector<int> input  = {13, 55, 12345, 42};
+            vector<int> output = {-1, -1, -1, -1};
+
+            const same_as<uninitialized_copy_n_result<iterator_t<vector<int>>, iterator_t<vector<int>>>> auto result =
+                uninitialized_copy_n(input.begin(), 3, output.begin(), output.end());
+            assert(next(result.in) == input.end());
+            assert(next(result.out) == output.end());
+            assert(equal(input, expected_input));
+            assert(equal(output, expected_output));
+        }
+
+        { // Validate shorter input
+            vector<int> input  = {13, 55};
+            vector<int> output = {-1, -1, -1, -1};
+
+            const same_as<uninitialized_copy_n_result<iterator_t<vector<int>>, iterator_t<vector<int>>>> auto result =
+                uninitialized_copy_n(input.begin(), 2, output.begin(), output.end());
+            assert(result.in == input.end());
+            assert(next(result.out, 2) == output.end());
+            assert(equal(input, expected_input_short));
+            assert(equal(output, expected_output_long));
+        }
+
+        { // Validate shorter output
+            vector<int> input  = {13, 55, 12345, 42};
+            vector<int> output = {-1, -1};
+
+            const same_as<uninitialized_copy_n_result<iterator_t<vector<int>>, iterator_t<vector<int>>>> auto result =
+                uninitialized_copy_n(input.begin(), 2, output.begin(), output.end());
+            assert(next(result.in, 2) == input.end());
+            assert(result.out == output.end());
+            assert(equal(input, expected_input));
+            assert(equal(output, expected_input_short));
+        }
+    }
+};
+
 template <test::ProxyRef IsProxy>
 using test_input  = test::range<test::input, int_wrapper, test::Sized::no, test::CanDifference::no, test::Common::no,
     test::CanCompare::yes, IsProxy>;
@@ -140,4 +226,5 @@ int main() {
     instantiator::call<test_input<test::ProxyRef::yes>, test_output>();
     throwing_test::call<test_input<test::ProxyRef::no>, test_output>();
     throwing_test::call<test_input<test::ProxyRef::yes>, test_output>();
+    memcpy_test::call();
 }