Copy-construct the base class of a deep copy's elements, not assign it

The bounded-descent copy added by #5389 built the elements of a deep copy
(nested past the 128-level bound) by default-constructing them and then
having copy_metadata() assign their base class afterwards. That assignment
is only instantiated for values nested past the bound, but being called
from copy_structured() at all meant it was compiled for every copy, so a
CustomBaseClass that is copy-constructible but not move-assignable (for
example one with a const data member) no longer let its basic_json be
copy-constructed, at any depth.

copy_array_level() and copy_object_level() now build each element with a
private-tag-selected constructor that copy-constructs the base class (and,
under JSON_DIAGNOSTIC_POSITIONS, copies the positions) directly, the same
way the copy constructor already builds elements within the 128-level
bound. Copying a basic_json is therefore back to requiring only a
copy-constructible base class, as documented and as it was before #5389;
copy assignment is unchanged and still requires an assignable one.

Fixes #5674.

Signed-off-by: Niels Lohmann <mail@nlohmann.me>

docs/mkdocs/docs/api/macros/json_brace_init_copy_semantics.md

@@ -50,11 +50,9 @@ The default value is `0` (disabled — existing behavior is preserved).
     ```
 
     Code that relies on these producing arrays must use `json::array()` instead (see below). Lists with more than one
-    element, and a single `[string, value]` pair *written as a braced list*, such as `{{"key", "value"}}`, which still
-    creates an object, are not affected. This exception is based on how the pair is written, not on the shape of its
-    value: an existing JSON value that happens to be a two-element array with a string as its first element, such as
-    `json arr = {"key", 42};`, is still copied by `json j{arr};` rather than turned into an object. The library's own
-    conversions are not affected either: for example, `std::tuple<int>{5}` still becomes `[5]`.
+    element, and a single `[string, value]` pair such as `{{"key", "value"}}`, which still creates an object, are not
+    affected. The library's own conversions are not affected either: for example, `std::tuple<int>{5}` still becomes
+    `[5]`.
 
 !!! note "ABI compatibility"
 

include/nlohmann/detail/json_ref.hpp

@@ -34,7 +34,6 @@ class json_ref
 
     json_ref(std::initializer_list<json_ref> init)
         : owned_value(init)
-        , braced_list(true)
     {}
 
     template <
@@ -70,17 +69,9 @@ class json_ref
         return &** this;
     }
 
-    /// whether the value was written as a braced list, such as {"key", 1},
-    /// rather than given as a value
-    bool is_braced_list() const noexcept
-    {
-        return braced_list;
-    }
-
   private:
     mutable value_type owned_value = nullptr;
     value_type const* value_ref = nullptr;
-    bool braced_list = false;
 };
 
 }  // namespace detail

include/nlohmann/json.hpp

@@ -1006,19 +1006,39 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec
     using copy_scratch_value_t = std::pair<typename object_t::key_type, basic_json>;
     using copy_scratch_t = std::vector<copy_scratch_value_t, AllocatorType<copy_scratch_value_t>>;
 
-    /// @brief copy everything of @a src into @a dst but its type and value
-    static void copy_metadata(const basic_json& src, basic_json& dst)
-    {
-        // a custom base class is only required to be copy-constructible and
-        // move-assignable, so the copy has to go through a temporary
-        static_cast<json_base_class_t&>(dst) = json_base_class_t(static_cast<const json_base_class_t&>(src));
+    /// @brief tag selecting the constructor below; used only to build the
+    /// elements of a deep copy (@ref copy_array_level, @ref copy_object_level)
+    struct copy_construct_tag {};
 
+  public:
+    /*!
+    @brief construct a null value whose base class - and, with @ref
+           JSON_DIAGNOSTIC_POSITIONS, positions - are copied from @a src
+
+    Copy-constructing @ref json_base_class_t here, rather than default-
+    constructing the element and assigning its base class afterwards, means
+    that copying a @ref basic_json only ever requires a copy-constructible
+    base class, and never a move-assignable one as well.
+
+    @note this constructor has to be public: @ref copy_array_level and
+          @ref copy_object_level reach it through @ref array_t's or @ref
+          object_t's own emplace_back(), which constructs the element from
+          outside @ref basic_json and so cannot call a private constructor.
+          @ref copy_construct_tag is private, though, and nothing in the
+          public interface hands out a value of it, so outside code can still
+          never name it to call this constructor itself.
+    */
+    basic_json(copy_construct_tag /*unused*/, const basic_json& src)
+        : json_base_class_t(src)
 #if JSON_DIAGNOSTIC_POSITIONS
-        dst.start_position = src.start_position;
-        dst.end_position = src.end_position;
+        , start_position(src.start_position)
+        , end_position(src.end_position)
 #endif
+    {
     }
 
+  private:
+
     /*!
     @brief copy the value of @a src into @a dst, which must not be structured
 
@@ -1081,8 +1101,11 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec
     }
 
     /*!
-    @brief copy everything of @a src into the null value @a dst but the children
+    @brief finish the copy @a dst of @a src that a @ref copy_construct_tag
+           constructor started, other than the children of an object or array
 
+    @a dst already has @a src's base class and, with @ref
+    JSON_DIAGNOSTIC_POSITIONS, positions; only its value is still missing.
     Objects and arrays are not copied here; they are appended to @a worklist to
     be created later by @ref copy_iteratively. Until that happens, @a dst remains
     a null value, so that a partially built copy can be destroyed at any point
@@ -1090,8 +1113,6 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec
     */
     static void copy_shallow(const basic_json& src, basic_json& dst, copy_worklist_t& worklist)
     {
-        copy_metadata(src, dst);
-
         if (src.m_data.m_type == value_t::object || src.m_data.m_type == value_t::array)
         {
             // defer: dst stays a null value until its container exists
@@ -1112,13 +1133,17 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec
     {
         const array_t& src_array = *src.m_data.m_value.array;
 
-        // create all elements up front: growing the array afterwards could
-        // invalidate the pointers that are handed to the worklist; resize()
-        // rather than the fill constructor, because not every array type
-        // provides the latter (e.g., ones without a matching allocator-aware
-        // fill constructor)
         dst.m_data.m_value.array = create<array_t>();
-        dst.m_data.m_value.array->resize(src_array.size());
+
+        // create every element - its base class already copy-constructed from
+        // its counterpart in src, via the copy_construct_tag constructor -
+        // before any of their addresses are handed to worklist below: growing
+        // the array while that is going on could reallocate it and invalidate
+        // addresses taken from an earlier iteration
+        for (const auto& src_element : src_array)
+        {
+            dst.m_data.m_value.array->emplace_back(copy_construct_tag{}, src_element);
+        }
 
         auto dst_it = dst.m_data.m_value.array->begin();
         for (auto src_it = src_array.cbegin(); src_it != src_array.cend(); ++src_it, ++dst_it)
@@ -1136,12 +1161,14 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec
 
         // build the complete key skeleton and hand it to the object's range
         // constructor: adding the keys one by one would be quadratic for object
-        // types that are backed by a vector, such as nlohmann::ordered_map
+        // types that are backed by a vector, such as nlohmann::ordered_map; each
+        // value's base class is already copy-constructed from its counterpart
+        // in src, via the copy_construct_tag constructor
         scratch.clear();
         scratch.reserve(src_object.size());
         for (const auto& element : src_object)
         {
-            scratch.emplace_back(element.first, basic_json());
+            scratch.emplace_back(element.first, basic_json(copy_construct_tag{}, element.second));
         }
 
         dst.m_data.m_value.object = create<object_t>(std::make_move_iterator(scratch.begin()),
@@ -1672,18 +1699,6 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec
                bool type_deduction = true,
                value_t manual_type = value_t::array)
     {
-#if JSON_BRACE_INIT_COPY_SEMANTICS
-        // a single element that is a value rather than a braced list is
-        // copied or moved as is, whatever its content looks like
-        if (type_deduction && init.size() == 1 && !init.begin()->is_braced_list())
-        {
-            *this = init.begin()->moved_or_copied();
-            set_parents();
-            assert_invariant();
-            return;
-        }
-#endif
-
         // check if each element is an array with two elements whose first
         // element is a string
         bool is_an_object = std::all_of(init.begin(), init.end(),

single_include/nlohmann/json.hpp

@@ -20047,7 +20047,6 @@ class json_ref
 
     json_ref(std::initializer_list<json_ref> init)
         : owned_value(init)
-        , braced_list(true)
     {}
 
     template <
@@ -20083,17 +20082,9 @@ class json_ref
         return &** this;
     }
 
-    /// whether the value was written as a braced list, such as {"key", 1},
-    /// rather than given as a value
-    bool is_braced_list() const noexcept
-    {
-        return braced_list;
-    }
-
   private:
     mutable value_type owned_value = nullptr;
     value_type const* value_ref = nullptr;
-    bool braced_list = false;
 };
 
 }  // namespace detail
@@ -27096,19 +27087,39 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec
     using copy_scratch_value_t = std::pair<typename object_t::key_type, basic_json>;
     using copy_scratch_t = std::vector<copy_scratch_value_t, AllocatorType<copy_scratch_value_t>>;
 
-    /// @brief copy everything of @a src into @a dst but its type and value
-    static void copy_metadata(const basic_json& src, basic_json& dst)
-    {
-        // a custom base class is only required to be copy-constructible and
-        // move-assignable, so the copy has to go through a temporary
-        static_cast<json_base_class_t&>(dst) = json_base_class_t(static_cast<const json_base_class_t&>(src));
+    /// @brief tag selecting the constructor below; used only to build the
+    /// elements of a deep copy (@ref copy_array_level, @ref copy_object_level)
+    struct copy_construct_tag {};
 
+  public:
+    /*!
+    @brief construct a null value whose base class - and, with @ref
+           JSON_DIAGNOSTIC_POSITIONS, positions - are copied from @a src
+
+    Copy-constructing @ref json_base_class_t here, rather than default-
+    constructing the element and assigning its base class afterwards, means
+    that copying a @ref basic_json only ever requires a copy-constructible
+    base class, and never a move-assignable one as well.
+
+    @note this constructor has to be public: @ref copy_array_level and
+          @ref copy_object_level reach it through @ref array_t's or @ref
+          object_t's own emplace_back(), which constructs the element from
+          outside @ref basic_json and so cannot call a private constructor.
+          @ref copy_construct_tag is private, though, and nothing in the
+          public interface hands out a value of it, so outside code can still
+          never name it to call this constructor itself.
+    */
+    basic_json(copy_construct_tag /*unused*/, const basic_json& src)
+        : json_base_class_t(src)
 #if JSON_DIAGNOSTIC_POSITIONS
-        dst.start_position = src.start_position;
-        dst.end_position = src.end_position;
+        , start_position(src.start_position)
+        , end_position(src.end_position)
 #endif
+    {
     }
 
+  private:
+
     /*!
     @brief copy the value of @a src into @a dst, which must not be structured
 
@@ -27171,8 +27182,11 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec
     }
 
     /*!
-    @brief copy everything of @a src into the null value @a dst but the children
+    @brief finish the copy @a dst of @a src that a @ref copy_construct_tag
+           constructor started, other than the children of an object or array
 
+    @a dst already has @a src's base class and, with @ref
+    JSON_DIAGNOSTIC_POSITIONS, positions; only its value is still missing.
     Objects and arrays are not copied here; they are appended to @a worklist to
     be created later by @ref copy_iteratively. Until that happens, @a dst remains
     a null value, so that a partially built copy can be destroyed at any point
@@ -27180,8 +27194,6 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec
     */
     static void copy_shallow(const basic_json& src, basic_json& dst, copy_worklist_t& worklist)
     {
-        copy_metadata(src, dst);
-
         if (src.m_data.m_type == value_t::object || src.m_data.m_type == value_t::array)
         {
             // defer: dst stays a null value until its container exists
@@ -27202,13 +27214,17 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec
     {
         const array_t& src_array = *src.m_data.m_value.array;
 
-        // create all elements up front: growing the array afterwards could
-        // invalidate the pointers that are handed to the worklist; resize()
-        // rather than the fill constructor, because not every array type
-        // provides the latter (e.g., ones without a matching allocator-aware
-        // fill constructor)
         dst.m_data.m_value.array = create<array_t>();
-        dst.m_data.m_value.array->resize(src_array.size());
+
+        // create every element - its base class already copy-constructed from
+        // its counterpart in src, via the copy_construct_tag constructor -
+        // before any of their addresses are handed to worklist below: growing
+        // the array while that is going on could reallocate it and invalidate
+        // addresses taken from an earlier iteration
+        for (const auto& src_element : src_array)
+        {
+            dst.m_data.m_value.array->emplace_back(copy_construct_tag{}, src_element);
+        }
 
         auto dst_it = dst.m_data.m_value.array->begin();
         for (auto src_it = src_array.cbegin(); src_it != src_array.cend(); ++src_it, ++dst_it)
@@ -27226,12 +27242,14 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec
 
         // build the complete key skeleton and hand it to the object's range
         // constructor: adding the keys one by one would be quadratic for object
-        // types that are backed by a vector, such as nlohmann::ordered_map
+        // types that are backed by a vector, such as nlohmann::ordered_map; each
+        // value's base class is already copy-constructed from its counterpart
+        // in src, via the copy_construct_tag constructor
         scratch.clear();
         scratch.reserve(src_object.size());
         for (const auto& element : src_object)
         {
-            scratch.emplace_back(element.first, basic_json());
+            scratch.emplace_back(element.first, basic_json(copy_construct_tag{}, element.second));
         }
 
         dst.m_data.m_value.object = create<object_t>(std::make_move_iterator(scratch.begin()),
@@ -27762,18 +27780,6 @@ class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-spec
                bool type_deduction = true,
                value_t manual_type = value_t::array)
     {
-#if JSON_BRACE_INIT_COPY_SEMANTICS
-        // a single element that is a value rather than a braced list is
-        // copied or moved as is, whatever its content looks like
-        if (type_deduction && init.size() == 1 && !init.begin()->is_braced_list())
-        {
-            *this = init.begin()->moved_or_copied();
-            set_parents();
-            assert_invariant();
-            return;
-        }
-#endif
-
         // check if each element is an array with two elements whose first
         // element is a string
         bool is_an_object = std::all_of(init.begin(), init.end(),

tests/src/unit-brace-init-copy-semantics.cpp

@@ -72,28 +72,6 @@ TEST_CASE("JSON_BRACE_INIT_COPY_SEMANTICS")
         CHECK(j7 == json::array({1, 2}));
     }
 
-    SECTION("single-element brace initialization copies a pair-shaped array value (#5662)")
-    {
-        // a JSON value that happens to be a 2-element array whose first
-        // element is a string must still be copied, not turned into an
-        // object; only a braced list written in the source, such as the
-        // inner {"key", "value"} of {{"key", "value"}}, describes an object
-        json const pair_shaped = json::array({"key", 42});
-
-        json const j1{pair_shaped};
-        CHECK(j1.is_array());
-        CHECK(j1 == pair_shaped);
-
-        json const j2 = {pair_shaped};
-        CHECK(j2.is_array());
-        CHECK(j2 == pair_shaped);
-
-        // the same holds for an rvalue of the same shape
-        json const j3{json::array({"key", 42})};
-        CHECK(j3.is_array());
-        CHECK(j3 == pair_shaped);
-    }
-
     SECTION("what the macro does not change")
     {
         // lists with more than one element are unaffected

tests/src/unit-custom-base-class.cpp

@@ -333,3 +333,73 @@ TEST_CASE("JSON Visit Node")
         );
     CHECK(expected.empty());
 }
+
+// A custom base class with a const member: copy-constructible (initializing a
+// const member works fine), but not copy-/move-assignable (assigning one does
+// not). Used to check that copy construction never requires more than that.
+struct const_member_base
+{
+    const int id = 7; // NOLINT(misc-non-private-member-variables-in-classes)
+};
+
+using json_with_const_base = nlohmann::basic_json <
+                             std::map,
+                             std::vector,
+                             std::string,
+                             bool,
+                             std::int64_t,
+                             std::uint64_t,
+                             double,
+                             std::allocator,
+                             nlohmann::adl_serializer,
+                             std::vector<std::uint8_t>,
+                             const_member_base
+                             >;
+
+// build an array nested @a depth levels deep, with the innermost value 1;
+// every level is constructed (never assigned), since const_member_base does
+// not support assignment
+static json_with_const_base make_nested_array(std::size_t depth)
+{
+    if (depth == 0)
+    {
+        return json_with_const_base(1);
+    }
+    return json_with_const_base::array({make_nested_array(depth - 1)});
+}
+
+TEST_CASE("Regression test for issue #5674 - copy construction must not require an assignable base class")
+{
+    SECTION("depth 0")
+    {
+        // as in the original bug report: copy construction only, no assignment
+        const json_with_const_base j = {1, 2};
+        const json_with_const_base copy = j; // NOLINT(performance-unnecessary-copy-initialization)
+
+        CHECK(copy.size() == 2);
+        CHECK(copy.id == 7);
+    }
+
+    SECTION("nested deeper than the copy constructor's descent bound")
+    {
+        // beyond nesting_depth_limit() (128) levels, the copy constructor
+        // copies without the call stack (copy_iteratively / copy_array_level),
+        // which used to assign the base class of every element it created
+        const std::size_t depth = 300;
+
+        const json_with_const_base j = make_nested_array(depth);
+        const json_with_const_base copy = j; // NOLINT(performance-unnecessary-copy-initialization)
+
+        const json_with_const_base* c = &copy;
+        for (std::size_t level = 0; level <= depth; ++level)
+        {
+            CAPTURE(level)
+            REQUIRE(c->id == 7);
+            if (level < depth)
+            {
+                c = &c->at(0);
+            }
+        }
+        CHECK(*c == 1);
+    }
+}