Compare the bytes of the binary writer sink tests without converting each one

std::string(begin, end) over a std::vector<std::uint8_t> converts every
byte to char, which Clang's -fsanitize=integer reports for bytes above
0x7F where char is signed (x86-64 with libstdc++). The string is now made
from the vector's data.

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

docs/mkdocs/docs/home/architecture.md

@@ -339,6 +339,11 @@ The `serializer` (used by [`dump`](../api/basic_json/dump.md) and [`operator<<`]
 - `output_stream_adapter` writes to a `std::ostream`.
 - `output_string_adapter` appends to a string.
 
+The `binary_writer` does not call an adapter for every byte. The `to_*` functions that return a `std::vector` write
+into the vector through a sink (`output_vector_sink`) that is not virtual. The functions that take an output adapter
+write through a sink (`output_adapter_sink`) that collects the output in a buffer of one kibibyte and passes it to the
+adapter in blocks, so that the adapter is called (and a stream locked) once per block instead of once per value.
+
 ## Value conversion
 
 Values are converted from and to other types with the `JSONSerializer` template parameter. The default,

include/nlohmann/detail/json_pointer.hpp

@@ -15,7 +15,7 @@
 #include <cstddef> // size_t
 #include <cstdint> // uint64_t
 #include <cstdlib> // strtoull
-#include <cstring> // memchr, memcmp, memcpy
+#include <cstring> // memcmp, memcpy
 #ifndef JSON_NO_IO
     #include <iosfwd> // ostream
 #endif  // JSON_NO_IO
@@ -303,12 +303,6 @@ class json_pointer
         exceeds_size_type  ///< @a s converts to an integer that exceeds size_type
     };
 
-    /// whether a reference token is "-" (a plain comparison with a string literal is not always inlined)
-    static bool is_dash(const string_t& s) noexcept
-    {
-        return s.size() == 1 && s[0] == '-';
-    }
-
     /*!
     @param[in] s    reference token to be converted into an array index
     @param[out] idx the integer representation of @a s if @ref array_index_status::ok
@@ -334,7 +328,7 @@ class json_pointer
         // error condition (cf. RFC 6901, Sect. 4); this also covers single-
         // character tokens, so "/x" and "/xy" fail alike; "-" and the empty
         // token are left to the conversion below and are reported as unresolved
-        if (JSON_HEDLEY_UNLIKELY(!s.empty() && !is_dash(s) && !(s[0] >= '0' && s[0] <= '9')))
+        if (JSON_HEDLEY_UNLIKELY(!s.empty() && s != "-" && !(s[0] >= '0' && s[0] <= '9')))
         {
             return array_index_status::not_a_number;
         }
@@ -481,40 +475,9 @@ class json_pointer
         return pos == size;
     }
 
-    /// reserve room for @a n elements of an array type that has reserve()
-    template<typename ArrayType>
-    static auto reserve_elements(ArrayType& array, const std::size_t n, detail::priority_tag<1> /*unused*/)
-    -> decltype(array.reserve(n), void())
-    {
-        array.reserve(n);
-    }
-
-    template<typename ArrayType>
-    static void reserve_elements(ArrayType& /*array*/, const std::size_t /*n*/, detail::priority_tag<0> /*unused*/)
-    {}
-
     /// @brief object member access with a key that is only copied if it is inserted
-    ///
-    /// Objects that can find the position of a key and insert there
-    /// (std::map) search once, others look the key up and add it if it is missing.
     template<typename BasicJsonType>
-    static auto object_member_impl(BasicJsonType& object, const string_t& key, detail::priority_tag<1> /*unused*/)
-    -> decltype(object.m_data.m_value.object->emplace_hint(object.m_data.m_value.object->lower_bound(key), key, nullptr), std::declval<BasicJsonType&>())
-    {
-        auto& members = *object.m_data.m_value.object;
-        auto it = members.lower_bound(key);
-        if (it != members.end() && !members.key_comp()(key, it->first))
-        {
-            return it->second;
-        }
-
-        const auto old_capacity = object.container_capacity();
-        it = members.emplace_hint(it, key, nullptr);
-        return object.set_parent_after_insert(it->second, old_capacity);
-    }
-
-    template<typename BasicJsonType>
-    static BasicJsonType& object_member_impl(BasicJsonType& object, const string_t& key, detail::priority_tag<0> /*unused*/)
+    static BasicJsonType& object_member(BasicJsonType& object, const string_t& key)
     {
         auto& members = *object.m_data.m_value.object;
         const auto it = members.find(key);
@@ -525,12 +488,6 @@ class json_pointer
         return object.operator[](key);
     }
 
-    template<typename BasicJsonType>
-    static BasicJsonType& object_member(BasicJsonType& object, const string_t& key)
-    {
-        return object_member_impl(object, key, detail::priority_tag<1> {});
-    }
-
   JSON_PRIVATE_UNLESS_TESTED:
     json_pointer top() const
     {
@@ -763,103 +720,48 @@ class json_pointer
         }
     }
 
-    /*!
-    @brief the values along the pointer last resolved by @ref resolve_plain
-
-    A patch from @ref basic_json::diff lists its operations in document order,
-    so consecutive pointers share most of their reference tokens; the values
-    along the previous pointer are kept to resume from the common prefix.
-    `steps[k]` is the value after the first `k` reference tokens together with
-    the length of the prefix of `*path` that holds them. A value stays valid as
-    long as no operation changes its container: callers keep only values up to
-    the parent of the pointer they modify, or call `reset()`.
-    */
-    template<typename BasicJsonType>
-    struct plain_steps
-    {
-        const string_t* path = nullptr;
-        std::vector<std::pair<std::size_t, BasicJsonType*>> steps{};
-
-        void reset() noexcept
-        {
-            path = nullptr;
-            steps.clear();
-        }
-    };
-
     /*!
     @brief resolve a plain JSON pointer without building its reference tokens
 
-    @param[in] root    the value the pointer starts at
+    @param[in] ptr     the value the pointer starts at
     @param[in] path    a JSON pointer string
     @param[in] length  the number of leading bytes of @a path to resolve; 0 or
                        the position of a slash
     @param[in,out] token  a string to hold the reference tokens, to reuse its storage
-    @param[in,out] cache  the values along the previously resolved pointer;
-                          @a path must outlive the next call
 
     @return the pointed to value, or `nullptr` if the pointer does not denote
-            an existing value, or does so in a way that is not plain: an array
-            index that is malformed or out of range, or a "-", or a reference
-            token with an escape. No exception is
-            thrown for these; callers use @ref json_pointer instead, which
-            handles them.
+            an existing value, or does so in a way that is not plain: a
+            reference token with an escape, an array index that is malformed or
+            out of range, or a "-". No exception is thrown for these; callers
+            use @ref json_pointer instead, which handles them.
     */
     template<typename BasicJsonType>
-    static BasicJsonType* resolve_plain(BasicJsonType* root, const string_t& path, const std::size_t length, string_t& token, plain_steps<BasicJsonType>& cache)
+    static BasicJsonType* resolve_plain(BasicJsonType* ptr, const string_t& path, const std::size_t length, string_t& token)
     {
-        auto& steps = cache.steps;
-
-        // resume from the deepest step that is also a prefix of this path
-        std::size_t depth = 1;
-        if (cache.path != nullptr && !steps.empty() && steps[0].second == root)
-        {
-            const string_t& previous = *cache.path;
-            for (depth = steps.size(); depth > 1; --depth)
-            {
-                // a step ends at a token boundary of the previous path
-                const std::size_t prefix = steps[depth - 1].first;
-                if (prefix <= length && (prefix == length || path[prefix] == '/') && std::memcmp(path.data(), previous.data(), prefix) == 0)
-                {
-                    break;
-                }
-            }
-            steps.resize(depth);
-        }
-        else
-        {
-            steps.assign(1, std::make_pair(std::size_t{0}, root));
-        }
-        cache.path = &path;
-
-        BasicJsonType* ptr = steps[depth - 1].second;
-        for (std::size_t pos = steps[depth - 1].first; pos < length;)
+        for (std::size_t pos = 0; pos < length;)
         {
             if (JSON_HEDLEY_UNLIKELY(path[pos] != '/'))
             {
                 return nullptr;
             }
 
-            // the reference token between the slash at pos and the next one,
-            // which must not have an escape
-            const char* const begin = path.data() + pos + 1;
-            const char* const limit = path.data() + length;
-            const char* end = begin;
-            while (end < limit && *end != '/')
+            // the reference token between the slash at pos and the next one
+            std::size_t end = path.find_first_of('/', pos + 1);
+            if (end == string_t::npos || end > length)
             {
-                if (JSON_HEDLEY_UNLIKELY(*end == '~'))
-                {
-                    return nullptr;
-                }
-                ++end;
+                end = length;
+            }
+            token.clear();
+            token.append(path.data() + pos + 1, end - pos - 1);
+            if (JSON_HEDLEY_UNLIKELY(token.find_first_of('~') != string_t::npos))
+            {
+                return nullptr;
             }
 
             switch (ptr->type())
             {
                 case detail::value_t::object:
                 {
-                    token.clear();
-                    token.append(begin, static_cast<std::size_t>(end - begin));
                     const auto it = ptr->m_data.m_value.object->find(token);
                     if (JSON_HEDLEY_UNLIKELY(it == ptr->m_data.m_value.object->end()))
                     {
@@ -871,24 +773,8 @@ class json_pointer
 
                 case detail::value_t::array:
                 {
-                    // an index of up to 18 digits without a leading zero fits into size_type
-                    // (9 digits for a 32-bit size_type); everything else is left to json_pointer
-                    constexpr std::size_t max_digits = sizeof(typename BasicJsonType::size_type) >= 8 ? 18 : 9;
-                    const std::size_t digits = static_cast<std::size_t>(end - begin);
-                    if (JSON_HEDLEY_UNLIKELY(digits == 0 || digits > max_digits || (digits > 1 && *begin == '0')))
-                    {
-                        return nullptr;
-                    }
-                    typename BasicJsonType::size_type idx = 0;
-                    for (const char* c = begin; c != end; ++c)
-                    {
-                        if (JSON_HEDLEY_UNLIKELY(*c < '0' || *c > '9'))
-                        {
-                            return nullptr;
-                        }
-                        idx = static_cast<typename BasicJsonType::size_type>((idx * 10) + static_cast<typename BasicJsonType::size_type>(*c - '0'));
-                    }
-                    if (JSON_HEDLEY_UNLIKELY(idx >= ptr->m_data.m_value.array->size()))
+                    typename BasicJsonType::size_type idx{};
+                    if (JSON_HEDLEY_UNLIKELY(parse_array_index<BasicJsonType>(token, idx) != array_index_status::ok || idx >= ptr->m_data.m_value.array->size()))
                     {
                         return nullptr;
                     }
@@ -907,8 +793,7 @@ class json_pointer
                 default:
                     return nullptr;
             }
-            pos = static_cast<std::size_t>(end - path.data());
-            steps.emplace_back(pos, ptr);
+            pos = end;
         }
 
         return ptr;
@@ -1022,7 +907,7 @@ class json_pointer
 
                 case detail::value_t::array:
                 {
-                    if (JSON_HEDLEY_UNLIKELY(is_dash(reference_token)))
+                    if (JSON_HEDLEY_UNLIKELY(reference_token == "-"))
                     {
                         // "-" always fails the range check
                         JSON_THROW(detail::out_of_range::create(detail::exception_id::pointer_past_the_end_index, detail::concat(
@@ -1093,7 +978,7 @@ class json_pointer
 
                 case detail::value_t::array:
                 {
-                    if (JSON_HEDLEY_UNLIKELY(is_dash(reference_token)))
+                    if (JSON_HEDLEY_UNLIKELY(reference_token == "-"))
                     {
                         // "-" cannot be used for const access
                         JSON_THROW(detail::out_of_range::create(detail::exception_id::pointer_past_the_end_index, detail::concat("array index '-' (", std::to_string(ptr->m_data.m_value.array->size()), ") is out of range"), ptr));
@@ -1155,7 +1040,7 @@ class json_pointer
 
                 case detail::value_t::array:
                 {
-                    if (JSON_HEDLEY_UNLIKELY(is_dash(reference_token)))
+                    if (JSON_HEDLEY_UNLIKELY(reference_token == "-"))
                     {
                         // "-" always fails the range check
                         return nullptr;
@@ -1232,7 +1117,7 @@ class json_pointer
 
                 case detail::value_t::array:
                 {
-                    if (JSON_HEDLEY_UNLIKELY(is_dash(reference_token)))
+                    if (JSON_HEDLEY_UNLIKELY(reference_token == "-"))
                     {
                         // "-" always fails the range check
                         return false;
@@ -1391,52 +1276,6 @@ class json_pointer
         obj[key] = std::forward<ValueType>(v);
     }
 
-    /// the number of entries that flattening @a value adds: its primitive
-    /// values and empty containers
-    template<typename BasicJsonType>
-    static std::size_t count_flattened(const BasicJsonType& value)
-    {
-        std::size_t count = 0;
-        std::vector<const BasicJsonType*> pending(1, &value);
-        while (!pending.empty())
-        {
-            const BasicJsonType* const v = pending.back();
-            pending.pop_back();
-            if (v->is_array() && !v->m_data.m_value.array->empty())
-            {
-                for (const auto& element : *v->m_data.m_value.array)
-                {
-                    pending.push_back(&element);
-                }
-            }
-            else if (v->is_object() && !v->m_data.m_value.object->empty())
-            {
-                for (const auto& member : *v->m_data.m_value.object)
-                {
-                    pending.push_back(&member.second);
-                }
-            }
-            else
-            {
-                ++count;
-            }
-        }
-        return count;
-    }
-
-    // Reserve the storage for @a value of objects that have a reserve(), so
-    // that it does not have to be moved while it grows
-    template<typename ObjectType, typename BasicJsonType>
-    static auto reserve_flattened(ObjectType& obj, const BasicJsonType& value, detail::priority_tag<1> /*unused*/)
-    -> decltype(obj.reserve(std::size_t {}), void())
-    {
-        obj.reserve(count_flattened(value));
-    }
-
-    template<typename ObjectType, typename BasicJsonType>
-    static void reserve_flattened(ObjectType& /*obj*/, const BasicJsonType& /*value*/, detail::priority_tag<0> /*unused*/)
-    {}
-
     /// append the decimal representation of @a i to @a out
     static void append_index(string_t& out, std::size_t i)
     {
@@ -1546,7 +1385,6 @@ class json_pointer
             }
         };
 
-        reserve_flattened(object, value, detail::priority_tag<1> {});
         enter(value);
         while (!stack.empty())
         {
@@ -1554,9 +1392,8 @@ class json_pointer
             // frame, which would invalidate a reference to it
             const BasicJsonType* const container = stack.back().container;
 
-            // the path of the previous child is replaced by the path of the
-            // next one; for an array element in one resize
-            const std::size_t path_length = stack.back().path_length;
+            // drop the path of the previous child
+            path.resize(stack.back().path_length);
 
             if (container->is_array())
             {
@@ -1570,20 +1407,8 @@ class json_pointer
 
                 // iterate array and use index as a reference string
                 ++stack.back().index;
-                std::array<char, 24> digits{};
-                std::size_t pos = digits.size();
-                std::size_t rest = i;
-                do
-                {
-                    digits[--pos] = static_cast<char>('0' + rest % 10);
-                    rest /= 10;
-                }
-                while (rest != 0);
-                const std::size_t count = digits.size() - pos;
-                path.resize(path_length + 1 + count);
-                char* const out = &path[path_length];
-                out[0] = '/';
-                std::memcpy(out + 1, digits.data() + pos, count);
+                path.push_back('/');
+                append_index(path, i);
                 enter(array[i]);
             }
             else
@@ -1597,7 +1422,6 @@ class json_pointer
 
                 // iterate object and use keys as reference string
                 ++stack.back().member;
-                path.resize(path_length);
                 path.push_back('/');
                 detail::escape_into(path, it->first);
                 enter(it->second);
@@ -1693,28 +1517,8 @@ class json_pointer
                 {
                     end = key.size();
                 }
-                const std::size_t token_size = end - pos - 1;
-
-                // a token that is a plain index of up to 9 digits, which is
-                // the case for almost all array elements, is not copied
-                std::size_t index = 0;
-                bool plain_index = current->type() == detail::value_t::array && token_size != 0 && token_size <= 9 && (token_size == 1 || data[pos + 1] != '0');
-                for (std::size_t k = pos + 1; plain_index && k < end; ++k)
-                {
-                    plain_index = data[k] >= '0' && data[k] <= '9';
-                    index = (index * 10) + static_cast<std::size_t>(data[k] - '0');
-                }
-                if (plain_index)
-                {
-                    current = &current->operator[](static_cast<typename BasicJsonType::size_type>(index));
-                    values.push_back(current);
-                    ends.push_back(end);
-                    pos = end;
-                    continue;
-                }
-
                 token.clear();
-                token.append(data + pos + 1, token_size);
+                token.append(data + pos + 1, end - pos - 1);
                 if (token.find_first_of('~') != string_t::npos)
                 {
                     detail::unescape(token);
@@ -1724,13 +1528,10 @@ class json_pointer
                 {
                     case detail::value_t::null:
                     {
-                        // a first child 0 makes the value an array without a look-up
-                        if ((token_size == 1 && data[pos + 1] == '0') || arrays.contains(data, pos))
+                        if (arrays.contains(data, pos))
                         {
                             // some reference token below this position is 0, so the
                             // value is an array
-                            *current = BasicJsonType(detail::value_t::array);
-                            reserve_elements(*current->m_data.m_value.array, 2, detail::priority_tag<1> {});
                             current = &current->operator[](array_index<BasicJsonType>(token));
                         }
                         else

include/nlohmann/detail/output/output_adapters.hpp

@@ -55,6 +55,66 @@ using output_adapter_t = std::shared_ptr<output_adapter_protocol<CharType>>;
 /// output_vector_sink also appends writes of at least this size with insert()
 constexpr std::size_t max_reserve_size = 256;
 
+/// @brief copy @a length characters, at most 32, without a call of memcpy()
+///
+/// Short strings are the typical payload of a binary value, and a call of
+/// memcpy() costs more than copying them. A copy of 16 to 32 characters is two
+/// possibly overlapping 16-byte copies, one of 8 to 15 two 8-byte copies, and
+/// so on down to a single character; each is a fixed-size load and store.
+/// Nothing outside of [src, src + length) is read or written.
+template<typename CharType>
+inline void copy_short(CharType* dst, const CharType* src, const std::size_t length, std::true_type /*is_byte*/) noexcept
+{
+    JSON_ASSERT(length <= 32);
+    if (length >= 16)
+    {
+        std::array<char, 16> first; // NOLINT(cppcoreguidelines-pro-type-member-init,hicpp-member-init): written before read
+        std::array<char, 16> last; // NOLINT(cppcoreguidelines-pro-type-member-init,hicpp-member-init): written before read
+        std::memcpy(first.data(), src, 16);
+        std::memcpy(last.data(), src + (length - 16), 16);
+        std::memcpy(dst, first.data(), 16);
+        std::memcpy(dst + (length - 16), last.data(), 16);
+    }
+    else if (length >= 8)
+    {
+        std::array<char, 8> first; // NOLINT(cppcoreguidelines-pro-type-member-init,hicpp-member-init): written before read
+        std::array<char, 8> last; // NOLINT(cppcoreguidelines-pro-type-member-init,hicpp-member-init): written before read
+        std::memcpy(first.data(), src, 8);
+        std::memcpy(last.data(), src + (length - 8), 8);
+        std::memcpy(dst, first.data(), 8);
+        std::memcpy(dst + (length - 8), last.data(), 8);
+    }
+    else if (length >= 4)
+    {
+        std::array<char, 4> first; // NOLINT(cppcoreguidelines-pro-type-member-init,hicpp-member-init): written before read
+        std::array<char, 4> last; // NOLINT(cppcoreguidelines-pro-type-member-init,hicpp-member-init): written before read
+        std::memcpy(first.data(), src, 4);
+        std::memcpy(last.data(), src + (length - 4), 4);
+        std::memcpy(dst, first.data(), 4);
+        std::memcpy(dst + (length - 4), last.data(), 4);
+    }
+    else if (length != 0)
+    {
+        // 1 to 3: the first, the middle, and the last character
+        const CharType a = src[0];
+        const CharType b = src[length >> 1];
+        const CharType c = src[length - 1];
+        dst[0] = a;
+        dst[length >> 1] = b;
+        dst[length - 1] = c;
+    }
+}
+
+/// @brief copy @a length characters, of a character type wider than a byte
+template<typename CharType>
+inline void copy_short(CharType* dst, const CharType* src, const std::size_t length, std::false_type /*is_byte*/) noexcept
+{
+    std::memcpy(dst, src, length * sizeof(CharType));
+}
+
+/// the number of characters an empty vector is grown to by output_vector_sink
+constexpr std::size_t min_vector_sink_capacity = 64;
+
 /// @brief non-virtual output sink writing into a std::vector
 ///
 /// This sink is not part of the virtual output_adapter_protocol hierarchy: it is
@@ -78,7 +138,7 @@ class output_vector_sink
 {
   public:
     explicit output_vector_sink(std::vector<CharType, AllocatorType>& vec) noexcept
-        : v(&vec), position(vec.size())
+        : v(&vec), cursor(vec.data() + vec.size()), limit(cursor)
     {}
 
     output_vector_sink(const output_vector_sink&) = delete;
@@ -86,7 +146,7 @@ class output_vector_sink
     output_vector_sink& operator=(output_vector_sink&&) = delete;
 
     output_vector_sink(output_vector_sink&& other) noexcept
-        : v(other.v), position(other.position)
+        : v(other.v), cursor(other.cursor), limit(other.limit)
     {
         other.v = nullptr;
     }
@@ -95,7 +155,7 @@ class output_vector_sink
     {
         if (v != nullptr)
         {
-            v->resize(position);
+            v->resize(position());
         }
     }
 
@@ -104,11 +164,11 @@ class output_vector_sink
     JSON_HEDLEY_ALWAYS_INLINE
     void write_character(CharType c)
     {
-        if (JSON_HEDLEY_UNLIKELY(position == v->size()))
+        if (JSON_HEDLEY_UNLIKELY(cursor == limit))
         {
             grow(1);
         }
-        (*v)[position++] = c;
+        *cursor++ = c;
     }
 
     // no JSON_HEDLEY_NON_NULL here: binary_writer legitimately passes a null
@@ -117,7 +177,7 @@ class output_vector_sink
     JSON_HEDLEY_ALWAYS_INLINE
     void write_characters(const CharType* s, std::size_t length)
     {
-        if (JSON_HEDLEY_UNLIKELY(v->size() - position < length))
+        if (JSON_HEDLEY_UNLIKELY(space() < length))
         {
             if (length >= max_reserve_size)
             {
@@ -126,10 +186,15 @@ class output_vector_sink
             }
             grow(length);
         }
-        if (length != 0)
+        if (length <= 32)
         {
-            std::memcpy(v->data() + position, s, length * sizeof(CharType));
-            position += length;
+            copy_short(cursor, s, length, std::integral_constant<bool, sizeof(CharType) == 1>());
+            cursor += length;
+        }
+        else
+        {
+            std::memcpy(cursor, s, length * sizeof(CharType));
+            cursor += length;
         }
     }
 
@@ -144,21 +209,37 @@ class output_vector_sink
     CharType* reserve(const std::size_t length)
     {
         JSON_ASSERT(length <= max_reserve_size);
-        if (JSON_HEDLEY_UNLIKELY(v->size() - position < length))
+        if (JSON_HEDLEY_UNLIKELY(space() < length))
         {
             grow(length);
         }
-        return v->data() + position;
+        return cursor;
     }
 
     /// keep the first @a length characters written to the space from reserve()
     JSON_HEDLEY_ALWAYS_INLINE
     void commit(const std::size_t length) noexcept
     {
-        position += length;
+        cursor += length;
     }
 
+    /// nothing to pass on: the characters are in the vector (see output_adapter_sink)
+    void flush() noexcept
+    {}
+
   private:
+    /// number of characters of *v written so far
+    std::size_t position() const noexcept
+    {
+        return static_cast<std::size_t>(cursor - v->data());
+    }
+
+    /// number of characters that fit the vector behind the write position
+    std::size_t space() const noexcept
+    {
+        return static_cast<std::size_t>(limit - cursor);
+    }
+
     /// make room for at least @a length more characters after the position;
     /// kept out of line so that the write functions stay small enough for
     /// GCC to inline them into the writer
@@ -166,12 +247,15 @@ class output_vector_sink
     void grow(const std::size_t length)
     {
         // use the capacity the vector already has, else at least double it;
-        // never less than 64 characters, so that a small value does not
+        // never less than min_vector_sink_capacity, so that a small value does not
         // reallocate for every few bytes
-        const std::size_t needed = position + length;
-        v->resize((std::max)(std::size_t{64}, needed <= v->capacity()
+        const std::size_t written = position();
+        const std::size_t needed = written + length;
+        v->resize((std::max)(min_vector_sink_capacity, needed <= v->capacity()
                              ? v->capacity()
                              : (std::max)(needed, v->capacity() * 2)));
+        cursor = v->data() + written;
+        limit = v->data() + v->size();
     }
 
     /// append @a length (at least max_reserve_size) characters that do not fit
@@ -181,20 +265,24 @@ class output_vector_sink
     JSON_HEDLEY_NEVER_INLINE
     void append(const CharType* s, const std::size_t length)
     {
-        v->resize(position);
-        if (v->capacity() - position < length)
+        const std::size_t written = position();
+        v->resize(written);
+        if (v->capacity() - written < length)
         {
             // grow geometrically, with room for the writes that follow, so
             // that the next grow() need not reallocate right away
-            v->reserve((std::max)(v->capacity() * 2, position + length + max_reserve_size));
+            v->reserve((std::max)(v->capacity() * 2, written + length + max_reserve_size));
         }
         v->insert(v->end(), s, s + length);
-        position = v->size();
+        cursor = v->data() + v->size();
+        limit = cursor;
     }
 
     std::vector<CharType, AllocatorType>* v;
-    /// number of characters of *v written so far
-    std::size_t position;
+    /// the write position: in *v, at the number of characters written so far
+    CharType* cursor;
+    /// the end of *v
+    CharType* limit;
 };
 
 /// output adapter for byte vectors
@@ -277,9 +365,17 @@ class output_string_adapter : public output_adapter_protocol<CharType>
 ///
 /// Wraps the polymorphic output_adapter_t so the same binary_writer template can
 /// also target arbitrary adapters (output streams, strings, user-provided
-/// adapters) via the `output_adapter`-based overloads. Each write still goes
-/// through one virtual call, exactly as before; only the concrete sinks above
-/// avoid it.
+/// adapters) via the `output_adapter`-based overloads.
+///
+/// A write through the adapter is a virtual call, and for a stream a call with
+/// a sentry on top, which costs more than encoding the few bytes of most values.
+/// The sink therefore collects what is written in a buffer and passes it on in
+/// one piece when it is full, a write is too large for it, or @ref flush is
+/// called. Everything written reaches the adapter in the same order and with
+/// the same bytes; the destructor passes on what is left, so that a write that
+/// throws leaves the adapter with exactly the bytes written before it. Only
+/// @ref flush can report an error of the adapter (the destructor must not
+/// throw), so it has to be called once the value is written.
 template<typename CharType>
 class output_adapter_sink
 {
@@ -290,36 +386,121 @@ class output_adapter_sink
         JSON_ASSERT(oa);
     }
 
+    output_adapter_sink(const output_adapter_sink&) = delete;
+    output_adapter_sink& operator=(const output_adapter_sink&) = delete;
+    output_adapter_sink& operator=(output_adapter_sink&&) = delete;
+
+    output_adapter_sink(output_adapter_sink&& other) noexcept
+        : oa(std::move(other.oa)), used(other.used)
+    {
+        std::memcpy(buffer.data(), other.buffer.data(), other.used * sizeof(CharType));
+        other.used = 0;
+    }
+
+    ~output_adapter_sink()
+    {
+        // a normal return has flushed already, so this is for a write that threw
+        if (used != 0)
+        {
+            JSON_TRY
+            {
+                flush();
+            }
+            JSON_CATCH(...)
+            {
+                // nothing can be done about an adapter that throws here
+            }
+        }
+    }
+
+    JSON_HEDLEY_ALWAYS_INLINE
     void write_character(CharType c)
     {
-        oa->write_character(c);
+        if (JSON_HEDLEY_UNLIKELY(used == buffer_size))
+        {
+            flush();
+        }
+        buffer[used++] = c;
     }
 
     // no JSON_HEDLEY_NON_NULL: forwards (null, 0) for empty payloads, exactly as
     // the type-erased path already did before this sink existed
+    JSON_HEDLEY_ALWAYS_INLINE
     void write_characters(const CharType* s, std::size_t length)
     {
-        oa->write_characters(s, length);
+        if (JSON_HEDLEY_UNLIKELY(buffer_size - used < length))
+        {
+            write_large(s, length);
+            return;
+        }
+        if (length <= 32)
+        {
+            copy_short(buffer.data() + used, s, length, std::integral_constant<bool, sizeof(CharType) == 1>());
+            used += length;
+        }
+        else
+        {
+            std::memcpy(buffer.data() + used, s, length * sizeof(CharType));
+            used += length;
+        }
     }
 
     /// space for writing up to @a length characters (see
     /// output_vector_sink::reserve)
-    CharType* reserve(const std::size_t length) noexcept
+    JSON_HEDLEY_ALWAYS_INLINE
+    CharType* reserve(const std::size_t length)
     {
         JSON_ASSERT(length <= max_reserve_size);
-        static_cast<void>(length);
-        return reserved.data();
+        if (JSON_HEDLEY_UNLIKELY(buffer_size - used < length))
+        {
+            flush();
+        }
+        return buffer.data() + used;
     }
 
-    /// write the first @a length characters written to the space from reserve()
-    void commit(const std::size_t length)
+    /// keep the first @a length characters written to the space from reserve()
+    JSON_HEDLEY_ALWAYS_INLINE
+    void commit(const std::size_t length) noexcept
     {
-        oa->write_characters(reserved.data(), length);
+        used += length;
+    }
+
+    /// pass everything written so far on to the adapter
+    void flush()
+    {
+        if (used != 0)
+        {
+            // reset first: the characters count as written even if the adapter throws
+            const std::size_t length = used;
+            used = 0;
+            oa->write_characters(buffer.data(), length);
+        }
     }
 
   private:
+    /// a write that does not fit the space left in the buffer
+    JSON_HEDLEY_NEVER_INLINE
+    void write_large(const CharType* s, const std::size_t length)
+    {
+        flush();
+        if (length >= buffer_size)
+        {
+            oa->write_characters(s, length);
+        }
+        else
+        {
+            std::memcpy(buffer.data(), s, length * sizeof(CharType));
+            used = length;
+        }
+    }
+
+    /// large enough for reserve() at any time
+    static constexpr std::size_t buffer_size = 4 * max_reserve_size;
+
     output_adapter_t<CharType> oa;
-    std::array<CharType, max_reserve_size> reserved{};
+    /// number of characters of @ref buffer not yet passed on
+    std::size_t used = 0;
+    std::array<CharType, buffer_size> buffer{};
 };
 
 /// @brief whether std::basic_string<CharType> has a non-deprecated std::char_traits

include/nlohmann/json.hpp

@@ -6321,7 +6321,7 @@ public:
             const error_handler_t error_handler = detail::binary_writer_default_error_handler())
     {
         std::vector<std::uint8_t> result;
-        result.reserve(detail::binary_reserve_hint(j));
+        result.reserve(detail::binary_initial_capacity(j));
         vector_writer(result, error_handler).write_cbor(j);
         return result;
     }
@@ -6331,7 +6331,9 @@ public:
     static void to_cbor(const basic_json& j, detail::output_adapter<std::uint8_t> o,
                         const error_handler_t error_handler = detail::binary_writer_default_error_handler())
     {
-        binary_writer<std::uint8_t>(o, error_handler).write_cbor(j);
+        binary_writer<std::uint8_t> writer(o, error_handler);
+        writer.write_cbor(j);
+        writer.flush();
     }
 
     /// @brief create a CBOR serialization of a given JSON value
@@ -6339,7 +6341,9 @@ public:
     static void to_cbor(const basic_json& j, detail::output_adapter<char> o,
                         const error_handler_t error_handler = detail::binary_writer_default_error_handler())
     {
-        binary_writer<char>(o, error_handler).write_cbor(j);
+        binary_writer<char> writer(o, error_handler);
+        writer.write_cbor(j);
+        writer.flush();
     }
 
     /// @brief create a MessagePack serialization of a given JSON value
@@ -6348,7 +6352,7 @@ public:
             const error_handler_t error_handler = error_handler_t::keep)
     {
         std::vector<std::uint8_t> result;
-        result.reserve(detail::binary_reserve_hint(j));
+        result.reserve(detail::binary_initial_capacity(j));
         vector_writer(result, error_handler).write_msgpack(j);
         return result;
     }
@@ -6358,7 +6362,9 @@ public:
     static void to_msgpack(const basic_json& j, detail::output_adapter<std::uint8_t> o,
                            const error_handler_t error_handler = error_handler_t::keep)
     {
-        binary_writer<std::uint8_t>(o, error_handler).write_msgpack(j);
+        binary_writer<std::uint8_t> writer(o, error_handler);
+        writer.write_msgpack(j);
+        writer.flush();
     }
 
     /// @brief create a MessagePack serialization of a given JSON value
@@ -6366,7 +6372,9 @@ public:
     static void to_msgpack(const basic_json& j, detail::output_adapter<char> o,
                            const error_handler_t error_handler = error_handler_t::keep)
     {
-        binary_writer<char>(o, error_handler).write_msgpack(j);
+        binary_writer<char> writer(o, error_handler);
+        writer.write_msgpack(j);
+        writer.flush();
     }
 
     /// @brief create a UBJSON serialization of a given JSON value
@@ -6377,7 +6385,7 @@ public:
             const error_handler_t error_handler = detail::binary_writer_default_error_handler())
     {
         std::vector<std::uint8_t> result;
-        result.reserve(detail::binary_reserve_hint(j));
+        result.reserve(detail::binary_initial_capacity(j));
         vector_writer(result, error_handler).write_ubjson(j, use_size, use_type);
         return result;
     }
@@ -6388,7 +6396,9 @@ public:
                           const bool use_size = false, const bool use_type = false,
                           const error_handler_t error_handler = detail::binary_writer_default_error_handler())
     {
-        binary_writer<std::uint8_t>(o, error_handler).write_ubjson(j, use_size, use_type);
+        binary_writer<std::uint8_t> writer(o, error_handler);
+        writer.write_ubjson(j, use_size, use_type);
+        writer.flush();
     }
 
     /// @brief create a UBJSON serialization of a given JSON value
@@ -6397,7 +6407,9 @@ public:
                           const bool use_size = false, const bool use_type = false,
                           const error_handler_t error_handler = detail::binary_writer_default_error_handler())
     {
-        binary_writer<char>(o, error_handler).write_ubjson(j, use_size, use_type);
+        binary_writer<char> writer(o, error_handler);
+        writer.write_ubjson(j, use_size, use_type);
+        writer.flush();
     }
 
     /// @brief create a BJData serialization of a given JSON value
@@ -6409,7 +6421,7 @@ public:
             const error_handler_t error_handler = detail::binary_writer_default_error_handler())
     {
         std::vector<std::uint8_t> result;
-        result.reserve(detail::binary_reserve_hint(j));
+        result.reserve(detail::binary_initial_capacity(j));
         vector_writer(result, error_handler).write_ubjson(j, use_size, use_type, true, true, version);
         return result;
     }
@@ -6421,7 +6433,9 @@ public:
                           const bjdata_version_t version = bjdata_version_t::draft2,
                           const error_handler_t error_handler = detail::binary_writer_default_error_handler())
     {
-        binary_writer<std::uint8_t>(o, error_handler).write_ubjson(j, use_size, use_type, true, true, version);
+        binary_writer<std::uint8_t> writer(o, error_handler);
+        writer.write_ubjson(j, use_size, use_type, true, true, version);
+        writer.flush();
     }
 
     /// @brief create a BJData serialization of a given JSON value
@@ -6431,7 +6445,9 @@ public:
                           const bjdata_version_t version = bjdata_version_t::draft2,
                           const error_handler_t error_handler = detail::binary_writer_default_error_handler())
     {
-        binary_writer<char>(o, error_handler).write_ubjson(j, use_size, use_type, true, true, version);
+        binary_writer<char> writer(o, error_handler);
+        writer.write_ubjson(j, use_size, use_type, true, true, version);
+        writer.flush();
     }
 
     /// @brief create a BSON serialization of a given JSON value
@@ -6440,7 +6456,7 @@ public:
             const error_handler_t error_handler = detail::binary_writer_default_error_handler())
     {
         std::vector<std::uint8_t> result;
-        result.reserve(detail::binary_reserve_hint(j));
+        result.reserve(detail::binary_initial_capacity(j));
         vector_writer(result, error_handler).write_bson(j);
         return result;
     }
@@ -6450,7 +6466,9 @@ public:
     static void to_bson(const basic_json& j, detail::output_adapter<std::uint8_t> o,
                         const error_handler_t error_handler = detail::binary_writer_default_error_handler())
     {
-        binary_writer<std::uint8_t>(o, error_handler).write_bson(j);
+        binary_writer<std::uint8_t> writer(o, error_handler);
+        writer.write_bson(j);
+        writer.flush();
     }
 
     /// @brief create a BSON serialization of a given JSON value
@@ -6458,7 +6476,9 @@ public:
     static void to_bson(const basic_json& j, detail::output_adapter<char> o,
                         const error_handler_t error_handler = detail::binary_writer_default_error_handler())
     {
-        binary_writer<char>(o, error_handler).write_bson(j);
+        binary_writer<char> writer(o, error_handler);
+        writer.write_bson(j);
+        writer.flush();
     }
 
     /// @brief create a BON8 serialization of a given JSON value
@@ -6466,7 +6486,7 @@ public:
     static std::vector<std::uint8_t> to_bon8(const basic_json& j)
     {
         std::vector<std::uint8_t> result;
-        result.reserve(detail::binary_reserve_hint(j));
+        result.reserve(detail::binary_initial_capacity(j));
         vector_writer(result).write_bon8(j);
         return result;
     }
@@ -6475,14 +6495,18 @@ public:
     /// @sa https://json.nlohmann.me/api/basic_json/to_bon8/
     static void to_bon8(const basic_json& j, detail::output_adapter<std::uint8_t> o)
     {
-        binary_writer<std::uint8_t>(o).write_bon8(j);
+        binary_writer<std::uint8_t> writer(o);
+        writer.write_bon8(j);
+        writer.flush();
     }
 
     /// @brief create a BON8 serialization of a given JSON value
     /// @sa https://json.nlohmann.me/api/basic_json/to_bon8/
     static void to_bon8(const basic_json& j, detail::output_adapter<char> o)
     {
-        binary_writer<char>(o).write_bon8(j);
+        binary_writer<char> writer(o);
+        writer.write_bon8(j);
+        writer.flush();
     }
 
     /// @brief create a JSON value from an input in CBOR format
@@ -6878,72 +6902,6 @@ public:
     /// @name JSON Patch functions
     /// @{
 
-  private:
-    /// the members of an operation of a JSON patch that are of interest
-    struct patch_members
-    {
-        basic_json* op = nullptr;
-        basic_json* path = nullptr;
-        basic_json* value = nullptr;
-        basic_json* from = nullptr;
-    };
-
-    /// @brief find the members "op", "path", "value", and "from" of the object @a val in one pass
-    static patch_members find_patch_members(const basic_json& val, std::true_type /*key_type is string_t*/)
-    {
-        patch_members members;
-        for (auto& member : *val.m_data.m_value.object)
-        {
-            const string_t& key = member.first;
-            switch (key.size())
-            {
-                case 2:
-                    if (key == "op")
-                    {
-                        members.op = &member.second;
-                    }
-                    break;
-                case 4:
-                    if (key == "path")
-                    {
-                        members.path = &member.second;
-                    }
-                    else if (key == "from")
-                    {
-                        members.from = &member.second;
-                    }
-                    break;
-                case 5:
-                    if (key == "value")
-                    {
-                        members.value = &member.second;
-                    }
-                    break;
-                default:
-                    break;
-            }
-        }
-        return members;
-    }
-
-    /// @brief find the members by lookup for key types other than string_t
-    static patch_members find_patch_members(const basic_json& val, std::false_type /*key_type is not string_t*/)
-    {
-        const auto find_member = [&val](const char* name) -> basic_json *
-        {
-            const auto it = val.m_data.m_value.object->find(string_t(name));
-            return it != val.m_data.m_value.object->end() ? &it->second : nullptr;
-        };
-
-        patch_members members;
-        members.op = find_member("op");
-        members.path = find_member("path");
-        members.value = find_member("value");
-        members.from = find_member("from");
-        return members;
-    }
-
-  public:
     /// @brief applies a JSON patch in-place without copying the object
     /// @sa https://json.nlohmann.me/api/basic_json/patch/
     void patch_inplace(const basic_json& json_patch)
@@ -7117,18 +7075,18 @@ public:
         // a diff creates them: apply the operation and return true; return
         // false, with nothing changed, for all others
         string_t token;
-        typename json_pointer::template plain_steps<basic_json> steps;
-        const auto try_plain = [&result, &token, &steps](patch_operations operation, const string_t& path, const patch_members & members, const basic_json & val)
+        const string_t value_key = "value";
+        const auto try_plain = [&result, &token, &value_key](patch_operations operation, const string_t& path, const basic_json & val)
         {
-            if (path.empty())
+            if (path.empty() || path.find_first_of('~') != string_t::npos)
             {
                 return false;
             }
 
             // a missing "value" is reported below
-            const basic_json* const value_ptr = members.value;
-            const basic_json& value = value_ptr != nullptr ? *value_ptr : val;
-            if (operation != patch_operations::remove && value_ptr == nullptr)
+            const auto value_it = val.m_data.m_value.object->find(value_key);
+            const basic_json& value = value_it != val.m_data.m_value.object->end() ? value_it->second : val;
+            if (operation != patch_operations::remove && value_it == val.m_data.m_value.object->end())
             {
                 return false;
             }
@@ -7137,7 +7095,7 @@ public:
             {
                 case patch_operations::replace:
                 {
-                    basic_json* const target = json_pointer::resolve_plain(&result, path, path.size(), token, steps);
+                    basic_json* const target = json_pointer::resolve_plain(&result, path, path.size(), token);
                     if (target == nullptr)
                     {
                         return false;
@@ -7148,7 +7106,7 @@ public:
 
                 case patch_operations::test:
                 {
-                    const basic_json* const target = json_pointer::resolve_plain(&result, path, path.size(), token, steps);
+                    const basic_json* const target = json_pointer::resolve_plain(&result, path, path.size(), token);
                     // a failing test is reported below
                     return target != nullptr && *target == value;
                 }
@@ -7168,17 +7126,13 @@ public:
                     }
                     --slash;
 
-                    basic_json* const parent = json_pointer::resolve_plain(&result, path, slash, token, steps);
+                    basic_json* const parent = json_pointer::resolve_plain(&result, path, slash, token);
                     if (parent == nullptr)
                     {
                         return false;
                     }
                     token.clear();
                     token.append(path.data() + slash + 1, path.size() - slash - 1);
-                    if (token.find_first_of('~') != string_t::npos)
-                    {
-                        return false;
-                    }
 
                     if (operation == patch_operations::add)
                     {
@@ -7234,32 +7188,31 @@ public:
             }
         };
 
-        // the name of the first member to check
-        const string_t op_member = "op";
-
         // iterate and apply the operations
         for (const auto& val : json_patch)
         {
             // wrapper to get a value for an operation
-            const auto get_value = [&val](basic_json * found,
-                                          const string_t& op,
-                                          const char* member,
+            const auto get_value = [&val](const string_t& op,
+                                          const string_t& member,
                                           bool string_type) -> basic_json &
             {
+                // find value
+                auto it = val.m_data.m_value.object->find(member);
+
                 // check if the desired value is present
-                if (JSON_HEDLEY_UNLIKELY(found == nullptr))
+                if (JSON_HEDLEY_UNLIKELY(it == val.m_data.m_value.object->end()))
                 {
-                    detail::throw_patch_missing_member(op, string_t(member), false, val);
+                    detail::throw_patch_missing_member(op, member, false, val);
                 }
 
                 // check if the result is of type string
-                if (JSON_HEDLEY_UNLIKELY(string_type && !found->is_string()))
+                if (JSON_HEDLEY_UNLIKELY(string_type && !it->second.is_string()))
                 {
-                    detail::throw_patch_missing_member(op, string_t(member), true, val);
+                    detail::throw_patch_missing_member(op, member, true, val);
                 }
 
                 // no error: return value
-                return *found;
+                return it->second;
             };
 
             // type check: every element of the array must be an object
@@ -7268,30 +7221,26 @@ public:
                 detail::throw_patch_not_an_array(val);
             }
 
-            const patch_members members = find_patch_members(val, std::is_same<typename object_t::key_type, string_t> {});
-
             // collect mandatory members
-            const string_t& op = *get_value(members.op, op_member, "op", true).m_data.m_value.string;
-            const string_t& path = *get_value(members.path, op, "path", true).m_data.m_value.string;
+            const string_t& op = *get_value("op", "op", true).m_data.m_value.string;
+            const string_t& path = *get_value(op, "path", true).m_data.m_value.string;
             const patch_operations operation = get_op(op);
 
             // the usual operations on a plain pointer are applied without
             // building the pointer's reference tokens; anything else (and
             // every error) is handled below
-            if (try_plain(operation, path, members, val))
+            if (try_plain(operation, path, val))
             {
                 continue;
             }
 
-            // the general operations may change any container
-            steps.reset();
             json_pointer ptr(path);
 
             switch (operation)
             {
                 case patch_operations::add:
                 {
-                    operation_add(ptr, get_value(members.value, "add", "value", false));
+                    operation_add(ptr, get_value("add", "value", false));
                     break;
                 }
 
@@ -7304,13 +7253,13 @@ public:
                 case patch_operations::replace:
                 {
                     // the "path" location must exist - use at()
-                    result.at(ptr) = get_value(members.value, "replace", "value", false);
+                    result.at(ptr) = get_value("replace", "value", false);
                     break;
                 }
 
                 case patch_operations::move:
                 {
-                    const string_t& from_path = *get_value(members.from, "move", "from", true).m_data.m_value.string;
+                    const string_t& from_path = *get_value("move", "from", true).m_data.m_value.string;
                     json_pointer from_ptr(from_path);
 
                     if (JSON_HEDLEY_UNLIKELY(is_proper_prefix(from_ptr, ptr)))
@@ -7333,7 +7282,7 @@ public:
 
                 case patch_operations::copy:
                 {
-                    const string_t& from_path = *get_value(members.from, "copy", "from", true).m_data.m_value.string;
+                    const string_t& from_path = *get_value("copy", "from", true).m_data.m_value.string;
                     const json_pointer from_ptr(from_path);
 
                     // The copy is functionally identical to an "add"
@@ -7351,7 +7300,7 @@ public:
                     {
                         // check if "value" matches the one at "path"
                         // the "path" location must exist - use at()
-                        success = (result.at(ptr) == get_value(members.value, "test", "value", false));
+                        success = (result.at(ptr) == get_value("test", "value", false));
                     }
                     JSON_INTERNAL_CATCH (out_of_range&)
                     {
@@ -7443,18 +7392,13 @@ public:
     /// @brief append an operation @a op for @a path to @a result, with @a value if it is not null
     static void diff_operation(basic_json& result, const char* op, const string_t& path, const basic_json* value)
     {
-        // the keys are distinct and in the order of std::map; the members are
-        // added without looking them up
         basic_json entry(value_t::object);
-        auto& members = *entry.m_data.m_value.object;
-        typename object_t::iterator hint = members.end();
-        json_pointer::add_flattened(members, hint, string_t("op"), string_t(op), detail::priority_tag<2> {});
-        json_pointer::add_flattened(members, hint, string_t("path"), path, detail::priority_tag<2> {});
+        entry["op"] = op;
+        entry["path"] = path;
         if (value != nullptr)
         {
-            json_pointer::add_flattened(members, hint, string_t("value"), *value, detail::priority_tag<2> {});
+            entry["value"] = *value;
         }
-        entry.set_parents();
         result.push_back(std::move(entry));
     }
 
@@ -7476,50 +7420,6 @@ public:
         diff_operation(result, "add", path, &value);
     }
 
-    /// @brief the values in the target of the members of a source object, for the
-    ///        objects of up to 16 members without allocating
-    ///
-    /// reserve() must be called with the number of values that are added.
-    class diff_target_values
-    {
-      public:
-        diff_target_values() = default;
-        diff_target_values(const diff_target_values&) = delete;
-        diff_target_values& operator=(const diff_target_values&) = delete;
-        ~diff_target_values() = default;
-
-        void reserve(const std::size_t n)
-        {
-            if (n > inline_size)
-            {
-                large.resize(n);
-                items = large.data();
-            }
-        }
-
-        void push_back(const basic_json* value)
-        {
-            items[count++] = value;
-        }
-
-        std::size_t size() const noexcept
-        {
-            return count;
-        }
-
-        const basic_json* operator[](const std::size_t i) const noexcept
-        {
-            return items[i];
-        }
-
-      private:
-        static constexpr std::size_t inline_size = 16;
-        std::size_t count = 0;
-        const basic_json** items = inline_items;
-        std::vector<const basic_json*> large{}; // NOLINT(readability-redundant-member-init)
-        const basic_json* inline_items[inline_size]; // NOLINT(cppcoreguidelines-pro-type-member-init,hicpp-member-init)
-    };
-
     /// @brief append "/" and the decimal representation of @a index to @a path
     static void diff_append_index(string_t& path, std::size_t index)
     {
@@ -7600,9 +7500,8 @@ public:
     operations that remove every member of @a source and add every member of
     @a target to @a result, and returns false.
     */
-    template<typename TargetValues>
     static bool diff_object_keys(basic_json& result, const basic_json& source, const basic_json& target,
-                                 const string_t& path, TargetValues& target_values,
+                                 const string_t& path, std::vector<const basic_json*>& target_values,
                                  std::vector<const_iterator>& added)
     {
         // first pass: one lookup per source key
@@ -7816,7 +7715,7 @@ public:
 
             case value_t::object:
             {
-                diff_target_values target_values;
+                std::vector<const basic_json*> target_values;
                 std::vector<const_iterator> added;
                 if (diff_object_keys(result, source, target, path, target_values, added))
                 {

tests/src/unit-binary_writer_sinks.cpp

@@ -492,3 +492,275 @@ TEST_CASE("binary_reserve_hint never over-reserves")
         CHECK(nlohmann::detail::binary_reserve_hint(json({{"a", 1}, {"b", 2}})) == 5);
     }
 }
+
+namespace
+{
+
+// numbers around every boundary of the integer encodings, and floats of each
+// kind (exact in float, needing double, special values)
+std::vector<json> boundary_numbers()
+{
+    std::vector<json> numbers;
+    for (int k = 0; k < 64; ++k)
+    {
+        const std::uint64_t p = std::uint64_t{1} << k;
+        for (const std::uint64_t d :
+                {
+                    p - 1, p, p + 1
+                })
+        {
+            numbers.emplace_back(d);
+            if (d <= static_cast<std::uint64_t>((std::numeric_limits<std::int64_t>::max)()))
+            {
+                numbers.emplace_back(static_cast<std::int64_t>(d));
+                numbers.emplace_back(-static_cast<std::int64_t>(d));
+                numbers.emplace_back(-static_cast<std::int64_t>(d) - 1);
+            }
+        }
+    }
+    numbers.emplace_back((std::numeric_limits<std::int64_t>::min)());
+    numbers.emplace_back((std::numeric_limits<std::uint64_t>::max)());
+    for (const double f :
+            {
+                0.0, -0.0, 1.0, -1.0, 0.5, 0.1, 3.25, 1e300, -1e-300, 3.4028234663852886e38, 1e39,
+                std::numeric_limits<double>::infinity(), -std::numeric_limits<double>::infinity(),
+                std::numeric_limits<double>::quiet_NaN()
+            })
+    {
+        numbers.emplace_back(f);
+    }
+    return numbers;
+}
+
+// the concatenation of the encodings of the elements of @a numbers[first, first + count)
+template<typename Single>
+std::vector<std::uint8_t> concat_singles(const std::vector<json>& numbers, const std::size_t first, const std::size_t count, Single single)
+{
+    std::vector<std::uint8_t> bytes;
+    for (std::size_t i = first; i < first + count; ++i)
+    {
+        const auto one = single(numbers[i]);
+        bytes.insert(bytes.end(), one.begin(), one.end());
+    }
+    return bytes;
+}
+
+} // namespace
+
+// Runs of numbers in arrays are encoded straight into space reserved from the
+// sink, by code that is separate from the encoding of a single value. Both have
+// to agree on every number: an array is its head and the encodings of its
+// elements, one after the other.
+TEST_CASE("runs of numbers in arrays are encoded as the numbers one by one")
+{
+    const std::vector<json> numbers = boundary_numbers();
+
+    // arrays of every length that a one-byte head covers, so that runs are as
+    // long as, and longer and shorter than, what fits one reservation
+    for (const std::size_t count :
+            {
+                1u, 2u, 3u, 5u, 14u, 15u, 23u
+            })
+    {
+        for (std::size_t first = 0; first + count <= numbers.size(); first += 7)
+        {
+            const json array(std::vector<json>(numbers.begin() + static_cast<std::ptrdiff_t>(first),
+                                               numbers.begin() + static_cast<std::ptrdiff_t>(first + count)));
+            CAPTURE(array.dump(-1, ' ', false, json::error_handler_t::replace))
+
+            std::vector<std::uint8_t> expected{static_cast<std::uint8_t>(0x80 + count)};
+            auto elements = concat_singles(numbers, first, count, [](const json & j)
+            {
+                return json::to_cbor(j);
+            });
+            expected.insert(expected.end(), elements.begin(), elements.end());
+            CHECK(json::to_cbor(array) == expected);
+
+            expected.assign(1, static_cast<std::uint8_t>(count <= 15 ? 0x90 + count : 0xDC));
+            if (count > 15)
+            {
+                expected.push_back(0);
+                expected.push_back(static_cast<std::uint8_t>(count));
+            }
+            elements = concat_singles(numbers, first, count, [](const json & j)
+            {
+                return json::to_msgpack(j);
+            });
+            expected.insert(expected.end(), elements.begin(), elements.end());
+            CHECK(json::to_msgpack(array) == expected);
+
+            for (const bool bjdata :
+                    {
+                        false, true
+                    })
+            {
+                expected.assign(1, static_cast<std::uint8_t>('['));
+                elements = concat_singles(numbers, first, count, [bjdata](const json & j)
+                {
+                    return bjdata ? json::to_bjdata(j) : json::to_ubjson(j);
+                });
+                expected.insert(expected.end(), elements.begin(), elements.end());
+                expected.push_back(static_cast<std::uint8_t>(']'));
+                CHECK((bjdata ? json::to_bjdata(array) : json::to_ubjson(array)) == expected);
+            }
+
+            // BON8 has no integers above the int64 range
+            bool representable = true;
+            for (std::size_t i = first; i < first + count; ++i)
+            {
+                representable = representable && bon8_representable(numbers[i]);
+            }
+            if (representable)
+            {
+                expected.assign(1, static_cast<std::uint8_t>(count <= 4 ? 0x80 + count : 0x85));
+                elements = concat_singles(numbers, first, count, [](const json & j)
+                {
+                    return json::to_bon8(j);
+                });
+                expected.insert(expected.end(), elements.begin(), elements.end());
+                if (count > 4)
+                {
+                    expected.push_back(0xFE);
+                }
+                CHECK(json::to_bon8(array) == expected);
+            }
+            else
+            {
+                CHECK_THROWS_AS(json::to_bon8(array), json::out_of_range&);
+            }
+        }
+    }
+}
+
+// An array that is written in more than one reservation (more numbers than fit
+// one), and arrays of an index with more digits, round trip for every format.
+TEST_CASE("large arrays and many digits of indices round trip")
+{
+    json array = json::array();
+    for (int i = 0; i < 12000; ++i)
+    {
+        switch (i % 4)
+        {
+            case 0:
+                array.push_back(i);
+                break;
+            case 1:
+                array.push_back(-i);
+                break;
+            case 2:
+                array.push_back(i * 0.25);
+                break;
+            default:
+                array.push_back(i % 8 == 3);
+                break;
+        }
+    }
+    json document = json::object();
+    document["numbers"] = array;
+    document["nested"] = json::array({array, "text", array});
+
+    CHECK(json::from_cbor(json::to_cbor(document)) == document);
+    CHECK(json::from_msgpack(json::to_msgpack(document)) == document);
+    CHECK(json::from_ubjson(json::to_ubjson(document)) == document);
+    CHECK(json::from_ubjson(json::to_ubjson(document, true, true)) == document);
+    CHECK(json::from_bjdata(json::to_bjdata(document)) == document);
+    CHECK(json::from_bson(json::to_bson(document)) == document);
+    CHECK(json::from_bon8(json::to_bon8(document)) == document);
+}
+
+// Strings and keys are written with their head in one reservation up to a
+// length; around that length, and with every kind of head, they must read back.
+TEST_CASE("strings of every head length round trip")
+{
+    json document = json::object();
+    json array = json::array();
+    for (const std::size_t length :
+            {
+                0u, 1u, 2u, 3u, 4u, 7u, 8u, 15u, 16u, 17u, 23u, 24u, 25u, 31u, 32u, 33u, 63u, 127u, 128u, 129u, 200u, 245u, 246u, 247u, 248u, 249u, 250u, 251u, 252u, 253u, 254u, 255u, 256u, 257u, 300u, 1000u, 65535u, 65536u, 70000u
+            })
+    {
+        const std::string text(length, 'x');
+        array.push_back(text);
+        document[std::string(length, 'k')] = text;
+    }
+    document["array"] = array;
+
+    CHECK(json::from_cbor(json::to_cbor(document)) == document);
+    CHECK(json::from_msgpack(json::to_msgpack(document)) == document);
+    CHECK(json::from_ubjson(json::to_ubjson(document)) == document);
+    CHECK(json::from_ubjson(json::to_ubjson(document, true, true)) == document);
+    CHECK(json::from_bjdata(json::to_bjdata(document)) == document);
+    CHECK(json::from_bson(json::to_bson(document)) == document);
+    CHECK(json::from_bon8(json::to_bon8(document)) == document);
+}
+
+// What a write that throws leaves behind in an output adapter is what was written
+// before it, whatever the adapter and however the adapter sink buffers.
+TEST_CASE("a write that throws leaves the same bytes in every kind of output")
+{
+    json array = json::array();
+    for (int i = 0; i < 700; ++i)
+    {
+        array.push_back(i * 1000);
+        array.push_back("text " + std::to_string(i));
+    }
+    array.push_back("bad \xFF byte");
+    json document = json::object();
+    document["first"] = array;
+    document["second"] = 1;
+
+    const auto strict = json::error_handler_t::strict;
+
+    SECTION("cbor")
+    {
+        std::vector<std::uint8_t> as_vector;
+        std::string as_string;
+        std::ostringstream as_stream;
+        CHECK_THROWS_AS(json::to_cbor(document, as_vector, strict), json::type_error&);
+        CHECK_THROWS_AS(json::to_cbor(document, as_string, strict), json::type_error&);
+        CHECK_THROWS_AS(json::to_cbor(document, as_stream, strict), json::type_error&);
+        CHECK(!as_vector.empty());
+        CHECK(std::string(reinterpret_cast<const char*>(as_vector.data()), as_vector.size()) == as_string); // NOLINT(cppcoreguidelines-pro-type-reinterpret-cast)
+        CHECK(as_stream.str() == as_string);
+    }
+
+    SECTION("msgpack")
+    {
+        std::vector<std::uint8_t> as_vector;
+        std::string as_string;
+        std::ostringstream as_stream;
+        CHECK_THROWS_AS(json::to_msgpack(document, as_vector, strict), json::type_error&);
+        CHECK_THROWS_AS(json::to_msgpack(document, as_string, strict), json::type_error&);
+        CHECK_THROWS_AS(json::to_msgpack(document, as_stream, strict), json::type_error&);
+        CHECK(!as_vector.empty());
+        CHECK(std::string(reinterpret_cast<const char*>(as_vector.data()), as_vector.size()) == as_string); // NOLINT(cppcoreguidelines-pro-type-reinterpret-cast)
+        CHECK(as_stream.str() == as_string);
+    }
+
+    SECTION("ubjson")
+    {
+        std::vector<std::uint8_t> as_vector;
+        std::string as_string;
+        std::ostringstream as_stream;
+        CHECK_THROWS_AS(json::to_ubjson(document, as_vector, false, false, strict), json::type_error&);
+        CHECK_THROWS_AS(json::to_ubjson(document, as_string, false, false, strict), json::type_error&);
+        CHECK_THROWS_AS(json::to_ubjson(document, as_stream, false, false, strict), json::type_error&);
+        CHECK(!as_vector.empty());
+        CHECK(std::string(reinterpret_cast<const char*>(as_vector.data()), as_vector.size()) == as_string); // NOLINT(cppcoreguidelines-pro-type-reinterpret-cast)
+        CHECK(as_stream.str() == as_string);
+    }
+
+    SECTION("bson")
+    {
+        // the size pass finds the string before anything is written
+        std::vector<std::uint8_t> as_vector;
+        std::string as_string;
+        std::ostringstream as_stream;
+        CHECK_THROWS_AS(json::to_bson(document, as_vector, strict), json::type_error&);
+        CHECK_THROWS_AS(json::to_bson(document, as_string, strict), json::type_error&);
+        CHECK_THROWS_AS(json::to_bson(document, as_stream, strict), json::type_error&);
+        CHECK(as_vector.empty());
+        CHECK(as_string.empty());
+        CHECK(as_stream.str().empty());
+    }
+}

tests/src/unit-custom-number-types.cpp

@@ -285,6 +285,55 @@ TEST_CASE("custom number types: integers beyond 64 bits")
         CHECK(wide_json::to_bjdata(wide_json::array({uint_big, uint_big}), true, true) == expected_bjdata_array);
     }
 
+    SECTION("integers in arrays and objects")
+    {
+        // runs of numbers and scalar object members are encoded by loops of
+        // their own for the number types up to 64 bits; a wider number type
+        // takes the generic path, which checks the range of every integer
+        const wide_json run = wide_json::array({make_int(1), make_int(-3), make_uint(7), wide_json(1.5), int_big, make_int(5)});
+        const wide_json run_negative = wide_json::array({make_int(1), make_int(-3), int_big_negative});
+        const wide_json run_unsigned = wide_json::array({make_uint(1), make_uint(7), uint_big});
+        const wide_json members = wide_json::object({{"a", make_int(1)}, {"b", wide_json(1.5)}, {"c", int_big}});
+
+        const std::vector<wide_json> failing = {run, run_negative, run_unsigned, members, wrap(run), wrap(members)};
+        for (const auto& value : failing)
+        {
+            CHECK_THROWS_AS(_ = wide_json::to_cbor(value), out_of_range&);
+            CHECK_THROWS_AS(_ = wide_json::to_msgpack(value), out_of_range&);
+            CHECK_THROWS_AS(_ = wide_json::to_bson(wrap(value)), out_of_range&);
+            CHECK_THROWS_AS(_ = wide_json::to_bon8(value), out_of_range&);
+        }
+
+        // a failing write leaves no more than the elements before the failing one
+        std::vector<std::uint8_t> out;
+        CHECK_THROWS_AS(wide_json::to_cbor(run, out), out_of_range&);
+        // (the head of the array, 1, -3, 7, and 1.5 as a single-precision float)
+        CHECK(out == std::vector<std::uint8_t>({0x86, 0x01, 0x22, 0x07, 0xFA, 0x3F, 0xC0, 0x00, 0x00}));
+        out.clear();
+        CHECK_THROWS_AS(wide_json::to_bon8(run, out), out_of_range&);
+        CHECK(!out.empty());
+
+        // the same values that fit are written exactly as with 64-bit number types
+        const wide_json fits = wide_json::array({make_int(1), make_int(-3), make_uint(7), wide_json(1.5), make_int(min_int64), make_int(max_int64), make_int(200), make_int(-70000)});
+        const nlohmann::json fits64 = nlohmann::json::array({1, -3, 7, 1.5, (std::numeric_limits<std::int64_t>::min)(), (std::numeric_limits<std::int64_t>::max)(), 200, -70000});
+        CHECK(wide_json::to_cbor(fits) == nlohmann::json::to_cbor(fits64));
+        CHECK(wide_json::to_msgpack(fits) == nlohmann::json::to_msgpack(fits64));
+        CHECK(wide_json::to_ubjson(fits) == nlohmann::json::to_ubjson(fits64));
+        CHECK(wide_json::to_ubjson(fits, true, true) == nlohmann::json::to_ubjson(fits64, true, true));
+        CHECK(wide_json::to_bjdata(fits) == nlohmann::json::to_bjdata(fits64));
+        CHECK(wide_json::to_bjdata(fits, true, true) == nlohmann::json::to_bjdata(fits64, true, true));
+        CHECK(wide_json::to_bon8(fits) == nlohmann::json::to_bon8(fits64));
+        CHECK(wide_json::to_bson(wrap(fits)) == nlohmann::json::to_bson(nlohmann::json::object({{"v", fits64}})));
+
+        const wide_json fit_members = wide_json::object({{"a", make_int(1)}, {"b", make_int(-300)}, {"c", make_uint(max_uint64)}, {"d", wide_json(2.5)}});
+        const nlohmann::json fit_members64 = nlohmann::json::object({{"a", 1}, {"b", -300}, {"c", (std::numeric_limits<std::uint64_t>::max)()}, {"d", 2.5}});
+        CHECK(wide_json::to_cbor(fit_members) == nlohmann::json::to_cbor(fit_members64));
+        CHECK(wide_json::to_msgpack(fit_members) == nlohmann::json::to_msgpack(fit_members64));
+        CHECK(wide_json::to_ubjson(fit_members) == nlohmann::json::to_ubjson(fit_members64));
+        CHECK(wide_json::to_bjdata(fit_members) == nlohmann::json::to_bjdata(fit_members64));
+        CHECK(wide_json::to_bson(fit_members) == nlohmann::json::to_bson(fit_members64));
+    }
+
     SECTION("BJData ND-array")
     {
         // an ND-array element beyond 64 bits does not fit any dtype, so the