Write binary formats through reserved sink space and number runs (#5885)

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/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

@@ -6322,7 +6322,7 @@ 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;
     }
@@ -6332,7 +6332,9 @@ 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
@@ -6340,7 +6342,9 @@ 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
@@ -6349,7 +6353,7 @@ 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;
     }
@@ -6359,7 +6363,9 @@ 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
@@ -6367,7 +6373,9 @@ 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
@@ -6378,7 +6386,7 @@ 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;
     }
@@ -6389,7 +6397,9 @@ 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
@@ -6398,7 +6408,9 @@ 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
@@ -6410,7 +6422,7 @@ 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;
     }
@@ -6422,7 +6434,9 @@ 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
@@ -6432,7 +6446,9 @@ 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
@@ -6441,7 +6457,7 @@ 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;
     }
@@ -6451,7 +6467,9 @@ 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
@@ -6459,7 +6477,9 @@ 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
@@ -6467,7 +6487,7 @@ 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;
     }
@@ -6476,14 +6496,18 @@ 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

tests/src/unit-binary_writer_sinks.cpp

@@ -492,3 +492,278 @@ 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);
+}
+
+#if !defined(JSON_NOEXCEPTION)
+// 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. (Without
+// exceptions, the writes in CHECK_THROWS_AS do not run.)
+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());
+    }
+}
+#endif

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

@@ -285,6 +285,58 @@ 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&);
+        }
+
+#if !defined(JSON_NOEXCEPTION)
+        // a failing write leaves no more than the elements before the failing one
+        // (without exceptions, the writes in CHECK_THROWS_AS do not run)
+        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());
+#endif
+
+        // 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