Merge branch 'develop' into cbor-integer-writer

Signed-off-by: Niels Lohmann <mail@nlohmann.me>
2026-10-16 19:18:12 +00:00 · 2026-10-15 20:26:44 +02:00
commit d60ac0aff1
@@ -9,11 +9,12 @@
 #pragma once

 #include <cstdint> // uint64_t
-#if !defined(__SIZEOF_INT128__) && defined(_MSC_VER) && (defined(_M_X64) || defined(_M_ARM64))
-    #include <intrin0.h> // __umulh, _umul128
+#include <cstring> // memcpy
+#if defined(_MSC_VER) && (defined(_M_X64) || defined(_M_ARM64)) && (!defined(__SIZEOF_INT128__) || (!defined(__GNUC__) && !defined(__clang__)))
+    #include <intrin0.h> // __umulh, _umul128, _BitScanForward64, _BitScanReverse64
 #endif

-#include <nlohmann/detail/abi_macros.hpp>
+#include <nlohmann/detail/macro_scope.hpp> // JSON_HEDLEY_ALWAYS_INLINE, NLOHMANN_JSON_NAMESPACE_BEGIN

 // Portable bit-level helpers for the number and string scanners. They use
 // compiler builtins or platform-specific intrinsics where available and plain
@@ -28,6 +29,10 @@ inline int count_leading_zeros(std::uint64_t x) noexcept
 {
 #if defined(__GNUC__) || defined(__clang__)
    return __builtin_clzll(x);
+#elif defined(_MSC_VER) && (defined(_M_X64) || defined(_M_ARM64))
+    unsigned long index = 0; // NOLINT(runtime/int): the type _BitScan*64 takes
+    _BitScanReverse64(&index, x);
+    return 63 - static_cast<int>(index);
 #else
    int n = 0;
    for (int shift = 32; shift != 0; shift >>= 1)
@@ -47,6 +52,10 @@ inline int count_trailing_zeros(std::uint64_t x) noexcept
 {
 #if defined(__GNUC__) || defined(__clang__)
    return __builtin_ctzll(x);
+#elif defined(_MSC_VER) && (defined(_M_X64) || defined(_M_ARM64))
+    unsigned long index = 0; // NOLINT(runtime/int): the type _BitScan*64 takes
+    _BitScanForward64(&index, x);
+    return static_cast<int>(index);
 #else
    int n = 0;
    for (int shift = 32; shift != 0; shift >>= 1)
@@ -94,15 +103,27 @@ inline uint128_parts full_multiplication(std::uint64_t a, std::uint64_t b) noexc
 #endif
 }

-/// eight bytes as a little-endian word (compilers fold this into one load on
-/// little-endian targets)
-inline std::uint64_t read_eight_bytes(const char* p) noexcept
+/// eight bytes as a little-endian word (a single load on little-endian
+/// targets; always inlined, as GCC otherwise calls it in the number loops)
+JSON_HEDLEY_ALWAYS_INLINE std::uint64_t read_eight_bytes(const unsigned char* b) noexcept
 {
-    const auto* b = reinterpret_cast<const unsigned char*>(p); // NOLINT(cppcoreguidelines-pro-type-reinterpret-cast)
+#if defined(_MSC_VER) || defined(__x86_64__) || defined(__i386__) || (defined(__BYTE_ORDER__) && defined(__ORDER_LITTLE_ENDIAN__) && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__)
+    // the byte order already matches (all MSVC targets are little-endian)
+    std::uint64_t result = 0;
+    std::memcpy(&result, b, sizeof(result));
+    return result;
+#else
    return static_cast<std::uint64_t>(b[0]) | (static_cast<std::uint64_t>(b[1]) << 8u)
           | (static_cast<std::uint64_t>(b[2]) << 16u) | (static_cast<std::uint64_t>(b[3]) << 24u)
           | (static_cast<std::uint64_t>(b[4]) << 32u) | (static_cast<std::uint64_t>(b[5]) << 40u)
           | (static_cast<std::uint64_t>(b[6]) << 48u) | (static_cast<std::uint64_t>(b[7]) << 56u);
+#endif
+}
+
+/// eight bytes as a little-endian word
+JSON_HEDLEY_ALWAYS_INLINE std::uint64_t read_eight_bytes(const char* p) noexcept
+{
+    return read_eight_bytes(reinterpret_cast<const unsigned char*>(p)); // NOLINT(cppcoreguidelines-pro-type-reinterpret-cast)
 }

 }  // namespace detail
@@ -80,7 +80,11 @@ void get_arithmetic_value(const BasicJsonType& j, ArithmeticType& val)
    {
        case value_t::number_unsigned:
        {
-            val = static_cast<ArithmeticType>(*j.template get_ptr<const typename BasicJsonType::number_unsigned_t*>());
+            // never null: a basic_json whose number_unsigned_t is number_integer_t
+            // stores no number_unsigned values (#2573)
+            const auto* ptr = j.template get_ptr<const typename BasicJsonType::number_unsigned_t*>();
+            JSON_ASSERT(ptr != nullptr);
+            val = static_cast<ArithmeticType>(*ptr);
            break;
        }
        case value_t::number_integer:
@@ -150,7 +154,7 @@ inline void from_json(const BasicJsonType& j, typename BasicJsonType::number_flo
 }

 template<typename BasicJsonType>
-inline void from_json(const BasicJsonType& j, typename BasicJsonType::number_unsigned_t& val)
+inline void from_json(const BasicJsonType& j, number_unsigned_overload_t<BasicJsonType>& val)
 {
    get_arithmetic_value(j, val);
 }
@@ -266,6 +270,22 @@ template<typename ConstructibleArrayType>
 void from_json_array_reserve(ConstructibleArrayType& /*arr*/, std::size_t /*size*/, priority_tag<0> /*unused*/)
 {}

+// append through push_back() where available; containers without it (such as
+// std::set) fall back to insert(end(), ...), which is what std::inserter did
+template<typename ConstructibleArrayType, typename ValueType>
+auto from_json_array_append(ConstructibleArrayType& arr, ValueType&& value, priority_tag<1> /*unused*/)
+-> decltype(arr.push_back(std::forward<ValueType>(value)), void())
+{
+    arr.push_back(std::forward<ValueType>(value));
+}
+
+template<typename ConstructibleArrayType, typename ValueType>
+void from_json_array_append(ConstructibleArrayType& arr, ValueType&& value, priority_tag<0> /*unused*/)
+{
+    using std::end;
+    arr.insert(end(arr), std::forward<ValueType>(value));
+}
+
 template<typename BasicJsonType, typename ConstructibleArrayType,
         enable_if_t<
             std::is_assignable<ConstructibleArrayType&, ConstructibleArrayType>::value,
@@ -275,17 +295,15 @@ auto from_json_array_impl(const BasicJsonType& j, ConstructibleArrayType& arr, p
    j.template get<typename ConstructibleArrayType::value_type>(),
    void())
 {
-    using std::end;
-
    ConstructibleArrayType ret;
-    from_json_array_reserve(ret, j.size(), priority_tag<1> {});
-    std::transform(j.begin(), j.end(),
-                   std::inserter(ret, end(ret)), [](const BasicJsonType & i)
+    const auto* inner_array = j.template get_ptr<const typename BasicJsonType::array_t*>();
+    from_json_array_reserve(ret, inner_array->size(), priority_tag<1> {});
+    for (const auto& elem : *inner_array)
    {
        // get<BasicJsonType>() returns *this, this won't call a from_json
        // method when value_type is BasicJsonType
-        return i.template get<typename ConstructibleArrayType::value_type>();
-    });
+        from_json_array_append(ret, elem.template get<typename ConstructibleArrayType::value_type>(), priority_tag<1> {});
+    }
    arr = std::move(ret);
 }

@@ -371,6 +389,23 @@ template<typename ConstructibleObjectType>
 inline void from_json_object_reserve(ConstructibleObjectType& /*obj*/, std::size_t /*size*/, priority_tag<0> /*unused*/)
 {}

+// the keys of the source are sorted for the usual object types, so hinting the
+// end makes every insertion constant-time (for containers that order their keys
+// the same way); containers without emplace_hint() (such as ordered_map) fall
+// back to emplace()
+template<typename ConstructibleObjectType, typename Key, typename Value>
+auto from_json_object_emplace(ConstructibleObjectType& obj, Key&& key, Value&& value, priority_tag<1> /*unused*/)
+-> decltype(obj.emplace_hint(obj.end(), std::forward<Key>(key), std::forward<Value>(value)), void())
+{
+    obj.emplace_hint(obj.end(), std::forward<Key>(key), std::forward<Value>(value));
+}
+
+template<typename ConstructibleObjectType, typename Key, typename Value>
+void from_json_object_emplace(ConstructibleObjectType& obj, Key&& key, Value&& value, priority_tag<0> /*unused*/)
+{
+    obj.emplace(std::forward<Key>(key), std::forward<Value>(value));
+}
+
 template<typename BasicJsonType, typename ConstructibleObjectType,
         enable_if_t<is_constructible_object_type<BasicJsonType, ConstructibleObjectType>::value, int> = 0>
 inline void from_json(const BasicJsonType& j, ConstructibleObjectType& obj)
@@ -385,7 +420,7 @@ inline void from_json(const BasicJsonType& j, ConstructibleObjectType& obj)
    from_json_object_reserve(ret, inner_object->size(), priority_tag<1> {});
    for (const auto& p : *inner_object)
    {
-        ret.emplace(p.first, p.second.template get<typename ConstructibleObjectType::mapped_type>());
+        from_json_object_emplace(ret, p.first, p.second.template get<typename ConstructibleObjectType::mapped_type>(), priority_tag<1> {});
    }
    obj = std::move(ret);
 }
@@ -406,7 +441,11 @@ inline void from_json(const BasicJsonType& j, ArithmeticType& val)
    {
        case value_t::number_unsigned:
        {
-            val = static_cast<ArithmeticType>(*j.template get_ptr<const typename BasicJsonType::number_unsigned_t*>());
+            // never null: a basic_json whose number_unsigned_t is number_integer_t
+            // stores no number_unsigned values (#2573)
+            const auto* ptr = j.template get_ptr<const typename BasicJsonType::number_unsigned_t*>();
+            JSON_ASSERT(ptr != nullptr);
+            val = static_cast<ArithmeticType>(*ptr);
            break;
        }
        case value_t::number_integer:
@@ -4,6 +4,7 @@
 // |_____|_____|_____|_|___|  https://github.com/nlohmann/json
 //
 // SPDX-FileCopyrightText: 2009 Florian Loitsch <https://florian.loitsch.com/>
+// SPDX-FileCopyrightText: 2025 Victor Zverovich <https://github.com/vitaut/zmij>
 // SPDX-FileCopyrightText: 2013-2026 Niels Lohmann <https://nlohmann.me>
 // SPDX-License-Identifier: MIT

@@ -11,11 +12,32 @@

 #include <array> // array
 #include <cmath>   // signbit, isfinite
+#include <cstddef> // size_t
 #include <cstdint> // intN_t, uintN_t
 #include <cstring> // memcpy, memmove
 #include <limits> // numeric_limits
 #include <type_traits> // conditional

+#ifdef _MSC_VER
+    #include <cstdlib> // _byteswap_uint64
+#endif
+
+// SSE2 (every x86-64 CPU) and NEON (every 64-bit Arm CPU) convert the 16
+// digits of a double at once
+#if defined(__x86_64__) || (defined(_M_X64) && !defined(_M_ARM64EC))
+    #include <emmintrin.h>
+    #define JSON_DTOA_SSE2 1
+    #define JSON_DTOA_NEON 0
+#elif (defined(__aarch64__) || defined(_M_ARM64)) && !defined(_M_ARM64EC) && !defined(__ARM_BIG_ENDIAN)
+    #include <arm_neon.h>
+    #define JSON_DTOA_SSE2 0
+    #define JSON_DTOA_NEON 1
+#else
+    #define JSON_DTOA_SSE2 0
+    #define JSON_DTOA_NEON 0
+#endif
+
+#include <nlohmann/detail/conversions/zmij.hpp>
 #include <nlohmann/detail/macro_scope.hpp>

 NLOHMANN_JSON_NAMESPACE_BEGIN
@@ -1047,6 +1069,381 @@ inline char* format_buffer(char* buf, int len, int decimal_exponent,
    return append_exponent(buf, n - 1);
 }

+/// eight decimal digits (a value below 10^8) as bytes 0..9, the first digit
+/// in the most significant byte: three steps that divide all lanes at once
+/// by a multiplication (the conversion of Xiang JunBo, as in Zmij)
+inline std::uint64_t eight_digit_bytes(std::uint64_t abcdefgh) noexcept
+{
+    const std::uint64_t abcd_efgh = abcdefgh + (((std::uint64_t{1} << 32u) - 10000u) * ((abcdefgh * (((std::uint64_t{1} << 40u) / 10000u) + 1u)) >> 40u));
+    const std::uint64_t ab_cd_ef_gh = abcd_efgh + (((std::uint64_t{1} << 16u) - 100u) * (((abcd_efgh * (((std::uint64_t{1} << 19u) / 100u) + 1u)) >> 19u) & 0x7F0000007Fu));
+    return ab_cd_ef_gh + (((std::uint64_t{1} << 8u) - 10u) * (((ab_cd_ef_gh * (((std::uint64_t{1} << 10u) / 10u) + 1u)) >> 10u) & 0x000F000F000F000Fu));
+}
+
+/// store the bytes of v, the most significant one first (one byte swap and
+/// one store where the byte order is known: compilers do not reliably merge
+/// the byte stores once this is inlined)
+inline void store_msb_first(char* p, std::uint64_t v) noexcept
+{
+#if defined(__BYTE_ORDER__) && defined(__ORDER_LITTLE_ENDIAN__) && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
+    v = __builtin_bswap64(v);
+    std::memcpy(p, &v, sizeof(v));
+#elif defined(__BYTE_ORDER__) && defined(__ORDER_BIG_ENDIAN__) && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__
+    std::memcpy(p, &v, sizeof(v));
+#elif defined(_MSC_VER) // (little-endian on all its targets)
+    v = _byteswap_uint64(v);
+    std::memcpy(p, &v, sizeof(v));
+#else
+    for (unsigned i = 0; i < 8; ++i)
+    {
+        p[i] = static_cast<char>(v >> (56u - (8u * i)));
+    }
+#endif
+}
+
+/*!
+@brief digits * 10^exp for a double, in the layout of format_buffer()
+
+The layout is that of format_buffer() with min_exp -4 and max_exp 15 (the
+digits10 of double). The digits are converted eight at a time and placed
+with fixed-size moves instead of per-digit loops and moves of the buffer.
+
+@param[in] digits  the digits (not 0, at most 17 digits; trailing zeros allowed)
+@param[in] exp     the decimal exponent of the last digit
+@return a pointer past the text; up to 41 bytes at @a first are written
+        (some beyond the returned end)
+*/
+JSON_HEDLEY_NON_NULL(1)
+JSON_HEDLEY_RETURNS_NON_NULL
+inline char* write_decimal(char* first, std::uint64_t digits, int exp) noexcept
+{
+    JSON_ASSERT(digits != 0 && digits < 100000000000000000u);
+    const std::uint64_t upper = digits / 100000000u;
+    const std::uint64_t b0 = upper / 100000000u; // (one digit: it is its own byte)
+    const std::uint64_t b1 = eight_digit_bytes(upper % 100000000u);
+    const std::uint64_t b2 = eight_digit_bytes(digits % 100000000u);
+    // leading and trailing zero digits: zero bytes, counted without division
+    int leading = 16;
+    int zeros = 16;
+    if (b0 != 0)
+    {
+        leading = count_leading_zeros(b0) / 8;
+    }
+    else if (b1 != 0)
+    {
+        leading = 8 + (count_leading_zeros(b1) / 8);
+    }
+    else
+    {
+        leading += count_leading_zeros(b2) / 8;
+    }
+    if (b2 != 0)
+    {
+        zeros = count_trailing_zeros(b2) / 8;
+    }
+    else if (b1 != 0)
+    {
+        zeros = 8 + (count_trailing_zeros(b1) / 8);
+    }
+    // (else: 16, b0 is the one digit that is not 0)
+    // the digits as text at text + leading, then '0's, so that fixed-size
+    // moves need not check how many digits there are
+    std::array<char, 64> text; // NOLINT(cppcoreguidelines-pro-type-member-init,hicpp-member-init): written before read
+    store_msb_first(text.data(), b0 + 0x3030303030303030u);
+    store_msb_first(text.data() + 8, b1 + 0x3030303030303030u);
+    store_msb_first(text.data() + 16, b2 + 0x3030303030303030u);
+    std::memset(text.data() + 24, '0', 40);
+    const int k = 24 - leading - zeros; // significant digits
+    const int n = k + exp + zeros;      // position of the decimal point after the first digit
+    const char* const s0 = text.data() + leading;
+
+    if (-4 < n && n <= 15)
+    {
+        // "0.[000]digits" (n <= 0) is the digits after 1 - n leading '0's
+        // with the point after the first; "digits[000].0" (n >= k) and
+        // "dig.its" put the point after n characters
+        const int pad = n <= 0 ? 1 - n : 0;
+        const char* const s = s0 - pad;
+        const int len = k + pad;
+        const int point = n + pad;
+        std::memcpy(first, s, 16);
+        std::memcpy(first + point + 1, s + point, 24);
+        first[point] = '.';
+        return first + (point >= len ? point + 2 : len + 1);
+    }
+
+    // d.igitse+XX, with at least two exponent digits (as append_exponent())
+    std::memcpy(first, s0, 16);
+    std::memcpy(first + 2, s0 + 1, 16);
+    first[1] = '.';
+    char* const end = first + (k == 1 ? 1 : k + 1);
+    const int e = n - 1;
+    const auto ea = e < 0 ? 0u - static_cast<unsigned>(e) : static_cast<unsigned>(e); // (unsigned: no signed overflow to assume)
+    const bool three = ea >= 100;
+    end[0] = 'e';
+    end[1] = e < 0 ? '-' : '+';
+    end[2] = static_cast<char>('0' + (three ? ea / 100 : (ea / 10) % 10));
+    end[3] = static_cast<char>('0' + (three ? (ea / 10) % 10 : ea % 10));
+    end[4] = static_cast<char>('0' + (ea % 10));
+    return end + (three ? 5 : 4);
+}
+
+/*!
+@brief the shortest decimal of a positive double (Zmij), as write_decimal()
+writes it
+
+For a normal double, the shorter candidate has 15 or 16 digits: they are
+converted at once (two halves of eight digits) and followed by the digit
+after them, if there is one, without the multiplication and division by 10
+that counting the digits of one number would take. The fixed layouts move
+the digits after the point by one byte.
+
+@return a pointer past the text; up to 41 bytes at @a first are written
+        (some beyond the returned end)
+*/
+JSON_HEDLEY_NON_NULL(1)
+JSON_HEDLEY_RETURNS_NON_NULL
+inline char* write_shortest(char* first, const zmij::shortest_decimal d) noexcept
+{
+    const std::uint64_t sig = d.integral;
+    if (JSON_HEDLEY_UNLIKELY(sig < 100000000000000u || sig >= 10000000000000000u))
+    {
+        // (subnormals)
+        return d.has_digit ? write_decimal(first, (sig * 10) + d.digit, d.exponent) : write_decimal(first, sig, d.exponent + 1);
+    }
+    const bool sixteen = sig >= 1000000000000000u; // (else 15 digits)
+    const int last = d.has_digit ? d.digit : 0;
+    const std::uint64_t upper = sig / 100000000u;
+#if JSON_DTOA_SSE2
+    // NOLINTBEGIN(portability-simd-intrinsics)
+    // the two halves in the 64-bit lanes, each as abcd * 2^32 + efgh, then as
+    // bytes (as eight_digit_bytes(), one lane each)
+    const __m128i x = _mm_set_epi64x(static_cast<long long>(sig - (upper * 100000000u)), static_cast<long long>(upper)); // NOLINT(runtime/int)
+    const __m128i abcd = _mm_srli_epi64(_mm_mul_epu32(x, _mm_set1_epi64x(109951163)), 40); // 2^40 / 10000 + 1
+    const __m128i abcd_efgh = _mm_add_epi64(x, _mm_mul_epu32(abcd, _mm_set1_epi64x(4294957296))); // 2^32 - 10000
+    // 32-bit lanes in the order of the text: abcd, efgh of both halves
+    const __m128i fours = _mm_shuffle_epi32(abcd_efgh, _MM_SHUFFLE(2, 3, 0, 1));
+    const __m128i ab = _mm_srli_epi16(_mm_mulhi_epu16(fours, _mm_set1_epi32(5243)), 3);
+    const __m128i ab_cd = _mm_or_si128(_mm_slli_epi32(_mm_sub_epi16(fours, _mm_mullo_epi16(ab, _mm_set1_epi32(100))), 16), ab);
+    // 16-bit lanes ab (< 100) -> bytes a, b: 256 * ab - 2559 * (ab / 10)
+    const __m128i bytes = _mm_sub_epi16(_mm_slli_epi16(ab_cd, 8), _mm_mullo_epi16(_mm_set1_epi16(2559), _mm_mulhi_epu16(ab_cd, _mm_set1_epi16(6554))));
+    // the last digit that is not 0 (sig is not 0)
+    const auto nonzero = static_cast<std::uint64_t>(_mm_movemask_epi8(_mm_cmpgt_epi8(bytes, _mm_setzero_si128())));
+    const int digits = 63 - count_leading_zeros(nonzero) + (sixteen ? 1 : 0); // without trailing zeros
+    const __m128i chars = _mm_add_epi8(bytes, _mm_set1_epi8('0'));
+    // the 16 characters from the first digit
+    const __m128i s = sixteen ? chars : _mm_or_si128(_mm_srli_si128(chars, 1), _mm_slli_si128(_mm_cvtsi32_si128('0' + last), 15));
+    const char s16 = static_cast<char>(sixteen ? '0' + last : '0'); // the 17th
+    const auto store_16 = [&s](char* p) noexcept
+    {
+        std::memcpy(p, &s, 16);
+    };
+    const char first_digit = static_cast<char>(_mm_cvtsi128_si32(s));
+    // NOLINTEND(portability-simd-intrinsics)
+#elif JSON_DTOA_NEON
+    // as with SSE2: the halves in 32-bit lanes, then abcd, efgh of both
+    const uint32x2_t halves = vcreate_u32(upper | ((sig - (upper * 100000000u)) << 32u));
+    const uint32x2_t abcd = vmovn_u64(vshrq_n_u64(vmull_n_u32(halves, static_cast<std::uint32_t>(((std::uint64_t{1} << 40u) / 10000u) + 1u)), 40));
+    const uint32x2_t efgh = vmls_n_u32(halves, abcd, 10000u);
+    const uint32x4_t fours = vcombine_u32(vzip1_u32(abcd, efgh), vzip2_u32(abcd, efgh));
+    const uint32x4_t ab = vshrq_n_u32(vmulq_n_u32(fours, 5243u), 19);
+    const uint16x8_t ab_cd = vreinterpretq_u16_u32(vorrq_u32(ab, vshlq_n_u32(vmlsq_n_u32(fours, ab, 100u), 16)));
+    const uint16x8_t tens = vshrq_n_u16(vmulq_n_u16(ab_cd, 103u), 10);
+    const uint8x16_t bytes = vreinterpretq_u8_u16(vorrq_u16(tens, vshlq_n_u16(vmlsq_n_u16(ab_cd, tens, 10u), 8)));
+    // the last digit that is not 0 (sig is not 0): a nibble per byte
+    const std::uint64_t nonzero = vget_lane_u64(vreinterpret_u64_u8(vshrn_n_u16(vreinterpretq_u16_u8(vtstq_u8(bytes, bytes)), 4)), 0);
+    const int digits = ((63 - count_leading_zeros(nonzero)) / 4) + (sixteen ? 1 : 0); // without trailing zeros
+    const uint8x16_t chars = vaddq_u8(bytes, vdupq_n_u8('0'));
+    // the 16 characters from the first digit
+    const uint8x16_t s = sixteen ? chars : vextq_u8(chars, vdupq_n_u8(static_cast<std::uint8_t>('0' + last)), 1);
+    const char s16 = static_cast<char>(sixteen ? '0' + last : '0'); // the 17th
+    const auto store_16 = [&s](char* p) noexcept
+    {
+        vst1q_u8(reinterpret_cast<std::uint8_t*>(p), s); // NOLINT(cppcoreguidelines-pro-type-reinterpret-cast)
+    };
+    const auto first_digit = static_cast<char>(vgetq_lane_u8(s, 0));
+#else
+    const std::uint64_t hi = eight_digit_bytes(upper);
+    const std::uint64_t lo = eight_digit_bytes(sig - (upper * 100000000u));
+    // trailing zero digits: zero bytes (sig is not 0)
+    const int zeros = lo != 0 ? count_trailing_zeros(lo) / 8 : 8 + (count_trailing_zeros(hi) / 8);
+    const int digits = 15 - zeros + (sixteen ? 1 : 0); // without trailing zeros
+    // the 16 characters from the first digit
+    const std::uint64_t s_hi = (sixteen ? hi : (hi << 8u) | (lo >> 56u)) + 0x3030303030303030u;
+    const std::uint64_t s_lo = (sixteen ? lo : (lo << 8u) | static_cast<std::uint64_t>(last)) + 0x3030303030303030u;
+    const char s16 = static_cast<char>(sixteen ? '0' + last : '0'); // the 17th
+    const auto store_16 = [s_hi, s_lo](char* p) noexcept
+    {
+        store_msb_first(p, s_hi);
+        store_msb_first(p + 8, s_lo);
+    };
+    const auto first_digit = static_cast<char>(s_hi >> 56u);
+#endif
+    const int len = d.has_digit ? 16 + (sixteen ? 1 : 0) : digits; // significant digits
+    const int n = 16 + (sixteen ? 1 : 0) + d.exponent;           // digits before the point
+
+    if (JSON_HEDLEY_LIKELY(n >= 1 && n <= 15))
+    {
+        // "dig.its" and "digits[000].0": the digits after the point move by
+        // one byte ('0's follow the digits)
+#if JSON_DTOA_SSE2
+        // NOLINTBEGIN(portability-simd-intrinsics)
+        // (in the register: reading the digits back from memory right after
+        // storing them waits until the stores are done)
+        const __m128i at = _mm_set1_epi8(static_cast<char>(n));
+        const __m128i index = _mm_setr_epi8(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15);
+        const __m128i before = _mm_cmpgt_epi8(at, index);
+        const __m128i after = _mm_cmpgt_epi8(index, at);
+        const __m128i text = _mm_or_si128(_mm_or_si128(_mm_and_si128(s, before), _mm_and_si128(_mm_slli_si128(s, 1), after)),
+                                          _mm_andnot_si128(_mm_or_si128(before, after), _mm_set1_epi8('.')));
+        std::memcpy(first, &text, 16);
+        first[16] = static_cast<char>(_mm_extract_epi16(s, 7) >> 8);
+        first[17] = s16;
+        // NOLINTEND(portability-simd-intrinsics)
+#elif JSON_DTOA_NEON
+        const uint8x16_t index = vcombine_u8(vcreate_u8(0x0706050403020100u), vcreate_u8(0x0F0E0D0C0B0A0908u));
+        const uint8x16_t at = vdupq_n_u8(static_cast<std::uint8_t>(n));
+        const uint8x16_t after_point = vbslq_u8(vcgtq_u8(index, at), vextq_u8(vdupq_n_u8(0), s, 15), vdupq_n_u8('.'));
+        vst1q_u8(reinterpret_cast<std::uint8_t*>(first), vbslq_u8(vcltq_u8(index, at), s, after_point)); // NOLINT(cppcoreguidelines-pro-type-reinterpret-cast)
+        first[16] = static_cast<char>(vgetq_lane_u8(s, 15));
+        first[17] = s16;
+#else
+        store_16(first);
+        first[16] = s16;
+        std::uint64_t after_point[2]; // NOLINT(cppcoreguidelines-avoid-c-arrays,hicpp-avoid-c-arrays,modernize-avoid-c-arrays,cppcoreguidelines-pro-type-member-init,hicpp-member-init): written before read
+        std::memcpy(after_point, first + n, 16);
+        std::memcpy(first + n + 1, after_point, 16);
+        first[n] = '.';
+#endif
+        return first + (n >= len ? n + 2 : len + 1);
+    }
+    if (n <= 0 && n > -4)
+    {
+        // "0.[000]digits"
+        std::memset(first, '0', 8);
+        first[1] = '.';
+        store_16(first + 2 - n);
+        first[18 - n] = s16;
+        return first + 2 - n + len;
+    }
+    // d.igitse+XX, with at least two exponent digits (as append_exponent())
+    store_16(first + 1);
+    first[17] = s16;
+    first[0] = first_digit;
+    first[1] = '.';
+    char* const end = first + (len == 1 ? 1 : len + 1);
+    const int e = n - 1;
+    const auto ea = e < 0 ? 0u - static_cast<unsigned>(e) : static_cast<unsigned>(e); // (unsigned: no signed overflow to assume)
+    const bool three = ea >= 100;
+    end[0] = 'e';
+    end[1] = e < 0 ? '-' : '+';
+    end[2] = static_cast<char>('0' + (three ? ea / 100 : (ea / 10) % 10));
+    end[3] = static_cast<char>('0' + (three ? (ea / 10) % 10 : ea % 10));
+    end[4] = static_cast<char>('0' + (ea % 10));
+    return end + (three ? 5 : 4);
+}
+
+/*!
+@brief whether FloatType is an IEEE 754 binary64 type (a double, or a long double
+that has the same format, as with MSVC and on Apple's Arm CPUs)
+
+These are the types the conversion of Zmij (see zmij.hpp) is used for; all
+others (binary32, or a format the library does not know) use Grisu2.
+*/
+template<typename FloatType>
+constexpr bool has_binary64_format() noexcept
+{
+    return std::numeric_limits<FloatType>::is_iec559
+           && std::numeric_limits<FloatType>::digits == 53
+           && std::numeric_limits<FloatType>::max_exponent == 1024
+           && sizeof(FloatType) == sizeof(std::uint64_t);
+}
+
+template<typename FloatType>
+struct is_binary64 : std::integral_constant<bool, has_binary64_format<FloatType>()> {};
+
+/// a positive finite float (other than binary64): Grisu2 and format_buffer()
+template<typename FloatType>
+JSON_HEDLEY_NON_NULL(1, 2)
+JSON_HEDLEY_RETURNS_NON_NULL
+char* write_positive_grisu2(char* first, const char* last, FloatType value)
+{
+    JSON_ASSERT(last - first >= std::numeric_limits<FloatType>::max_digits10);
+    static_cast<void>(last); // (only used in the assertion)
+
+    // Compute v = buffer * 10^decimal_exponent.
+    // The decimal digits are stored in the buffer, which needs to be interpreted
+    // as an unsigned decimal integer.
+    // len is the length of the buffer, i.e., the number of decimal digits.
+    int len = 0;
+    int decimal_exponent = 0;
+    grisu2(first, len, decimal_exponent, value);
+
+    JSON_ASSERT(len <= std::numeric_limits<FloatType>::max_digits10);
+
+    // Format the buffer like printf("%.*g", prec, value)
+    constexpr int kMinExp = -4;
+    // Use digits10 here to increase compatibility with version 2.
+    constexpr int kMaxExp = std::numeric_limits<FloatType>::digits10;
+
+    JSON_ASSERT(last - first >= kMaxExp + 2);
+    JSON_ASSERT(last - first >= 2 + (-kMinExp - 1) + std::numeric_limits<FloatType>::max_digits10);
+    JSON_ASSERT(last - first >= std::numeric_limits<FloatType>::max_digits10 + 6);
+
+    return format_buffer(first, len, decimal_exponent, kMinExp, kMaxExp);
+}
+
+/// a positive finite binary64 number: the shortest digits (Zmij), laid out by
+/// write_shortest() (through a local buffer if [first, last) is shorter than
+/// the 41 bytes it may write)
+template<typename FloatType>
+JSON_HEDLEY_NON_NULL(1, 2)
+JSON_HEDLEY_RETURNS_NON_NULL
+char* write_positive_zmij(char* first, const char* last, FloatType value)
+{
+    static_assert(is_binary64<FloatType>::value,
+                  "internal error: the conversion of Zmij needs IEEE 754 binary64 numbers");
+    std::uint64_t bits = 0;
+    std::memcpy(&bits, &value, sizeof(bits));
+    const zmij::shortest_decimal d = zmij::to_shortest(bits);
+    if (JSON_HEDLEY_LIKELY(last - first >= 41))
+    {
+        return write_shortest(first, d);
+    }
+    std::array<char, 64> buf; // NOLINT(cppcoreguidelines-pro-type-member-init,hicpp-member-init): written before read
+    const auto len = static_cast<std::size_t>(write_shortest(buf.data(), d) - buf.data());
+    JSON_ASSERT(last - first >= static_cast<std::ptrdiff_t>(len));
+    std::memcpy(first, buf.data(), len);
+    return first + len;
+}
+
+/// a positive finite binary64 number: Zmij (as a long double has the format of
+/// a double here, its bits are those of the double of the same value)
+template<typename FloatType>
+JSON_HEDLEY_NON_NULL(1, 2)
+JSON_HEDLEY_RETURNS_NON_NULL
+char* write_positive(char* first, const char* last, FloatType value, std::true_type /*is_binary64*/)
+{
+    return write_positive_zmij(first, last, value);
+}
+
+/// a positive finite float of any other format: Grisu2
+template<typename FloatType>
+JSON_HEDLEY_NON_NULL(1, 2)
+JSON_HEDLEY_RETURNS_NON_NULL
+char* write_positive(char* first, const char* last, FloatType value, std::false_type /*is_binary64*/)
+{
+    return write_positive_grisu2(first, last, value);
+}
+
+/// a positive finite float: Zmij for binary64 numbers, Grisu2 otherwise
+template<typename FloatType>
+JSON_HEDLEY_NON_NULL(1, 2)
+JSON_HEDLEY_RETURNS_NON_NULL
+char* write_positive(char* first, const char* last, FloatType value)
+{
+    return write_positive(first, last, value, is_binary64<FloatType> {});
+}
+
 }  // namespace dtoa_impl

 /*!
@@ -1064,7 +1461,6 @@ JSON_HEDLEY_NON_NULL(1, 2)
 JSON_HEDLEY_RETURNS_NON_NULL
 char* to_chars(char* first, const char* last, FloatType value)
 {
-    static_cast<void>(last); // maybe unused - fix warning
    JSON_ASSERT(std::isfinite(value));

    // Use signbit(value) instead of (value < 0) since signbit works for -0.
@@ -1090,28 +1486,7 @@ char* to_chars(char* first, const char* last, FloatType value)
    JSON_HEDLEY_DIAGNOSTIC_POP
 #endif

-    JSON_ASSERT(last - first >= std::numeric_limits<FloatType>::max_digits10);
-
-    // Compute v = buffer * 10^decimal_exponent.
-    // The decimal digits are stored in the buffer, which needs to be interpreted
-    // as an unsigned decimal integer.
-    // len is the length of the buffer, i.e., the number of decimal digits.
-    int len = 0;
-    int decimal_exponent = 0;
-    dtoa_impl::grisu2(first, len, decimal_exponent, value);
-
-    JSON_ASSERT(len <= std::numeric_limits<FloatType>::max_digits10);
-
-    // Format the buffer like printf("%.*g", prec, value)
-    constexpr int kMinExp = -4;
-    // Use digits10 here to increase compatibility with version 2.
-    constexpr int kMaxExp = std::numeric_limits<FloatType>::digits10;
-
-    JSON_ASSERT(last - first >= kMaxExp + 2);
-    JSON_ASSERT(last - first >= 2 + (-kMinExp - 1) + std::numeric_limits<FloatType>::max_digits10);
-    JSON_ASSERT(last - first >= std::numeric_limits<FloatType>::max_digits10 + 6);
-
-    return dtoa_impl::format_buffer(first, len, decimal_exponent, kMinExp, kMaxExp);
+    return dtoa_impl::write_positive(first, last, value);
 }

 }  // namespace detail
@@ -167,6 +167,17 @@ struct external_constructor<value_t::array>
        j.replace_value(value_t::array, value);
    }

+    /// an rvalue container: move its elements
+    template<typename BasicJsonType, typename CompatibleArrayType>
+    static void construct_moving(BasicJsonType& j, CompatibleArrayType& arr)
+    {
+        using std::begin;
+        using std::end;
+
+        const typename BasicJsonType::json_value value(j.template create<typename BasicJsonType::array_t>(std::make_move_iterator(begin(arr)), std::make_move_iterator(end(arr))));
+        j.replace_value(value_t::array, value);
+    }
+
    template<typename BasicJsonType>
    static void construct(BasicJsonType& j, const std::vector<bool>& arr)
    {
@@ -217,6 +228,17 @@ struct external_constructor<value_t::object>
        j.replace_value(value_t::object, value);
    }

+    /// an rvalue container: move its values
+    template<typename BasicJsonType, typename CompatibleObjectType>
+    static void construct_moving(BasicJsonType& j, CompatibleObjectType& obj)
+    {
+        using std::begin;
+        using std::end;
+
+        const typename BasicJsonType::json_value value(j.template create<typename BasicJsonType::object_t>(std::make_move_iterator(begin(obj)), std::make_move_iterator(end(obj))));
+        j.replace_value(value_t::object, value);
+    }
+
    template < typename BasicJsonType, typename CompatibleObjectType,
               enable_if_t < !std::is_same<CompatibleObjectType, typename BasicJsonType::object_t>::value, int > = 0 >
    static void construct(BasicJsonType& j, const CompatibleObjectType& obj)
@@ -303,14 +325,14 @@ inline void to_json(BasicJsonType& j, FloatType val) noexcept
 }

 template<typename BasicJsonType, typename CompatibleNumberUnsignedType,
-         enable_if_t<is_compatible_integer_type<typename BasicJsonType::number_unsigned_t, CompatibleNumberUnsignedType>::value, int> = 0>
+         enable_if_t<is_compatible_integer_type<number_unsigned_overload_t<BasicJsonType>, CompatibleNumberUnsignedType>::value, int> = 0>
 inline void to_json(BasicJsonType& j, CompatibleNumberUnsignedType val) noexcept
 {
    external_constructor<value_t::number_unsigned>::construct(j, static_cast<typename BasicJsonType::number_unsigned_t>(val));
 }

 template<typename BasicJsonType, typename CompatibleNumberIntegerType,
-         enable_if_t<is_compatible_integer_type<typename BasicJsonType::number_integer_t, CompatibleNumberIntegerType>::value, int> = 0>
+         enable_if_t<is_compatible_number_integer_type<BasicJsonType, CompatibleNumberIntegerType>::value, int> = 0>
 inline void to_json(BasicJsonType& j, CompatibleNumberIntegerType val) noexcept
 {
    external_constructor<value_t::number_integer>::construct(j, static_cast<typename BasicJsonType::number_integer_t>(val));
@@ -354,6 +376,66 @@ inline void to_json(BasicJsonType& j, const CompatibleArrayType& arr)
    external_constructor<value_t::array>::construct(j, arr);
 }

+// types that have a dedicated to_json overload and must not be treated as generic containers
+template<typename T>
+struct has_dedicated_to_json : std::false_type {};
+
+#if JSON_HAS_FILESYSTEM || JSON_HAS_EXPERIMENTAL_FILESYSTEM
+template<>
+struct has_dedicated_to_json<std_fs::path> : std::true_type {};
+#endif
+
+#if defined(__cpp_lib_char8_t)
+template<typename Tr, typename Allocator>
+struct has_dedicated_to_json<std::basic_string<char8_t, Tr, Allocator>> : std::true_type {};
+#endif
+
+#if JSON_HAS_RANGE_VIEW_CONVERSION
+template<typename T>
+struct is_range_view_type : is_compatible_range_view<T> {};
+#else
+template<typename T>
+struct is_range_view_type : std::false_type {};
+#endif
+
+#if JSON_USE_OBJECTS_FOR_ENUM_KEYED_MAPS
+template<typename T>
+struct is_enum_keyed_map_type : is_enum_keyed_map<T> {};
+#else
+template<typename T>
+struct is_enum_keyed_map_type : std::false_type {};
+#endif
+
+// rvalue (non-const) container: the generic to_json takes its elements by move
+template<typename T>
+using is_movable_container_rvalue = negation<disjunction<
+                                    std::is_lvalue_reference<T>,
+                                    std::is_const<T>,
+                                    has_dedicated_to_json<T>,
+                                    is_basic_json<T>,
+                                    is_range_view_type<T>,
+                                    is_enum_keyed_map_type<T>,
+                                    std::is_same<T, std::vector<bool>>>>;
+
+// rvalue array-like containers
+template<typename BasicJsonType, typename T>
+using is_rvalue_array_to_json = conjunction<is_movable_container_rvalue<T>,
+      is_compatible_array_type<BasicJsonType, T>,
+      negation<disjunction<
+      std::is_same<T, typename BasicJsonType::array_t>,
+      std::is_same<T, typename BasicJsonType::binary_t>,
+      is_compatible_object_type<BasicJsonType, T>,
+      is_compatible_string_type<BasicJsonType, T>,
+      is_compatible_binary_type<BasicJsonType, T>>>>;
+
+// rvalue containers: their elements are moved instead of copied
+template<typename BasicJsonType, typename CompatibleArrayType,
+         enable_if_t<is_rvalue_array_to_json<BasicJsonType, CompatibleArrayType>::value, int> = 0>
+inline void to_json(BasicJsonType& j, CompatibleArrayType && arr)
+{
+    external_constructor<value_t::array>::construct_moving(j, arr);
+}
+
 #if JSON_HAS_RANGE_VIEW_CONVERSION
 template < typename BasicJsonType, typename T,
           enable_if_t < is_compatible_range_view<std::remove_cvref_t<T>>::value
@@ -427,6 +509,19 @@ inline void to_json(BasicJsonType& j, const EnumKeyedMap& map)
 }
 #endif

+// rvalue object-like containers
+template<typename BasicJsonType, typename T>
+using is_rvalue_object_to_json = conjunction<is_movable_container_rvalue<T>,
+      negation<std::is_same<T, typename BasicJsonType::object_t>>,
+      is_compatible_object_type<BasicJsonType, T>>;
+
+template<typename BasicJsonType, typename CompatibleObjectType,
+         enable_if_t<is_rvalue_object_to_json<BasicJsonType, CompatibleObjectType>::value, int> = 0>
+inline void to_json(BasicJsonType& j, CompatibleObjectType && obj)
+{
+    external_constructor<value_t::object>::construct_moving(j, obj);
+}
+
 template<typename BasicJsonType>
 inline void to_json(BasicJsonType& j, typename BasicJsonType::object_t&& obj)
 {
@@ -0,0 +1,222 @@
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.12.0
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2025 Victor Zverovich <https://github.com/vitaut/zmij>
+// SPDX-FileCopyrightText: 2013-2026 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+#pragma once
+
+#include <array> // array
+#include <cstddef> // size_t
+#include <cstdint> // uint32_t, uint64_t
+
+#include <nlohmann/detail/abi_macros.hpp>
+#include <nlohmann/detail/bit_ops.hpp>
+#include <nlohmann/detail/input/pow5_table.hpp>
+#include <nlohmann/detail/macro_scope.hpp>
+
+NLOHMANN_JSON_NAMESPACE_BEGIN
+namespace detail
+{
+
+/*!
+@brief the shortest decimal representation of a double
+
+A C++11 port of the conversion of Zmij by Victor Zverovich
+(https://github.com/vitaut/zmij, MIT license): the shortest decimal in the
+rounding interval of a double, the closest one if there are several. Zmij
+credits Xiang JunBo (producing the shorter candidate without a division) and
+Dougall Johnson (the compressed powers of ten). The powers of ten are taken
+from the table for number parsing (pow5_table.hpp) where it holds them, and
+computed from the compressed tables of Zmij beyond it.
+*/
+namespace zmij
+{
+
+/// the compressed powers of ten of Zmij
+inline const std::array<std::uint64_t, 28>& pow10_minor() noexcept
+{
+    static const std::array<std::uint64_t, 28> table =
+    {
+        {
+            0x8000000000000000u, 0xa000000000000000u, 0xc800000000000000u, 0xfa00000000000000u, 0x9c40000000000000u,
+            0xc350000000000000u, 0xf424000000000000u, 0x9896800000000000u, 0xbebc200000000000u, 0xee6b280000000000u,
+            0x9502f90000000000u, 0xba43b74000000000u, 0xe8d4a51000000000u, 0x9184e72a00000000u, 0xb5e620f480000000u,
+            0xe35fa931a0000000u, 0x8e1bc9bf04000000u, 0xb1a2bc2ec5000000u, 0xde0b6b3a76400000u, 0x8ac7230489e80000u,
+            0xad78ebc5ac620000u, 0xd8d726b7177a8000u, 0x878678326eac9000u, 0xa968163f0a57b400u, 0xd3c21bcecceda100u,
+            0x84595161401484a0u, 0xa56fa5b99019a5c8u, 0xcecb8f27f4200f3au
+        }
+    };
+    return table;
+}
+
+/// (high, low) pairs
+inline const std::array<std::uint64_t, 50>& pow10_major() noexcept
+{
+    static const std::array<std::uint64_t, 50> table =
+    {
+        {
+            0xaddcb9e83c6b1793u, 0xdf4abe242a1bbf3eu, 0xaf8e5410288e1b6fu, 0x07ecf0ae5ee44ddau, 0xb1442798f49ffb4au, 0x99cd11cfdf41779du,
+            0xb2fe3f0b8599ef07u, 0x861fa7e6dcb4aa15u, 0xb4bca50b065abe63u, 0x0fed077a756b53aau, 0xb67f6455292cbf08u, 0x1a3bc84c17b1d543u,
+            0xb84687c269ef3bfbu, 0x3d5d514f40eea742u, 0xba121a4650e4ddebu, 0x92f34d62616ce413u, 0xbbe226efb628afeau, 0x890489f70a55368cu,
+            0xbdb6b8e905cb600fu, 0x5400e987bbc1c921u, 0xbf8fdb78849a5f96u, 0xde98520472bdd034u, 0xc16d9a0095928a27u, 0x75b7053c0f178294u,
+            0xc350000000000000u, 0x0000000000000000u, 0xc5371912364ce305u, 0x6c28000000000000u, 0xc722f0ef9d80aad6u, 0x424d3ad2b7b97ef6u,
+            0xc913936dd571c84cu, 0x03bc3a19cd1e38eau, 0xcb090c8001ab551cu, 0x5cadf5bfd3072cc6u, 0xcd036837130890a1u, 0x36dba887c37a8c10u,
+            0xcf02b2c21207ef2eu, 0x94f967e45e03f4bcu, 0xd106f86e69d785c7u, 0xe13336d701beba52u, 0xd31045a8341ca07cu, 0x1ede48111209a051u,
+            0xd51ea6fa85785631u, 0x552a74227f3ea566u, 0xd732290fbacaf133u, 0xa97c177947ad4096u, 0xd94ad8b1c7380874u, 0x18375281ae7822bdu,
+            0xdb68c2ca82ed2a05u, 0xa67398db9f6820e1u
+        }
+    };
+    return table;
+}
+
+/// one bit per power: whether the computed value is one unit too large
+inline const std::array<std::uint32_t, 21>& pow10_fixups() noexcept
+{
+    static const std::array<std::uint32_t, 21> table =
+    {
+        {
+            0x8d8fc810u, 0x06100293u, 0x19000000u, 0x00100000u, 0x00000908u, 0x00000000u, 0x04e00300u, 0x3807e0b2u, 0x3d83d793u, 0x0006f5ccu,
+            0x00000000u, 0xffff0000u, 0x8076337du, 0x4ff45ba0u, 0x09405033u, 0x034376d9u, 0x09000000u, 0x4e100501u, 0x076d14dcu, 0xf964f45eu,
+            0x0000003du
+        }
+    };
+    return table;
+}
+
+/// the 128-bit significand of 10^k, rounded down, for k in [-307, 341]
+/// (compute_pow10 of Zmij)
+inline uint128_parts compute_pow10(int k) noexcept
+{
+    const auto i = static_cast<unsigned>(k + 307);
+    const std::uint64_t m = pow10_minor()[(i + 24) % 28];
+    const std::size_t j = 2 * static_cast<std::size_t>((i + 24) / 28);
+    const std::uint64_t h_hi = pow10_major()[j];
+    const std::uint64_t h_lo = pow10_major()[j + 1];
+    const std::uint64_t h1 = full_multiplication(h_lo, m).high;
+    const std::uint64_t c0 = h_lo * m;
+    const std::uint64_t c1 = h1 + (h_hi * m);
+    const std::uint64_t c2 = (c1 < h1 ? 1u : 0u) + full_multiplication(h_hi, m).high;
+    uint128_parts r{};
+    if ((c2 >> 63u) != 0)
+    {
+        r.high = c2;
+        r.low = c1;
+    }
+    else
+    {
+        r.high = (c2 << 1u) | (c1 >> 63u);
+        r.low = (c1 << 1u) | (c0 >> 63u);
+    }
+    r.low -= (pow10_fixups()[i >> 5u] >> (i & 31u)) & 1u;
+    return r;
+}
+
+/// The 128-bit significand of 10^k, rounded down, for k in [-342, 341].
+/// Up to 10^308, the table for number parsing holds the same significands
+/// (those of 5^k), except for k in [-27, -1], where it holds them one unit
+/// larger (as the Eisel-Lemire algorithm needs them).
+inline uint128_parts pow10(int k) noexcept
+{
+    if (k > pow5_128_largest_power)
+    {
+        return compute_pow10(k); // (only for the smallest doubles)
+    }
+    const auto i = 2 * static_cast<std::size_t>(k - pow5_128_smallest_power);
+    uint128_parts r{pow5_128()[i + 1], pow5_128()[i]};
+    const std::uint64_t adjust = static_cast<unsigned>(k + 27) < 27u ? 1u : 0u;
+    r.high -= r.low < adjust ? 1u : 0u;
+    r.low -= adjust;
+    return r;
+}
+
+/// (x_hi * 2^64 + x_lo) * y >> 64, as 128 bits
+inline uint128_parts umul192_hi128(std::uint64_t x_hi, std::uint64_t x_lo, std::uint64_t y) noexcept
+{
+    const uint128_parts p = full_multiplication(x_hi, y);
+    uint128_parts r{};
+    r.low = p.low + full_multiplication(x_lo, y).high;
+    r.high = p.high + (r.low < p.low ? 1u : 0u);
+    return r;
+}
+
+/// (x * y + c) >> 64
+inline std::uint64_t umul128_add_hi64(std::uint64_t x, std::uint64_t y, std::uint64_t c) noexcept
+{
+    const uint128_parts p = full_multiplication(x, y);
+    return p.high + (p.low + c < p.low ? 1u : 0u);
+}
+
+/// the result of Zmij: the shorter candidate and, if that is outside the
+/// rounding interval, the digit after it (16 bytes: returned in registers)
+struct shortest_decimal
+{
+    std::uint64_t integral;  ///< the shorter candidate (15 or 16 digits for normal doubles)
+    int exponent;            ///< the decimal exponent of the digit after it
+    unsigned char digit;     ///< the digit after it (if has_digit)
+    bool has_digit;          ///< whether the shortest decimal is integral * 10 + digit
+};
+
+/// The shortest decimal in the rounding interval of a positive finite double
+/// given by its bits, the closest one if there are several (to_decimal of
+/// Zmij, which keeps the last digit apart: the 15 or 16 digits before it can be
+/// converted without a multiplication by 10 first). Always inlined: GCC
+/// otherwise calls it, and its result goes through memory.
+JSON_HEDLEY_ALWAYS_INLINE shortest_decimal to_shortest(std::uint64_t bits) noexcept
+{
+    constexpr int extra_shift = 9;
+    const auto raw_exp = static_cast<int>((bits >> 52u) & 0x7FFu);
+    std::uint64_t bin_sig = bits & ((std::uint64_t{1} << 52u) - 1);
+    // a power of two has a narrower interval below (except the smallest normal)
+    const bool regular = bin_sig != 0 || raw_exp <= 1;
+    const int bin_exp = (raw_exp == 0 ? 1 : raw_exp) - 1075;
+    if (raw_exp != 0)
+    {
+        bin_sig |= std::uint64_t{1} << 52u;
+    }
+    // floor(log10(2^bin_exp)), or floor(log10(3/4 * 2^bin_exp)) for the irregular case
+    const int dec_exp = ((bin_exp * 315653) - (regular ? 0 : 131072)) >> 20;
+    // scaled by 10^(-dec_exp - 1): the integral part is the shorter candidate
+    const int shift = bin_exp + ((-(dec_exp + 1) * 217707) >> 16) + 1 + extra_shift;
+    const uint128_parts p10 = pow10(-dec_exp - 1);
+    const uint128_parts p = umul192_hi128(p10.high, p10.low, bin_sig << static_cast<unsigned>(shift));
+    std::uint64_t integral = p.high >> static_cast<unsigned>(extra_shift);
+    // (no casts in the shifts: MSVC 2015 warns C4554 on them despite the parentheses)
+    constexpr unsigned fractional_low_shift = static_cast<unsigned>(extra_shift);
+    constexpr unsigned fractional_high_shift = 64u - fractional_low_shift;
+    const std::uint64_t fractional = (p.high << fractional_high_shift) | (p.low >> fractional_low_shift);
+    std::uint64_t digit = 0;
+    bool round_up = false;
+    bool round_down = false;
+    if (JSON_HEDLEY_LIKELY(regular))
+    {
+        const std::uint64_t half_ulp = (p10.high >> static_cast<unsigned>(extra_shift + 1 - shift)) + (1 - (bin_sig & 1u));
+        round_up = fractional + half_ulp < fractional;
+        round_down = half_ulp > fractional;
+        // the last digit of the longer candidate, rounded to nearest
+        digit = umul128_add_hi64(fractional, 10, (std::uint64_t{1} << 63u) + 6);
+        if (fractional == (std::uint64_t{1} << 62u))
+        {
+            digit = 2; // 2.5 rounds to 2
+        }
+    }
+    else
+    {
+        const std::uint64_t half_ulp = p10.high >> static_cast<unsigned>(extra_shift + 1 - shift);
+        round_up = half_ulp > ~std::uint64_t{0} - fractional;
+        round_down = (half_ulp >> 1u) > fractional;
+        digit = umul128_add_hi64(fractional, 10, (std::uint64_t{1} << 63u) - 1);
+        const std::uint64_t lowest = umul128_add_hi64(fractional - (half_ulp >> 1u), 10, ~std::uint64_t{0});
+        digit = digit < lowest ? lowest : digit;
+    }
+    integral += round_up ? 1u : 0u;
+    // if the shorter candidate is outside the rounding interval: one digit more
+    return shortest_decimal{integral, dec_exp, static_cast<unsigned char>(digit), !round_up && !round_down};
+}
+
+}  // namespace zmij
+}  // namespace detail
+NLOHMANN_JSON_NAMESPACE_END
@@ -428,7 +428,7 @@ void templated_json_throw(ExceptionType exception)
@throw type_error.302 always
 */
 template<typename BasicJsonType>
-JSON_HEDLEY_NO_RETURN inline void throw_type_must_be(const char* expected, const BasicJsonType& j)
+JSON_HEDLEY_NO_RETURN JSON_HEDLEY_NEVER_INLINE void throw_type_must_be(const char* expected, const BasicJsonType& j)
 {
    static_cast<void>(expected); // unused when JSON_NOEXCEPTION is defined
    static_cast<void>(j);
@@ -443,7 +443,7 @@ JSON_HEDLEY_NO_RETURN inline void throw_type_must_be(const char* expected, const
@throw type_error always
 */
 template<typename BasicJsonType>
-JSON_HEDLEY_NO_RETURN inline void throw_cannot_use_with(const exception_id id_, const char* operation, const BasicJsonType& j)
+JSON_HEDLEY_NO_RETURN JSON_HEDLEY_NEVER_INLINE void throw_cannot_use_with(const exception_id id_, const char* operation, const BasicJsonType& j)
 {
    static_cast<void>(id_); // unused when JSON_NOEXCEPTION is defined
    static_cast<void>(operation);
@@ -451,6 +451,77 @@ JSON_HEDLEY_NO_RETURN inline void throw_cannot_use_with(const exception_id id_,
    JSON_THROW(type_error::create(id_, concat("cannot use ", operation, " with ", j.type_name()), &j));
 }

+/*!
+@brief throws ExceptionType with a fixed message; kept out of line so that callers do not carry the exception construction
+*/
+template<typename ExceptionType, typename BasicJsonContext>
+JSON_HEDLEY_NO_RETURN JSON_HEDLEY_NEVER_INLINE void throw_literal(const exception_id id_, const char* what_arg, BasicJsonContext context)
+{
+    static_cast<void>(id_); // unused when JSON_NOEXCEPTION is defined
+    static_cast<void>(what_arg);
+    static_cast<void>(context);
+    JSON_THROW(ExceptionType::create(id_, what_arg, context));
+}
+
+// JSON Patch errors; constructed out of line so that patch_inplace() does not carry them
+
+template<typename BasicJsonType>
+JSON_HEDLEY_NO_RETURN JSON_HEDLEY_NEVER_INLINE void throw_patch_add_parent(const BasicJsonType& parent)
+{
+    static_cast<void>(parent);
+    JSON_THROW(out_of_range::create(exception_id::patch_add_parent_not_container, concat("cannot add value: the JSON Patch 'add' target's parent is of type ", parent.type_name(), ", but must be an object or array"), &parent));
+}
+
+template<typename BasicJsonType>
+JSON_HEDLEY_NO_RETURN JSON_HEDLEY_NEVER_INLINE void throw_patch_remove_parent(const BasicJsonType& parent)
+{
+    static_cast<void>(parent);
+    JSON_THROW(out_of_range::create(exception_id::patch_remove_parent_not_container, concat("cannot remove value: the JSON Patch 'remove' target's parent is of type ", parent.type_name(), ", but must be an object or array"), &parent));
+}
+
+template<typename BasicJsonType>
+JSON_HEDLEY_NO_RETURN JSON_HEDLEY_NEVER_INLINE void throw_patch_not_an_array(const BasicJsonType& j)
+{
+    static_cast<void>(j);
+    JSON_THROW(parse_error::create(exception_id::patch_not_an_array, 0, "JSON patch must be an array of objects", &j));
+}
+
+template<typename StringType, typename BasicJsonType>
+JSON_HEDLEY_NO_RETURN JSON_HEDLEY_NEVER_INLINE void throw_patch_missing_member(const StringType& op, const StringType& member, const bool string_type, const BasicJsonType& val)
+{
+    static_cast<void>(op);
+    static_cast<void>(member);
+    static_cast<void>(string_type);
+    static_cast<void>(val);
+    // context-sensitive error message
+    const std::string error_msg = (op == "op") ? std::string("operation") : concat("operation '", op, '\'');
+    JSON_THROW(parse_error::create(exception_id::patch_invalid_operation, 0, concat(error_msg, string_type ? " must have string member '" : " must have member '", member, "'"), &val));
+}
+
+template<typename StringType, typename BasicJsonType>
+JSON_HEDLEY_NO_RETURN JSON_HEDLEY_NEVER_INLINE void throw_patch_move_into_child(const StringType& from_path, const StringType& path, const BasicJsonType& result)
+{
+    static_cast<void>(from_path);
+    static_cast<void>(path);
+    static_cast<void>(result);
+    JSON_THROW(out_of_range::create(exception_id::patch_move_into_child, concat("cannot move value: 'from' path '", from_path, "' is a proper prefix of 'path' '", path, "'"), &result));
+}
+
+template<typename BasicJsonType>
+JSON_HEDLEY_NO_RETURN JSON_HEDLEY_NEVER_INLINE void throw_patch_test_failed(const BasicJsonType& val)
+{
+    static_cast<void>(val);
+    JSON_THROW(other_error::create(exception_id::patch_test_failed, concat("unsuccessful: ", val.dump()), &val));
+}
+
+template<typename StringType, typename BasicJsonType>
+JSON_HEDLEY_NO_RETURN JSON_HEDLEY_NEVER_INLINE void throw_patch_invalid_operation(const StringType& op, const BasicJsonType& val)
+{
+    static_cast<void>(op);
+    static_cast<void>(val);
+    JSON_THROW(parse_error::create(exception_id::patch_invalid_operation, 0, concat("operation value '", op, "' is invalid"), &val));
+}
+
 }  // namespace detail
 NLOHMANN_JSON_NAMESPACE_END

@@ -195,7 +195,7 @@ class binary_reader

            if (JSON_HEDLEY_UNLIKELY(current != char_traits<char_type>::eof()))
            {
-                return last_byte_error(exception_id::unexpected_end_of_input, concat("expected end of input; last byte: 0x", get_token_string()), "value");
+                return last_byte_error(exception_id::unexpected_end_of_input, "expected end of input", "value", "");
            }
        }

@@ -317,7 +317,7 @@ class binary_reader
    {
        if (JSON_HEDLEY_UNLIKELY(document_size < 0 || static_cast<std::size_t>(document_size) != chars_read - document_start))
        {
-            return last_byte_error(exception_id::unexpected_byte, concat("document size ", std::to_string(document_size), " does not match the number of bytes read (", std::to_string(chars_read - document_start), ")"), "document");
+            return bson_document_size_error(document_size, chars_read - document_start);
        }
        return true;
    }
@@ -509,7 +509,7 @@ class binary_reader
    {
        if (JSON_HEDLEY_UNLIKELY(len < 1))
        {
-            return last_byte_error(exception_id::unexpected_byte, concat("string length must be at least 1, is ", std::to_string(len)), "string");
+            return bson_length_error("string length must be at least 1, is ", len, "string");
        }

        if (JSON_HEDLEY_UNLIKELY(!get_string(len - static_cast<NumberType>(1), result)))
@@ -539,7 +539,7 @@ class binary_reader
    {
        if (JSON_HEDLEY_UNLIKELY(len < 0))
        {
-            return last_byte_error(exception_id::unexpected_byte, concat("byte array length cannot be negative, is ", std::to_string(len)), "binary");
+            return bson_length_error("byte array length cannot be negative, is ", len, "binary");
        }

        // All BSON binary values have a subtype
@@ -629,9 +629,7 @@ class binary_reader

            default: // anything else is not supported (yet)
            {
-                const std::string cr_str = hex_byte(static_cast<std::uint8_t>(element_type));
-                return sax->parse_error(element_type_parse_position, cr_str,
-                                        parse_error::create(exception_id::bson_unsupported_type, element_type_parse_position, concat("Unsupported BSON record type 0x", cr_str), nullptr));
+                return bson_unsupported_type_error(element_type_parse_position, element_type);
            }
        }
    }
@@ -1060,6 +1058,7 @@ class binary_reader
    @param[in] context  parsing context for the error message
    @return whether the SAX consumer accepts the parse error
    */
+    JSON_HEDLEY_NEVER_INLINE
    bool cbor_indefinite_string_error(const char* type_name, const char* context)
    {
        return last_byte_error(exception_id::invalid_string_or_size, concat("indefinite-length ", type_name, " is not allowed inside indefinite-length ", type_name, "; last byte: 0x", get_token_string()), context);
@@ -1139,7 +1138,7 @@ class binary_reader

            default:
            {
-                return last_byte_error(exception_id::invalid_string_or_size, concat("expected length specification (0x60-0x7B)", inside_indefinite ? "" : " or indefinite string type (0x7F)", "; last byte: 0x", get_token_string()), "string");
+                return last_byte_error(exception_id::invalid_string_or_size, "expected length specification (0x60-0x7B)", "string", inside_indefinite ? "" : " or indefinite string type (0x7F)");
            }
        }
    }
@@ -1269,7 +1268,7 @@ class binary_reader
                break;
        }

-        return last_byte_error(exception_id::invalid_string_or_size, concat("only string keys are supported, but found ", found, "; last byte: 0x", get_token_string()), "object key");
+        return last_byte_error(exception_id::invalid_string_or_size, "only string keys are supported, but found ", "object key", found);
    }

    /*!
@@ -1350,7 +1349,7 @@ class binary_reader

            default:
            {
-                return last_byte_error(exception_id::invalid_string_or_size, concat("expected length specification (0x40-0x5B)", inside_indefinite ? "" : " or indefinite binary array type (0x5F)", "; last byte: 0x", get_token_string()), "binary");
+                return last_byte_error(exception_id::invalid_string_or_size, "expected length specification (0x40-0x5B)", "binary", inside_indefinite ? "" : " or indefinite binary array type (0x5F)");
            }
        }
    }
@@ -1496,8 +1495,7 @@ class binary_reader
    {
        if (JSON_HEDLEY_UNLIKELY(!value_in_range_of<std::size_t>(len) || len == detail::unknown_size()))
        {
-            return sax->parse_error(chars_read, get_token_string(), out_of_range::create(exception_id::container_too_large,
-                                    exception_message(concat("excessive ", context, " size"), "size"), nullptr));
+            return container_size_error(context);
        }
        result = conditional_static_cast<std::size_t>(len);
        return true;
@@ -2065,7 +2063,7 @@ class binary_reader

            default:
            {
-                return last_byte_error(exception_id::invalid_string_or_size, concat("expected length specification (0xA0-0xBF, 0xD9-0xDB); last byte: 0x", get_token_string()), "string");
+                return last_byte_error(exception_id::invalid_string_or_size, "expected length specification (0xA0-0xBF, 0xD9-0xDB)", "string", "");
            }
        }
    }
@@ -2157,7 +2155,7 @@ class binary_reader
                break;
        }

-        return last_byte_error(exception_id::invalid_string_or_size, concat("only string keys are supported, but found ", found, "; last byte: 0x", get_token_string()), "object key");
+        return last_byte_error(exception_id::invalid_string_or_size, "only string keys are supported, but found ", "object key", found);
    }

    /*!
@@ -2655,8 +2653,7 @@ class binary_reader
        }
        if (JSON_HEDLEY_UNLIKELY(!value_in_range_of<std::size_t>(number)))
        {
-            return sax->parse_error(chars_read, get_token_string(), out_of_range::create(exception_id::container_too_large,
-                                    exception_message("integer value overflow", "size"), nullptr));
+            return out_of_range_error(exception_id::container_too_large, "integer value overflow", "size");
        }
        result = static_cast<std::size_t>(number); // NOLINT(bugprone-signed-char-misuse,cert-str34-c): number is not a char
        return true;
@@ -2753,8 +2750,7 @@ class binary_reader
                }
                if (!value_in_range_of<std::size_t>(number))
                {
-                    return sax->parse_error(chars_read, get_token_string(), out_of_range::create(exception_id::container_too_large,
-                                            exception_message("integer value overflow", "size"), nullptr));
+                    return out_of_range_error(exception_id::container_too_large, "integer value overflow", "size");
                }
                result = detail::conditional_static_cast<std::size_t>(number);
                return true;
@@ -2832,7 +2828,7 @@ class binary_reader
                        // or SIZE_MAX.
                        if (JSON_HEDLEY_UNLIKELY(result > (std::numeric_limits<std::size_t>::max)() / i))
                        {
-                            return sax->parse_error(chars_read, get_token_string(), out_of_range::create(exception_id::container_too_large, exception_message("excessive ndarray size caused overflow", "size"), nullptr));
+                            return out_of_range_error(exception_id::container_too_large, "excessive ndarray size caused overflow", "size");
                        }
                        result *= i;
                        // the pre-check above already rules out result becoming 0
@@ -2841,7 +2837,7 @@ class binary_reader
                        // unknown-size container (see get_ubjson_size_type())
                        if (result == npos)
                        {
-                            return sax->parse_error(chars_read, get_token_string(), out_of_range::create(exception_id::container_too_large, exception_message("excessive ndarray size caused overflow", "size"), nullptr));
+                            return out_of_range_error(exception_id::container_too_large, "excessive ndarray size caused overflow", "size");
                        }
                        if (JSON_HEDLEY_UNLIKELY(!emit_unsigned(i)))
                        {
@@ -2891,7 +2887,7 @@ class binary_reader
            if (input_format == input_format_t::bjdata
                    && JSON_HEDLEY_UNLIKELY(is_bjd_excluded_optimized_type(result.second)))
            {
-                return last_byte_error(exception_id::unexpected_byte, concat("marker 0x", get_token_string(), " is not a permitted optimized array type"), "type");
+                return optimized_type_error();
            }

            if (JSON_HEDLEY_UNLIKELY(!unexpect_eof("type")))
@@ -2906,7 +2902,7 @@ class binary_reader
                {
                    return false;
                }
-                return last_byte_error(exception_id::unexpected_byte, concat("expected '#' after type information; last byte: 0x", get_token_string()), "size");
+                return last_byte_error(exception_id::unexpected_byte, "expected '#' after type information", "size", "");
            }

            const bool is_error = get_ubjson_size_value(result.first, is_ndarray, 0, result.second);
@@ -3057,7 +3053,7 @@ class binary_reader
                }
                if (JSON_HEDLEY_UNLIKELY(current > 127))
                {
-                    return last_byte_error(exception_id::invalid_string_or_size, concat("byte after 'C' must be in range 0x00..0x7F; last byte: 0x", get_token_string()), "char");
+                    return last_byte_error(exception_id::invalid_string_or_size, "byte after 'C' must be in range 0x00..0x7F", "char", "");
                }
                string_t s(1, static_cast<typename string_t::value_type>(current));
                return sax->string(s);
@@ -3141,8 +3137,7 @@ class binary_reader
            if (JSON_HEDLEY_UNLIKELY((size_and_type.second == 'Z' || size_and_type.second == 'T' || size_and_type.second == 'F')
                                     && size_and_type.first > max_valueless_container_size))
            {
-                return sax->parse_error(chars_read, get_token_string(), out_of_range::create(exception_id::container_too_large,
-                                        exception_message("excessive array size", "size"), nullptr));
+                return out_of_range_error(exception_id::container_too_large, "excessive array size", "size");
            }

            if (JSON_HEDLEY_UNLIKELY(!enter_array(size_and_type.first, size_and_type.second)))
@@ -3854,7 +3849,7 @@ class binary_reader
        {
            // in case of failure, advance position by 1 to report the failing location
            ++chars_read;
-            sax->parse_error(chars_read, "<end of file>", parse_error::create(exception_id::unexpected_end_of_input, chars_read, exception_message("unexpected end of input", context), nullptr));
+            static_cast<void>(report_eof(context));
            return false;
        }
        return true;
@@ -4007,8 +4002,7 @@ class binary_reader
        const auto result = static_cast<number_float_t>(number);
        if (JSON_HEDLEY_UNLIKELY(std::isfinite(number) && !std::isfinite(result)))
        {
-            return sax->parse_error(chars_read, get_token_string(),
-                                    out_of_range::create(exception_id::number_overflow, exception_message("number overflow", "value"), nullptr));
+            return out_of_range_error(exception_id::number_overflow, "number overflow", "value");
        }
        return sax->number_float(result, "");
    }
@@ -4223,22 +4217,82 @@ class binary_reader
    {
        if (JSON_HEDLEY_UNLIKELY(current == char_traits<char_type>::eof()))
        {
-            return sax->parse_error(chars_read, "<end of file>",
-                                    parse_error::create(exception_id::unexpected_end_of_input, chars_read, exception_message("unexpected end of input", context), nullptr));
+            return report_eof(context);
        }
        return true;
    }

+    /*!
+    @brief reports an unexpected end of input (out of line to keep the exception
+           construction out of the hot reading functions)
+    @param[in] context  further context information (for diagnostics)
+    @return the result of the SAX parser's parse_error()
+    */
+    JSON_HEDLEY_NON_NULL(2)
+    JSON_HEDLEY_NEVER_INLINE
+    bool report_eof(const char* context) const
+    {
+        return sax->parse_error(chars_read, "<end of file>",
+                                parse_error::create(exception_id::unexpected_end_of_input, chars_read, exception_message("unexpected end of input", context), nullptr));
+    }
+
+    /*!
+    @brief reports an out_of_range error at the last read byte (out of line)
+    */
+    JSON_HEDLEY_NEVER_INLINE
+    bool out_of_range_error(const exception_id id_, const char* detail, const char* context) const
+    {
+        return sax->parse_error(chars_read, get_token_string(), out_of_range::create(id_, exception_message(detail, context), nullptr));
+    }
+
+    JSON_HEDLEY_NEVER_INLINE
+    bool container_size_error(const char* what) const
+    {
+        return sax->parse_error(chars_read, get_token_string(), out_of_range::create(exception_id::container_too_large,
+                                exception_message(concat("excessive ", what, " size"), "size"), nullptr));
+    }
+
+    JSON_HEDLEY_NEVER_INLINE
+    bool optimized_type_error() const
+    {
+        return last_byte_error(exception_id::unexpected_byte, concat("marker 0x", get_token_string(), " is not a permitted optimized array type"), "type");
+    }
+
+    template<typename NumberType>
+    JSON_HEDLEY_NEVER_INLINE
+    bool bson_length_error(const char* prefix, const NumberType len, const char* context) const
+    {
+        return last_byte_error(exception_id::unexpected_byte, concat(prefix, std::to_string(len)), context);
+    }
+
+    JSON_HEDLEY_NEVER_INLINE
+    bool bson_document_size_error(const std::int32_t document_size, const std::size_t bytes_read) const
+    {
+        return last_byte_error(exception_id::unexpected_byte, concat("document size ", std::to_string(document_size), " does not match the number of bytes read (", std::to_string(bytes_read), ")"), "document");
+    }
+
+    JSON_HEDLEY_NEVER_INLINE
+    bool bson_unsupported_type_error(const std::size_t position, const char element_type) const
+    {
+        const std::string cr_str = hex_byte(static_cast<std::uint8_t>(element_type));
+        return sax->parse_error(position, cr_str,
+                                parse_error::create(exception_id::bson_unsupported_type, position, concat("Unsupported BSON record type 0x", cr_str), nullptr));
+    }
+
    /*!
    @brief reports a parse error at the last read byte
    @param[in] id_      the id of the parse_error exception
    @param[in] detail   a detailed error message
    @param[in] context  further context information
+    @param[in] tail     optional text appended to the detail, followed by "; last byte: 0x<token>"
    @return the result of the SAX parser's parse_error()
    */
-    bool last_byte_error(const exception_id id_, const std::string& detail, const std::string& context) const
+    JSON_HEDLEY_NEVER_INLINE
+    bool last_byte_error(const exception_id id_, const std::string& detail, const char* context, const char* tail = nullptr) const
    {
-        return sax->parse_error(chars_read, get_token_string(), parse_error::create(id_, chars_read, exception_message(detail, context), nullptr));
+        // with a tail, the message reads "<detail><tail>; last byte: 0x<token>"
+        return sax->parse_error(chars_read, get_token_string(), parse_error::create(id_, chars_read,
+                                exception_message(tail == nullptr ? detail : concat(detail, tail, "; last byte: 0x", get_token_string()), context), nullptr));
    }

    /*!
@@ -4247,7 +4301,8 @@ class binary_reader
    @param[in] context  further context information
    @return the result of the SAX parser's parse_error()
    */
-    bool unexpected_byte(const char* detail, const std::string& context) const
+    JSON_HEDLEY_NEVER_INLINE
+    bool unexpected_byte(const char* detail, const char* context) const
    {
        return last_byte_error(exception_id::unexpected_byte, concat(detail, ": 0x", get_token_string()), context);
    }
@@ -4258,6 +4313,7 @@ class binary_reader
    @param[in] context   further context information
    @return the result of the SAX parser's parse_error()
    */
+    JSON_HEDLEY_NEVER_INLINE
    bool length_type_error(const char* position, const char* context) const
    {
        const char* types = input_format == input_format_t::bjdata ? "U, i, u, I, m, l, M, L" : "U, i, I, l, L";
@@ -187,12 +187,19 @@ Shared by json_sax_dom_parser and json_sax_dom_callback_parser.
@param[in] ref   the object or array that was just created
@return whether parsing should continue (false after reporting out_of_range.408)
 */
+template<typename SAX, typename BasicJsonType>
+JSON_HEDLEY_NEVER_INLINE
+bool container_too_large_error(SAX& sax, std::size_t len, const char* kind, BasicJsonType* ref)
+{
+    return sax.parse_error(0, "", out_of_range::create(exception_id::container_too_large, concat("excessive ", kind, " size: ", std::to_string(len)), ref));
+}
+
 template<typename SAX, typename BasicJsonType>
 bool check_container_size(SAX& sax, std::size_t len, const char* kind, BasicJsonType* ref)
 {
    if (JSON_HEDLEY_UNLIKELY(len != detail::unknown_size() && len > ref->max_size()))
    {
-        return sax.parse_error(0, "", out_of_range::create(exception_id::container_too_large, concat("excessive ", kind, " size: ", std::to_string(len)), ref));
+        return container_too_large_error(sax, len, kind, ref);
    }
    return true;
 }
@@ -674,16 +681,19 @@ class json_sax_dom_callback_parser
        if (keep && ref_stack.back())
        {
            auto& obj = *ref_stack.back()->m_data.m_value.object;
-            const auto it = obj.find(val);
-            if (it != obj.end())
+            // a single lookup: insert the placeholder, or find the existing member
+            const auto inserted = obj.emplace(val, discarded);
+            const auto it = inserted.first;
+            if (!inserted.second)
            {
                // this is a duplicate key (legal in JSON); remember its
                // current value so it can be restored later if the new
                // value is rejected by the callback, instead of being
                // erased together with the discarded placeholder
                duplicate_key_stash.emplace_back(&(it->second), it->second);
+                it->second = discarded;
            }
-            object_element = &(obj[val] = discarded);
+            object_element = &(it->second);
        }

        return true;
@@ -233,6 +233,41 @@ class lexer : public lexer_base<BasicJsonType>
    // scan functions
    /////////////////////

+    /// contiguous input: try to decode the 4 hex digits following `\\u`
+    /// directly from the input buffer via hex_codepoint(), instead of 4 calls
+    /// to get(). On success, advances the adapter exactly as those 4 get()
+    /// calls would (bulk_scan implies lazy_position, so there are no position
+    /// counters to advance) and leaves @a current holding the last of
+    /// the 4 digits, just as the last such get() would; the codepoint is
+    /// written to @a out. Makes no state change and returns false for a
+    /// pending unget, fewer than 4 remaining bytes, or any of the 4 bytes not
+    /// being a hex digit, so the caller falls back unchanged to the
+    /// per-character loop, which then reports the same diagnostic (stopping
+    /// at the first invalid digit) as before this optimization.
+    bool get_codepoint_bulk(std::true_type /*bulk*/, int& out)
+    {
+        if (next_unget || ia.bulk_remaining() < 4)
+        {
+            return false;
+        }
+        const char_type* const raw = ia.bulk_data();
+        const int codepoint = hex_codepoint(reinterpret_cast<const unsigned char*>(raw));
+        if (codepoint < 0)
+        {
+            return false;
+        }
+        ia.bulk_skip(4);
+        current = char_traits<char_type>::to_int_type(raw[3]);
+        out = codepoint;
+        return true;
+    }
+
+    /// streaming input: no bulk fast path
+    bool get_codepoint_bulk(std::false_type /*bulk*/, int& /*out*/) const noexcept
+    {
+        return false;
+    }
+
    /*!
    @brief get codepoint from 4 hex characters following `\\u`

@@ -252,6 +287,14 @@ class lexer : public lexer_base<BasicJsonType>
    {
        // this function only makes sense after reading `\u`
        JSON_ASSERT(current == 'u');
+
+        // contiguous input: decode all 4 hex digits directly from the buffer
+        int fast_codepoint = 0;
+        if (get_codepoint_bulk(std::integral_constant<bool, bulk_scan> {}, fast_codepoint))
+        {
+            return fast_codepoint;
+        }
+
        int codepoint = 0;

        const auto factors = { 12u, 8u, 4u, 0u };
@@ -1052,9 +1095,11 @@ class lexer : public lexer_base<BasicJsonType>
            token_type::parse_error otherwise

    @note The scanner is independent of the current locale: token_buffer
-          always holds `.`. Only the std::strtod fallback of convert_number()
-          depends on the locale, and it looks up the decimal point right
-          before converting (see detail::convert_float_locale_aware()).
+          always holds `.`. The conversion of float and double does not use
+          the locale either. Only the std::strtold fallback of
+          convert_number() for long double formats other than binary64
+          depends on it, and it looks up the decimal point right before
+          converting (see detail::convert_float_locale_aware()).
    */
    token_type scan_number()  // lgtm [cpp/use-of-goto] `goto` is used in this function to implement the number-parsing state machine described above. By design, any finite input will eventually reach the "done" state or return token_type::parse_error. In each intermediate state, 1 byte of the input is appended to the token_buffer vector, and only the already initialized variables token_buffer, number_type, and error_message are manipulated.
    {
@@ -1067,7 +1112,7 @@ class lexer : public lexer_base<BasicJsonType>

        // offset just past the last mantissa byte in token_buffer (i.e. the
        // index of 'e'/'E', or the whole token when there is no exponent).
-        // convert_number() uses it to count significant digits; npos means
+        // convert_number() uses it to split the token; npos means
        // "not seen an exponent yet" and is resolved at scan_number_done
        std::size_t mantissa_end = std::string::npos;

@@ -1397,8 +1442,8 @@ scan_number_done:
    @param[in] mantissa_end  offset just past the last mantissa byte in
                             token_buffer (the index of 'e'/'E', or
                             token_buffer.size() when there is no exponent);
-                             used to skip Clinger's fast path when it cannot
-                             possibly succeed - see detail::mantissa_fits_clinger()
+                             with decimal_point_position, it locates the parts
+                             of a float token without scanning it again
    */
    token_type convert_number(token_type number_type, std::size_t mantissa_end)
    {
@@ -1452,10 +1497,11 @@ scan_number_done:
        }

        // this code is reached if we parse a floating-point number or if an
-        // integer conversion above overflowed. Prefer std::from_chars
-        // (Eisel-Lemire, locale-independent, correctly rounded) when available;
-        // otherwise the exact Clinger fast path (double only); otherwise the
-        // locale-aware strtof/strtod/strtold.
+        // integer conversion above overflowed. float and double (and long
+        // double where it is binary64) are converted by the library itself,
+        // correctly rounded and independent of the locale; other long double
+        // formats use std::from_chars when available, otherwise the
+        // locale-aware strtold.
        if (convert_float_fast(num_begin, num_end, decimal_point_position, mantissa_end, value_float))
        {
            return token_type::value_float;
@@ -22,6 +22,10 @@ namespace detail
 constexpr std::int64_t pow5_128_smallest_power = -342;
 constexpr std::int64_t pow5_128_largest_power = 308;

+// every entry of pow5_128() holds two 64-bit halves of 5^q, one per covered power of 5
+static_assert((pow5_128_largest_power - pow5_128_smallest_power + 1) * 2 == 1302,
+              "pow5_128_smallest_power/pow5_128_largest_power must match the size of the pow5_128() table");
+
 /*!
@brief 128-bit approximations of 5^q for q in [-342, 308]

@@ -8,10 +8,12 @@

 #pragma once

+#include <array> // array
 #include <cstddef> // size_t
-#include <cstdint> // uint64_t
+#include <cstdint> // uint64_t, uint8_t
 #include <cstring> // memcpy

+#include <nlohmann/detail/bit_ops.hpp>
 #include <nlohmann/detail/macro_scope.hpp>

 // Optional SIMD backend for bulk UTF-8 validation. This is an opt-in external
@@ -69,18 +71,12 @@ inline std::size_t find_string_special(const unsigned char* data, std::size_t n)
    std::size_t i = 0;
    for (; i + 8 <= n; i += 8)
    {
-        std::uint64_t word = 0;
-        std::memcpy(&word, data + i, sizeof(word));
-        if (swar_string_special(word) != 0)
+        const std::uint64_t special = swar_string_special(read_eight_bytes(data + i));
+        if (special != 0)
        {
-            // a special byte is in this word; locate it (endian-agnostic)
-            for (std::size_t j = 0; j < 8; ++j)
-            {
-                if (is_string_special(data[i + j]))
-                {
-                    return i + j;
-                }
-            }
+            // the lowest flagged byte is the first special one: the borrows of
+            // the subtractions can only flag bytes above a true hit
+            return i + (static_cast<std::size_t>(count_trailing_zeros(special)) / 8);
        }
    }
    for (; i < n; ++i)
@@ -114,8 +110,7 @@ inline std::size_t find_ascii_copyable_run(const unsigned char* data, std::size_
    std::size_t i = 0;
    for (; i + 8 <= n; i += 8)
    {
-        std::uint64_t v = 0;
-        std::memcpy(&v, data + i, sizeof(v));
+        const std::uint64_t v = read_eight_bytes(data + i);
        const std::uint64_t q = v ^ 0x2222222222222222ull; // '"'  (0x22)
        const std::uint64_t b = v ^ 0x5C5C5C5C5C5C5C5Cull; // '\\' (0x5C)
        const std::uint64_t d = v ^ 0x7F7F7F7F7F7F7F7Full; // DEL  (0x7F)
@@ -126,7 +121,9 @@ inline std::size_t find_ascii_copyable_run(const unsigned char* data, std::size_
                                   | (v & high);               // >= 0x80
        if (stop != 0)
        {
-            break;
+            // the lowest flagged byte is the first one to stop at (see
+            // find_string_special())
+            return i + (static_cast<std::size_t>(count_trailing_zeros(stop)) / 8);
        }
    }
    for (; i < n; ++i)
@@ -253,12 +250,18 @@ inline std::size_t scalar_string_bulk_run(const unsigned char* data, std::size_t
        {
            break; // end of buffer, or a quote/escape/control byte
        }
-        const std::size_t seq = validate_one_utf8(data + pos, n - pos);
-        if (seq == 0)
+        // a run of multi-byte sequences (e.g. CJK text) is validated sequence
+        // by sequence without searching for the next special byte in between
+        do
        {
-            break; // ill-formed or truncated: let the byte path diagnose it
+            const std::size_t seq = validate_one_utf8(data + pos, n - pos);
+            if (seq == 0)
+            {
+                return pos; // ill-formed or truncated: let the byte path diagnose it
+            }
+            pos += seq;
        }
-        pos += seq;
+        while (pos < n && data[pos] >= 0x80u);
    }
    return pos;
 }
@@ -273,8 +276,7 @@ inline std::size_t find_string_delimiter(const unsigned char* data, std::size_t
    std::size_t i = 0;
    for (; i + 8 <= n; i += 8)
    {
-        std::uint64_t v = 0;
-        std::memcpy(&v, data + i, sizeof(v));
+        const std::uint64_t v = read_eight_bytes(data + i);
        const std::uint64_t q = v ^ 0x2222222222222222ull;
        const std::uint64_t b = v ^ 0x5C5C5C5C5C5C5C5Cull;
        const std::uint64_t hit = ((q - ones) & ~q & high)
@@ -282,14 +284,8 @@ inline std::size_t find_string_delimiter(const unsigned char* data, std::size_t
                                  | ((v - 0x2020202020202020ull) & ~v & high);
        if (hit != 0)
        {
-            for (std::size_t j = 0; j < 8; ++j)
-            {
-                const unsigned char c = data[i + j];
-                if (c == '\"' || c == '\\' || c < 0x20u)
-                {
-                    return i + j;
-                }
-            }
+            // the lowest flagged byte is the first delimiter (see find_string_special())
+            return i + (static_cast<std::size_t>(count_trailing_zeros(hit)) / 8);
        }
    }
    for (; i < n; ++i)
@@ -320,5 +316,50 @@ inline std::size_t string_bulk_run(const unsigned char* data, std::size_t n) noe
    return scalar_string_bulk_run(data, n);
 }

+// Decode the 4 hex digits at [data, data+4) - the digits following a `\u`
+// escape - into a codepoint 0x0000..0xFFFF via one table lookup per byte
+// (after yyjson's read_hex_u16), or return -1 if any of the 4 bytes is not a
+// hex digit ('0'..'9', 'A'..'F', 'a'..'f'). The caller must already have
+// checked that 4 bytes are available; used by lexer::get_codepoint()'s
+// contiguous fast path. On -1 it falls back to the byte-at-a-time loop, which
+// stops at the first invalid digit, so the reported error and position are
+// unaffected by this fast path.
+inline int hex_codepoint(const unsigned char* data) noexcept
+{
+    static const std::array<std::uint8_t, 256> hex_digit_table = // NOLINT(cppcoreguidelines-avoid-non-const-global-variables)
+    {
+        {
+            0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, // 00..0F
+            0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, // 10..1F
+            0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, // 20..2F
+            0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07, 0x08, 0x09, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, // 30..3F ('0'..'9')
+            0xFF, 0x0A, 0x0B, 0x0C, 0x0D, 0x0E, 0x0F, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, // 40..4F ('A'..'F')
+            0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, // 50..5F
+            0xFF, 0x0A, 0x0B, 0x0C, 0x0D, 0x0E, 0x0F, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, // 60..6F ('a'..'f')
+            0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, // 70..7F
+            0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, // 80..8F
+            0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, // 90..9F
+            0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, // A0..AF
+            0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, // B0..BF
+            0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, // C0..CF
+            0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, // D0..DF
+            0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, // E0..EF
+            0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF  // F0..FF
+        }
+    };
+
+    const std::uint8_t d0 = hex_digit_table[data[0]];
+    const std::uint8_t d1 = hex_digit_table[data[1]];
+    const std::uint8_t d2 = hex_digit_table[data[2]];
+    const std::uint8_t d3 = hex_digit_table[data[3]];
+    // every valid digit is <= 0xF; the combined OR only exceeds it if at
+    // least one of the four bytes was not a hex digit (looked up as 0xFF)
+    if ((d0 | d1 | d2 | d3) > 0x0F)
+    {
+        return -1;
+    }
+    return (d0 << 12) | (d1 << 8) | (d2 << 4) | d3;
+}
+
 }  // namespace detail
 NLOHMANN_JSON_NAMESPACE_END
@@ -280,7 +280,7 @@ class iter_impl // NOLINT(cppcoreguidelines-special-member-functions,hicpp-speci
    }

    /// @throw invalid_iterator.214 always
-    JSON_HEDLEY_NO_RETURN void throw_cannot_get_value() const
+    JSON_HEDLEY_NO_RETURN JSON_HEDLEY_NEVER_INLINE void throw_cannot_get_value() const
    {
        JSON_THROW(invalid_iterator::create(exception_id::iterator_value_unavailable, "cannot get value", m_object));
    }
@@ -294,40 +294,25 @@ class iter_impl // NOLINT(cppcoreguidelines-special-member-functions,hicpp-speci
    {
        JSON_ASSERT(m_object != nullptr);

-        switch (m_object->m_data.m_type)
+        if (JSON_HEDLEY_LIKELY(m_object->m_data.m_type == value_t::array))
        {
-            case value_t::object:
-            {
-                JSON_ASSERT(m_it.object_iterator != m_object->m_data.m_value.object->end());
-                return m_it.object_iterator->second;
-            }
-
-            case value_t::array:
-            {
-                JSON_ASSERT(m_it.array_iterator != m_object->m_data.m_value.array->end());
-                return *m_it.array_iterator;
-            }
-
-            case value_t::null:
-                throw_cannot_get_value();
-
-            case value_t::string:
-            case value_t::boolean:
-            case value_t::number_integer:
-            case value_t::number_unsigned:
-            case value_t::number_float:
-            case value_t::binary:
-            case value_t::discarded:
-            default:
-            {
-                if (JSON_HEDLEY_LIKELY(m_it.primitive_iterator.is_begin()))
-                {
-                    return *m_object;
-                }
-
-                throw_cannot_get_value();
-            }
+            JSON_ASSERT(m_it.array_iterator != m_object->m_data.m_value.array->end());
+            return *m_it.array_iterator;
        }
+
+        if (JSON_HEDLEY_LIKELY(m_object->m_data.m_type == value_t::object))
+        {
+            JSON_ASSERT(m_it.object_iterator != m_object->m_data.m_value.object->end());
+            return m_it.object_iterator->second;
+        }
+
+        // neither array nor object: a primitive value (null has no value to point to)
+        if (JSON_HEDLEY_LIKELY(m_object->m_data.m_type != value_t::null && m_it.primitive_iterator.is_begin()))
+        {
+            return *m_object;
+        }
+
+        throw_cannot_get_value();
    }

    /*!
@@ -338,38 +323,25 @@ class iter_impl // NOLINT(cppcoreguidelines-special-member-functions,hicpp-speci
    {
        JSON_ASSERT(m_object != nullptr);

-        switch (m_object->m_data.m_type)
+        if (JSON_HEDLEY_LIKELY(m_object->m_data.m_type == value_t::array))
        {
-            case value_t::object:
-            {
-                JSON_ASSERT(m_it.object_iterator != m_object->m_data.m_value.object->end());
-                return &(m_it.object_iterator->second);
-            }
-
-            case value_t::array:
-            {
-                JSON_ASSERT(m_it.array_iterator != m_object->m_data.m_value.array->end());
-                return &*m_it.array_iterator;
-            }
-
-            case value_t::null:
-            case value_t::string:
-            case value_t::boolean:
-            case value_t::number_integer:
-            case value_t::number_unsigned:
-            case value_t::number_float:
-            case value_t::binary:
-            case value_t::discarded:
-            default:
-            {
-                if (JSON_HEDLEY_LIKELY(m_it.primitive_iterator.is_begin()))
-                {
-                    return m_object;
-                }
-
-                throw_cannot_get_value();
-            }
+            JSON_ASSERT(m_it.array_iterator != m_object->m_data.m_value.array->end());
+            return &*m_it.array_iterator;
        }
+
+        if (JSON_HEDLEY_LIKELY(m_object->m_data.m_type == value_t::object))
+        {
+            JSON_ASSERT(m_it.object_iterator != m_object->m_data.m_value.object->end());
+            return &(m_it.object_iterator->second);
+        }
+
+        // neither array nor object: a primitive value
+        if (JSON_HEDLEY_LIKELY(m_it.primitive_iterator.is_begin()))
+        {
+            return m_object;
+        }
+
+        throw_cannot_get_value();
    }

    /*!
@@ -391,35 +363,20 @@ class iter_impl // NOLINT(cppcoreguidelines-special-member-functions,hicpp-speci
    {
        JSON_ASSERT(m_object != nullptr);

-        switch (m_object->m_data.m_type)
+        if (JSON_HEDLEY_LIKELY(m_object->m_data.m_type == value_t::array))
        {
-            case value_t::object:
-            {
-                std::advance(m_it.object_iterator, 1);
-                break;
-            }
-
-            case value_t::array:
-            {
-                std::advance(m_it.array_iterator, 1);
-                break;
-            }
-
-            case value_t::null:
-            case value_t::string:
-            case value_t::boolean:
-            case value_t::number_integer:
-            case value_t::number_unsigned:
-            case value_t::number_float:
-            case value_t::binary:
-            case value_t::discarded:
-            default:
-            {
-                ++m_it.primitive_iterator;
-                break;
-            }
+            ++m_it.array_iterator;
+            return *this;
        }

+        if (JSON_HEDLEY_LIKELY(m_object->m_data.m_type == value_t::object))
+        {
+            std::advance(m_it.object_iterator, 1);
+            return *this;
+        }
+
+        // neither array nor object: a primitive value
+        ++m_it.primitive_iterator;
        return *this;
    }

@@ -442,35 +399,20 @@ class iter_impl // NOLINT(cppcoreguidelines-special-member-functions,hicpp-speci
    {
        JSON_ASSERT(m_object != nullptr);

-        switch (m_object->m_data.m_type)
+        if (JSON_HEDLEY_LIKELY(m_object->m_data.m_type == value_t::array))
        {
-            case value_t::object:
-            {
-                std::advance(m_it.object_iterator, -1);
-                break;
-            }
-
-            case value_t::array:
-            {
-                std::advance(m_it.array_iterator, -1);
-                break;
-            }
-
-            case value_t::null:
-            case value_t::string:
-            case value_t::boolean:
-            case value_t::number_integer:
-            case value_t::number_unsigned:
-            case value_t::number_float:
-            case value_t::binary:
-            case value_t::discarded:
-            default:
-            {
-                --m_it.primitive_iterator;
-                break;
-            }
+            --m_it.array_iterator;
+            return *this;
        }

+        if (JSON_HEDLEY_LIKELY(m_object->m_data.m_type == value_t::object))
+        {
+            std::advance(m_it.object_iterator, -1);
+            return *this;
+        }
+
+        // neither array nor object: a primitive value
+        --m_it.primitive_iterator;
        return *this;
    }

@@ -484,7 +426,7 @@ class iter_impl // NOLINT(cppcoreguidelines-special-member-functions,hicpp-speci
        // if objects are not the same, the comparison is undefined
        if (JSON_HEDLEY_UNLIKELY(m_object != other.m_object))
        {
-            JSON_THROW(invalid_iterator::create(exception_id::iterators_compare_different_values, "cannot compare iterators of different containers", m_object));
+            throw_literal<invalid_iterator>(exception_id::iterators_compare_different_values, "cannot compare iterators of different containers", m_object);
        }

        // value-initialized forward iterators can be compared, and must compare equal to other value-initialized iterators of the same type #4493
@@ -493,25 +435,18 @@ class iter_impl // NOLINT(cppcoreguidelines-special-member-functions,hicpp-speci
            return true;
        }

-        switch (m_object->m_data.m_type)
+        if (JSON_HEDLEY_LIKELY(m_object->m_data.m_type == value_t::array))
        {
-            case value_t::object:
-                return (m_it.object_iterator == other.m_it.object_iterator);
-
-            case value_t::array:
-                return (m_it.array_iterator == other.m_it.array_iterator);
-
-            case value_t::null:
-            case value_t::string:
-            case value_t::boolean:
-            case value_t::number_integer:
-            case value_t::number_unsigned:
-            case value_t::number_float:
-            case value_t::binary:
-            case value_t::discarded:
-            default:
-                return (m_it.primitive_iterator == other.m_it.primitive_iterator);
+            return (m_it.array_iterator == other.m_it.array_iterator);
        }
+
+        if (JSON_HEDLEY_LIKELY(m_object->m_data.m_type == value_t::object))
+        {
+            return (m_it.object_iterator == other.m_it.object_iterator);
+        }
+
+        // neither array nor object: a primitive value
+        return (m_it.primitive_iterator == other.m_it.primitive_iterator);
    }

    /*!
@@ -533,7 +468,7 @@ class iter_impl // NOLINT(cppcoreguidelines-special-member-functions,hicpp-speci
        // if objects are not the same, the comparison is undefined
        if (JSON_HEDLEY_UNLIKELY(m_object != other.m_object))
        {
-            JSON_THROW(invalid_iterator::create(exception_id::iterators_compare_different_values, "cannot compare iterators of different containers", m_object));
+            throw_literal<invalid_iterator>(exception_id::iterators_compare_different_values, "cannot compare iterators of different containers", m_object);
        }

        // value-initialized forward iterators can be compared, and must compare equal to other value-initialized iterators of the same type #4493
@@ -546,7 +481,7 @@ class iter_impl // NOLINT(cppcoreguidelines-special-member-functions,hicpp-speci
        switch (m_object->m_data.m_type)
        {
            case value_t::object:
-                JSON_THROW(invalid_iterator::create(exception_id::iterator_order_on_object, "cannot compare order of object iterators", m_object));
+                throw_literal<invalid_iterator>(exception_id::iterator_order_on_object, "cannot compare order of object iterators", m_object);

            case value_t::array:
                return (m_it.array_iterator < other.m_it.array_iterator);
@@ -602,7 +537,7 @@ class iter_impl // NOLINT(cppcoreguidelines-special-member-functions,hicpp-speci
        switch (m_object->m_data.m_type)
        {
            case value_t::object:
-                JSON_THROW(invalid_iterator::create(exception_id::iterator_arithmetic_on_object, "cannot use offsets with object iterators", m_object));
+                throw_literal<invalid_iterator>(exception_id::iterator_arithmetic_on_object, "cannot use offsets with object iterators", m_object);

            case value_t::array:
            {
@@ -681,7 +616,7 @@ class iter_impl // NOLINT(cppcoreguidelines-special-member-functions,hicpp-speci
        switch (m_object->m_data.m_type)
        {
            case value_t::object:
-                JSON_THROW(invalid_iterator::create(exception_id::iterator_arithmetic_on_object, "cannot use offsets with object iterators", m_object));
+                throw_literal<invalid_iterator>(exception_id::iterator_arithmetic_on_object, "cannot use offsets with object iterators", m_object);

            case value_t::array:
                return m_it.array_iterator - other.m_it.array_iterator;
@@ -710,7 +645,7 @@ class iter_impl // NOLINT(cppcoreguidelines-special-member-functions,hicpp-speci
        switch (m_object->m_data.m_type)
        {
            case value_t::object:
-                JSON_THROW(invalid_iterator::create(exception_id::iterator_subscript_on_object, "cannot use operator[] for object iterators", m_object));
+                throw_literal<invalid_iterator>(exception_id::iterator_subscript_on_object, "cannot use operator[] for object iterators", m_object);

            case value_t::array:
                return *std::next(m_it.array_iterator, n);
@@ -750,7 +685,7 @@ class iter_impl // NOLINT(cppcoreguidelines-special-member-functions,hicpp-speci
            return m_it.object_iterator->first;
        }

-        JSON_THROW(invalid_iterator::create(exception_id::iterator_key_not_object, "cannot use key() for non-object iterators", m_object));
+        throw_literal<invalid_iterator>(exception_id::iterator_key_not_object, "cannot use key() for non-object iterators", m_object);
    }

    /*!
@@ -16,14 +16,13 @@
    #include <iosfwd> // ostream
 #endif  // JSON_NO_IO
 #include <limits> // max
-#include <map> // map
-#include <numeric> // accumulate
 #include <string> // string
 #include <utility> // move
 #include <vector> // vector

 #include <nlohmann/detail/exceptions.hpp>
 #include <nlohmann/detail/macro_scope.hpp>
+#include <nlohmann/detail/meta/cpp_future.hpp>
 #include <nlohmann/detail/string_concat.hpp>
 #include <nlohmann/detail/string_escape.hpp>
 #include <nlohmann/detail/string_utils.hpp>
@@ -69,12 +68,21 @@ class json_pointer
    /// @sa https://json.nlohmann.me/api/json_pointer/to_string/
    string_t to_string() const
    {
-        return std::accumulate(reference_tokens.begin(), reference_tokens.end(),
-                               string_t{},
-                               [](const string_t& a, const string_t& b)
+        // one buffer: every token adds its escaped form (at least its length) and a slash
+        std::size_t length = 0;
+        for (const auto& reference_token : reference_tokens)
        {
-            return detail::concat<string_t>(a, '/', detail::escape(b));
-        });
+            length += reference_token.size() + 1;
+        }
+
+        string_t result;
+        result.reserve(length);
+        for (const auto& reference_token : reference_tokens)
+        {
+            result.append("/", 1);
+            detail::escape_into(result, reference_token);
+        }
+        return result;
    }

    /// @brief return a string representation of the JSON pointer
@@ -129,21 +137,29 @@ class json_pointer
    friend json_pointer operator/(const json_pointer& lhs,
                                  const json_pointer& rhs)
    {
-        return json_pointer(lhs) /= rhs;
+        json_pointer result;
+        result.reference_tokens.reserve(lhs.reference_tokens.size() + rhs.reference_tokens.size());
+        result.reference_tokens.insert(result.reference_tokens.end(), lhs.reference_tokens.begin(), lhs.reference_tokens.end());
+        result.reference_tokens.insert(result.reference_tokens.end(), rhs.reference_tokens.begin(), rhs.reference_tokens.end());
+        return result;
    }

    /// @brief create a new JSON pointer by appending the unescaped token at the end of the JSON pointer
    /// @sa https://json.nlohmann.me/api/json_pointer/operator_slash/
    friend json_pointer operator/(const json_pointer& lhs, string_t token) // NOLINT(performance-unnecessary-value-param)
    {
-        return json_pointer(lhs) /= std::move(token);
+        json_pointer result;
+        result.reference_tokens.reserve(lhs.reference_tokens.size() + 1);
+        result.reference_tokens.insert(result.reference_tokens.end(), lhs.reference_tokens.begin(), lhs.reference_tokens.end());
+        result.reference_tokens.push_back(std::move(token));
+        return result;
    }

    /// @brief create a new JSON pointer by appending the array-index-token at the end of the JSON pointer
    /// @sa https://json.nlohmann.me/api/json_pointer/operator_slash/
    friend json_pointer operator/(const json_pointer& lhs, std::size_t array_idx)
    {
-        return json_pointer(lhs) /= array_idx;
+        return lhs / detail::to_string<string_t>(array_idx);
    }

    /// @brief returns the parent of this JSON pointer
@@ -155,8 +171,8 @@ class json_pointer
            return *this;
        }

-        json_pointer res = *this;
-        res.pop_back();
+        json_pointer res;
+        res.reference_tokens = std::vector<string_t>(reference_tokens.begin(), reference_tokens.end() - 1);
        return res;
    }

@@ -245,14 +261,14 @@ class json_pointer

  private:
    /// @throw out_of_range.405 always
-    JSON_HEDLEY_NO_RETURN static void throw_no_parent()
+    JSON_HEDLEY_NO_RETURN JSON_HEDLEY_NEVER_INLINE static void throw_no_parent()
    {
        JSON_THROW(detail::out_of_range::create(detail::exception_id::patch_on_root, "JSON pointer has no parent", nullptr));
    }

    /// @throw out_of_range.404 always
    template<typename BasicJsonContext>
-    JSON_HEDLEY_NO_RETURN static void throw_unresolved(const string_t& reference_token, BasicJsonContext context)
+    JSON_HEDLEY_NO_RETURN JSON_HEDLEY_NEVER_INLINE static void throw_unresolved(const string_t& reference_token, BasicJsonContext context)
    {
        static_cast<void>(reference_token); // unused when JSON_NOEXCEPTION is defined
        static_cast<void>(context);
@@ -305,6 +321,28 @@ class json_pointer
            return array_index_status::not_a_number;
        }

+        // fast path: up to 18 decimal digits cannot overflow unsigned long long or
+        // size_type (9 digits for a 32-bit size_type), so strtoull and errno are not needed
+        constexpr std::size_t max_fast_digits = sizeof(size_type) >= 8 ? 18 : 9;
+        if (!s.empty() && s.size() <= max_fast_digits)
+        {
+            size_type acc = 0;
+            for (std::size_t i = 0; i < s.size(); ++i)
+            {
+                const char c = s[i];
+                if (JSON_HEDLEY_UNLIKELY(c < '0' || c > '9'))
+                {
+                    break; // not a plain number: let strtoull decide
+                }
+                acc = static_cast<size_type>(acc * 10 + static_cast<size_type>(c - '0'));
+                if (JSON_HEDLEY_UNLIKELY(i == s.size() - 1))
+                {
+                    idx = acc;
+                    return array_index_status::ok;
+                }
+            }
+        }
+
        const char* p = s.data();
        char* p_end = nullptr; // NOLINT(misc-const-correctness)
        errno = 0; // strtoull doesn't reset errno
@@ -361,6 +399,83 @@ class json_pointer
        return idx;
    }

+    /*!
+    @brief compare with a JSON pointer string without splitting it into tokens
+
+    @throw parse_error.107  if the string is not empty or begins with '/'
+    @throw parse_error.108  if character '~' is not followed by '0' or '1'
+    */
+    bool equals_string(const string_t& reference_string) const
+    {
+        if (reference_string.empty())
+        {
+            return reference_tokens.empty();
+        }
+
+        if (JSON_HEDLEY_UNLIKELY(reference_string[0] != '/'))
+        {
+            JSON_THROW(detail::parse_error::create(detail::exception_id::pointer_missing_slash, 1, detail::concat("JSON pointer must be empty or begin with '/' - was: '", reference_string, "'"), nullptr));
+        }
+
+        // the whole string is validated, whether or not it matches
+        const std::size_t size = reference_string.size();
+        for (std::size_t pos = reference_string.find_first_of('~');
+                pos != string_t::npos;
+                pos = reference_string.find_first_of('~', pos + 1))
+        {
+            if (JSON_HEDLEY_UNLIKELY(pos == size - 1 || (reference_string[pos + 1] != '0' && reference_string[pos + 1] != '1')))
+            {
+                JSON_THROW(detail::parse_error::create(detail::exception_id::pointer_invalid_escape, 0, "escape character '~' must be followed with '0' or '1'", nullptr));
+            }
+        }
+
+        // compare token by token, unescaping the string on the fly
+        std::size_t pos = 0; // the slash in front of the next token
+        for (const auto& reference_token : reference_tokens)
+        {
+            if (pos >= size)
+            {
+                return false;
+            }
+
+            std::size_t i = pos + 1;
+            std::size_t j = 0;
+            const std::size_t token_size = reference_token.size();
+            while (i < size && reference_string[i] != '/')
+            {
+                char c = reference_string[i++];
+                if (c == '~')
+                {
+                    c = (reference_string[i++] == '0') ? '~' : '/';
+                }
+                if (j == token_size || reference_token[j] != c)
+                {
+                    return false;
+                }
+                ++j;
+            }
+            if (j != token_size)
+            {
+                return false;
+            }
+            pos = i;
+        }
+        return pos == size;
+    }
+
+    /// @brief object member access with a key that is only copied if it is inserted
+    template<typename BasicJsonType>
+    static BasicJsonType& object_member(BasicJsonType& object, const string_t& key)
+    {
+        auto& members = *object.m_data.m_value.object;
+        const auto it = members.find(key);
+        if (it != members.end())
+        {
+            return it->second;
+        }
+        return object.operator[](key);
+    }
+
  JSON_PRIVATE_UNLESS_TESTED:
    json_pointer top() const
    {
@@ -384,84 +499,141 @@ class json_pointer
    the result independent of the order in which the flattened object is
    iterated, which is unspecified for some object types.

-    The prefixes form a tree and are numbered, so each of them is stored only
-    once (as a node) rather than as a copy of all of its reference tokens.
+    The prefixes form a tree and are numbered (number 0 is the empty prefix),
+    so each of them is stored only once (as a node) rather than as a copy of
+    all of its reference tokens. The tree does not own the reference tokens:
+    it refers to the tokens of the parsed keys, which must outlive it.
    */
    struct prefix_tree
    {
-        // children[id] maps a reference token to the number of the prefix
-        // extended by that token; number 0 is the empty prefix
-        std::vector<std::map<string_t, std::size_t>> children;
+        // open-addressing hash table from (parent, last reference token) to
+        // the number of the node; empty slots have token == nullptr
+        struct slot
+        {
+            std::size_t hash;
+            std::size_t parent;
+            const string_t* token;
+            std::size_t id;
+
+            slot() noexcept
+                : hash(0), parent(0), token(nullptr), id(0)
+            {}
+
+            slot(std::size_t h, std::size_t p, const string_t* t, std::size_t i) noexcept
+                : hash(h), parent(p), token(t), id(i)
+            {}
+        };
+
+        std::vector<slot> slots;
+        std::size_t used = 0;
+
+        static std::size_t hash_of(std::size_t parent, const string_t& token) noexcept
+        {
+            std::size_t h = parent * static_cast<std::size_t>(0x9E3779B97F4A7C15ULL);
+            for (std::size_t i = 0; i < token.size(); ++i)
+            {
+                h = (h ^ static_cast<unsigned char>(token[i])) * static_cast<std::size_t>(0x100000001B3ULL);
+            }
+            return h ^ (h >> 29u);
+        }
+
+        void grow()
+        {
+            std::vector<slot> old(slots.size() * 2);
+            old.swap(slots);
+            const std::size_t mask = slots.size() - 1;
+            for (const auto& s : old)
+            {
+                if (s.token != nullptr)
+                {
+                    std::size_t i = s.hash & mask;
+                    while (slots[i].token != nullptr)
+                    {
+                        i = (i + 1) & mask;
+                    }
+                    slots[i] = s;
+                }
+            }
+        }
+
+        // parent[id] is the number of the prefix without its last token
+        std::vector<std::size_t> parent;
        // is_array[id] is true iff some flattened key has the reference token
        // 0 directly below the prefix with number id
        std::vector<bool> is_array;

        // start with the empty prefix only
        prefix_tree()
-            : children(1)
+            : slots(16)
+            , parent(1, 0)
            , is_array(1, false)
        {}

        // return the number of the prefix with number id extended by
        // reference_token, adding it if it is new
-        std::size_t add_child(std::size_t id, string_t&& reference_token)
+        std::size_t add_child(std::size_t id, const string_t& reference_token)
        {
+            const std::size_t hash = hash_of(id, reference_token);
+            const std::size_t mask = slots.size() - 1;
+            std::size_t i = hash & mask;
+            while (slots[i].token != nullptr)
+            {
+                if (slots[i].hash == hash && slots[i].parent == id && *slots[i].token == reference_token)
+                {
+                    return slots[i].id;
+                }
+                i = (i + 1) & mask;
+            }
+
+            // new prefix
+            const std::size_t child = parent.size();
+            slots[i] = slot(hash, id, &reference_token, child);
            if (reference_token == "0")
            {
                is_array[id] = true;
            }
+            parent.push_back(id);
+            is_array.push_back(false);

-            // read the number before the emplace_back below, which may
-            // reallocate children and invalidate the iterator
-            const std::size_t next = children.size();
-            const auto inserted = children[id].emplace(std::move(reference_token), next);
-            const std::size_t child = inserted.first->second;
-            if (inserted.second)
+            // keep the load factor below 1/2
+            if (++used * 2 > slots.size())
            {
-                children.emplace_back();
-                is_array.push_back(false);
+                grow();
            }
            return child;
        }
-
-        // return the number of the prefix with number id extended by
-        // reference_token, which must have been added before
-        std::size_t find_child(std::size_t id, const string_t& reference_token) const
-        {
-            const auto it = children[id].find(reference_token);
-            JSON_ASSERT(it != children[id].end());
-            return it->second;
-        }
    };

    /*!
    @brief create and return a reference to the pointed to value

-    Complexity: Linear in the number of reference tokens (times the logarithm
-    of the number of siblings for the prefix lookup).
+    @param[in] j  the value to create the pointed to value in
+    @param[in] tokens  the reference tokens of the pointer
+    @param[in] prefixes  prefixes[i] is the number in @a tree of the prefix
+                         consisting of the first i reference tokens
+
+    Complexity: Linear in the number of reference tokens.

    @throw parse_error.106 if an array index begins with '0'
    @throw parse_error.109 if array index is not a number
    @throw type_error.313 if value cannot be unflattened
    */
    template<typename BasicJsonType>
-    BasicJsonType& get_and_create(BasicJsonType& j, const prefix_tree& tree) const
+    static BasicJsonType& get_and_create(BasicJsonType& j, const std::vector<string_t>& tokens,
+                                         const std::vector<std::size_t>& prefixes, const prefix_tree& tree)
    {
        auto* result = &j;

-        // the number of the prefix consumed so far; used to look up whether
-        // the value to be created below is an array or an object
-        std::size_t id = 0;
-
        // in case no reference tokens exist, return a reference to the JSON value
        // j which will be overwritten by a primitive value
-        for (const auto& reference_token : reference_tokens)
+        for (std::size_t i = 0; i < tokens.size(); ++i)
        {
+            const auto& reference_token = tokens[i];
            switch (result->type())
            {
                case detail::value_t::null:
                {
-                    if (tree.is_array[id])
+                    if (tree.is_array[prefixes[i]])
                    {
                        // some reference token below this position is 0, so the
                        // value is an array
@@ -478,7 +650,7 @@ class json_pointer
                case detail::value_t::object:
                {
                    // create an entry in the object
-                    result = &result->operator[](reference_token);
+                    result = &object_member(*result, reference_token);
                    break;
                }

@@ -505,8 +677,6 @@ class json_pointer
                default:
                    JSON_THROW(detail::type_error::create(detail::exception_id::unflatten_invalid_value, "invalid value to unflatten", &j));
            }
-
-            id = tree.find_child(id, reference_token);
        }

        return *result;
@@ -558,7 +728,7 @@ class json_pointer
                case detail::value_t::object:
                {
                    // use unchecked object access
-                    ptr = &ptr->operator[](reference_token);
+                    ptr = &object_member(*ptr, reference_token);
                    break;
                }

@@ -608,8 +778,13 @@ class json_pointer
            {
                case detail::value_t::object:
                {
-                    // note: at performs range check
-                    ptr = &ptr->at(reference_token);
+                    const auto it = ptr->m_data.m_value.object->find(reference_token);
+                    if (JSON_HEDLEY_UNLIKELY(it == ptr->m_data.m_value.object->end()))
+                    {
+                        // at() throws out_of_range.403 with the usual message
+                        ptr->at(reference_token);
+                    }
+                    ptr = &it->second;
                    break;
                }

@@ -630,7 +805,7 @@ class json_pointer
                        JSON_THROW(detail::out_of_range::create(detail::exception_id::array_index_out_of_range, detail::concat(
                                "array index ", std::to_string(idx), " is out of range"), ptr));
                    }
-                    ptr = &ptr->operator[](idx);
+                    ptr = &(*ptr->m_data.m_value.array)[idx];
                    break;
                }

@@ -678,7 +853,9 @@ class json_pointer
                {
                    // use unchecked object access; the const operator[]
                    // asserts that the key exists
-                    ptr = &ptr->operator[](reference_token);
+                    const auto it = ptr->m_data.m_value.object->find(reference_token);
+                    JSON_ASSERT(it != ptr->m_data.m_value.object->end());
+                    ptr = &it->second;
                    break;
                }

@@ -692,7 +869,9 @@ class json_pointer

                    // use unchecked array access; the const operator[]
                    // asserts that the index exists
-                    ptr = &ptr->operator[](array_index<BasicJsonType>(reference_token));
+                    const auto idx = array_index<BasicJsonType>(reference_token);
+                    JSON_ASSERT(idx < ptr->m_data.m_value.array->size());
+                    ptr = &(*ptr->m_data.m_value.array)[idx];
                    break;
                }

@@ -733,12 +912,12 @@ class json_pointer
            {
                case detail::value_t::object:
                {
-                    const auto it = ptr->find(reference_token);
-                    if (JSON_HEDLEY_UNLIKELY(it == ptr->end()))
+                    const auto it = ptr->m_data.m_value.object->find(reference_token);
+                    if (JSON_HEDLEY_UNLIKELY(it == ptr->m_data.m_value.object->end()))
                    {
                        return nullptr;
                    }
-                    ptr = &*it;
+                    ptr = &it->second;
                    break;
                }

@@ -775,7 +954,7 @@ class json_pointer
                    {
                        return nullptr;
                    }
-                    ptr = &ptr->operator[](idx);
+                    ptr = &(*ptr->m_data.m_value.array)[idx];
                    break;
                }

@@ -808,13 +987,14 @@ class json_pointer
            {
                case detail::value_t::object:
                {
-                    if (!ptr->contains(reference_token))
+                    const auto it = ptr->m_data.m_value.object->find(reference_token);
+                    if (it == ptr->m_data.m_value.object->end())
                    {
                        // we did not find the key in the object
                        return false;
                    }

-                    ptr = &ptr->operator[](reference_token);
+                    ptr = &it->second;
                    break;
                }

@@ -836,13 +1016,13 @@ class json_pointer
                        return false;
                    }

-                    if (idx >= ptr->size())
+                    if (idx >= ptr->m_data.m_value.array->size())
                    {
                        // index out of range
                        return false;
                    }

-                    ptr = &ptr->operator[](idx);
+                    ptr = &(*ptr->m_data.m_value.array)[idx];
                    break;
                }

@@ -892,6 +1072,15 @@ class json_pointer
            JSON_THROW(detail::parse_error::create(detail::exception_id::pointer_missing_slash, 1, detail::concat("JSON pointer must be empty or begin with '/' - was: '", reference_string, "'"), nullptr));
        }

+        // every slash starts a reference token
+        const std::size_t size = reference_string.size();
+        std::size_t token_count = 0;
+        for (std::size_t i = 0; i < size; ++i)
+        {
+            token_count += (reference_string[i] == '/') ? 1 : 0;
+        }
+        result.reserve(token_count);
+
        // extract the reference tokens:
        // - slash: position of the last read slash (or end of string)
        // - start: position after the previous slash
@@ -913,31 +1102,52 @@ class json_pointer
            const auto count = (slash == string_t::npos ? reference_string.size() : slash) - start;
            auto reference_token = string_t(reference_string.data() + start, count);

-            // check reference tokens are properly escaped
-            for (std::size_t pos = reference_token.find_first_of('~');
-                    pos != string_t::npos;
-                    pos = reference_token.find_first_of('~', pos + 1))
+            // check reference tokens are properly escaped; a token without
+            // a '~' (the common case) needs neither check nor unescaping
+            std::size_t pos = reference_token.find_first_of('~');
+            if (pos != string_t::npos)
            {
-                JSON_ASSERT(reference_token[pos] == '~');
-
-                // ~ must be followed by 0 or 1
-                if (JSON_HEDLEY_UNLIKELY(pos == reference_token.size() - 1 ||
-                                         (reference_token[pos + 1] != '0' &&
-                                          reference_token[pos + 1] != '1')))
+                for (; pos != string_t::npos; pos = reference_token.find_first_of('~', pos + 1))
                {
-                    JSON_THROW(detail::parse_error::create(detail::exception_id::pointer_invalid_escape, 0, "escape character '~' must be followed with '0' or '1'", nullptr));
+                    JSON_ASSERT(reference_token[pos] == '~');
+
+                    // ~ must be followed by 0 or 1
+                    if (JSON_HEDLEY_UNLIKELY(pos == reference_token.size() - 1 ||
+                                             (reference_token[pos + 1] != '0' &&
+                                              reference_token[pos + 1] != '1')))
+                    {
+                        JSON_THROW(detail::parse_error::create(detail::exception_id::pointer_invalid_escape, 0, "escape character '~' must be followed with '0' or '1'", nullptr));
+                    }
                }
+
+                detail::unescape(reference_token);
            }

            // finally, store the reference token
-            detail::unescape(reference_token);
-            result.push_back(reference_token);
+            result.push_back(std::move(reference_token));
        }

        return result;
    }

  private:
+    // Add the entry (key, v) to the object `obj`. Flattened paths are pairwise
+    // distinct, so objects that can append a key known to be absent
+    // (ordered_map) skip the lookup, which is linear and would make flatten
+    // quadratic; all other object types use operator[].
+    template<typename ObjectType, typename ValueType>
+    static auto add_flattened(ObjectType& obj, const string_t& key, ValueType&& v, detail::priority_tag<1> /*unused*/)
+    -> decltype(obj.emplace_new_key(key, std::forward<ValueType>(v)), void())
+    {
+        obj.emplace_new_key(key, std::forward<ValueType>(v));
+    }
+
+    template<typename ObjectType, typename ValueType>
+    static void add_flattened(ObjectType& obj, const string_t& key, ValueType&& v, detail::priority_tag<0> /*unused*/)
+    {
+        obj[key] = std::forward<ValueType>(v);
+    }
+
    /*!
    @param[in] reference_string  the reference string to the current value
    @param[in] value             the value to consider
@@ -982,7 +1192,8 @@ class json_pointer

        // flatten `v`, whose path is `path`: primitives and empty containers
        // are added to the result right away; other containers get a frame
-        const auto enter = [&stack, &path, &result](const BasicJsonType & v)
+        auto& object = *result.m_data.m_value.object;
+        const auto enter = [&stack, &path, &object](const BasicJsonType & v)
        {
            switch (v.type())
            {
@@ -991,7 +1202,7 @@ class json_pointer
                    if (v.m_data.m_value.array->empty())
                    {
                        // flatten empty array as null
-                        result[path] = nullptr;
+                        add_flattened(object, path, nullptr, detail::priority_tag<1> {});
                    }
                    else
                    {
@@ -1005,7 +1216,7 @@ class json_pointer
                    if (v.m_data.m_value.object->empty())
                    {
                        // flatten empty object as null
-                        result[path] = nullptr;
+                        add_flattened(object, path, nullptr, detail::priority_tag<1> {});
                    }
                    else
                    {
@@ -1025,7 +1236,7 @@ class json_pointer
                default:
                {
                    // add a primitive value with its reference string
-                    result[path] = v;
+                    add_flattened(object, path, v, detail::priority_tag<1> {});
                    return;
                }
            }
@@ -1071,6 +1282,10 @@ class json_pointer
                enter(it->second);
            }
        }
+
+        // the entries were added without going through basic_json, so the
+        // parent pointers (JSON_DIAGNOSTICS) have to be set now
+        result.set_parents();
    }

    /*!
@@ -1094,21 +1309,53 @@ class json_pointer

        BasicJsonType result;

+        // parse every key once; the reference tokens are kept for the second
+        // pass (the vectors are never resized, so the tree may refer to them)
+        std::vector<std::vector<string_t>> keys;
+        keys.reserve(value.m_data.m_value.object->size());
+        for (const auto& element : *value.m_data.m_value.object)
+        {
+            json_pointer ptr(element.first);
+            keys.push_back(std::move(ptr.reference_tokens));
+        }
+
        // collect the pointer prefixes that have a reference token 0 among
        // their children; the values below them are arrays, all others are
        // objects (see prefix_tree)
        prefix_tree tree;
-        for (const auto& element : *value.m_data.m_value.object)
+        // the number of the prefix with i + 1 tokens of the previous key; keys
+        // are often sorted, so most of the prefix is the same as for the
+        // previous key and need not be looked up again
+        std::vector<std::size_t> path;
+        const std::vector<string_t>* previous = nullptr;
+        // the number of the complete prefix (i.e., the key) of each key
+        std::vector<std::size_t> leaves;
+        leaves.reserve(keys.size());
+        for (const auto& tokens : keys)
        {
-            json_pointer ptr(element.first);
-            std::size_t id = 0;
-            for (auto& reference_token : ptr.reference_tokens)
+            std::size_t common = 0;
+            if (previous != nullptr)
            {
-                id = tree.add_child(id, std::move(reference_token));
+                const std::size_t limit = (std::min)(tokens.size(), previous->size());
+                while (common < limit && tokens[common] == (*previous)[common])
+                {
+                    ++common;
+                }
            }
+            path.resize(common);
+            std::size_t id = common == 0 ? 0 : path[common - 1];
+            for (std::size_t i = common; i < tokens.size(); ++i)
+            {
+                id = tree.add_child(id, tokens[i]);
+                path.push_back(id);
+            }
+            leaves.push_back(id);
+            previous = &tokens;
        }

        // iterate the JSON object values
+        std::vector<std::size_t> prefixes;
+        std::size_t key_index = 0;
        for (const auto& element : *value.m_data.m_value.object)
        {
            if (JSON_HEDLEY_UNLIKELY(!element.second.is_primitive()))
@@ -1116,11 +1363,21 @@ class json_pointer
                JSON_THROW(detail::type_error::create(detail::exception_id::unflatten_value_not_primitive, "values in object must be primitive", &element.second));
            }

+            const auto& tokens = keys[key_index];
+            // number the prefixes of this key from its last one to the first
+            prefixes.resize(tokens.size());
+            for (std::size_t i = tokens.size(), id = leaves[key_index]; i > 0; --i)
+            {
+                id = tree.parent[id];
+                prefixes[i - 1] = id;
+            }
+            ++key_index;
+
            // Assign the value to the reference pointed to by JSON pointer. Note
            // that if the JSON pointer is "" (i.e., points to the whole value),
            // function get_and_create returns a reference to the result itself.
            // An assignment will then create a primitive value.
-            json_pointer(element.first).get_and_create(result, tree) = element.second;
+            get_and_create(result, tokens, prefixes, tree) = element.second;
        }

        return result;
@@ -1159,7 +1416,7 @@ class json_pointer
        = delete;
 #else
    {
-        return *this == json_pointer(rhs);
+        return equals_string(rhs);
    }
 #endif

@@ -1243,7 +1500,7 @@ inline bool operator==(const json_pointer<RefStringTypeLhs>& lhs,
    = delete;
 #else
 {
-    return lhs == json_pointer<RefStringTypeLhs>(rhs);
+    return lhs.equals_string(rhs);
 }
 #endif

@@ -1256,7 +1513,7 @@ inline bool operator==(const StringType& lhs,
    = delete;
 #else
 {
-    return json_pointer<RefStringTypeRhs>(lhs) == rhs;
+    return rhs.equals_string(lhs);
 }
 #endif

@@ -21,6 +21,8 @@
 #undef JSON_NO_UNIQUE_ADDRESS
 #undef JSON_DISABLE_ENUM_SERIALIZATION
 #undef JSON_DISABLE_TUPLE_REFERENCE_CONVERSION
+#undef JSON_DTOA_SSE2
+#undef JSON_DTOA_NEON

 #ifndef JSON_TEST_KEEP_MACROS
    #undef JSON_CATCH
@@ -9,9 +9,11 @@
 #pragma once

 #include <limits> // numeric_limits
+#include <map> // map, multimap
 #include <string> // char_traits
 #include <tuple> // tuple
 #include <type_traits> // false_type, is_constructible, is_integral, is_same, true_type
+#include <unordered_map> // unordered_map, unordered_multimap
 #include <utility> // declval
 #include <vector> // vector
 #if defined(__cpp_lib_byte) && __cpp_lib_byte >= 201603L
@@ -683,6 +685,22 @@ struct is_compatible_integer_type
    : is_compatible_integer_type_impl<RealIntegerType,
      CompatibleNumberIntegerType> {};

+// whether to_json() stores an integer of type CompatibleNumberIntegerType as
+// number_integer
+template<typename BasicJsonType, typename CompatibleNumberIntegerType, typename = void>
+struct is_compatible_number_integer_type
+    : is_compatible_integer_type<typename BasicJsonType::number_integer_t, CompatibleNumberIntegerType> {};
+
+// if number_unsigned_t is the same type as number_integer_t, unsigned integers
+// are stored as number_integer as well (#2573)
+template<typename BasicJsonType, typename CompatibleNumberIntegerType>
+struct is_compatible_number_integer_type<BasicJsonType, CompatibleNumberIntegerType,
+           enable_if_t<conjunction<std::is_same<typename BasicJsonType::number_integer_t, typename BasicJsonType::number_unsigned_t>,
+           std::is_integral<CompatibleNumberIntegerType>,
+           std::is_unsigned<CompatibleNumberIntegerType>,
+           negation<std::is_same<bool, CompatibleNumberIntegerType>>>::value>>
+           : std::is_constructible<typename BasicJsonType::number_integer_t, CompatibleNumberIntegerType> {};
+
 template<typename BasicJsonType, typename CompatibleType, typename = void>
 struct is_compatible_type_impl: std::false_type {};

@@ -897,6 +915,31 @@ struct has_capacity : std::integral_constant<bool, is_detected<detect_capacity,
 template <typename T>
 struct is_ordered_map : has_capacity<T> {};

+// whether T is a specialization of nlohmann::ordered_map, which only ever
+// appends members, so they can only move when its capacity changes
+template<typename T>
+struct is_nlohmann_ordered_map : std::false_type {};
+
+template<class Key, class T, class IgnoredLess, class Allocator>
+struct is_nlohmann_ordered_map<ordered_map<Key, T, IgnoredLess, Allocator>> : std::true_type {};
+
+// whether adding or erasing elements never moves the other elements of the
+// container T; only node-based standard containers are known to guarantee this
+template<typename T>
+struct has_stable_elements : std::false_type {};
+
+template<class Key, class T, class Compare, class Allocator>
+struct has_stable_elements<std::map<Key, T, Compare, Allocator>> : std::true_type {};
+
+template<class Key, class T, class Compare, class Allocator>
+struct has_stable_elements<std::multimap<Key, T, Compare, Allocator>> : std::true_type {};
+
+template<class Key, class T, class Hash, class KeyEqual, class Allocator>
+struct has_stable_elements<std::unordered_map<Key, T, Hash, KeyEqual, Allocator>> : std::true_type {};
+
+template<class Key, class T, class Hash, class KeyEqual, class Allocator>
+struct has_stable_elements<std::unordered_multimap<Key, T, Hash, KeyEqual, Allocator>> : std::true_type {};
+
 // to avoid useless casts (see https://github.com/nlohmann/json/issues/2893#issuecomment-889152324)
 template < typename T, typename U, enable_if_t < !std::is_same<T, U>::value, int > = 0 >
 T conditional_static_cast(U value)
@@ -1020,6 +1063,25 @@ constexpr bool value_in_range_of(T val)
 template<bool Value>
 using bool_constant = std::integral_constant<bool, Value>;

+// Stands in for number_unsigned_t in overload sets if it is the same type as
+// number_integer_t, so that the unsigned overloads do not collide with the
+// integer ones. Never constructed. Such a basic_json stores every integer as
+// number_integer (#2573).
+struct same_as_number_integer_type {};
+
+// NumberUnsignedType, or same_as_number_integer_type if it is the same type as
+// NumberIntegerType; basic_json uses this one, as it cannot name its own member
+// types through an incomplete basic_json (MSVC reports a recursive alias, and
+// old GCC versions crash)
+template<typename NumberIntegerType, typename NumberUnsignedType>
+using number_unsigned_overload_type = typename std::conditional<std::is_same<NumberIntegerType, NumberUnsignedType>::value,
+      same_as_number_integer_type, NumberUnsignedType>::type;
+
+// number_unsigned_t of BasicJsonType, or same_as_number_integer_type if it is
+// the same type as number_integer_t
+template<typename BasicJsonType>
+using number_unsigned_overload_t = number_unsigned_overload_type<typename BasicJsonType::number_integer_t, typename BasicJsonType::number_unsigned_t>;
+
 ///////////////////////////////////////////////////////////////////////////////
 // is_c_string
 ///////////////////////////////////////////////////////////////////////////////
@@ -1536,7 +1536,7 @@ class binary_writer
                return calc_bson_integer_size(j.m_data.m_value.number_integer);

            case value_t::number_unsigned:
-                return calc_bson_unsigned_size(j.m_data.m_value.number_unsigned);
+                return calc_bson_unsigned_size(conditional_static_cast<std::uint64_t>(j.m_data.m_value.number_unsigned));

            case value_t::string:
                return calc_bson_string_size(*j.m_data.m_value.string, j);
@@ -1578,7 +1578,7 @@ class binary_writer
                return write_bson_integer(name, j.m_data.m_value.number_integer);

            case value_t::number_unsigned:
-                return write_bson_unsigned(name, j.m_data.m_value.number_unsigned);
+                return write_bson_unsigned(name, conditional_static_cast<std::uint64_t>(j.m_data.m_value.number_unsigned));

            case value_t::string:
                return write_bson_string(name, *j.m_data.m_value.string);
@@ -49,6 +49,12 @@ class serializer
    using number_float_t = typename BasicJsonType::number_float_t;
    using number_integer_t = typename BasicJsonType::number_integer_t;
    using number_unsigned_t = typename BasicJsonType::number_unsigned_t;
+    /// type of the absolute value of an integer: number_unsigned_t, or its
+    /// unsigned counterpart if it is signed (it may be number_integer_t, #2573)
+    using number_magnitude_t = typename std::conditional <
+                               std::is_integral<number_unsigned_t>::value && std::is_signed<number_unsigned_t>::value,
+                               std::make_unsigned<number_unsigned_t>,
+                               std::enable_if<true, number_unsigned_t >>::type::type;
    using binary_char_t = typename BasicJsonType::binary_t::value_type;

  public:
@@ -772,6 +778,33 @@ class serializer
                        break;
                    }
                }
+
+                // A byte below 0x60 is a complete code point and needs no
+                // UTF-8 decoding. The scanner stops only at a control
+                // character, a quote, or a backslash in that range, so look up
+                // its escape sequence and write all 8 bytes of the table entry
+                // straight into the write buffer; only the length stored in the
+                // last byte of the entry is kept.
+                if (data[i] < 0x60)
+                {
+                    // emit any bytes still pending in string_buffer first to
+                    // preserve output order
+                    if (bytes != 0)
+                    {
+                        put_buffer(string_buffer, bytes);
+                        bytes = 0;
+                    }
+                    const auto& escape = ascii_escapes()[data[i]];
+                    if (JSON_HEDLEY_UNLIKELY(write_buffer.size() - write_buffer_pos < escape.size()))
+                    {
+                        flush();
+                    }
+                    std::memcpy(write_buffer.data() + write_buffer_pos, escape.data(), escape.size());
+                    write_buffer_pos += static_cast<std::size_t>(escape[7]);
+                    bytes_after_last_accept = 0;
+                    undumped_chars = 0;
+                    continue;
+                }
            }

            const auto byte = static_cast<std::uint8_t>(s[i]);
@@ -780,92 +813,39 @@ class serializer
            {
                case UTF8_ACCEPT:  // decode found a new code point
                {
-                    switch (codepoint)
+                    // the code point is at least 0x60 (smaller ones are handled
+                    // above) and is not printable ASCII (which the scanner
+                    // copies): it is DEL (0x7F) under EnsureAscii or a non-ASCII
+                    // character
+                    if (EnsureAscii)
                    {
-                        case 0x08: // backspace
+                        JSON_ASSERT(codepoint >= 0x7F);
+                        if (error_handler == error_handler_t::keep)
                        {
-                            string_buffer[bytes++] = '\\';
-                            string_buffer[bytes++] = 'b';
-                            break;
+                            // this character was buffered as raw bytes below
+                            // in case it turned out to be part of an
+                            // ill-formed sequence (which is kept as is); now
+                            // that it decoded to a well-formed code point,
+                            // undo that and \u-escape it like any other
+                            // character instead
+                            bytes = bytes_after_last_accept;
                        }
-
-                        case 0x09: // horizontal tab
+                        if (codepoint <= 0xFFFF)
                        {
-                            string_buffer[bytes++] = '\\';
-                            string_buffer[bytes++] = 't';
-                            break;
+                            write_u_escape(bytes, static_cast<std::uint16_t>(codepoint));
                        }
-
-                        case 0x0A: // newline
+                        else
                        {
-                            string_buffer[bytes++] = '\\';
-                            string_buffer[bytes++] = 'n';
-                            break;
-                        }
-
-                        case 0x0C: // formfeed
-                        {
-                            string_buffer[bytes++] = '\\';
-                            string_buffer[bytes++] = 'f';
-                            break;
-                        }
-
-                        case 0x0D: // carriage return
-                        {
-                            string_buffer[bytes++] = '\\';
-                            string_buffer[bytes++] = 'r';
-                            break;
-                        }
-
-                        case 0x22: // quotation mark
-                        {
-                            string_buffer[bytes++] = '\\';
-                            string_buffer[bytes++] = '"';
-                            break;
-                        }
-
-                        case 0x5C: // reverse solidus
-                        {
-                            string_buffer[bytes++] = '\\';
-                            string_buffer[bytes++] = '\\';
-                            break;
-                        }
-
-                        default:
-                        {
-                            // escape control characters (0x00..0x1F) or, if
-                            // EnsureAscii parameter is used, non-ASCII characters
-                            if ((codepoint <= 0x1F) || (EnsureAscii && (codepoint >= 0x7F)))
-                            {
-                                if (EnsureAscii && error_handler == error_handler_t::keep)
-                                {
-                                    // this character was buffered as raw bytes
-                                    // below in case it turned out to be part of
-                                    // an ill-formed sequence (which is kept as
-                                    // is); now that it decoded to a well-formed
-                                    // code point, undo that and \u-escape it
-                                    // like any other character instead
-                                    bytes = bytes_after_last_accept;
-                                }
-                                if (codepoint <= 0xFFFF)
-                                {
-                                    write_u_escape(bytes, static_cast<std::uint16_t>(codepoint));
-                                }
-                                else
-                                {
-                                    write_u_escape(bytes, static_cast<std::uint16_t>(0xD7C0u + (codepoint >> 10u)));
-                                    write_u_escape(bytes, static_cast<std::uint16_t>(0xDC00u + (codepoint & 0x3FFu)));
-                                }
-                            }
-                            else
-                            {
-                                // copy byte to buffer (all previous bytes
-                                // been copied have in default case above)
-                                string_buffer[bytes++] = s[i];
-                            }
-                            break;
+                            write_u_escape(bytes, static_cast<std::uint16_t>(0xD7C0u + (codepoint >> 10u)));
+                            write_u_escape(bytes, static_cast<std::uint16_t>(0xDC00u + (codepoint & 0x3FFu)));
                        }
                    }
+                    else
+                    {
+                        // copy byte to buffer (all previous bytes of the
+                        // sequence have been copied in the default case below)
+                        string_buffer[bytes++] = s[i];
+                    }

                    // write buffer and reset index; there must be 13 bytes
                    // left, as this is the maximal number of bytes to be
@@ -1227,7 +1207,7 @@ class serializer
    @param[in] x  unsigned integer number to count its digits
    @return    number of decimal digits
    */
-    unsigned int count_digits(number_unsigned_t x) noexcept
+    unsigned int count_digits(number_magnitude_t x) noexcept
    {
        unsigned int n_digits = 1;
        for (;;)
@@ -1253,6 +1233,121 @@ class serializer
        }
    }

+    /*!
+    @brief the JSON form of each character below 0x60
+
+    Entry c holds the escape sequence of character c - the short form for
+    \\b, \\t, \\n, \\f, \\r, \\" and \\\\, \\u00xx for the other control characters -
+    or c itself if it needs no escaping, padded with zeros to 7 bytes. The last
+    byte is the length of the sequence. The 8 bytes of an entry are copied at
+    once, so writing an escape needs neither a branch nor a byte-wise loop.
+    */
+    static const std::array<std::array<char, 8>, 0x60>& ascii_escapes() noexcept
+    {
+        static const std::array<std::array<char, 8>, 0x60> escapes =
+        {
+            {
+                {{'\\', 'u', '0', '0', '0', '0', 0, 6}}, // 0x00
+                {{'\\', 'u', '0', '0', '0', '1', 0, 6}}, // 0x01
+                {{'\\', 'u', '0', '0', '0', '2', 0, 6}}, // 0x02
+                {{'\\', 'u', '0', '0', '0', '3', 0, 6}}, // 0x03
+                {{'\\', 'u', '0', '0', '0', '4', 0, 6}}, // 0x04
+                {{'\\', 'u', '0', '0', '0', '5', 0, 6}}, // 0x05
+                {{'\\', 'u', '0', '0', '0', '6', 0, 6}}, // 0x06
+                {{'\\', 'u', '0', '0', '0', '7', 0, 6}}, // 0x07
+                {{'\\', 'b', 0, 0, 0, 0, 0, 2}}, // 0x08
+                {{'\\', 't', 0, 0, 0, 0, 0, 2}}, // 0x09
+                {{'\\', 'n', 0, 0, 0, 0, 0, 2}}, // 0x0A
+                {{'\\', 'u', '0', '0', '0', 'b', 0, 6}}, // 0x0B
+                {{'\\', 'f', 0, 0, 0, 0, 0, 2}}, // 0x0C
+                {{'\\', 'r', 0, 0, 0, 0, 0, 2}}, // 0x0D
+                {{'\\', 'u', '0', '0', '0', 'e', 0, 6}}, // 0x0E
+                {{'\\', 'u', '0', '0', '0', 'f', 0, 6}}, // 0x0F
+                {{'\\', 'u', '0', '0', '1', '0', 0, 6}}, // 0x10
+                {{'\\', 'u', '0', '0', '1', '1', 0, 6}}, // 0x11
+                {{'\\', 'u', '0', '0', '1', '2', 0, 6}}, // 0x12
+                {{'\\', 'u', '0', '0', '1', '3', 0, 6}}, // 0x13
+                {{'\\', 'u', '0', '0', '1', '4', 0, 6}}, // 0x14
+                {{'\\', 'u', '0', '0', '1', '5', 0, 6}}, // 0x15
+                {{'\\', 'u', '0', '0', '1', '6', 0, 6}}, // 0x16
+                {{'\\', 'u', '0', '0', '1', '7', 0, 6}}, // 0x17
+                {{'\\', 'u', '0', '0', '1', '8', 0, 6}}, // 0x18
+                {{'\\', 'u', '0', '0', '1', '9', 0, 6}}, // 0x19
+                {{'\\', 'u', '0', '0', '1', 'a', 0, 6}}, // 0x1A
+                {{'\\', 'u', '0', '0', '1', 'b', 0, 6}}, // 0x1B
+                {{'\\', 'u', '0', '0', '1', 'c', 0, 6}}, // 0x1C
+                {{'\\', 'u', '0', '0', '1', 'd', 0, 6}}, // 0x1D
+                {{'\\', 'u', '0', '0', '1', 'e', 0, 6}}, // 0x1E
+                {{'\\', 'u', '0', '0', '1', 'f', 0, 6}}, // 0x1F
+                {{' ', 0, 0, 0, 0, 0, 0, 1}}, // 0x20
+                {{'!', 0, 0, 0, 0, 0, 0, 1}}, // 0x21
+                {{'\\', '"', 0, 0, 0, 0, 0, 2}}, // 0x22
+                {{'#', 0, 0, 0, 0, 0, 0, 1}}, // 0x23
+                {{'$', 0, 0, 0, 0, 0, 0, 1}}, // 0x24
+                {{'%', 0, 0, 0, 0, 0, 0, 1}}, // 0x25
+                {{'&', 0, 0, 0, 0, 0, 0, 1}}, // 0x26
+                {{'\'', 0, 0, 0, 0, 0, 0, 1}}, // 0x27
+                {{'(', 0, 0, 0, 0, 0, 0, 1}}, // 0x28
+                {{')', 0, 0, 0, 0, 0, 0, 1}}, // 0x29
+                {{'*', 0, 0, 0, 0, 0, 0, 1}}, // 0x2A
+                {{'+', 0, 0, 0, 0, 0, 0, 1}}, // 0x2B
+                {{',', 0, 0, 0, 0, 0, 0, 1}}, // 0x2C
+                {{'-', 0, 0, 0, 0, 0, 0, 1}}, // 0x2D
+                {{'.', 0, 0, 0, 0, 0, 0, 1}}, // 0x2E
+                {{'/', 0, 0, 0, 0, 0, 0, 1}}, // 0x2F
+                {{'0', 0, 0, 0, 0, 0, 0, 1}}, // 0x30
+                {{'1', 0, 0, 0, 0, 0, 0, 1}}, // 0x31
+                {{'2', 0, 0, 0, 0, 0, 0, 1}}, // 0x32
+                {{'3', 0, 0, 0, 0, 0, 0, 1}}, // 0x33
+                {{'4', 0, 0, 0, 0, 0, 0, 1}}, // 0x34
+                {{'5', 0, 0, 0, 0, 0, 0, 1}}, // 0x35
+                {{'6', 0, 0, 0, 0, 0, 0, 1}}, // 0x36
+                {{'7', 0, 0, 0, 0, 0, 0, 1}}, // 0x37
+                {{'8', 0, 0, 0, 0, 0, 0, 1}}, // 0x38
+                {{'9', 0, 0, 0, 0, 0, 0, 1}}, // 0x39
+                {{':', 0, 0, 0, 0, 0, 0, 1}}, // 0x3A
+                {{';', 0, 0, 0, 0, 0, 0, 1}}, // 0x3B
+                {{'<', 0, 0, 0, 0, 0, 0, 1}}, // 0x3C
+                {{'=', 0, 0, 0, 0, 0, 0, 1}}, // 0x3D
+                {{'>', 0, 0, 0, 0, 0, 0, 1}}, // 0x3E
+                {{'?', 0, 0, 0, 0, 0, 0, 1}}, // 0x3F
+                {{'@', 0, 0, 0, 0, 0, 0, 1}}, // 0x40
+                {{'A', 0, 0, 0, 0, 0, 0, 1}}, // 0x41
+                {{'B', 0, 0, 0, 0, 0, 0, 1}}, // 0x42
+                {{'C', 0, 0, 0, 0, 0, 0, 1}}, // 0x43
+                {{'D', 0, 0, 0, 0, 0, 0, 1}}, // 0x44
+                {{'E', 0, 0, 0, 0, 0, 0, 1}}, // 0x45
+                {{'F', 0, 0, 0, 0, 0, 0, 1}}, // 0x46
+                {{'G', 0, 0, 0, 0, 0, 0, 1}}, // 0x47
+                {{'H', 0, 0, 0, 0, 0, 0, 1}}, // 0x48
+                {{'I', 0, 0, 0, 0, 0, 0, 1}}, // 0x49
+                {{'J', 0, 0, 0, 0, 0, 0, 1}}, // 0x4A
+                {{'K', 0, 0, 0, 0, 0, 0, 1}}, // 0x4B
+                {{'L', 0, 0, 0, 0, 0, 0, 1}}, // 0x4C
+                {{'M', 0, 0, 0, 0, 0, 0, 1}}, // 0x4D
+                {{'N', 0, 0, 0, 0, 0, 0, 1}}, // 0x4E
+                {{'O', 0, 0, 0, 0, 0, 0, 1}}, // 0x4F
+                {{'P', 0, 0, 0, 0, 0, 0, 1}}, // 0x50
+                {{'Q', 0, 0, 0, 0, 0, 0, 1}}, // 0x51
+                {{'R', 0, 0, 0, 0, 0, 0, 1}}, // 0x52
+                {{'S', 0, 0, 0, 0, 0, 0, 1}}, // 0x53
+                {{'T', 0, 0, 0, 0, 0, 0, 1}}, // 0x54
+                {{'U', 0, 0, 0, 0, 0, 0, 1}}, // 0x55
+                {{'V', 0, 0, 0, 0, 0, 0, 1}}, // 0x56
+                {{'W', 0, 0, 0, 0, 0, 0, 1}}, // 0x57
+                {{'X', 0, 0, 0, 0, 0, 0, 1}}, // 0x58
+                {{'Y', 0, 0, 0, 0, 0, 0, 1}}, // 0x59
+                {{'Z', 0, 0, 0, 0, 0, 0, 1}}, // 0x5A
+                {{'[', 0, 0, 0, 0, 0, 0, 1}}, // 0x5B
+                {{'\\', '\\', 0, 0, 0, 0, 0, 2}}, // 0x5C
+                {{']', 0, 0, 0, 0, 0, 0, 1}}, // 0x5D
+                {{'^', 0, 0, 0, 0, 0, 0, 1}}, // 0x5E
+                {{'_', 0, 0, 0, 0, 0, 0, 1}}, // 0x5F
+            }
+        };
+        return escapes;
+    }
+
    /*!
     * @brief write a lowercase "\\uXXXX" escape sequence into @a string_buffer
     *
@@ -1366,8 +1461,9 @@ class serializer
    /*!
    @brief dump an integer

-    Dump a given integer, appending it to @ref write_buffer. Works internally with
-    @a number_buffer.
+    Dump a given integer, appending it to @ref write_buffer (directly: copying
+    the digits from another buffer right after writing them waits until the
+    stores are done).

    @param[in] x  integer number (signed or unsigned) to dump
    @tparam NumberType either @a number_integer_t or @a number_unsigned_t
@@ -1402,33 +1498,57 @@ class serializer
            return;
        }

-        // use a pointer to fill the buffer
-        auto buffer_ptr = number_buffer.begin(); // NOLINT(llvm-qualified-auto,readability-qualified-auto)
+        // use a pointer to fill the buffer (room for as much as number_buffer holds)
+        if (JSON_HEDLEY_UNLIKELY(write_buffer_pos + number_buffer.size() > write_buffer.size()))
+        {
+            flush();
+        }
+        auto* buffer_ptr = write_buffer.data() + write_buffer_pos;

-        number_unsigned_t abs_value;
+        number_magnitude_t abs_value;

-        unsigned int n_chars{};
+        // one byte for the minus sign
+        unsigned int n_chars = 0;

        if (is_negative_number(x))
        {
            *buffer_ptr = '-';
            abs_value = remove_sign(static_cast<number_integer_t>(x));
-
-            // account one more byte for the minus sign
-            n_chars = 1 + count_digits(abs_value);
+            n_chars = 1;
        }
        else
        {
-            abs_value = static_cast<number_unsigned_t>(x);
-            n_chars = count_digits(abs_value);
+            abs_value = static_cast<number_magnitude_t>(x);
        }

+        // up to 16 digits: eight at a time (as the digits of floats), written
+        // without leading zeros
+        if (abs_value < 10000000000000000u)
+        {
+            const std::uint64_t value = abs_value;
+            const std::uint64_t upper = value / 100000000u;
+            const std::uint64_t first = dtoa_impl::eight_digit_bytes(upper != 0 ? upper : value);
+            const auto leading = static_cast<unsigned>(count_leading_zeros(first) / 8); // (first is not 0)
+            char* const p = buffer_ptr + n_chars;
+            dtoa_impl::store_msb_first(p, (first << (8 * leading)) + 0x3030303030303030u);
+            n_chars += 8 - leading;
+            if (upper != 0)
+            {
+                dtoa_impl::store_msb_first(p + 8 - leading, dtoa_impl::eight_digit_bytes(value - (upper * 100000000u)) + 0x3030303030303030u);
+                n_chars += 8;
+            }
+            write_buffer_pos += n_chars;
+            return;
+        }
+
+        n_chars += count_digits(abs_value);
+
        // spare 1 byte for '\0'
        JSON_ASSERT(n_chars < number_buffer.size() - 1);

        // jump to the end to generate the string from backward,
        // so we later avoid reversing the result
-        buffer_ptr += static_cast<typename decltype(number_buffer)::difference_type>(n_chars);
+        buffer_ptr += n_chars;

        // Fast int2ascii implementation inspired by "Fastware" talk by Andrei Alexandrescu
        // See: https://www.youtube.com/watch?v=o4-CwDo2zpg
@@ -1451,14 +1571,13 @@ class serializer
            *(--buffer_ptr) = static_cast<char>('0' + abs_value);
        }

-        put_buffer(number_buffer, n_chars);
+        write_buffer_pos += n_chars;
    }

    /*!
    @brief dump a floating-point number

-    Dump a given floating-point number, appending it to @ref write_buffer. Works internally
-    with @a number_buffer.
+    Dump a given floating-point number, appending it to @ref write_buffer.

    @param[in] x  floating-point number to dump
    */
@@ -1485,10 +1604,15 @@ class serializer

    void dump_float(number_float_t x, std::true_type /*is_ieee_single_or_double*/)
    {
-        auto* begin = number_buffer.data();
+        // directly into the write buffer: copying the text from number_buffer
+        // right after to_chars() wrote it waits until its stores are done
+        if (JSON_HEDLEY_UNLIKELY(write_buffer_pos + number_buffer.size() > write_buffer.size()))
+        {
+            flush();
+        }
+        auto* begin = write_buffer.data() + write_buffer_pos;
        auto* end = ::nlohmann::detail::to_chars(begin, begin + number_buffer.size(), x);
-
-        put_buffer(number_buffer, static_cast<std::size_t>(end - begin));
+        write_buffer_pos += static_cast<std::size_t>(end - begin);
    }

    JSON_HEDLEY_NON_NULL(1)
@@ -1562,17 +1686,6 @@ class serializer
        }
    }

-    /*
-     * Overload to make the compiler happy while it is instantiating
-     * dump_integer for number_unsigned_t.
-     * Must never be called.
-     */
-    number_unsigned_t remove_sign(number_unsigned_t x)
-    {
-        JSON_ASSERT(false); // NOLINT(cert-dcl03-c,hicpp-static-assert,misc-static-assert) LCOV_EXCL_LINE
-        return x; // LCOV_EXCL_LINE
-    }
-
    /*
     * Helper function for dump_integer
     *
@@ -1582,10 +1695,10 @@ class serializer
     * absolute values of INT_MIN and INT_MAX are usually not the same. See
     * #1708 for details.
     */
-    number_unsigned_t remove_sign(number_integer_t x) noexcept
+    number_magnitude_t remove_sign(number_integer_t x) noexcept
    {
        JSON_ASSERT(x < 0);
-        return static_cast<number_unsigned_t>(-(x + 1)) + 1;
+        return static_cast<number_magnitude_t>(-(x + 1)) + 1;
    }

  private:
@@ -9,6 +9,7 @@
 #pragma once

 #include <cstddef> // size_t
+#include <utility> // move

 #include <nlohmann/detail/abi_macros.hpp>

@@ -24,39 +25,46 @@ namespace detail
 * Note the order of escaping "~" to "~0" and "/" to "~1" is important.
 *
 * The string is rebuilt in a single pass, appending whole runs between the
- * characters that need escaping. Scanning with find_first_of() keeps the
- * common case -- nothing to escape -- as fast as a single search, while
- * repeated replace() calls would move the tail of the string once per
- * escaped character.
+ * characters that need escaping, while repeated replace() calls would move
+ * the tail of the string once per escaped character. @ref escape_into appends
+ * the escaped string to an existing one.
 */
+template<typename StringType>
+inline void escape_into(StringType& out, const StringType& s)
+{
+    const std::size_t size = s.size();
+    std::size_t run = 0;
+    for (std::size_t i = 0; i < size; ++i)
+    {
+        const char c = s[i];
+        if (c == '~' || c == '/')
+        {
+            out.append(s.data() + run, i - run);
+            out.append(c == '~' ? "~0" : "~1", 2);
+            run = i + 1;
+        }
+    }
+    out.append(s.data() + run, size - run);
+}
+
 template<typename StringType>
 inline StringType escape(const StringType& s)
 {
-    auto next_special = [&s](std::size_t from)
+    // a single scan for both special characters; nothing to escape is the common case
+    const std::size_t size = s.size();
+    std::size_t i = 0;
+    while (i < size && s[i] != '~' && s[i] != '/')
    {
-        const auto tilde = s.find_first_of('~', from);
-        const auto slash = s.find_first_of('/', from);
-        return tilde < slash ? tilde : slash; // npos is the largest value
-    };
-
-    auto pos = next_special(0);
-    if (pos == StringType::npos)
+        ++i;
+    }
+    if (i == size)
    {
        return s;
    }

    StringType result;
-    result.reserve(s.size() + 2);
-
-    std::size_t run = 0;
-    while (pos != StringType::npos)
-    {
-        result.append(s.data() + run, pos - run);
-        result.append(s[pos] == '~' ? "~0" : "~1", 2);
-        run = pos + 1;
-        pos = next_special(run);
-    }
-    result.append(s.data() + run, s.size() - run);
+    result.reserve(size + 2);
+    escape_into(result, s);
    return result;
 }

@@ -101,7 +109,7 @@ inline void unescape(StringType& s)
        pos = s.find_first_of('~', run);
    }
    result.append(s.data() + run, s.size() - run);
-    s = result;
+    s = std::move(result);
 }

 }  // namespace detail
@@ -117,6 +117,34 @@ inline bool operator<(const value_t lhs, const value_t rhs) noexcept
 }
 #endif

+/*!
+@brief whether a value of the given type owns heap memory
+
+Strings, binaries, objects, and arrays are stored as pointers to allocated
+values; all other types are stored directly in the union. Every enumerator is
+listed so that -Wswitch-enum flags this function when a new type is added.
+*/
+inline bool owns_memory(const value_t t) noexcept
+{
+    switch (t)
+    {
+        case value_t::object:
+        case value_t::array:
+        case value_t::string:
+        case value_t::binary:
+            return true;
+
+        case value_t::null:
+        case value_t::boolean:
+        case value_t::number_integer:
+        case value_t::number_unsigned:
+        case value_t::number_float:
+        case value_t::discarded:
+        default:
+            return false;
+    }
+}
+

 /*!
@brief compare an integer with a floating point number without precision loss
@@ -90,6 +90,15 @@ private:
        return self.end();
    }

+    /// @brief throw the exception of a failed at() lookup
+    /// @note kept out of line so that at_impl() stays small enough to inline
+    /// @throw std::out_of_range always
+    JSON_HEDLEY_NO_RETURN JSON_HEDLEY_NEVER_INLINE
+    static void throw_key_not_found()
+    {
+        JSON_THROW(std::out_of_range("key not found"));
+    }
+
    /// @brief shared implementation of the const and non-const at() overloads
    /// @throw std::out_of_range if @a key is not found
    template<typename Self, typename KeyType>
@@ -98,7 +107,7 @@ private:
        const auto it = find_impl(self, key);
        if (it == self.end())
        {
-            JSON_THROW(std::out_of_range("key not found"));
+            throw_key_not_found();
        }
        return it->second;
    }
@@ -143,6 +152,17 @@ public:
        return {std::prev(this->end()), true};
    }

+    /// @brief add an element at the end without checking whether the key exists
+    /// @note The caller must guarantee that @a key is not yet contained, otherwise
+    ///       the container holds duplicate keys. This avoids the linear search of
+    ///       emplace() and operator[] when many keys known to be new are added.
+    template<class V, detail::enable_if_t<
+                 detail::is_constructible<T, V>::value, int> = 0>
+    void emplace_new_key(const key_type& key, V && t)
+    {
+        append(key, std::forward<V>(t));
+    }
+
    T& operator[](const key_type& key)
    {
        return emplace(key, T{}).first->second;