Scan strings vector-first on x86-64 and avoid a stall when nesting

- With SSE2, string runs are checked 16 bytes at a time from their first
  byte: one compare finds the end of most keys and short values, faster on
  x86-64 than a branch per byte. AArch64 keeps the byte steps (there, a
  NEON mask costs more and the branches predict well; vector-first was 20%
  slower on Apple M1).
- open() stores the parent's frame field by field. Built on the stack and
  copied, it was read back by loads wider than its stores, which waited for
  them (store forwarding fails): 18% of the time on citm_catalog.json.

Parse on x86-64 (GCC 13 / Clang 18, us): twitter 352 -> 306 / 316 -> 288,
citm_catalog 1000 -> 770 / 843 -> 692, canada 1941 -> 1687 / 1978 -> 1722.
Unchanged on Apple M1.

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

BUILD.bazel

@@ -26,6 +26,7 @@ cc_library(
         "include/nlohmann/detail/conversions/from_json.hpp",
         "include/nlohmann/detail/conversions/to_chars.hpp",
         "include/nlohmann/detail/conversions/to_json.hpp",
+        "include/nlohmann/detail/conversions/zmij.hpp",
         "include/nlohmann/detail/exceptions.hpp",
         "include/nlohmann/detail/hash.hpp",
         "include/nlohmann/detail/input/binary_reader.hpp",

README.md

@@ -1399,6 +1399,7 @@ THE SOFTWARE IS PROVIDED “AS IS”, WITHOUT WARRANTY OF ANY KIND, EXPRESS OR I
 
 - The class contains the UTF-8 Decoder from Bjoern Hoehrmann which is licensed under the [MIT License](https://opensource.org/licenses/MIT) (see above). Copyright &copy; 2008-2009 [Björn Hoehrmann](https://bjoern.hoehrmann.de/) <bjoern@hoehrmann.de>
 - The class contains a slightly modified version of the Grisu2 algorithm from Florian Loitsch which is licensed under the [MIT License](https://opensource.org/licenses/MIT) (see above). Copyright &copy; 2009 [Florian Loitsch](https://florian.loitsch.com/)
+- The class contains a port of the shortest double-to-decimal conversion of [Żmij](https://github.com/vitaut/zmij) by Victor Zverovich, which is licensed under the [MIT License](https://opensource.org/licenses/MIT) (see above). Copyright &copy; 2025 [Victor Zverovich](https://github.com/vitaut)
 - The class contains a copy of [Hedley](https://nemequ.github.io/hedley/) from Evan Nemerson which is licensed as [CC0-1.0](https://creativecommons.org/publicdomain/zero/1.0/).
 - The class contains parts of [Google Abseil](https://github.com/abseil/abseil-cpp) which is licensed under the [Apache 2.0 License](https://opensource.org/licenses/Apache-2.0).
 - The class contains an adapted version of the Eisel-Lemire algorithm, its table of powers of five, and its digit comparison for long numbers from [fast_float](https://github.com/fastfloat/fast_float) by Daniel Lemire and contributors, which is available under the [MIT License](https://opensource.org/licenses/MIT) (used here), the Apache 2.0 License, and the Boost Software License. Copyright &copy; 2021 The fast_float authors

docs/mkdocs/docs/api/basic_json/dump.md

@@ -62,6 +62,9 @@ Linear.
 
 ## Notes
 
+Floating-point numbers are written with the fewest digits that read back as the same value (for `#!cpp double`; see
+[number handling](../../features/types/number_handling.md#number-serialization)).
+
 Binary values are serialized as an object containing two keys:
 
 - "bytes": an array of bytes as integers
@@ -99,3 +102,5 @@ Binary values are serialized as an object containing two keys:
 - Error handlers added in version 3.4.0.
 - Serialization of binary values added in version 3.8.0.
 - Error handler `keep` added in version 3.13.0.
+- Doubles are written with the shortest digits (Żmij instead of Grisu2) since version 3.13.0; about 0.1% of doubles are
+  written differently, most of them with fewer digits.

docs/mkdocs/docs/api/basic_json_document/read.md

@@ -49,6 +49,11 @@ whether or not the new parse succeeds; take fresh views from [`root()`](root.md)
 `input` is borrowed or owned by the same rules as [`parse()`](parse.md#notes); a document can borrow on one call and
 own on the next, since ownership is decided freshly each time.
 
+Reusing a document matters most for large inputs: the operating system provides the memory of a fresh node index one
+page at a time, and every page costs a page fault the first time it is written. On x86-64 Linux (4 KiB pages), parsing
+a 55 MB document into a reused document took about 40 % less time than parsing it into a fresh one. Programs that parse
+many documents of similar size should therefore keep one document and call `read()`.
+
 ## Examples
 
 ??? example

docs/mkdocs/docs/api/macros/json_view_use_ssse3.md

@@ -5,13 +5,17 @@
 ```
 
 When defined on x86-64, the parser of [`basic_json_document`](../basic_json_document/index.md)
-(`<nlohmann/json_view.hpp>`) validates non-ASCII text in strings with SSSE3, 16 bytes at a time, using the "lookup4"
+(`<nlohmann/json_view.hpp>`) validates non-ASCII text in strings with SSSE3 without asking the CPU first.
-algorithm of [simdjson](https://github.com/simdjson/simdjson). Without it, non-ASCII text is validated one UTF-8
-sequence at a time on x86-64; on AArch64, the vector check uses NEON and is always on.
 
-SSSE3 is not part of the x86-64 baseline, so the code must be compiled for it: define the macro only together with a
+By default, the parser checks once at run time whether the CPU has SSSE3 (all x86-64 CPUs since about 2011 have it)
-compiler option that enables SSSE3 (e.g. `-mssse3`, or `-march=` with a CPU that has it), and only for programs that
+and then validates non-ASCII text 16 bytes at a time, using the "lookup4" algorithm of
-run on such CPUs. The same input is accepted or rejected either way; only the speed of non-ASCII text differs.
+[simdjson](https://github.com/simdjson/simdjson); on CPUs without SSSE3, it validates one UTF-8 sequence at a time.
+The vector check is compiled for SSSE3 with a function attribute (GCC 4.9 and later, Clang), so this needs no compiler
+option. With MSVC, the check uses `__cpuid`. On AArch64, the vector check uses NEON and is always on.
+
+Define the macro only together with a compiler option that enables SSSE3 (e.g. `-mssse3`, or `-march=` with a CPU that
+has it), and only for programs that run on such CPUs. It saves the check of the CPU, which costs little. The same
+input is accepted or rejected either way; only the speed of non-ASCII text differs.
 
 !!! warning "Define consistently"
 

docs/mkdocs/docs/features/types/number_handling.md

@@ -134,9 +134,10 @@ That is, `-0` is stored as a signed integer, but the serialization does not repr
 ### Number serialization
 
 - Integer numbers are serialized as is; that is, no scientific notation is used.
-- Floating-point numbers are serialized as specified by the `#!c %g` printf modifier with 
+- Floating-point numbers are serialized with the fewest digits that read back as the same value (the closest such
-  [`std::numeric_limits<double>::max_digits10`](https://en.cppreference.com/w/cpp/types/numeric_limits/max_digits10)
+  digits if there are several), in the layout of the `#!c %g` printf modifier: `#!c 1.5`, `#!c 100.0`, `#!c 1e+100`.
-  significant digits. The rationale is to use the shortest representation while still allowing round-tripping.
+  Doubles are converted with the algorithm of [Żmij](https://github.com/vitaut/zmij), floats with Grisu2, which
+  can write more digits than necessary.
 
 !!! hint "Notes regarding precision of floating-point numbers"
 

docs/mkdocs/docs/features/types/template_parameters.md

@@ -545,9 +545,10 @@ therefore silently changes parse results rather than raising an error. See
   specifiers, for which the library likewise provides only `#!cpp double` and `#!cpp long double` overloads
   (`#!cpp float` is promoted to `#!cpp double`).
 
-If `#!cpp std::numeric_limits<NumberFloatType>` describes an IEEE 754 binary32 or binary64 number, `dump` uses the
+If `#!cpp std::numeric_limits<NumberFloatType>` describes an IEEE 754 binary64 number, `dump` uses the algorithm of
-Grisu2 algorithm, which produces the shortest representation that round-trips. Otherwise the `snprintf` fallback with
+Żmij, which produces the shortest representation that round-trips. For IEEE 754 binary32 numbers, it uses Grisu2,
-`max_digits10` digits is used.
+which produces a short representation that round-trips. Otherwise the `snprintf` fallback with `max_digits10` digits is
+used.
 
 ### Required for the binary formats
 
@@ -559,7 +560,7 @@ binary32 or binary64 field and have no encoding for `#!cpp long double`.
 
 | Type                     | Support                                                                                                               |
 |--------------------------|-----------------------------------------------------------------------------------------------------------------------|
-| `#!cpp double` (default) | full; short round-trip output through Grisu2                                                                          |
+| `#!cpp double` (default) | full; shortest round-trip output through Żmij                                                                         |
 | `#!cpp float`            | full; short round-trip output through Grisu2                                                                          |
 | `#!cpp long double`      | `dump` and `parse` only; the binary format writers do not compile, as they only handle IEEE 754 binary32 and binary64 |
 | any other type           | not usable                                                                                                            |

docs/mkdocs/docs/home/license.md

@@ -18,6 +18,8 @@ The class contains the UTF-8 Decoder from Bjoern Hoehrmann which is licensed und
 
 The class contains a slightly modified version of the Grisu2 algorithm from Florian Loitsch which is licensed under the [MIT License](https://opensource.org/licenses/MIT) (see above). Copyright © 2009 [Florian Loitsch](https://florian.loitsch.com/)
 
+The class contains a port of the shortest double-to-decimal conversion of [Żmij](https://github.com/vitaut/zmij) by Victor Zverovich, which is licensed under the [MIT License](https://opensource.org/licenses/MIT) (see above). Copyright © 2025 [Victor Zverovich](https://github.com/vitaut)
+
 The class contains a copy of [Hedley](https://nemequ.github.io/hedley/) from Evan Nemerson which is licensed as [CC0-1.0](https://creativecommons.org/publicdomain/zero/1.0/).
 
 The class contains an adapted version of the Eisel-Lemire algorithm, its table of powers of five, and its digit comparison for long numbers from [fast_float](https://github.com/fastfloat/fast_float) by Daniel Lemire and contributors, which is available under the [MIT License](https://opensource.org/licenses/MIT) (used here), the Apache 2.0 License, and the Boost Software License. Copyright © 2021 The fast_float authors

include/nlohmann/detail/bit_ops.hpp

@@ -86,8 +86,9 @@ inline uint128_parts full_multiplication(std::uint64_t a, std::uint64_t b) noexc
 }
 
 /// eight bytes as a little-endian word (compilers fold this into one load on
-/// little-endian targets)
+/// little-endian targets; always inlined, as GCC otherwise calls it in the
-inline std::uint64_t read_eight_bytes(const unsigned char* b) noexcept
+/// number loops)
+JSON_HEDLEY_ALWAYS_INLINE std::uint64_t read_eight_bytes(const unsigned char* b) noexcept
 {
     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)
@@ -96,7 +97,7 @@ inline std::uint64_t read_eight_bytes(const unsigned char* b) noexcept
 }
 
 /// eight bytes as a little-endian word
-inline std::uint64_t read_eight_bytes(const char* p) noexcept
+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)
 }

include/nlohmann/detail/conversions/to_chars.hpp

@@ -11,11 +11,32 @@
 
 #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
@@ -918,6 +939,88 @@ void grisu2(char* buf, int& len, int& decimal_exponent, FloatType value)
     grisu2(buf, len, decimal_exponent, w.minus, w.w, w.plus);
 }
 
+/*!
+@brief the shortest digits of a positive finite float (other than double): Grisu2
+*/
+template<typename FloatType>
+JSON_HEDLEY_NON_NULL(1)
+void shortest_digits(char* buf, int& len, int& decimal_exponent, FloatType value)
+{
+    grisu2(buf, len, decimal_exponent, value);
+}
+
+/*!
+@brief the shortest digits of a positive finite double: the conversion of
+Zmij (see zmij.hpp), which always finds the shortest digits that read back as
+the same value (Grisu2 does not for about one double in a thousand), and the
+closest of them if there are several
+
+v = buf * 10^decimal_exponent, as for grisu2()
+*/
+JSON_HEDLEY_NON_NULL(1)
+inline void shortest_digits(char* buf, int& len, int& decimal_exponent, double value)
+{
+    static_assert(std::numeric_limits<double>::is_iec559 && std::numeric_limits<double>::digits == 53,
+                  "internal error: the conversion of Zmij needs IEEE 754 binary64 doubles");
+    JSON_ASSERT(std::isfinite(value));
+    JSON_ASSERT(value > 0);
+
+    std::uint64_t bits = 0;
+    std::memcpy(&bits, &value, sizeof(bits));
+    zmij::decimal d = zmij::to_decimal(bits);
+    // without trailing zeros (up to 16): 8, 4, 2, 1 at a time
+    while (d.significand % 100000000 == 0)
+    {
+        d.significand /= 100000000;
+        d.exponent += 8;
+    }
+    if (d.significand % 10000 == 0)
+    {
+        d.significand /= 10000;
+        d.exponent += 4;
+    }
+    if (d.significand % 100 == 0)
+    {
+        d.significand /= 100;
+        d.exponent += 2;
+    }
+    if (d.significand % 10 == 0)
+    {
+        d.significand /= 10;
+        d.exponent += 1;
+    }
+    // at most 17 digits, written from the back two at a time
+    static constexpr const char* pairs =
+        "00010203040506070809101112131415161718192021222324252627282930313233343536373839"
+        "40414243444546474849505152535455565758596061626364656667686970717273747576777879"
+        "8081828384858687888990919293949596979899";
+    std::array<char, 20> digits{};
+    std::size_t n = digits.size();
+    while (d.significand >= 100)
+    {
+        const std::uint64_t two_digits = d.significand % 100; // a variable: GCC calls a cast of the remainder useless where std::uint64_t is std::size_t
+        const auto i = static_cast<std::size_t>(two_digits) * 2;
+        d.significand /= 100;
+        n -= 2;
+        digits[n] = pairs[i];
+        digits[n + 1] = pairs[i + 1];
+    }
+    if (d.significand >= 10)
+    {
+        const auto i = static_cast<std::size_t>(d.significand) * 2;
+        n -= 2;
+        digits[n] = pairs[i];
+        digits[n + 1] = pairs[i + 1];
+    }
+    else
+    {
+        digits[--n] = static_cast<char>('0' + d.significand);
+    }
+    len = static_cast<int>(digits.size() - n);
+    std::memcpy(buf, digits.data() + n, static_cast<std::size_t>(len));
+    decimal_exponent = d.exponent;
+}
+
 /*!
 @brief appends a decimal representation of e to buf
 @return a pointer to the element following the exponent.
@@ -1047,6 +1150,374 @@ 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 = static_cast<unsigned>(e < 0 ? -e : e);
+    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));
+    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 = static_cast<unsigned>(e < 0 ? -e : e);
+    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);
+}
+
+/// the powers of ten up to 10^16
+inline const std::array<std::uint64_t, 17>& powers_of_ten_16() noexcept
+{
+    static const std::array<std::uint64_t, 17> powers =
+    {
+        {
+            1u, 10u, 100u, 1000u, 10000u, 100000u, 1000000u, 10000000u, 100000000u, 1000000000u, 10000000000u,
+            100000000000u, 1000000000000u, 10000000000000u, 100000000000000u, 1000000000000000u, 10000000000000000u
+        }
+    };
+    return powers;
+}
+
+/*!
+@brief digits * 10^exp, as write_decimal() writes it, for the digits of a
+double that need no conversion (count digits, at most 15, the first not 0;
+trailing zeros allowed): extended to 16 digits and written by write_shortest()
+
+@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_short_decimal(char* first, std::uint64_t digits, int count, int exp) noexcept
+{
+    JSON_ASSERT(digits >= powers_of_ten_16()[static_cast<std::size_t>(count - 1)] && count <= 15);
+    const int scale = 16 - count;
+    return write_shortest(first, zmij::shortest_decimal{digits * powers_of_ten_16()[static_cast<std::size_t>(scale)], exp - scale - 1, 0, false});
+}
+
+/// as write_short_decimal(), counting the digits (not 0, less than 10^15)
+JSON_HEDLEY_NON_NULL(1)
+JSON_HEDLEY_RETURNS_NON_NULL
+inline char* write_short_decimal(char* first, std::uint64_t digits, int exp) noexcept
+{
+    JSON_ASSERT(digits != 0 && digits < 1000000000000000u);
+    // floor(log10(2^bits)) + 1 digits, or one less
+    const int log2_bound = ((64 - count_leading_zeros(digits)) * 1233) >> 12;
+    const int count = log2_bound + (digits >= powers_of_ten_16()[static_cast<std::size_t>(log2_bound)] ? 1 : 0);
+    return write_short_decimal(first, digits, count, exp);
+}
+
+/// a positive finite float (other than double): Grisu2 and format_buffer()
+template<typename FloatType>
+JSON_HEDLEY_NON_NULL(1, 2)
+JSON_HEDLEY_RETURNS_NON_NULL
+char* write_positive(char* first, const char* last, FloatType value)
+{
+    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;
+    shortest_digits(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 double: 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)
+JSON_HEDLEY_NON_NULL(1, 2)
+JSON_HEDLEY_RETURNS_NON_NULL
+inline char* write_positive(char* first, const char* last, double value)
+{
+    static_assert(std::numeric_limits<double>::is_iec559 && std::numeric_limits<double>::digits == 53,
+                  "internal error: the conversion of Zmij needs IEEE 754 binary64 doubles");
+    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(static_cast<std::size_t>(last - first) >= len);
+    std::memcpy(first, buf.data(), len);
+    return first + len;
+}
+
 }  // namespace dtoa_impl
 
 /*!
@@ -1064,7 +1535,6 @@ 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 +1560,7 @@ 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);
+    return dtoa_impl::write_positive(first, last, value);
-
-    // 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);
 }
 
 }  // namespace detail

include/nlohmann/detail/conversions/zmij.hpp

@@ -0,0 +1,238 @@
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  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
+{
+
+/// significand * 10^exponent
+struct decimal
+{
+    std::uint64_t significand;
+    int exponent;
+};
+
+/// 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);
+    const std::uint64_t fractional = (p.high << static_cast<unsigned>(64 - extra_shift)) | (p.low >> static_cast<unsigned>(extra_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};
+}
+
+/// The shortest decimal in the rounding interval of a positive finite double
+/// given by its bits, as one number. The significand can end in zeros.
+inline decimal to_decimal(std::uint64_t bits) noexcept
+{
+    const shortest_decimal d = to_shortest(bits);
+    if (d.has_digit)
+    {
+        return decimal{(d.integral * 10) + d.digit, d.exponent};
+    }
+    return decimal{d.integral, d.exponent + 1};
+}
+
+}  // namespace zmij
+}  // namespace detail
+NLOHMANN_JSON_NAMESPACE_END

include/nlohmann/detail/input/number_parse.hpp

@@ -249,8 +249,9 @@ template<typename FloatType>
 using native_float_t = typename std::conditional<std::numeric_limits<FloatType>::digits == 24, float, double>::type;
 
 /// the value of the eight ASCII digits in @a v (see read_eight_bytes()), three
-/// multiplications instead of eight (after simdjson and fast_float)
+/// multiplications instead of eight (after simdjson and fast_float); always
-inline std::uint32_t parse_eight_digits(std::uint64_t v) noexcept
+/// inlined, as GCC otherwise calls it in the number loops
+JSON_HEDLEY_ALWAYS_INLINE std::uint32_t parse_eight_digits(std::uint64_t v) noexcept
 {
     v = ((v & 0x0F0F0F0F0F0F0F0Fu) * 2561u) >> 8u;
     v = ((v & 0x00FF00FF00FF00FFu) * 6553601u) >> 16u;

include/nlohmann/detail/macro_unscope.hpp

@@ -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

include/nlohmann/detail/output/serializer.hpp

@@ -1366,8 +1366,9 @@ class serializer
     /*!
     @brief dump an integer
 
-    Dump a given integer, appending it to @ref write_buffer. Works internally with
+    Dump a given integer, appending it to @ref write_buffer (directly: copying
-    @a number_buffer.
+    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 +1403,57 @@ class serializer
             return;
         }
 
-        // use a pointer to fill the buffer
+        // use a pointer to fill the buffer (room for as much as number_buffer holds)
-        auto buffer_ptr = number_buffer.begin(); // NOLINT(llvm-qualified-auto,readability-qualified-auto)
+        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;
 
-        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));
-
+            n_chars = 1;
-            // account one more byte for the minus sign
-            n_chars = 1 + count_digits(abs_value);
         }
         else
         {
             abs_value = static_cast<number_unsigned_t>(x);
-            n_chars = count_digits(abs_value);
         }
 
+        // 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 +1476,13 @@ 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
+    Dump a given floating-point number, appending it to @ref write_buffer.
-    with @a number_buffer.
 
     @param[in] x  floating-point number to dump
     */
@@ -1485,10 +1509,15 @@ 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);
-
+        write_buffer_pos += static_cast<std::size_t>(end - begin);
-        put_buffer(number_buffer, static_cast<std::size_t>(end - begin));
     }
 
     JSON_HEDLEY_NON_NULL(1)

include/nlohmann/detail/view/builder.hpp

@@ -828,14 +828,19 @@ indent_done:
             const auto idx = static_cast<std::uint32_t>(emit(k, 0, 0, static_cast<std::size_t>(p - b), 0) - base);
             if (depth != 0)
             {
-                const frame f = {cur_idx, cur_count, cur_is_object};
                 if (NLOHMANN_VIEW_LIKELY(depth <= 64))
                 {
-                    cold.shallow[depth - 1] = f;
+                    // field by field: a frame put together on the stack and
+                    // copied would be read back wider than it was written,
+                    // and that load waits until the stores are done
+                    frame& f = cold.shallow[depth - 1];
+                    f.idx = cur_idx;
+                    f.count = cur_count;
+                    f.is_object = cur_is_object;
                 }
                 else
                 {
-                    cold.deep.push_back(f);
+                    cold.deep.push_back(frame{cur_idx, cur_count, cur_is_object});
                 }
             }
             ++depth;
@@ -851,20 +856,22 @@ indent_done:
             n.next = static_cast<std::uint32_t>(out - base) - cur_idx;
             if (--depth != 0)
             {
-                frame f{};
                 if (NLOHMANN_VIEW_LIKELY(depth <= 64))
                 {
-                    f = cold.shallow[depth - 1];
+                    const frame& f = cold.shallow[depth - 1];
-                }
-                else
-                {
-                    f = cold.deep.back();
-                    cold.deep.pop_back();
-                }
                     cur_idx = f.idx;
                     cur_count = f.count;
                     cur_is_object = f.is_object;
                 }
+                else
+                {
+                    const frame f = cold.deep.back();
+                    cold.deep.pop_back();
+                    cur_idx = f.idx;
+                    cur_count = f.count;
+                    cur_is_object = f.is_object;
+                }
+            }
         }
 
         NLOHMANN_VIEW_ALWAYS_INLINE bool literal(const char* text, std::size_t n, value_t k, std::uint8_t flags)

include/nlohmann/detail/view/macro_unscope.hpp

@@ -22,5 +22,7 @@
 #undef NLOHMANN_VIEW_NEON
 #undef NLOHMANN_VIEW_SSE2
 #undef NLOHMANN_VIEW_SSSE3
+#undef NLOHMANN_VIEW_SSSE3_DISPATCH
+#undef NLOHMANN_VIEW_SSSE3_TARGET
 #undef NLOHMANN_VIEW_VECTOR
 #undef NLOHMANN_VIEW_VECTOR_UTF8

include/nlohmann/detail/view/scan.hpp

@@ -67,13 +67,19 @@ NLOHMANN_VIEW_ALWAYS_INLINE std::uint16_t load16(const unsigned char* p) noexcep
 /// in predicted branches: 16 for keys, whose lengths repeat from record to
 /// record, and 8 for string values (Value) where a vector loop follows, as
 /// their lengths vary more. Longer runs continue 16 bytes at a time with NEON
-/// or SSE2, else eight bytes at a time.
+/// or SSE2, else eight bytes at a time. With SSE2, the run is checked 16 bytes
+/// at a time from its first byte instead: on x86-64, one compare that finds
+/// the end of most keys and short values is faster than a branch per byte (on
+/// AArch64, where a NEON mask costs more and branches predict well, slower).
 template<bool Value = false>
 NLOHMANN_VIEW_ALWAYS_INLINE const unsigned char* scan_string_run(const unsigned char* p, const unsigned char* e) noexcept
 {
     const std::uint8_t* plain = string_plain();
     for (;;)
     {
+#if NLOHMANN_VIEW_SSE2
+        p = vector_plain_run(p, e);
+#else
         if (e - p >= 16)
         {
 #define NLOHMANN_VIEW_STEP(i) if (NLOHMANN_VIEW_LIKELY(plain[p[i]] != 0)) {} else { p += (i); goto stop; }
@@ -107,6 +113,7 @@ NLOHMANN_VIEW_ALWAYS_INLINE const unsigned char* scan_string_run(const unsigned
 #endif
             continue;
         }
+#endif
         while (p != e && plain[*p] != 0)
         {
             ++p;
@@ -115,15 +122,23 @@ NLOHMANN_VIEW_ALWAYS_INLINE const unsigned char* scan_string_run(const unsigned
         {
             return p;
         }
+#if !NLOHMANN_VIEW_SSE2
 stop:
+#endif
         if (*p < 0x80)
         {
             return p; // quote, backslash, or control character
         }
 #if NLOHMANN_VIEW_VECTOR_UTF8
+#if NLOHMANN_VIEW_SSSE3_DISPATCH
+        if (NLOHMANN_VIEW_LIKELY(cpu_has_ssse3()))
+#endif
+        {
             // non-ASCII: the vector check, out of line
             return scan_string_vector(p, e, plain);
-#else
+        }
+#endif
+#if !NLOHMANN_VIEW_VECTOR_UTF8 || NLOHMANN_VIEW_SSSE3_DISPATCH
         // non-ASCII: a run of well-formed sequences (the library's check, so
         // that exactly what json::parse accepts is accepted)
         do

include/nlohmann/detail/view/serializer.hpp

@@ -496,10 +496,15 @@ class view_serializer
                         room(n->len);
                         copy(src + n->off, n->len);
                     }
+                    else if (std::is_same<number_float_t, double>::value)
+                    {
+                        room(64);
+                        w = write_double_at(w, *n);
+                    }
                     else
                     {
                         m_out.set_cursor(w);
-                        write_float(float_value<number_float_t>(m_doc, *n));
+                        write_float_node(*n);
                         w = m_out.cursor();
                         lim = m_out.limit();
                     }
@@ -666,7 +671,7 @@ class view_serializer
                 }
                 else
                 {
-                    write_float(float_value<number_float_t>(m_doc, n));
+                    write_float_node(n);
                 }
                 break;
             case value_t::object:    // LCOV_EXCL_LINE (containers are written by dump())
@@ -678,6 +683,83 @@ class view_serializer
         }
     }
 
+    /// a float node as dump() writes it
+    void write_float_node(const node& n)
+    {
+        write_float_node(n, std::is_same<number_float_t, double> {});
+    }
+
+    void write_float_node(const node& n, std::false_type /*other*/)
+    {
+        write_float(float_value<number_float_t>(m_doc, n));
+    }
+
+    void write_float_node(const node& n, std::true_type /*double*/)
+    {
+        m_out.reserve(64);
+        m_out.set_cursor(write_double_at(m_out.cursor(), n));
+    }
+
+    /*!
+    @brief (doubles) the float at n as dump() writes it, at w (64 bytes of room)
+
+    A token of at most 15 significant digits is written from its digits,
+    without a conversion: two decimals of at most 15 digits are farther
+    apart than the rounding interval of a (normal) double (the argument
+    behind DBL_DIG), so the token's digits are the shortest ones of its
+    double, which the library's conversion writes (Zmij). Other tokens are
+    converted from the digits already read.
+    */
+    char* write_double_at(char* w, const node& n)
+    {
+        const unsigned int_digits = n.extra & 0xFFu;
+        const unsigned frac_digits = n.extra >> 8u;
+        if ((n.flags & node_flags::storage) != node_flags::edited && int_digits + frac_digits <= 19)
+        {
+            const auto* const first = reinterpret_cast<const unsigned char*>(m_doc.src + n.off); // NOLINT(cppcoreguidelines-pro-type-reinterpret-cast)
+            const float_significand d = layout_decimal(first, first + n.len, int_digits, frac_digits, reinterpret_cast<const unsigned char*>(m_doc.src + m_doc.size)); // NOLINT(cppcoreguidelines-pro-type-reinterpret-cast)
+            // (the exponent keeps the value far from subnormals and overflow)
+            if (d.w != 0 && d.w < 1000000000000000u && d.exponent >= -290 && d.exponent <= 290)
+            {
+                *w = '-';
+                w += d.negative ? 1 : 0;
+                // (without leading zeros, all digits of the token count)
+                const unsigned char lead = first[d.negative ? 1 : 0];
+                return lead != '0' ? ::nlohmann::detail::dtoa_impl::write_short_decimal(w, d.w, static_cast<int>(int_digits + frac_digits), static_cast<int>(d.exponent))
+                       : ::nlohmann::detail::dtoa_impl::write_short_decimal(w, d.w, static_cast<int>(d.exponent));
+            }
+            return write_double_value_at(w, decimal_to_float<double>(d)); // (without reading the token again)
+        }
+        return write_double_value_at(w, static_cast<double>(float_value<number_float_t>(m_doc, n)));
+    }
+
+    /// n bytes of text at w
+    static char* write_text_at(char* w, const char* text, std::size_t n) noexcept
+    {
+        std::memcpy(w, text, n);
+        return w + n;
+    }
+
+    /// a double as dump() writes it, at w (64 bytes of room)
+    static char* write_double_value_at(char* w, double x)
+    {
+        // (from the bits: without the checks of to_chars())
+        std::uint64_t bits = 0;
+        std::memcpy(&bits, &x, sizeof(bits));
+        if (NLOHMANN_VIEW_UNLIKELY((bits & 0x7FF0000000000000u) == 0x7FF0000000000000u))
+        {
+            return write_text_at(w, "null", 4);
+        }
+        *w = '-';
+        w += bits >> 63u;
+        bits &= ~(std::uint64_t{1} << 63u);
+        if (bits == 0)
+        {
+            return write_text_at(w, "0.0", 3);
+        }
+        return ::nlohmann::detail::dtoa_impl::write_shortest(w, ::nlohmann::detail::zmij::to_shortest(bits));
+    }
+
     /// as serializer::dump_float()
     void write_float(number_float_t x)
     {

include/nlohmann/detail/view/simd.hpp

@@ -10,6 +10,7 @@
 #pragma once
 
 #include <array> // array
+#include <atomic> // atomic
 #include <cstddef> // size_t
 #include <cstdint> // uint8_t, uint64_t
 
@@ -18,11 +19,13 @@
 
 // Vector code for long runs of string bytes. NEON (AArch64) and SSE2 (x86-64)
 // belong to the baseline instruction sets and are used by default. The vector
-// UTF-8 check needs NEON, or SSSE3 if JSON_VIEW_USE_SSSE3 is defined: SSSE3 is
+// UTF-8 check needs NEON or SSSE3. SSSE3 is not part of x86-64, and the code
-// not part of x86-64, so it must not depend on the flags of a translation unit
+// must not depend on the flags of a translation unit (two translation units
-// (two translation units with different flags would have different
+// with different flags would have different definitions of the same inline
-// definitions of the same inline functions). JSON_VIEW_NO_SIMD selects the
+// functions): the check is compiled for SSSE3 with a function attribute and
-// portable code.
+// used where the CPU has SSSE3 (all x86-64 CPUs since about 2011), else the
+// portable check. JSON_VIEW_USE_SSSE3 skips the CPU check (for code compiled
+// for SSSE3 anyway); JSON_VIEW_NO_SIMD selects the portable code.
 #if !defined(JSON_VIEW_NO_SIMD) && defined(__aarch64__) && (defined(__GNUC__) || defined(__clang__)) && NLOHMANN_VIEW_LITTLE_ENDIAN
     #include <arm_neon.h>
     #define NLOHMANN_VIEW_NEON 1
@@ -41,8 +44,24 @@
 #else
     #define NLOHMANN_VIEW_SSSE3 0 // NOLINT(cppcoreguidelines-macro-to-enum,modernize-macro-to-enum)
 #endif
+#if NLOHMANN_VIEW_SSE2 && !NLOHMANN_VIEW_SSSE3 && ((defined(__clang__) && __clang_major__ >= 4) || (defined(__GNUC__) && !defined(__clang__) && (__GNUC__ > 4 || (__GNUC__ == 4 && __GNUC_MINOR__ >= 9))))
+    // (GCC before 4.9 has no SSSE3 intrinsics without -mssse3)
+    #include <cpuid.h>
+    #include <tmmintrin.h>
+    #define NLOHMANN_VIEW_SSSE3_DISPATCH 1 // NOLINT(cppcoreguidelines-macro-to-enum,modernize-macro-to-enum)
+    #define NLOHMANN_VIEW_SSSE3_TARGET __attribute__((target("ssse3")))
+#elif NLOHMANN_VIEW_SSE2 && !NLOHMANN_VIEW_SSSE3 && defined(_MSC_VER)
+    // (MSVC compiles intrinsics of any instruction set)
+    #include <intrin.h>
+    #include <tmmintrin.h>
+    #define NLOHMANN_VIEW_SSSE3_DISPATCH 1 // NOLINT(cppcoreguidelines-macro-to-enum,modernize-macro-to-enum)
+    #define NLOHMANN_VIEW_SSSE3_TARGET
+#else
+    #define NLOHMANN_VIEW_SSSE3_DISPATCH 0 // NOLINT(cppcoreguidelines-macro-to-enum,modernize-macro-to-enum)
+    #define NLOHMANN_VIEW_SSSE3_TARGET
+#endif
 #define NLOHMANN_VIEW_VECTOR (NLOHMANN_VIEW_NEON || NLOHMANN_VIEW_SSE2)
-#define NLOHMANN_VIEW_VECTOR_UTF8 (NLOHMANN_VIEW_NEON || NLOHMANN_VIEW_SSSE3)
+#define NLOHMANN_VIEW_VECTOR_UTF8 (NLOHMANN_VIEW_NEON || NLOHMANN_VIEW_SSSE3 || NLOHMANN_VIEW_SSSE3_DISPATCH)
 
 NLOHMANN_JSON_NAMESPACE_BEGIN
 namespace detail
@@ -90,6 +109,40 @@ NLOHMANN_VIEW_ALWAYS_INLINE const unsigned char* vector_plain_run(const unsigned
 }
 #endif
 
+#if NLOHMANN_VIEW_SSSE3_DISPATCH
+/// whether the CPU has SSSE3 (CPUID leaf 1, ECX bit 9)
+inline bool cpu_ssse3() noexcept
+{
+#if defined(_MSC_VER) && !defined(__clang__)
+    std::array<int, 4> regs {{}};
+    __cpuid(regs.data(), 1);
+    return (static_cast<unsigned>(regs[2]) & (1u << 9u)) != 0;
+#else
+    unsigned eax = 0;
+    unsigned ebx = 0;
+    unsigned ecx = 0;
+    unsigned edx = 0;
+    return __get_cpuid(1, &eax, &ebx, &ecx, &edx) != 0 && (ecx & (1u << 9u)) != 0;
+#endif
+}
+
+/// whether the CPU has SSSE3, asked once: the answer is kept in an atomic
+/// that is initialized at compile time, so that neither a guard of a local
+/// static nor a global constructor is needed (threads that ask at the same
+/// time all store the same answer)
+NLOHMANN_VIEW_ALWAYS_INLINE bool cpu_has_ssse3() noexcept
+{
+    static std::atomic<int> known{0}; // 0: not asked yet, 1: no, 2: yes
+    int state = known.load(std::memory_order_relaxed);
+    if (NLOHMANN_VIEW_UNLIKELY(state == 0))
+    {
+        state = cpu_ssse3() ? 2 : 1;
+        known.store(state, std::memory_order_relaxed);
+    }
+    return state == 2;
+}
+#endif
+
 #if NLOHMANN_VIEW_VECTOR_UTF8
 /// Tables of the UTF-8 check of J. Keiser and D. Lemire, "Validating UTF-8 In
 /// Less Than One Instruction Per Byte" (2021), as in simdjson ("lookup4"): each
@@ -192,8 +245,9 @@ compares, and the UTF-8 check covers the bytes up to it. Returns where the
 string scan stops, like scan_string_run: before ill-formed UTF-8 and for the
 last bytes of the input, the bytes are checked one sequence at a time. Out of
 line, so that no constants of the check occupy registers in the parse loop.
+On x86-64, it is compiled for SSSE3 (see cpu_has_ssse3()).
 */
-NLOHMANN_VIEW_NOINLINE inline const unsigned char* scan_string_vector(const unsigned char* p, const unsigned char* e, const std::uint8_t* plain) noexcept
+NLOHMANN_VIEW_SSSE3_TARGET NLOHMANN_VIEW_NOINLINE inline const unsigned char* scan_string_vector(const unsigned char* p, const unsigned char* e, const std::uint8_t* plain) noexcept
 {
     using lookup = utf8_lookup4<>;
     const unsigned char* block = p;

single_include/nlohmann/json.hpp

@@ -8860,8 +8860,9 @@ inline uint128_parts full_multiplication(std::uint64_t a, std::uint64_t b) noexc
 }
 
 /// eight bytes as a little-endian word (compilers fold this into one load on
-/// little-endian targets)
+/// little-endian targets; always inlined, as GCC otherwise calls it in the
-inline std::uint64_t read_eight_bytes(const unsigned char* b) noexcept
+/// number loops)
+JSON_HEDLEY_ALWAYS_INLINE std::uint64_t read_eight_bytes(const unsigned char* b) noexcept
 {
     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)
@@ -8870,7 +8871,7 @@ inline std::uint64_t read_eight_bytes(const unsigned char* b) noexcept
 }
 
 /// eight bytes as a little-endian word
-inline std::uint64_t read_eight_bytes(const char* p) noexcept
+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)
 }
@@ -9479,8 +9480,9 @@ template<typename FloatType>
 using native_float_t = typename std::conditional<std::numeric_limits<FloatType>::digits == 24, float, double>::type;
 
 /// the value of the eight ASCII digits in @a v (see read_eight_bytes()), three
-/// multiplications instead of eight (after simdjson and fast_float)
+/// multiplications instead of eight (after simdjson and fast_float); always
-inline std::uint32_t parse_eight_digits(std::uint64_t v) noexcept
+/// inlined, as GCC otherwise calls it in the number loops
+JSON_HEDLEY_ALWAYS_INLINE std::uint32_t parse_eight_digits(std::uint64_t v) noexcept
 {
     v = ((v & 0x0F0F0F0F0F0F0F0Fu) * 2561u) >> 8u;
     v = ((v & 0x00FF00FF00FF00FFu) * 6553601u) >> 16u;
@@ -24416,11 +24418,275 @@ NLOHMANN_JSON_NAMESPACE_END
 
 #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>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  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
+
+
+
+#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
+{
+
+/// significand * 10^exponent
+struct decimal
+{
+    std::uint64_t significand;
+    int exponent;
+};
+
+/// 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);
+    const std::uint64_t fractional = (p.high << static_cast<unsigned>(64 - extra_shift)) | (p.low >> static_cast<unsigned>(extra_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};
+}
+
+/// The shortest decimal in the rounding interval of a positive finite double
+/// given by its bits, as one number. The significand can end in zeros.
+inline decimal to_decimal(std::uint64_t bits) noexcept
+{
+    const shortest_decimal d = to_shortest(bits);
+    if (d.has_digit)
+    {
+        return decimal{(d.integral * 10) + d.digit, d.exponent};
+    }
+    return decimal{d.integral, d.exponent + 1};
+}
+
+}  // namespace zmij
+}  // namespace detail
+NLOHMANN_JSON_NAMESPACE_END
+
 // #include <nlohmann/detail/macro_scope.hpp>
 
 
@@ -25324,6 +25590,88 @@ void grisu2(char* buf, int& len, int& decimal_exponent, FloatType value)
     grisu2(buf, len, decimal_exponent, w.minus, w.w, w.plus);
 }
 
+/*!
+@brief the shortest digits of a positive finite float (other than double): Grisu2
+*/
+template<typename FloatType>
+JSON_HEDLEY_NON_NULL(1)
+void shortest_digits(char* buf, int& len, int& decimal_exponent, FloatType value)
+{
+    grisu2(buf, len, decimal_exponent, value);
+}
+
+/*!
+@brief the shortest digits of a positive finite double: the conversion of
+Zmij (see zmij.hpp), which always finds the shortest digits that read back as
+the same value (Grisu2 does not for about one double in a thousand), and the
+closest of them if there are several
+
+v = buf * 10^decimal_exponent, as for grisu2()
+*/
+JSON_HEDLEY_NON_NULL(1)
+inline void shortest_digits(char* buf, int& len, int& decimal_exponent, double value)
+{
+    static_assert(std::numeric_limits<double>::is_iec559 && std::numeric_limits<double>::digits == 53,
+                  "internal error: the conversion of Zmij needs IEEE 754 binary64 doubles");
+    JSON_ASSERT(std::isfinite(value));
+    JSON_ASSERT(value > 0);
+
+    std::uint64_t bits = 0;
+    std::memcpy(&bits, &value, sizeof(bits));
+    zmij::decimal d = zmij::to_decimal(bits);
+    // without trailing zeros (up to 16): 8, 4, 2, 1 at a time
+    while (d.significand % 100000000 == 0)
+    {
+        d.significand /= 100000000;
+        d.exponent += 8;
+    }
+    if (d.significand % 10000 == 0)
+    {
+        d.significand /= 10000;
+        d.exponent += 4;
+    }
+    if (d.significand % 100 == 0)
+    {
+        d.significand /= 100;
+        d.exponent += 2;
+    }
+    if (d.significand % 10 == 0)
+    {
+        d.significand /= 10;
+        d.exponent += 1;
+    }
+    // at most 17 digits, written from the back two at a time
+    static constexpr const char* pairs =
+        "00010203040506070809101112131415161718192021222324252627282930313233343536373839"
+        "40414243444546474849505152535455565758596061626364656667686970717273747576777879"
+        "8081828384858687888990919293949596979899";
+    std::array<char, 20> digits{};
+    std::size_t n = digits.size();
+    while (d.significand >= 100)
+    {
+        const std::uint64_t two_digits = d.significand % 100; // a variable: GCC calls a cast of the remainder useless where std::uint64_t is std::size_t
+        const auto i = static_cast<std::size_t>(two_digits) * 2;
+        d.significand /= 100;
+        n -= 2;
+        digits[n] = pairs[i];
+        digits[n + 1] = pairs[i + 1];
+    }
+    if (d.significand >= 10)
+    {
+        const auto i = static_cast<std::size_t>(d.significand) * 2;
+        n -= 2;
+        digits[n] = pairs[i];
+        digits[n + 1] = pairs[i + 1];
+    }
+    else
+    {
+        digits[--n] = static_cast<char>('0' + d.significand);
+    }
+    len = static_cast<int>(digits.size() - n);
+    std::memcpy(buf, digits.data() + n, static_cast<std::size_t>(len));
+    decimal_exponent = d.exponent;
+}
+
 /*!
 @brief appends a decimal representation of e to buf
 @return a pointer to the element following the exponent.
@@ -25453,6 +25801,374 @@ 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 = static_cast<unsigned>(e < 0 ? -e : e);
+    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));
+    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 = static_cast<unsigned>(e < 0 ? -e : e);
+    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);
+}
+
+/// the powers of ten up to 10^16
+inline const std::array<std::uint64_t, 17>& powers_of_ten_16() noexcept
+{
+    static const std::array<std::uint64_t, 17> powers =
+    {
+        {
+            1u, 10u, 100u, 1000u, 10000u, 100000u, 1000000u, 10000000u, 100000000u, 1000000000u, 10000000000u,
+            100000000000u, 1000000000000u, 10000000000000u, 100000000000000u, 1000000000000000u, 10000000000000000u
+        }
+    };
+    return powers;
+}
+
+/*!
+@brief digits * 10^exp, as write_decimal() writes it, for the digits of a
+double that need no conversion (count digits, at most 15, the first not 0;
+trailing zeros allowed): extended to 16 digits and written by write_shortest()
+
+@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_short_decimal(char* first, std::uint64_t digits, int count, int exp) noexcept
+{
+    JSON_ASSERT(digits >= powers_of_ten_16()[static_cast<std::size_t>(count - 1)] && count <= 15);
+    const int scale = 16 - count;
+    return write_shortest(first, zmij::shortest_decimal{digits * powers_of_ten_16()[static_cast<std::size_t>(scale)], exp - scale - 1, 0, false});
+}
+
+/// as write_short_decimal(), counting the digits (not 0, less than 10^15)
+JSON_HEDLEY_NON_NULL(1)
+JSON_HEDLEY_RETURNS_NON_NULL
+inline char* write_short_decimal(char* first, std::uint64_t digits, int exp) noexcept
+{
+    JSON_ASSERT(digits != 0 && digits < 1000000000000000u);
+    // floor(log10(2^bits)) + 1 digits, or one less
+    const int log2_bound = ((64 - count_leading_zeros(digits)) * 1233) >> 12;
+    const int count = log2_bound + (digits >= powers_of_ten_16()[static_cast<std::size_t>(log2_bound)] ? 1 : 0);
+    return write_short_decimal(first, digits, count, exp);
+}
+
+/// a positive finite float (other than double): Grisu2 and format_buffer()
+template<typename FloatType>
+JSON_HEDLEY_NON_NULL(1, 2)
+JSON_HEDLEY_RETURNS_NON_NULL
+char* write_positive(char* first, const char* last, FloatType value)
+{
+    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;
+    shortest_digits(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 double: 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)
+JSON_HEDLEY_NON_NULL(1, 2)
+JSON_HEDLEY_RETURNS_NON_NULL
+inline char* write_positive(char* first, const char* last, double value)
+{
+    static_assert(std::numeric_limits<double>::is_iec559 && std::numeric_limits<double>::digits == 53,
+                  "internal error: the conversion of Zmij needs IEEE 754 binary64 doubles");
+    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(static_cast<std::size_t>(last - first) >= len);
+    std::memcpy(first, buf.data(), len);
+    return first + len;
+}
+
 }  // namespace dtoa_impl
 
 /*!
@@ -25470,7 +26186,6 @@ 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.
@@ -25496,28 +26211,7 @@ 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);
+    return dtoa_impl::write_positive(first, last, value);
-
-    // 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);
 }
 
 }  // namespace detail
@@ -26876,8 +27570,9 @@ class serializer
     /*!
     @brief dump an integer
 
-    Dump a given integer, appending it to @ref write_buffer. Works internally with
+    Dump a given integer, appending it to @ref write_buffer (directly: copying
-    @a number_buffer.
+    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
@@ -26912,33 +27607,57 @@ class serializer
             return;
         }
 
-        // use a pointer to fill the buffer
+        // use a pointer to fill the buffer (room for as much as number_buffer holds)
-        auto buffer_ptr = number_buffer.begin(); // NOLINT(llvm-qualified-auto,readability-qualified-auto)
+        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;
 
-        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));
-
+            n_chars = 1;
-            // account one more byte for the minus sign
-            n_chars = 1 + count_digits(abs_value);
         }
         else
         {
             abs_value = static_cast<number_unsigned_t>(x);
-            n_chars = count_digits(abs_value);
         }
 
+        // 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
@@ -26961,14 +27680,13 @@ 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
+    Dump a given floating-point number, appending it to @ref write_buffer.
-    with @a number_buffer.
 
     @param[in] x  floating-point number to dump
     */
@@ -26995,10 +27713,15 @@ 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);
-
+        write_buffer_pos += static_cast<std::size_t>(end - begin);
-        put_buffer(number_buffer, static_cast<std::size_t>(end - begin));
     }
 
     JSON_HEDLEY_NON_NULL(1)
@@ -34681,6 +35404,8 @@ struct formatter<nlohmann::NLOHMANN_BASIC_JSON_TPL, char> // NOLINT(cert-dcl58-c
 #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

single_include/nlohmann/json_view.hpp

@@ -542,6 +542,7 @@ NLOHMANN_JSON_NAMESPACE_END
 
 
 #include <array> // array
+#include <atomic> // atomic
 #include <cstddef> // size_t
 #include <cstdint> // uint8_t, uint64_t
 
@@ -551,11 +552,13 @@ NLOHMANN_JSON_NAMESPACE_END
 
 // Vector code for long runs of string bytes. NEON (AArch64) and SSE2 (x86-64)
 // belong to the baseline instruction sets and are used by default. The vector
-// UTF-8 check needs NEON, or SSSE3 if JSON_VIEW_USE_SSSE3 is defined: SSSE3 is
+// UTF-8 check needs NEON or SSSE3. SSSE3 is not part of x86-64, and the code
-// not part of x86-64, so it must not depend on the flags of a translation unit
+// must not depend on the flags of a translation unit (two translation units
-// (two translation units with different flags would have different
+// with different flags would have different definitions of the same inline
-// definitions of the same inline functions). JSON_VIEW_NO_SIMD selects the
+// functions): the check is compiled for SSSE3 with a function attribute and
-// portable code.
+// used where the CPU has SSSE3 (all x86-64 CPUs since about 2011), else the
+// portable check. JSON_VIEW_USE_SSSE3 skips the CPU check (for code compiled
+// for SSSE3 anyway); JSON_VIEW_NO_SIMD selects the portable code.
 #if !defined(JSON_VIEW_NO_SIMD) && defined(__aarch64__) && (defined(__GNUC__) || defined(__clang__)) && NLOHMANN_VIEW_LITTLE_ENDIAN
     #include <arm_neon.h>
     #define NLOHMANN_VIEW_NEON 1
@@ -574,8 +577,24 @@ NLOHMANN_JSON_NAMESPACE_END
 #else
     #define NLOHMANN_VIEW_SSSE3 0 // NOLINT(cppcoreguidelines-macro-to-enum,modernize-macro-to-enum)
 #endif
+#if NLOHMANN_VIEW_SSE2 && !NLOHMANN_VIEW_SSSE3 && ((defined(__clang__) && __clang_major__ >= 4) || (defined(__GNUC__) && !defined(__clang__) && (__GNUC__ > 4 || (__GNUC__ == 4 && __GNUC_MINOR__ >= 9))))
+    // (GCC before 4.9 has no SSSE3 intrinsics without -mssse3)
+    #include <cpuid.h>
+    #include <tmmintrin.h>
+    #define NLOHMANN_VIEW_SSSE3_DISPATCH 1 // NOLINT(cppcoreguidelines-macro-to-enum,modernize-macro-to-enum)
+    #define NLOHMANN_VIEW_SSSE3_TARGET __attribute__((target("ssse3")))
+#elif NLOHMANN_VIEW_SSE2 && !NLOHMANN_VIEW_SSSE3 && defined(_MSC_VER)
+    // (MSVC compiles intrinsics of any instruction set)
+    #include <intrin.h>
+    #include <tmmintrin.h>
+    #define NLOHMANN_VIEW_SSSE3_DISPATCH 1 // NOLINT(cppcoreguidelines-macro-to-enum,modernize-macro-to-enum)
+    #define NLOHMANN_VIEW_SSSE3_TARGET
+#else
+    #define NLOHMANN_VIEW_SSSE3_DISPATCH 0 // NOLINT(cppcoreguidelines-macro-to-enum,modernize-macro-to-enum)
+    #define NLOHMANN_VIEW_SSSE3_TARGET
+#endif
 #define NLOHMANN_VIEW_VECTOR (NLOHMANN_VIEW_NEON || NLOHMANN_VIEW_SSE2)
-#define NLOHMANN_VIEW_VECTOR_UTF8 (NLOHMANN_VIEW_NEON || NLOHMANN_VIEW_SSSE3)
+#define NLOHMANN_VIEW_VECTOR_UTF8 (NLOHMANN_VIEW_NEON || NLOHMANN_VIEW_SSSE3 || NLOHMANN_VIEW_SSSE3_DISPATCH)
 
 NLOHMANN_JSON_NAMESPACE_BEGIN
 namespace detail
@@ -623,6 +642,40 @@ NLOHMANN_VIEW_ALWAYS_INLINE const unsigned char* vector_plain_run(const unsigned
 }
 #endif
 
+#if NLOHMANN_VIEW_SSSE3_DISPATCH
+/// whether the CPU has SSSE3 (CPUID leaf 1, ECX bit 9)
+inline bool cpu_ssse3() noexcept
+{
+#if defined(_MSC_VER) && !defined(__clang__)
+    std::array<int, 4> regs {{}};
+    __cpuid(regs.data(), 1);
+    return (static_cast<unsigned>(regs[2]) & (1u << 9u)) != 0;
+#else
+    unsigned eax = 0;
+    unsigned ebx = 0;
+    unsigned ecx = 0;
+    unsigned edx = 0;
+    return __get_cpuid(1, &eax, &ebx, &ecx, &edx) != 0 && (ecx & (1u << 9u)) != 0;
+#endif
+}
+
+/// whether the CPU has SSSE3, asked once: the answer is kept in an atomic
+/// that is initialized at compile time, so that neither a guard of a local
+/// static nor a global constructor is needed (threads that ask at the same
+/// time all store the same answer)
+NLOHMANN_VIEW_ALWAYS_INLINE bool cpu_has_ssse3() noexcept
+{
+    static std::atomic<int> known{0}; // 0: not asked yet, 1: no, 2: yes
+    int state = known.load(std::memory_order_relaxed);
+    if (NLOHMANN_VIEW_UNLIKELY(state == 0))
+    {
+        state = cpu_ssse3() ? 2 : 1;
+        known.store(state, std::memory_order_relaxed);
+    }
+    return state == 2;
+}
+#endif
+
 #if NLOHMANN_VIEW_VECTOR_UTF8
 /// Tables of the UTF-8 check of J. Keiser and D. Lemire, "Validating UTF-8 In
 /// Less Than One Instruction Per Byte" (2021), as in simdjson ("lookup4"): each
@@ -725,8 +778,9 @@ compares, and the UTF-8 check covers the bytes up to it. Returns where the
 string scan stops, like scan_string_run: before ill-formed UTF-8 and for the
 last bytes of the input, the bytes are checked one sequence at a time. Out of
 line, so that no constants of the check occupy registers in the parse loop.
+On x86-64, it is compiled for SSSE3 (see cpu_has_ssse3()).
 */
-NLOHMANN_VIEW_NOINLINE inline const unsigned char* scan_string_vector(const unsigned char* p, const unsigned char* e, const std::uint8_t* plain) noexcept
+NLOHMANN_VIEW_SSSE3_TARGET NLOHMANN_VIEW_NOINLINE inline const unsigned char* scan_string_vector(const unsigned char* p, const unsigned char* e, const std::uint8_t* plain) noexcept
 {
     using lookup = utf8_lookup4<>;
     const unsigned char* block = p;
@@ -863,13 +917,19 @@ NLOHMANN_VIEW_ALWAYS_INLINE std::uint16_t load16(const unsigned char* p) noexcep
 /// in predicted branches: 16 for keys, whose lengths repeat from record to
 /// record, and 8 for string values (Value) where a vector loop follows, as
 /// their lengths vary more. Longer runs continue 16 bytes at a time with NEON
-/// or SSE2, else eight bytes at a time.
+/// or SSE2, else eight bytes at a time. With SSE2, the run is checked 16 bytes
+/// at a time from its first byte instead: on x86-64, one compare that finds
+/// the end of most keys and short values is faster than a branch per byte (on
+/// AArch64, where a NEON mask costs more and branches predict well, slower).
 template<bool Value = false>
 NLOHMANN_VIEW_ALWAYS_INLINE const unsigned char* scan_string_run(const unsigned char* p, const unsigned char* e) noexcept
 {
     const std::uint8_t* plain = string_plain();
     for (;;)
     {
+#if NLOHMANN_VIEW_SSE2
+        p = vector_plain_run(p, e);
+#else
         if (e - p >= 16)
         {
 #define NLOHMANN_VIEW_STEP(i) if (NLOHMANN_VIEW_LIKELY(plain[p[i]] != 0)) {} else { p += (i); goto stop; }
@@ -903,6 +963,7 @@ NLOHMANN_VIEW_ALWAYS_INLINE const unsigned char* scan_string_run(const unsigned
 #endif
             continue;
         }
+#endif
         while (p != e && plain[*p] != 0)
         {
             ++p;
@@ -911,15 +972,23 @@ NLOHMANN_VIEW_ALWAYS_INLINE const unsigned char* scan_string_run(const unsigned
         {
             return p;
         }
+#if !NLOHMANN_VIEW_SSE2
 stop:
+#endif
         if (*p < 0x80)
         {
             return p; // quote, backslash, or control character
         }
 #if NLOHMANN_VIEW_VECTOR_UTF8
+#if NLOHMANN_VIEW_SSSE3_DISPATCH
+        if (NLOHMANN_VIEW_LIKELY(cpu_has_ssse3()))
+#endif
+        {
             // non-ASCII: the vector check, out of line
             return scan_string_vector(p, e, plain);
-#else
+        }
+#endif
+#if !NLOHMANN_VIEW_VECTOR_UTF8 || NLOHMANN_VIEW_SSSE3_DISPATCH
         // non-ASCII: a run of well-formed sequences (the library's check, so
         // that exactly what json::parse accepts is accepted)
         do
@@ -1813,14 +1882,19 @@ indent_done:
             const auto idx = static_cast<std::uint32_t>(emit(k, 0, 0, static_cast<std::size_t>(p - b), 0) - base);
             if (depth != 0)
             {
-                const frame f = {cur_idx, cur_count, cur_is_object};
                 if (NLOHMANN_VIEW_LIKELY(depth <= 64))
                 {
-                    cold.shallow[depth - 1] = f;
+                    // field by field: a frame put together on the stack and
+                    // copied would be read back wider than it was written,
+                    // and that load waits until the stores are done
+                    frame& f = cold.shallow[depth - 1];
+                    f.idx = cur_idx;
+                    f.count = cur_count;
+                    f.is_object = cur_is_object;
                 }
                 else
                 {
-                    cold.deep.push_back(f);
+                    cold.deep.push_back(frame{cur_idx, cur_count, cur_is_object});
                 }
             }
             ++depth;
@@ -1836,20 +1910,22 @@ indent_done:
             n.next = static_cast<std::uint32_t>(out - base) - cur_idx;
             if (--depth != 0)
             {
-                frame f{};
                 if (NLOHMANN_VIEW_LIKELY(depth <= 64))
                 {
-                    f = cold.shallow[depth - 1];
+                    const frame& f = cold.shallow[depth - 1];
-                }
-                else
-                {
-                    f = cold.deep.back();
-                    cold.deep.pop_back();
-                }
                     cur_idx = f.idx;
                     cur_count = f.count;
                     cur_is_object = f.is_object;
                 }
+                else
+                {
+                    const frame f = cold.deep.back();
+                    cold.deep.pop_back();
+                    cur_idx = f.idx;
+                    cur_count = f.count;
+                    cur_is_object = f.is_object;
+                }
+            }
         }
 
         NLOHMANN_VIEW_ALWAYS_INLINE bool literal(const char* text, std::size_t n, value_t k, std::uint8_t flags)
@@ -5798,10 +5874,15 @@ class view_serializer
                         room(n->len);
                         copy(src + n->off, n->len);
                     }
+                    else if (std::is_same<number_float_t, double>::value)
+                    {
+                        room(64);
+                        w = write_double_at(w, *n);
+                    }
                     else
                     {
                         m_out.set_cursor(w);
-                        write_float(float_value<number_float_t>(m_doc, *n));
+                        write_float_node(*n);
                         w = m_out.cursor();
                         lim = m_out.limit();
                     }
@@ -5968,7 +6049,7 @@ class view_serializer
                 }
                 else
                 {
-                    write_float(float_value<number_float_t>(m_doc, n));
+                    write_float_node(n);
                 }
                 break;
             case value_t::object:    // LCOV_EXCL_LINE (containers are written by dump())
@@ -5980,6 +6061,83 @@ class view_serializer
         }
     }
 
+    /// a float node as dump() writes it
+    void write_float_node(const node& n)
+    {
+        write_float_node(n, std::is_same<number_float_t, double> {});
+    }
+
+    void write_float_node(const node& n, std::false_type /*other*/)
+    {
+        write_float(float_value<number_float_t>(m_doc, n));
+    }
+
+    void write_float_node(const node& n, std::true_type /*double*/)
+    {
+        m_out.reserve(64);
+        m_out.set_cursor(write_double_at(m_out.cursor(), n));
+    }
+
+    /*!
+    @brief (doubles) the float at n as dump() writes it, at w (64 bytes of room)
+
+    A token of at most 15 significant digits is written from its digits,
+    without a conversion: two decimals of at most 15 digits are farther
+    apart than the rounding interval of a (normal) double (the argument
+    behind DBL_DIG), so the token's digits are the shortest ones of its
+    double, which the library's conversion writes (Zmij). Other tokens are
+    converted from the digits already read.
+    */
+    char* write_double_at(char* w, const node& n)
+    {
+        const unsigned int_digits = n.extra & 0xFFu;
+        const unsigned frac_digits = n.extra >> 8u;
+        if ((n.flags & node_flags::storage) != node_flags::edited && int_digits + frac_digits <= 19)
+        {
+            const auto* const first = reinterpret_cast<const unsigned char*>(m_doc.src + n.off); // NOLINT(cppcoreguidelines-pro-type-reinterpret-cast)
+            const float_significand d = layout_decimal(first, first + n.len, int_digits, frac_digits, reinterpret_cast<const unsigned char*>(m_doc.src + m_doc.size)); // NOLINT(cppcoreguidelines-pro-type-reinterpret-cast)
+            // (the exponent keeps the value far from subnormals and overflow)
+            if (d.w != 0 && d.w < 1000000000000000u && d.exponent >= -290 && d.exponent <= 290)
+            {
+                *w = '-';
+                w += d.negative ? 1 : 0;
+                // (without leading zeros, all digits of the token count)
+                const unsigned char lead = first[d.negative ? 1 : 0];
+                return lead != '0' ? ::nlohmann::detail::dtoa_impl::write_short_decimal(w, d.w, static_cast<int>(int_digits + frac_digits), static_cast<int>(d.exponent))
+                       : ::nlohmann::detail::dtoa_impl::write_short_decimal(w, d.w, static_cast<int>(d.exponent));
+            }
+            return write_double_value_at(w, decimal_to_float<double>(d)); // (without reading the token again)
+        }
+        return write_double_value_at(w, static_cast<double>(float_value<number_float_t>(m_doc, n)));
+    }
+
+    /// n bytes of text at w
+    static char* write_text_at(char* w, const char* text, std::size_t n) noexcept
+    {
+        std::memcpy(w, text, n);
+        return w + n;
+    }
+
+    /// a double as dump() writes it, at w (64 bytes of room)
+    static char* write_double_value_at(char* w, double x)
+    {
+        // (from the bits: without the checks of to_chars())
+        std::uint64_t bits = 0;
+        std::memcpy(&bits, &x, sizeof(bits));
+        if (NLOHMANN_VIEW_UNLIKELY((bits & 0x7FF0000000000000u) == 0x7FF0000000000000u))
+        {
+            return write_text_at(w, "null", 4);
+        }
+        *w = '-';
+        w += bits >> 63u;
+        bits &= ~(std::uint64_t{1} << 63u);
+        if (bits == 0)
+        {
+            return write_text_at(w, "0.0", 3);
+        }
+        return ::nlohmann::detail::dtoa_impl::write_shortest(w, ::nlohmann::detail::zmij::to_shortest(bits));
+    }
+
     /// as serializer::dump_float()
     void write_float(number_float_t x)
     {
@@ -7744,6 +7902,8 @@ class tuple_element<N, ::nlohmann::detail::view::view_item<View>> // NOLINT(cert
 #undef NLOHMANN_VIEW_NEON
 #undef NLOHMANN_VIEW_SSE2
 #undef NLOHMANN_VIEW_SSSE3
+#undef NLOHMANN_VIEW_SSSE3_DISPATCH
+#undef NLOHMANN_VIEW_SSSE3_TARGET
 #undef NLOHMANN_VIEW_VECTOR
 #undef NLOHMANN_VIEW_VECTOR_UTF8
 

tests/benchmarks/json_view/README.md

@@ -60,7 +60,10 @@ outputs are checked to describe the same value.
 
 Before anything is timed, all engines must accept each document and agree on the traversal: the number of values, the
 bytes of all strings and keys, and the sum of all numbers. All engines run interleaved in every round, and the best
-round is reported, as time and as a factor of the `json_view` time (below 1 means faster than `json_view`).
+round is reported, as time and as a factor of the `json_view` time (below 1 means faster than `json_view`). Each timed
+call follows an untimed call of the same engine: otherwise the engine after `json::parse` pays for the allocator
+cleaning up the tens of thousands of nodes `json::parse` just freed (with glibc, this made `json_view` look 1.7 times
+slower on citm_catalog traverse).
 
 The engines do not all offer the same features, which the numbers should be read with:
 
@@ -68,11 +71,18 @@ The engines do not all offer the same features, which the numbers should be read
 |---|---|---|---|---|
 | `json_view` | immutable index into the text | yes | no | a fresh document per parse; "reused" parses into the same document |
 | yyjson | immutable (`yyjson_read`) | yes | via a mutable copy | |
-| simdjson DOM | immutable, parser reused | yes | no | |
+| simdjson DOM | immutable, parser reused | yes | no | "fresh" uses a new parser per parse |
 | simdjson On-Demand | none: forward-only, lazy | no | no | only traverse and select |
 | Boost.JSON | owning, mutable DOM | yes | yes | monotonic resource |
 | `json::parse` | owning, mutable DOM | yes | yes | |
 
+Reusing memory matters as much as the parser. simdjson DOM reuses its parser, so it writes into memory it already
+touched; a fresh `json_view` document or yyjson document gets new memory for every parse. On Linux, glibc returns large
+blocks to the system when they are freed, so every fresh parse of a large document pays a page fault per 4 KiB page:
+on x86-64 Linux, a fresh `json_view` parse of jeopardy took about twice as long as a reused one. On macOS on Apple
+silicon, with 16 KiB pages, the difference is much smaller. Compare "json_view (reused)" with "simdjson DOM", and the
+fresh `json_view` with "simdjson DOM (fresh)" and yyjson.
+
 ## Published results
 
 Results are only published with the file `compare.py` wrote, which names the machine and the versions; see

tests/benchmarks/json_view/bench_corpus.cpp

@@ -29,6 +29,7 @@
 #include <chrono>
 #include <cmath>
 #include <cstdio>
+#include <cstdlib>
 #include <cstring>
 #include <fstream>
 #include <functional>
@@ -195,6 +196,11 @@ static void walk(const boost::json::value& v, stats& st)
 static std::string slurp(const std::string& p)
 {
     std::ifstream f(p, std::ios::binary);
+    if (!f)
+    {
+        std::fprintf(stderr, "cannot open %s\n", p.c_str());
+        std::exit(1);
+    }
     std::stringstream ss;
     ss << f.rdbuf();
     return ss.str();
@@ -222,6 +228,7 @@ int main(int argc, char** argv)
     }
     std::FILE* csv = std::fopen("bench_corpus.csv", "w");
     std::fprintf(csv, "file,bytes,workload,engine,ns\n");
+    json_document reused;
     simdjson::dom::parser sj;
     for (const auto& path : files)
     {
@@ -272,8 +279,10 @@ int main(int argc, char** argv)
             {
                 "parse", {
                     {"json_view", [&] { auto x = json_document::parse(s); g_sink = static_cast<double>(x.node_count()); }},
+                    {"json_view (reused)", [&] { reused.read(s); g_sink = static_cast<double>(reused.node_count()); }},
                     {"yyjson", [&] { yyjson_doc* x = yyjson_read(s.data(), s.size(), 0); g_sink = static_cast<double>(yyjson_doc_get_val_count(x)); yyjson_doc_free(x); }},
                     {"simdjson DOM", [&] { auto e = sj.parse(ps).value_unsafe(); g_sink = e.is_object(); }},
+                    {"simdjson DOM (fresh)", [&] { simdjson::dom::parser p; auto e = p.parse(ps).value_unsafe(); g_sink = e.is_object(); }},
 #if JSON_VIEW_BENCH_BOOST
                     {"Boost.JSON", [&] { boost::json::monotonic_resource mr; auto v = boost::json::parse(s, &mr); g_sink = v.is_object(); }},
 #endif
@@ -306,6 +315,9 @@ int main(int argc, char** argv)
             {
                 for (std::size_t k = 0; k < wl.second.size(); ++k)
                 {
+                    // an untimed call first: whatever the previous engine left to the allocator
+                    // (e.g. thousands of freed json nodes) is cleaned up here, not in the timing
+                    wl.second[k].fn();
                     const auto t0 = std::chrono::steady_clock::now();
                     wl.second[k].fn();
                     best[k] = std::min(best[k], std::chrono::duration<double, std::nano>(std::chrono::steady_clock::now() - t0).count());

tests/benchmarks/json_view/bench_edit.cpp

@@ -479,6 +479,11 @@ static std::string edit_boost(const std::string& name, const std::string& s, boo
 static std::string slurp(const std::string& p)
 {
     std::ifstream f(p, std::ios::binary);
+    if (!f)
+    {
+        std::fprintf(stderr, "cannot open %s\n", p.c_str());
+        std::exit(1);
+    }
     std::stringstream ss;
     ss << f.rdbuf();
     return ss.str();
@@ -550,6 +555,9 @@ int main(int argc, char** argv)
             {
                 for (std::size_t k = 0; k < engines.size(); ++k)
                 {
+                    // an untimed call first: whatever the previous engine left to the allocator
+                    // (e.g. thousands of freed json nodes) is cleaned up here, not in the timing
+                    g_sink = engines[k].second(dc.name, dc.text, update).size();
                     const auto t0 = std::chrono::steady_clock::now();
                     for (int b = 0; b < dc.batch; ++b)
                     {

tests/benchmarks/json_view/bench_view.cpp

@@ -10,7 +10,8 @@
 //
 //   json_view      nlohmann/json_view.hpp (fresh document per parse / reused)
 //   yyjson         yyjson_read(): immutable document, random access
-//   simdjson DOM   dom::parser (reused, as recommended): immutable, random access
+//   simdjson DOM   dom::parser (reused, as recommended; "fresh": a new parser
+//                  per parse): immutable, random access
 // references (different feature sets):
 //   simdjson OD    On-Demand: forward-only, lazy
 //   Boost.JSON     owning, mutable DOM (monotonic resource)
@@ -19,7 +20,8 @@
 // Workloads: parse (build + free), traverse (visit everything, convert every
 // number, touch every string and key), select (a few fields per document),
 // dump (compact serialization of the parsed document).
-// All engines run interleaved in every round; the best round is reported.
+// All engines run interleaved in every round, each timed call after an untimed
+// one of the same engine; the best round is reported.
 #include <nlohmann/json_view.hpp>
 
 #if JSON_VIEW_BENCH_BOOST
@@ -33,6 +35,7 @@
 #include <chrono>
 #include <cmath>
 #include <cstdio>
+#include <cstdlib>
 #include <fstream>
 #include <functional>
 #include <map>
@@ -581,6 +584,11 @@ static double pick_od(const std::string& name, simdjson::ondemand::document& d)
 static std::string slurp(const std::string& p)
 {
     std::ifstream f(p, std::ios::binary);
+    if (!f)
+    {
+        std::fprintf(stderr, "cannot open %s\n", p.c_str());
+        std::exit(1);
+    }
     std::stringstream ss;
     ss << f.rdbuf();
     return ss.str();
@@ -657,6 +665,7 @@ int main(int argc, char** argv)
                 {"json_view (reused)", [&] { reused.read(s); g_sink = static_cast<double>(reused.node_count()); }},
                 {"yyjson", [&] { yyjson_doc* d = yyjson_read(s.data(), s.size(), 0); g_sink = static_cast<double>(yyjson_doc_get_val_count(d)); yyjson_doc_free(d); }},
                 {"simdjson DOM", [&] { auto e = sj.parse(ps).value_unsafe(); g_sink = e.is_object(); }},
+                {"simdjson DOM (fresh)", [&] { simdjson::dom::parser p; auto e = p.parse(ps).value_unsafe(); g_sink = e.is_object(); }},
 #if JSON_VIEW_BENCH_BOOST
                 {"Boost.JSON", [&] { boost::json::monotonic_resource mr; auto v = boost::json::parse(s, &mr); g_sink = v.is_object(); }},
 #endif
@@ -664,6 +673,7 @@ int main(int argc, char** argv)
             }});
         workloads.push_back({"traverse", {
                 {"json_view", [&] { auto d = json_document::parse(s); stats st; walk(d.root(), st); g_sink = st.num; }},
+                {"json_view (reused)", [&] { reused.read(s); stats st; walk(reused.root(), st); g_sink = st.num; }},
                 {"yyjson", [&] { yyjson_doc* d = yyjson_read(s.data(), s.size(), 0); stats st; walk(yyjson_doc_get_root(d), st); g_sink = st.num; yyjson_doc_free(d); }},
                 {"simdjson DOM", [&] { stats st; walk(sj.parse(ps).value_unsafe(), st); g_sink = st.num; }},
                 {"simdjson OD", [&] { auto d = od.iterate(ps).value_unsafe(); stats st; walk_od(d.get_value().value_unsafe(), st); g_sink = st.num; }},
@@ -674,6 +684,7 @@ int main(int argc, char** argv)
             }});
         workloads.push_back({"select", {
                 {"json_view", [&] { auto d = json_document::parse(s); g_sink = pick(name, d.root()); }},
+                {"json_view (reused)", [&] { reused.read(s); g_sink = pick(name, reused.root()); }},
                 {"yyjson", [&] { yyjson_doc* d = yyjson_read(s.data(), s.size(), 0); g_sink = pick(name, yyjson_doc_get_root(d)); yyjson_doc_free(d); }},
                 {"simdjson DOM", [&] { g_sink = pick(name, sj.parse(ps).value_unsafe()); }},
                 {"simdjson OD", [&] { auto d = od.iterate(ps).value_unsafe(); g_sink = pick_od(name, d); }},
@@ -710,6 +721,9 @@ int main(int argc, char** argv)
             {
                 for (std::size_t k = 0; k < wl.second.size(); ++k)
                 {
+                    // an untimed call first: whatever the previous engine left to the allocator
+                    // (e.g. thousands of freed json nodes) is cleaned up here, not in the timing
+                    wl.second[k].fn();
                     const auto t0 = std::chrono::steady_clock::now();
                     for (int b = 0; b < dc.batch; ++b)
                     {

tests/benchmarks/json_view/compare.py

@@ -154,11 +154,14 @@ def download_library(name, work):
     if not os.path.isfile(archive):
         print(f'downloading {pin["url"]}', flush=True)
         # the URLs are the https constants in PINNED, and the SHA-256 is checked below
-        urllib.request.urlretrieve(pin['url'], archive)  # nosec B310
+        # (into a .part file first, so that an interrupted download is not kept)
+        urllib.request.urlretrieve(pin['url'], archive + '.part')  # nosec B310
+        os.replace(archive + '.part', archive)
     with open(archive, 'rb') as f:
         digest = hashlib.sha256(f.read()).hexdigest()
     if digest != pin['sha256']:
-        sys.exit(f'error: SHA-256 of {archive} is {digest}, expected {pin["sha256"]}')
+        os.remove(archive)  # downloaded again by the next run
+        sys.exit(f'error: SHA-256 of {archive} is {digest}, expected {pin["sha256"]} (removed)')
     src = os.path.join(work, 'download', pin['dir'])
     if not os.path.isdir(src):
         with tarfile.open(archive) as t:
@@ -214,6 +217,10 @@ def main():
     ap.add_argument('--corpus', nargs='*', default=[], help='more files for bench_corpus')
     ap.add_argument('--build-dir', default=os.path.join(HERE, 'build'), help='where to build (default: build/ next to this script)')
     args = ap.parse_args()
+    # the benchmarks run in the build directory: make the paths absolute
+    args.data = os.path.abspath(args.data)
+    args.corpus = [os.path.abspath(f) for f in args.corpus]
+    args.build_dir = os.path.abspath(args.build_dir)
 
     cxx = os.environ.get('CXX', 'c++')
     cc = os.environ.get('CC', 'cc')

tests/src/unit-binary_formats.cpp

@@ -33,7 +33,7 @@ TEST_CASE("Binary Formats" * doctest::skip())
         const auto ubjson_2_size = json::to_ubjson(j, true).size();
         const auto ubjson_3_size = json::to_ubjson(j, true, true).size();
 
-        CHECK(json_size == 2090303);
+        CHECK(json_size == 2090234);
         CHECK(bjdata_1_size == 1112030);
         CHECK(bjdata_2_size == 1224148);
         CHECK(bjdata_3_size == 1224148);
@@ -46,16 +46,16 @@ TEST_CASE("Binary Formats" * doctest::skip())
         CHECK(ubjson_3_size == 1169069);
 
         CHECK((100.0 * double(json_size) / double(json_size)) == Approx(100.0));
-        CHECK((100.0 * double(bjdata_1_size) / double(json_size)) == Approx(53.199));
+        CHECK((100.0 * double(bjdata_1_size) / double(json_size)) == Approx(53.201));
-        CHECK((100.0 * double(bjdata_2_size) / double(json_size)) == Approx(58.563));
+        CHECK((100.0 * double(bjdata_2_size) / double(json_size)) == Approx(58.565));
-        CHECK((100.0 * double(bjdata_3_size) / double(json_size)) == Approx(58.563));
+        CHECK((100.0 * double(bjdata_3_size) / double(json_size)) == Approx(58.565));
-        CHECK((100.0 * double(bon8_size) / double(json_size)) == Approx(50.509));
+        CHECK((100.0 * double(bon8_size) / double(json_size)) == Approx(50.511));
-        CHECK((100.0 * double(bson_size) / double(json_size)) == Approx(85.849));
+        CHECK((100.0 * double(bson_size) / double(json_size)) == Approx(85.853));
-        CHECK((100.0 * double(cbor_size) / double(json_size)) == Approx(50.497));
+        CHECK((100.0 * double(cbor_size) / double(json_size)) == Approx(50.499));
-        CHECK((100.0 * double(msgpack_size) / double(json_size)) == Approx(50.526));
+        CHECK((100.0 * double(msgpack_size) / double(json_size)) == Approx(50.528));
-        CHECK((100.0 * double(ubjson_1_size) / double(json_size)) == Approx(53.199));
+        CHECK((100.0 * double(ubjson_1_size) / double(json_size)) == Approx(53.201));
-        CHECK((100.0 * double(ubjson_2_size) / double(json_size)) == Approx(58.563));
+        CHECK((100.0 * double(ubjson_2_size) / double(json_size)) == Approx(58.565));
-        CHECK((100.0 * double(ubjson_3_size) / double(json_size)) == Approx(55.928));
+        CHECK((100.0 * double(ubjson_3_size) / double(json_size)) == Approx(55.930));
     }
 
     SECTION("twitter.json")

tests/src/unit-json_view.cpp

@@ -1151,6 +1151,54 @@ TEST_CASE("json_view dump")
         CHECK(d.root().dump(0, ' ', false, json_view::number_format::source) == "[\n1.50,\n1E2,\n-0,\n-0.0,\n123456789012345678901234567890,\n18446744073709551615,\n-9223372036854775808,\n0.1,\n1e-7,\n5e-324\n]");
         CHECK(d.root().dump(-1, ' ', true, json_view::number_format::source) == "[1.50,1E2,-0,-0.0,123456789012345678901234567890,18446744073709551615,-9223372036854775808,0.1,1e-7,5e-324]");
 
+        // float tokens of up to 17 significant digits in every spelling: those
+        // of at most 15 digits are written from their digits, the others
+        // through the conversion; both as dump() writes them
+        {
+            std::mt19937_64 tokens(1170); // NOLINT(cert-msc32-c,cert-msc51-cpp,bugprone-random-generator-seed)
+            // a number below n; the remainder is a std::uint64_t, which is
+            // std::size_t on some platforms and wider on others
+            const auto draw = [&tokens](std::size_t n)
+            {
+                const std::uint64_t r = tokens() % n;
+                return static_cast<std::size_t>(r);
+            };
+            std::string many_tokens = "[";
+            for (int i = 0; i < 20000; ++i)
+            {
+                const std::size_t length = 1 + draw(17);
+                std::string digits(1, static_cast<char>('1' + draw(9)));
+                for (std::size_t k = 1; k < length; ++k)
+                {
+                    digits += static_cast<char>('0' + draw(10));
+                }
+                digits += std::string(draw(4), '0'); // trailing zeros
+                std::string token = draw(3) == 0 ? "-" : "";
+                const std::size_t point = draw(digits.size() + 1);
+                if (point == 0)
+                {
+                    token += "0." + std::string(draw(5), '0') + digits;
+                }
+                else
+                {
+                    token += digits.substr(0, point) + (point < digits.size() ? "." + digits.substr(point) : "");
+                }
+                // an exponent that keeps the value between about 1e-320 and 1e300
+                const int exponent = static_cast<int>(draw(600)) - 300 - static_cast<int>(point);
+                if (draw(4) != 0)
+                {
+                    token += (draw(2) == 0 ? "e" : "E") + std::string(exponent >= 0 && draw(2) == 0 ? "+" : "") + std::to_string(exponent);
+                }
+                else if (point == digits.size())
+                {
+                    token += ".0"; // (a float, not an integer)
+                }
+                many_tokens += (i != 0 ? "," : "") + token;
+            }
+            many_tokens += ']';
+            CHECK(json_document::parse(many_tokens).root().dump() == json::parse(many_tokens).dump());
+        }
+
         // random doubles, written as parse() and dump() would
         std::mt19937_64 rng(1170); // NOLINT(cert-msc32-c,cert-msc51-cpp,bugprone-random-generator-seed)
         std::string many = "[";

tests/src/unit-json_view_edit.cpp

@@ -26,6 +26,7 @@ using ptr_t = ordered_json::json_pointer;
 #include <functional>
 #include <iterator>
 #include <limits>
+#include <map>
 #include <random>
 #include <string>
 #include <vector>
@@ -481,6 +482,19 @@ TEST_CASE("json_view edits: views and values")
         CHECK(d.root().materialize().dump() == json::parse(R"([1.5, 100.0, 0.1, null, null, 18446744073709551615, -9223372036854775808])").dump());
     }
 
+    SECTION("numbers of other float types")
+    {
+        // doubles have their own path to the output; other float types are
+        // written as basic_json writes them, non-finite values as null
+        using json_float = nlohmann::basic_json<std::map, std::vector, std::string, bool, std::int64_t, std::uint64_t, float>;
+        using document_float = nlohmann::basic_json_document<json_float, true>;
+        document_float d = document_float::parse("[1.5]");
+        d.push_back(d.root(), std::numeric_limits<float>::quiet_NaN());
+        d.push_back(d.root(), -std::numeric_limits<float>::infinity());
+        CHECK(d.root().dump() == "[1.5,null,null]");
+        CHECK(d.root().dump(2) == json_float::parse("[1.5, null, null]").dump(2));
+    }
+
     SECTION("nulls become containers, and the root can be replaced")
     {
         json_editable_document d = json_editable_document::parse("[null, null]");

tests/src/unit-to_chars.cpp

@@ -15,6 +15,23 @@
 #include <nlohmann/json.hpp>
 using nlohmann::detail::dtoa_impl::reinterpret_bits;
 
+#include <array>
+#include <cmath>
+#include <cstdint>
+#include <cstdio>
+#include <cstdlib>
+#include <iomanip>
+#include <limits>
+#include <locale>
+#include <random>
+#include <sstream>
+#include <string>
+#include <utility>
+#include <vector>
+#if defined(JSON_HAS_CPP_17)
+    #include <charconv>
+#endif
+
 namespace
 {
 float make_float(uint32_t sign_bit, uint32_t biased_exponent, uint32_t significand)
@@ -450,7 +467,7 @@ TEST_CASE("formatting")
         check_double(  1.2345e+18,   "1.2345e+18"             ); //  1.2345e+18                1.2345e+18                1.2345e18
         check_double(  1.2345e+19,   "1.2345e+19"             ); //  1.2345e+19                1.2345e+19                1.2345e19
         check_double(  1.2345e+20,   "1.2345e+20"             ); //  1.2345e+20                1.2345e+20                1.2345e20
-        check_double(  1.2345e+21,   "1.2344999999999999e+21" ); //  1.2345e+21                1.2344999999999999e+21    1.2345e21
+        check_double(  1.2345e+21,   "1.2345e+21"             ); //  1.2345e+21                1.2344999999999999e+21    1.2345e21
         check_double(  1.2345e+22,   "1.2345e+22"             ); //  1.2345e+22                1.2345e+22                1.2345e22
     }
 
@@ -514,3 +531,245 @@ TEST_CASE("formatting")
         check_integer(1000000000000000000LL, "1000000000000000000");
     }
 }
+
+namespace
+{
+// a small unsigned big integer (32-bit limbs, least significant first), to
+// recompute the powers of ten of the shortest double conversion
+using big = std::vector<std::uint32_t>;
+
+void big_mul_small(big& x, std::uint32_t m)
+{
+    std::uint64_t carry = 0;
+    for (auto& limb : x)
+    {
+        const std::uint64_t v = (static_cast<std::uint64_t>(limb) * m) + carry;
+        limb = static_cast<std::uint32_t>(v);
+        carry = v >> 32u;
+    }
+    if (carry != 0)
+    {
+        x.push_back(static_cast<std::uint32_t>(carry));
+    }
+}
+
+void big_div_small(big& x, std::uint32_t d)
+{
+    std::uint64_t rest = 0;
+    for (std::size_t i = x.size(); i-- > 0;)
+    {
+        const std::uint64_t v = (rest << 32u) | x[i];
+        x[i] = static_cast<std::uint32_t>(v / d);
+        rest = v % d;
+    }
+    while (!x.empty() && x.back() == 0)
+    {
+        x.pop_back();
+    }
+}
+
+std::size_t big_bit_length(const big& x)
+{
+    std::size_t n = 32 * x.size();
+    for (std::uint32_t top = x.back(); (top & 0x80000000u) == 0; top <<= 1u)
+    {
+        --n;
+    }
+    return n;
+}
+
+bool big_bit(const big& x, std::size_t i)
+{
+    return ((x[i / 32] >> (i % 32)) & 1u) != 0;
+}
+
+/// the 128 most significant bits of x (floor), shifted left if x has fewer bits
+std::pair<std::uint64_t, std::uint64_t> big_top128(const big& x)
+{
+    const std::size_t n = big_bit_length(x);
+    std::uint64_t high = 0;
+    std::uint64_t low = 0;
+    for (std::size_t k = 0; k < 128; ++k)
+    {
+        const bool bit = k < n && big_bit(x, n - 1 - k);
+        if (k < 64)
+        {
+            high = (high << 1u) | (bit ? 1u : 0u);
+        }
+        else
+        {
+            low = (low << 1u) | (bit ? 1u : 0u);
+        }
+    }
+    return {high, low};
+}
+
+/// the digits (without trailing zeros) and the decimal exponent of a
+/// representation "[-]d[.ddd][e[+-]x]"
+std::pair<std::string, int> digits_and_exponent(const std::string& s)
+{
+    std::string digits;
+    int point = -1;
+    int exponent = 0;
+    for (std::size_t i = 0; i < s.size(); ++i)
+    {
+        const char c = s[i];
+        if (c >= '0' && c <= '9')
+        {
+            digits += c;
+        }
+        else if (c == '.')
+        {
+            point = static_cast<int>(digits.size());
+        }
+        else if (c == 'e' || c == 'E')
+        {
+            exponent = std::stoi(s.substr(i + 1));
+            break;
+        }
+    }
+    int e = exponent + (point < 0 ? static_cast<int>(digits.size()) : point) - static_cast<int>(digits.size());
+    const std::size_t first = digits.find_first_not_of('0');
+    digits = first == std::string::npos ? "0" : digits.substr(first);
+    while (digits.size() > 1 && digits.back() == '0')
+    {
+        digits.pop_back();
+        ++e;
+    }
+    return {digits, e};
+}
+
+/// whether the decimal digits * 10^e reads back as v
+bool reads_back(const std::string& digits, int e, double v)
+{
+    const std::string text = digits + "e" + std::to_string(e);
+    return std::strtod(text.c_str(), nullptr) == v;
+}
+
+/// Check the representation of a positive finite double: it reads back as
+/// the same value, and no representation with fewer digits does.
+void check_shortest(double v)
+{
+    std::array<char, 33> buf{};
+    char* end = nlohmann::detail::to_chars(buf.data(), buf.data() + 32, v);
+    const std::string text(buf.data(), end);
+    CAPTURE(text);
+    CHECK(std::strtod(text.c_str(), nullptr) == v);
+    // the layout is that of format_buffer() for the same digits
+    std::array<char, 64> reference{};
+    int len = 0;
+    int exponent = 0;
+    nlohmann::detail::dtoa_impl::shortest_digits(reference.data(), len, exponent, v);
+    const char* const reference_end = nlohmann::detail::dtoa_impl::format_buffer(reference.data(), len, exponent, -4, 15);
+    CHECK(text == std::string(reference.data(), static_cast<std::size_t>(reference_end - reference.data())));
+    const auto de = digits_and_exponent(text);
+    const std::string& digits = de.first;
+    if (digits.size() > 1)
+    {
+        // the decimals of one digit fewer next to the value
+        // (a stream rather than snprintf("%.*e"), whose output GCC cannot bound)
+        std::ostringstream shorter;
+        shorter.imbue(std::locale::classic());
+        shorter << std::scientific << std::setprecision(static_cast<int>(digits.size()) - 2) << v;
+        const auto near = digits_and_exponent(shorter.str());
+        // as an integer with digits.size() - 1 digits
+        std::string m = near.first;
+        int e = near.second;
+        while (m.size() < digits.size() - 1)
+        {
+            m += '0';
+            --e;
+        }
+        const std::uint64_t mid = std::stoull(m);
+        for (const std::uint64_t candidate :
+                {
+                    mid - 1, mid, mid + 1
+                })
+        {
+            CAPTURE(candidate);
+            CHECK(!reads_back(std::to_string(candidate), e, v));
+        }
+    }
+#if defined(JSON_HAS_CPP_17) && defined(__cpp_lib_to_chars)
+    // the closest of the shortest representations, as std::to_chars finds it
+    std::array<char, 64> std_text{};
+    const auto r = std::to_chars(std_text.data(), std_text.data() + std_text.size(), v, std::chars_format::scientific);
+    CHECK(digits_and_exponent(std::string(std_text.data(), r.ptr)) == de);
+#endif
+}
+} // namespace
+
+TEST_CASE("shortest digits of doubles")
+{
+    SECTION("powers of ten")
+    {
+        // the 128-bit significands of 10^k, rounded down, recomputed
+        for (int k = -342; k <= 341; ++k)
+        {
+            CAPTURE(k);
+            big x{1};
+            if (k >= 0)
+            {
+                for (int i = 0; i < k; ++i)
+                {
+                    big_mul_small(x, 10);
+                }
+            }
+            else
+            {
+                // floor(2^b / 10^-k) for a b that leaves more than 128 bits
+                const int b = 128 + 64 + (4 * -k);
+                x.assign(static_cast<std::size_t>(b / 32) + 1, 0);
+                x.back() = 1u << (b % 32);
+                for (int i = 0; i < -k; ++i)
+                {
+                    big_div_small(x, 10);
+                }
+            }
+            const auto expected = big_top128(x);
+            const auto actual = nlohmann::detail::zmij::pow10(k);
+            CHECK(actual.high == expected.first);
+            CHECK(actual.low == expected.second);
+        }
+    }
+
+    SECTION("boundary values")
+    {
+        for (const double v :
+                {
+                    std::numeric_limits<double>::min(), std::numeric_limits<double>::max(), std::numeric_limits<double>::denorm_min(),
+                    std::nextafter(std::numeric_limits<double>::min(), 0.0), 1.0, 2.0, 0.1, 0.3, 1e21, 1e22, 1e23, 5e-324, 9007199254740993.0,
+                    1.2345e+21, 2.2250738585072014e-308, 1.7976931348623157e308, 4.9406564584124654e-324, 123456789012345680.0
+                })
+        {
+            check_shortest(v);
+        }
+        // all powers of two (their rounding interval is narrower below)
+        for (int e = -1074; e <= 1023; ++e)
+        {
+            check_shortest(std::ldexp(1.0, e));
+        }
+        // powers of ten and their neighbors
+        for (int e = -323; e <= 308; ++e)
+        {
+            const double p = std::strtod(("1e" + std::to_string(e)).c_str(), nullptr);
+            check_shortest(p);
+            check_shortest(std::nextafter(p, 0.0));
+            check_shortest(std::nextafter(p, std::numeric_limits<double>::infinity()));
+        }
+    }
+
+    SECTION("random doubles")
+    {
+        std::mt19937_64 rng(5295); // NOLINT(cert-msc32-c,cert-msc51-cpp,bugprone-random-generator-seed): reproducible
+        for (int i = 0; i < 100000; ++i)
+        {
+            const std::uint64_t bits = rng() & 0x7FFFFFFFFFFFFFFFu;
+            const auto v = reinterpret_bits<double>(bits);
+            if (std::isfinite(v) && v != 0)
+            {
+                check_shortest(v);
+            }
+        }
+    }
+}