Grow the json_view dump's output in steps and compare raw-number dumps

- The output string reserves the size estimate (the source extent) and grows
  in steps of 64 KiB within it, instead of being resized to the estimate at
  once: resize() zero-fills, and for a pretty-printed source the estimate is
  far larger than the compact output. citm_catalog dump: 342 -> 283 us on
  x86-64, 144 -> 134 us on Apple M1.
- bench_corpus compares the dump with source numbers with yyjson writing
  numbers read as raw text (YYJSON_READ_NUMBER_AS_RAW); both write the same
  bytes.

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

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"
-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.
+(`<nlohmann/json_view.hpp>`) validates non-ASCII text in strings with SSSE3 without asking the CPU first.
 
-SSSE3 is not part of the x86-64 baseline, so the code must be compiled for it: 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. The same input is accepted or rejected either way; only the speed of non-ASCII text differs.
+By default, the parser checks once at run time whether the CPU has SSSE3 (all x86-64 CPUs since about 2011 have it)
+and then validates non-ASCII text 16 bytes at a time, using the "lookup4" algorithm of
+[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"
 

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)
-inline std::uint64_t read_eight_bytes(const unsigned char* b) noexcept
+/// little-endian targets; always inlined, as GCC otherwise calls it in the
+/// number loops)
+JSON_HEDLEY_ALWAYS_INLINE std::uint64_t read_eight_bytes(const unsigned char* b) noexcept
 {
     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

@@ -21,6 +21,21 @@
     #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>
 
@@ -1253,6 +1268,203 @@ inline char* write_decimal(char* first, std::uint64_t digits, int exp) noexcept
     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)
@@ -1284,7 +1496,7 @@ char* write_positive(char* first, const char* last, FloatType value)
 }
 
 /// a positive finite double: the shortest digits (Zmij), laid out by
-/// write_decimal() (through a local buffer if [first, last) is shorter than
+/// 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
@@ -1294,13 +1506,13 @@ inline char* write_positive(char* first, const char* last, double value)
                   "internal error: the conversion of Zmij needs IEEE 754 binary64 doubles");
     std::uint64_t bits = 0;
     std::memcpy(&bits, &value, sizeof(bits));
-    const zmij::decimal d = zmij::to_decimal(bits);
+    const zmij::shortest_decimal d = zmij::to_shortest(bits);
     if (JSON_HEDLEY_LIKELY(last - first >= 41))
     {
-        return write_decimal(first, d.significand, d.exponent);
+        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_decimal(buf.data(), d.significand, d.exponent) - buf.data());
+    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;

include/nlohmann/detail/conversions/zmij.hpp

@@ -157,10 +157,22 @@ inline std::uint64_t umul128_add_hi64(std::uint64_t x, std::uint64_t y, std::uin
     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). The significand can end in zeros.
-inline decimal to_decimal(std::uint64_t bits) noexcept
+/// 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);
@@ -205,12 +217,20 @@ inline decimal to_decimal(std::uint64_t bits) noexcept
         digit = digit < lowest ? lowest : digit;
     }
     integral += round_up ? 1u : 0u;
-    if (!round_up && !round_down)
+    // 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)
     {
-        // the shorter candidate is outside the rounding interval: one digit more
-        return decimal{(integral * 10) + digit, dec_exp};
+        return decimal{(d.integral * 10) + d.digit, d.exponent};
     }
-    return decimal{integral, dec_exp + 1};
+    return decimal{d.integral, d.exponent + 1};
 }
 
 }  // namespace zmij

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)
-inline std::uint32_t parse_eight_digits(std::uint64_t v) noexcept
+/// multiplications instead of eight (after simdjson and fast_float); always
+/// 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
-    @a number_buffer.
+    Dump a given integer, appending it to @ref write_buffer (directly: copying
+    the digits from another buffer right after writing them waits until the
+    stores are done).
 
     @param[in] x  integer number (signed or unsigned) to dump
     @tparam NumberType either @a number_integer_t or @a number_unsigned_t
@@ -1402,33 +1403,57 @@ class serializer
             return;
         }
 
-        // use a pointer to fill the buffer
-        auto buffer_ptr = number_buffer.begin(); // NOLINT(llvm-qualified-auto,readability-qualified-auto)
+        // use a pointer to fill the buffer (room for as much as number_buffer holds)
+        if (JSON_HEDLEY_UNLIKELY(write_buffer_pos + number_buffer.size() > write_buffer.size()))
+        {
+            flush();
+        }
+        auto* buffer_ptr = write_buffer.data() + write_buffer_pos;
 
         number_unsigned_t abs_value;
 
-        unsigned int n_chars{};
+        // one byte for the minus sign
+        unsigned int n_chars = 0;
 
         if (is_negative_number(x))
         {
             *buffer_ptr = '-';
             abs_value = remove_sign(static_cast<number_integer_t>(x));
-
-            // account one more byte for the minus sign
-            n_chars = 1 + count_digits(abs_value);
+            n_chars = 1;
         }
         else
         {
             abs_value = static_cast<number_unsigned_t>(x);
-            n_chars = count_digits(abs_value);
         }
 
+        // 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
-    with @a number_buffer.
+    Dump a given floating-point number, appending it to @ref write_buffer.
 
     @param[in] x  floating-point number to dump
     */
@@ -1485,10 +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);
-
-        put_buffer(number_buffer, static_cast<std::size_t>(end - begin));
+        write_buffer_pos += static_cast<std::size_t>(end - begin);
     }
 
     JSON_HEDLEY_NON_NULL(1)

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;
@@ -851,19 +856,21 @@ 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];
+                    cur_idx = f.idx;
+                    cur_count = f.count;
+                    cur_is_object = f.is_object;
                 }
                 else
                 {
-                    f = cold.deep.back();
+                    const frame f = cold.deep.back();
                     cold.deep.pop_back();
+                    cur_idx = f.idx;
+                    cur_count = f.count;
+                    cur_is_object = f.is_object;
                 }
-                cur_idx = f.idx;
-                cur_count = f.count;
-                cur_is_object = f.is_object;
             }
         }
 

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
-        // non-ASCII: the vector check, out of line
-        return scan_string_vector(p, e, plain);
-#else
+#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);
+        }
+#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

@@ -23,7 +23,6 @@
 #include <nlohmann/detail/view/macro_scope.hpp>
 #include <nlohmann/detail/view/node.hpp>
 #include <nlohmann/detail/view/number.hpp>
-#include <nlohmann/detail/view/simd.hpp>
 
 NLOHMANN_JSON_NAMESPACE_BEGIN
 namespace detail
@@ -32,7 +31,11 @@ namespace view
 {
 
 /// append-only output buffer: writes through a raw pointer into a string that
-/// is resized ahead, and trimmed by finish()
+/// is resized ahead, and trimmed by finish(). The estimate is reserved, and the
+/// string grows in steps of 64 KiB within it: resize() fills the new bytes
+/// with zeros (before C++23, a string cannot grow without), and a small step
+/// is filled while the writer is about to use it, in the cache, instead of
+/// filling the whole estimate in memory first.
 template<typename StringType>
 class output_buffer
 {
@@ -94,16 +97,31 @@ class output_buffer
     }
 
   private:
+    /// the size of a growth step (a function: std::min() takes a reference,
+    /// which a static constexpr member does not have before C++17)
+    static constexpr std::size_t step() noexcept
+    {
+        return 65536;
+    }
+
     static StringType& sized(StringType& out, std::size_t estimate)
     {
-        out.resize((std::max)(estimate, static_cast<std::size_t>(64)));
+        out.reserve(estimate);
+        out.resize((std::min)((std::max)(estimate, static_cast<std::size_t>(64)), step()));
         return out;
     }
 
     NLOHMANN_VIEW_NOINLINE void grow(std::size_t n)
     {
         const auto used = static_cast<std::size_t>(m_pos - m_out.data());
-        m_out.resize((std::max)(m_out.size() * 2, used + n + 256));
+        // (a step does not go beyond the reserved estimate, so that a good
+        // estimate is never copied to a larger allocation)
+        const std::size_t size = (std::max)((std::min)(m_out.size() + step(), m_out.capacity()), used + n + 256);
+        if (size > m_out.capacity())
+        {
+            m_out.reserve((std::max)(m_out.capacity() * 2, size));
+        }
+        m_out.resize(size);
         m_pos = &m_out[0] + used;
         m_end = &m_out[0] + m_out.size();
     }
@@ -222,69 +240,6 @@ struct dump_style
     bool source_numbers = false;   ///< copy number tokens from the source
 };
 
-/*!
-@brief digits * 10^exp as dtoa_impl::write_decimal() writes it (digits not 0,
-at most 17 digits; up to 41 bytes are written at first)
-
-With NEON, the fixed layouts ("0.00123", "12.5", "100.0") are put together in
-vector registers: output byte i is byte s + i of the digits (after '0's)
-before the point, and byte s + i - 1 after it. The portable code writes the
-digits to a buffer and copies them from there at another offset, and a load
-that spans several recent stores waits until they reach the cache.
-*/
-NLOHMANN_VIEW_ALWAYS_INLINE char* write_decimal(char* first, std::uint64_t digits, int exp) noexcept
-{
-#if NLOHMANN_VIEW_NEON
-    namespace dtoa = ::nlohmann::detail::dtoa_impl;
-    const std::uint64_t upper = digits / 100000000u;
-    const std::uint64_t b0 = upper / 100000000u; // one digit
-    const std::uint64_t b1 = dtoa::eight_digit_bytes(upper % 100000000u);
-    const std::uint64_t b2 = dtoa::eight_digit_bytes(digits % 100000000u);
-    // leading and trailing zero digits (as dtoa_impl::write_decimal())
-    int leading = 7;
-    if (b0 == 0)
-    {
-        leading = b1 != 0 ? 8 + (count_leading_zeros(b1) / 8) : 16 + (count_leading_zeros(b2) / 8);
-    }
-    int zeros = 16;
-    if (b2 != 0)
-    {
-        zeros = count_trailing_zeros(b2) / 8;
-    }
-    else if (b1 != 0)
-    {
-        zeros = 8 + (count_trailing_zeros(b1) / 8);
-    }
-    const int k = 24 - leading - zeros; // significant digits
-    const int n = k + exp + zeros;      // position of the point after the first digit
-    if (NLOHMANN_VIEW_LIKELY(-4 < n && n <= 15))
-    {
-        static const std::array<std::uint8_t, 32> iota = {{0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31}};
-        const int pad = n <= 0 ? 1 - n : 0;
-        const int len = k + pad;
-        const int point = n + pad;
-        // the 24 digit bytes in memory order, then '0's
-        const std::uint64_t zero_chars = 0x3030303030303030u;
-        const uint8x16x2_t table = {{
-                vcombine_u8(vcreate_u8(__builtin_bswap64(b0 + zero_chars)), vcreate_u8(__builtin_bswap64(b1 + zero_chars))),
-                vcombine_u8(vcreate_u8(__builtin_bswap64(b2 + zero_chars)), vdup_n_u8('0'))
-            }
-        };
-        const uint8x16_t s = vdupq_n_u8(static_cast<std::uint8_t>(leading - pad));
-        const uint8x16_t at_point = vdupq_n_u8(static_cast<std::uint8_t>(point));
-        for (std::size_t half = 0; half < 2; ++half)
-        {
-            const uint8x16_t i = vld1q_u8(iota.data() + (16 * half));
-            // (+ 0xFF is - 1 after the point; indexes past the digits read a '0')
-            const uint8x16_t index = vminq_u8(vaddq_u8(vaddq_u8(i, s), vcgtq_u8(i, at_point)), vdupq_n_u8(31));
-            vst1q_u8(reinterpret_cast<std::uint8_t*>(first) + (16 * half), vbslq_u8(vceqq_u8(i, at_point), vdupq_n_u8('.'), vqtbl2q_u8(table, index))); // NOLINT(cppcoreguidelines-pro-type-reinterpret-cast)
-        }
-        return first + (point >= len ? point + 2 : len + 1); // "digits[000].0" ends after ".0"
-    }
-#endif
-    return ::nlohmann::detail::dtoa_impl::write_decimal(first, digits, exp);
-}
-
 /*!
 @brief write a view's subtree as basic_json::dump() writes the value
 
@@ -786,7 +741,11 @@ class view_serializer
             if (d.w != 0 && d.w < 1000000000000000u && d.exponent >= -290 && d.exponent <= 290)
             {
                 *w = '-';
-                return write_decimal(w + (d.negative ? 1 : 0), d.w, static_cast<int>(d.exponent));
+                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)
         }
@@ -803,13 +762,13 @@ class view_serializer
     /// a double as dump() writes it, at w (64 bytes of room)
     static char* write_double_value_at(char* w, double x)
     {
-        if (NLOHMANN_VIEW_UNLIKELY(!std::isfinite(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);
         }
-#if NLOHMANN_VIEW_NEON
-        std::uint64_t bits = 0;
-        std::memcpy(&bits, &x, sizeof(bits));
         *w = '-';
         w += bits >> 63u;
         bits &= ~(std::uint64_t{1} << 63u);
@@ -817,11 +776,7 @@ class view_serializer
         {
             return write_text_at(w, "0.0", 3);
         }
-        const ::nlohmann::detail::zmij::decimal d = ::nlohmann::detail::zmij::to_decimal(bits);
-        return write_decimal(w, d.significand, d.exponent);
-#else
-        return ::nlohmann::detail::to_chars(w, w + 64, x);
-#endif
+        return ::nlohmann::detail::dtoa_impl::write_shortest(w, ::nlohmann::detail::zmij::to_shortest(bits));
     }
 
     /// as serializer::dump_float()

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
-// not part of x86-64, so it must not depend on the flags of a translation unit
-// (two translation units with different flags would have different
-// definitions of the same inline functions). JSON_VIEW_NO_SIMD selects the
-// portable code.
+// UTF-8 check needs NEON or SSSE3. SSSE3 is not part of x86-64, and the code
+// must not depend on the flags of a translation unit (two translation units
+// with different flags would have different definitions of the same inline
+// functions): the check is compiled for SSSE3 with a function attribute and
+// 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)
-inline std::uint64_t read_eight_bytes(const unsigned char* b) noexcept
+/// little-endian targets; always inlined, as GCC otherwise calls it in the
+/// number loops)
+JSON_HEDLEY_ALWAYS_INLINE std::uint64_t read_eight_bytes(const unsigned char* b) noexcept
 {
     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)
-inline std::uint32_t parse_eight_digits(std::uint64_t v) noexcept
+/// multiplications instead of eight (after simdjson and fast_float); always
+/// 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;
@@ -24426,6 +24428,21 @@ NLOHMANN_JSON_NAMESPACE_END
     #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++
@@ -24590,10 +24607,22 @@ inline std::uint64_t umul128_add_hi64(std::uint64_t x, std::uint64_t y, std::uin
     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). The significand can end in zeros.
-inline decimal to_decimal(std::uint64_t bits) noexcept
+/// 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);
@@ -24638,12 +24667,20 @@ inline decimal to_decimal(std::uint64_t bits) noexcept
         digit = digit < lowest ? lowest : digit;
     }
     integral += round_up ? 1u : 0u;
-    if (!round_up && !round_down)
+    // 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)
     {
-        // the shorter candidate is outside the rounding interval: one digit more
-        return decimal{(integral * 10) + digit, dec_exp};
+        return decimal{(d.integral * 10) + d.digit, d.exponent};
     }
-    return decimal{integral, dec_exp + 1};
+    return decimal{d.integral, d.exponent + 1};
 }
 
 }  // namespace zmij
@@ -25882,6 +25919,203 @@ inline char* write_decimal(char* first, std::uint64_t digits, int exp) noexcept
     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)
@@ -25913,7 +26147,7 @@ char* write_positive(char* first, const char* last, FloatType value)
 }
 
 /// a positive finite double: the shortest digits (Zmij), laid out by
-/// write_decimal() (through a local buffer if [first, last) is shorter than
+/// 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
@@ -25923,13 +26157,13 @@ inline char* write_positive(char* first, const char* last, double value)
                   "internal error: the conversion of Zmij needs IEEE 754 binary64 doubles");
     std::uint64_t bits = 0;
     std::memcpy(&bits, &value, sizeof(bits));
-    const zmij::decimal d = zmij::to_decimal(bits);
+    const zmij::shortest_decimal d = zmij::to_shortest(bits);
     if (JSON_HEDLEY_LIKELY(last - first >= 41))
     {
-        return write_decimal(first, d.significand, d.exponent);
+        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_decimal(buf.data(), d.significand, d.exponent) - buf.data());
+    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;
@@ -27336,8 +27570,9 @@ class serializer
     /*!
     @brief dump an integer
 
-    Dump a given integer, appending it to @ref write_buffer. Works internally with
-    @a number_buffer.
+    Dump a given integer, appending it to @ref write_buffer (directly: copying
+    the digits from another buffer right after writing them waits until the
+    stores are done).
 
     @param[in] x  integer number (signed or unsigned) to dump
     @tparam NumberType either @a number_integer_t or @a number_unsigned_t
@@ -27372,33 +27607,57 @@ class serializer
             return;
         }
 
-        // use a pointer to fill the buffer
-        auto buffer_ptr = number_buffer.begin(); // NOLINT(llvm-qualified-auto,readability-qualified-auto)
+        // use a pointer to fill the buffer (room for as much as number_buffer holds)
+        if (JSON_HEDLEY_UNLIKELY(write_buffer_pos + number_buffer.size() > write_buffer.size()))
+        {
+            flush();
+        }
+        auto* buffer_ptr = write_buffer.data() + write_buffer_pos;
 
         number_unsigned_t abs_value;
 
-        unsigned int n_chars{};
+        // one byte for the minus sign
+        unsigned int n_chars = 0;
 
         if (is_negative_number(x))
         {
             *buffer_ptr = '-';
             abs_value = remove_sign(static_cast<number_integer_t>(x));
-
-            // account one more byte for the minus sign
-            n_chars = 1 + count_digits(abs_value);
+            n_chars = 1;
         }
         else
         {
             abs_value = static_cast<number_unsigned_t>(x);
-            n_chars = count_digits(abs_value);
         }
 
+        // 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
@@ -27421,14 +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
-    with @a number_buffer.
+    Dump a given floating-point number, appending it to @ref write_buffer.
 
     @param[in] x  floating-point number to dump
     */
@@ -27455,10 +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);
-
-        put_buffer(number_buffer, static_cast<std::size_t>(end - begin));
+        write_buffer_pos += static_cast<std::size_t>(end - begin);
     }
 
     JSON_HEDLEY_NON_NULL(1)
@@ -35141,6 +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
-// not part of x86-64, so it must not depend on the flags of a translation unit
-// (two translation units with different flags would have different
-// definitions of the same inline functions). JSON_VIEW_NO_SIMD selects the
-// portable code.
+// UTF-8 check needs NEON or SSSE3. SSSE3 is not part of x86-64, and the code
+// must not depend on the flags of a translation unit (two translation units
+// with different flags would have different definitions of the same inline
+// functions): the check is compiled for SSSE3 with a function attribute and
+// 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
-        // non-ASCII: the vector check, out of line
-        return scan_string_vector(p, e, plain);
-#else
+#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);
+        }
+#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;
@@ -1836,19 +1910,21 @@ 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];
+                    cur_idx = f.idx;
+                    cur_count = f.count;
+                    cur_is_object = f.is_object;
                 }
                 else
                 {
-                    f = cold.deep.back();
+                    const frame f = cold.deep.back();
                     cold.deep.pop_back();
+                    cur_idx = f.idx;
+                    cur_count = f.count;
+                    cur_is_object = f.is_object;
                 }
-                cur_idx = f.idx;
-                cur_count = f.count;
-                cur_is_object = f.is_object;
             }
         }
 
@@ -5325,8 +5401,6 @@ NLOHMANN_JSON_NAMESPACE_END
 
 // #include <nlohmann/detail/view/number.hpp>
 
-// #include <nlohmann/detail/view/simd.hpp>
-
 
 NLOHMANN_JSON_NAMESPACE_BEGIN
 namespace detail
@@ -5335,7 +5409,11 @@ namespace view
 {
 
 /// append-only output buffer: writes through a raw pointer into a string that
-/// is resized ahead, and trimmed by finish()
+/// is resized ahead, and trimmed by finish(). The estimate is reserved, and the
+/// string grows in steps of 64 KiB within it: resize() fills the new bytes
+/// with zeros (before C++23, a string cannot grow without), and a small step
+/// is filled while the writer is about to use it, in the cache, instead of
+/// filling the whole estimate in memory first.
 template<typename StringType>
 class output_buffer
 {
@@ -5397,16 +5475,31 @@ class output_buffer
     }
 
   private:
+    /// the size of a growth step (a function: std::min() takes a reference,
+    /// which a static constexpr member does not have before C++17)
+    static constexpr std::size_t step() noexcept
+    {
+        return 65536;
+    }
+
     static StringType& sized(StringType& out, std::size_t estimate)
     {
-        out.resize((std::max)(estimate, static_cast<std::size_t>(64)));
+        out.reserve(estimate);
+        out.resize((std::min)((std::max)(estimate, static_cast<std::size_t>(64)), step()));
         return out;
     }
 
     NLOHMANN_VIEW_NOINLINE void grow(std::size_t n)
     {
         const auto used = static_cast<std::size_t>(m_pos - m_out.data());
-        m_out.resize((std::max)(m_out.size() * 2, used + n + 256));
+        // (a step does not go beyond the reserved estimate, so that a good
+        // estimate is never copied to a larger allocation)
+        const std::size_t size = (std::max)((std::min)(m_out.size() + step(), m_out.capacity()), used + n + 256);
+        if (size > m_out.capacity())
+        {
+            m_out.reserve((std::max)(m_out.capacity() * 2, size));
+        }
+        m_out.resize(size);
         m_pos = &m_out[0] + used;
         m_end = &m_out[0] + m_out.size();
     }
@@ -5525,69 +5618,6 @@ struct dump_style
     bool source_numbers = false;   ///< copy number tokens from the source
 };
 
-/*!
-@brief digits * 10^exp as dtoa_impl::write_decimal() writes it (digits not 0,
-at most 17 digits; up to 41 bytes are written at first)
-
-With NEON, the fixed layouts ("0.00123", "12.5", "100.0") are put together in
-vector registers: output byte i is byte s + i of the digits (after '0's)
-before the point, and byte s + i - 1 after it. The portable code writes the
-digits to a buffer and copies them from there at another offset, and a load
-that spans several recent stores waits until they reach the cache.
-*/
-NLOHMANN_VIEW_ALWAYS_INLINE char* write_decimal(char* first, std::uint64_t digits, int exp) noexcept
-{
-#if NLOHMANN_VIEW_NEON
-    namespace dtoa = ::nlohmann::detail::dtoa_impl;
-    const std::uint64_t upper = digits / 100000000u;
-    const std::uint64_t b0 = upper / 100000000u; // one digit
-    const std::uint64_t b1 = dtoa::eight_digit_bytes(upper % 100000000u);
-    const std::uint64_t b2 = dtoa::eight_digit_bytes(digits % 100000000u);
-    // leading and trailing zero digits (as dtoa_impl::write_decimal())
-    int leading = 7;
-    if (b0 == 0)
-    {
-        leading = b1 != 0 ? 8 + (count_leading_zeros(b1) / 8) : 16 + (count_leading_zeros(b2) / 8);
-    }
-    int zeros = 16;
-    if (b2 != 0)
-    {
-        zeros = count_trailing_zeros(b2) / 8;
-    }
-    else if (b1 != 0)
-    {
-        zeros = 8 + (count_trailing_zeros(b1) / 8);
-    }
-    const int k = 24 - leading - zeros; // significant digits
-    const int n = k + exp + zeros;      // position of the point after the first digit
-    if (NLOHMANN_VIEW_LIKELY(-4 < n && n <= 15))
-    {
-        static const std::array<std::uint8_t, 32> iota = {{0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31}};
-        const int pad = n <= 0 ? 1 - n : 0;
-        const int len = k + pad;
-        const int point = n + pad;
-        // the 24 digit bytes in memory order, then '0's
-        const std::uint64_t zero_chars = 0x3030303030303030u;
-        const uint8x16x2_t table = {{
-                vcombine_u8(vcreate_u8(__builtin_bswap64(b0 + zero_chars)), vcreate_u8(__builtin_bswap64(b1 + zero_chars))),
-                vcombine_u8(vcreate_u8(__builtin_bswap64(b2 + zero_chars)), vdup_n_u8('0'))
-            }
-        };
-        const uint8x16_t s = vdupq_n_u8(static_cast<std::uint8_t>(leading - pad));
-        const uint8x16_t at_point = vdupq_n_u8(static_cast<std::uint8_t>(point));
-        for (std::size_t half = 0; half < 2; ++half)
-        {
-            const uint8x16_t i = vld1q_u8(iota.data() + (16 * half));
-            // (+ 0xFF is - 1 after the point; indexes past the digits read a '0')
-            const uint8x16_t index = vminq_u8(vaddq_u8(vaddq_u8(i, s), vcgtq_u8(i, at_point)), vdupq_n_u8(31));
-            vst1q_u8(reinterpret_cast<std::uint8_t*>(first) + (16 * half), vbslq_u8(vceqq_u8(i, at_point), vdupq_n_u8('.'), vqtbl2q_u8(table, index))); // NOLINT(cppcoreguidelines-pro-type-reinterpret-cast)
-        }
-        return first + (point >= len ? point + 2 : len + 1); // "digits[000].0" ends after ".0"
-    }
-#endif
-    return ::nlohmann::detail::dtoa_impl::write_decimal(first, digits, exp);
-}
-
 /*!
 @brief write a view's subtree as basic_json::dump() writes the value
 
@@ -6089,7 +6119,11 @@ class view_serializer
             if (d.w != 0 && d.w < 1000000000000000u && d.exponent >= -290 && d.exponent <= 290)
             {
                 *w = '-';
-                return write_decimal(w + (d.negative ? 1 : 0), d.w, static_cast<int>(d.exponent));
+                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)
         }
@@ -6106,13 +6140,13 @@ class view_serializer
     /// a double as dump() writes it, at w (64 bytes of room)
     static char* write_double_value_at(char* w, double x)
     {
-        if (NLOHMANN_VIEW_UNLIKELY(!std::isfinite(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);
         }
-#if NLOHMANN_VIEW_NEON
-        std::uint64_t bits = 0;
-        std::memcpy(&bits, &x, sizeof(bits));
         *w = '-';
         w += bits >> 63u;
         bits &= ~(std::uint64_t{1} << 63u);
@@ -6120,11 +6154,7 @@ class view_serializer
         {
             return write_text_at(w, "0.0", 3);
         }
-        const ::nlohmann::detail::zmij::decimal d = ::nlohmann::detail::zmij::to_decimal(bits);
-        return write_decimal(w, d.significand, d.exponent);
-#else
-        return ::nlohmann::detail::to_chars(w, w + 64, x);
-#endif
+        return ::nlohmann::detail::dtoa_impl::write_shortest(w, ::nlohmann::detail::zmij::to_shortest(bits));
     }
 
     /// as serializer::dump_float()
@@ -7891,6 +7921,8 @@ 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

@@ -50,7 +50,8 @@ JSON-RPC request (`rpc`):
 | dump | serialize a parsed document (compact) |
 
 `bench_corpus.cpp` runs parse, traverse, and dump on any list of files, so that no library is tuned to a handful of
-documents.
+documents. Its dump also writes the numbers as they are in the input: `json_view` with `number_format::source`, and
+yyjson with numbers read as raw text (`YYJSON_READ_NUMBER_AS_RAW`), without converting them.
 
 `bench_edit.cpp` measures read-modify-write: parse, apply the same logical edits with each library's own API, and
 serialize (compact). Workloads: `patch` (a handful of edits at fixed places) and `update` (edits in every record).
@@ -60,7 +61,10 @@ 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 +72,18 @@ 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

@@ -15,7 +15,8 @@
 // count, string bytes, sum of numbers) before anything is timed. Workloads:
 // parse (build and free a document), traverse (visit every value, convert
 // every number), dump (compact), and for json_view also dump with the source
-// number text. Results go to bench_corpus.csv.
+// number text, compared with yyjson writing numbers read as raw text
+// (YYJSON_READ_NUMBER_AS_RAW). Results go to bench_corpus.csv.
 #include <nlohmann/json_view.hpp>
 
 #if JSON_VIEW_BENCH_BOOST
@@ -29,6 +30,7 @@
 #include <chrono>
 #include <cmath>
 #include <cstdio>
+#include <cstdlib>
 #include <cstring>
 #include <fstream>
 #include <functional>
@@ -195,6 +197,11 @@ 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 +229,7 @@ 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)
     {
@@ -265,6 +273,7 @@ int main(int argc, char** argv)
         };
         json_document vd = json_document::parse(s);
         yyjson_doc* yd = yyjson_read(s.data(), s.size(), 0);
+        yyjson_doc* yd_raw = yyjson_read(s.data(), s.size(), YYJSON_READ_NUMBER_AS_RAW);
         simdjson::dom::parser sjd;
         const simdjson::dom::element se = sjd.parse(ps).value_unsafe();
         const std::vector<std::pair<std::string, std::vector<engine>>> workloads =
@@ -272,8 +281,10 @@ 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
@@ -295,6 +306,7 @@ int main(int argc, char** argv)
                     {"yyjson", [&] { std::size_t n = 0; char* o = yyjson_write(yd, 0, &n); g_sink = static_cast<double>(n); std::free(o); }},
                     {"simdjson DOM", [&] { std::string o = simdjson::to_string(se); g_sink = static_cast<double>(o.size()); }},
                     {"json_view (source numbers)", [&] { std::string o = vd.root().dump(-1, ' ', false, json_view::number_format::source); g_sink = static_cast<double>(o.size()); }},
+                    {"yyjson (raw numbers)", [&] { std::size_t n = 0; char* o = yyjson_write(yd_raw, 0, &n); g_sink = static_cast<double>(n); std::free(o); }},
                 }
             },
         };
@@ -306,6 +318,9 @@ 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());
@@ -321,6 +336,7 @@ int main(int argc, char** argv)
             std::fflush(stdout);
         }
         yyjson_doc_free(yd);
+        yyjson_doc_free(yd_raw);
     }
     std::fclose(csv);
 }

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')