json 2.6.3 → 2.13.0

data/ext/json/ext/simd/conf.rb

@@ -0,0 +1,20 @@
+case RbConfig::CONFIG['host_cpu']
+when /^(arm|aarch64)/
+  # Try to compile a small program using NEON instructions
+  header, type, init = 'arm_neon.h', 'uint8x16_t', 'vdupq_n_u8(32)'
+when /^(x86_64|x64)/
+  header, type, init = 'x86intrin.h', '__m128i', '_mm_set1_epi8(32)'
+end
+if header
+  have_header(header) && try_compile(<<~SRC)
+    #{cpp_include(header)}
+    int main(int argc, char **argv) {
+      #{type} test = #{init};
+      if (argc > 100000) printf("%p", &test);
+      return 0;
+    }
+  SRC
+  $defs.push("-DJSON_ENABLE_SIMD")
+end
+
+have_header('cpuid.h')

data/ext/json/ext/simd/simd.h

@@ -0,0 +1,187 @@
+typedef enum {
+    SIMD_NONE,
+    SIMD_NEON,
+    SIMD_SSE2
+} SIMD_Implementation;
+
+#ifdef JSON_ENABLE_SIMD
+
+#ifdef __clang__
+  #if __has_builtin(__builtin_ctzll)
+    #define HAVE_BUILTIN_CTZLL 1
+  #else
+    #define HAVE_BUILTIN_CTZLL 0
+  #endif
+#elif defined(__GNUC__) && (__GNUC__ > 4 || (__GNUC__ == 4 && __GNUC_MINOR__ >= 3))
+  #define HAVE_BUILTIN_CTZLL 1
+#else
+  #define HAVE_BUILTIN_CTZLL 0
+#endif
+
+static inline uint32_t trailing_zeros64(uint64_t input)
+{
+#if HAVE_BUILTIN_CTZLL
+  return __builtin_ctzll(input);
+#else
+  uint32_t trailing_zeros = 0;
+  uint64_t temp = input;
+  while ((temp & 1) == 0 && temp > 0) {
+    trailing_zeros++;
+    temp >>= 1;
+  }
+  return trailing_zeros;
+#endif
+}
+
+static inline int trailing_zeros(int input)
+{
+  #if HAVE_BUILTIN_CTZLL
+    return __builtin_ctz(input);
+  #else
+    int trailing_zeros = 0;
+    int temp = input;
+    while ((temp & 1) == 0 && temp > 0) {
+      trailing_zeros++;
+      temp >>= 1;
+    }
+    return trailing_zeros;
+  #endif
+}
+
+#if (defined(__GNUC__ ) || defined(__clang__))
+#define FORCE_INLINE __attribute__((always_inline))
+#else
+#define FORCE_INLINE
+#endif
+
+
+#define SIMD_MINIMUM_THRESHOLD 6
+
+#if defined(__ARM_NEON) || defined(__ARM_NEON__) || defined(__aarch64__) || defined(_M_ARM64)
+#include <arm_neon.h>
+
+#define FIND_SIMD_IMPLEMENTATION_DEFINED 1
+static inline SIMD_Implementation find_simd_implementation(void)
+{
+    return SIMD_NEON;
+}
+
+#define HAVE_SIMD 1
+#define HAVE_SIMD_NEON 1
+
+// See: https://community.arm.com/arm-community-blogs/b/servers-and-cloud-computing-blog/posts/porting-x86-vector-bitmask-optimizations-to-arm-neon
+static inline FORCE_INLINE uint64_t neon_match_mask(uint8x16_t matches)
+{
+    const uint8x8_t res = vshrn_n_u16(vreinterpretq_u16_u8(matches), 4);
+    const uint64_t mask = vget_lane_u64(vreinterpret_u64_u8(res), 0);
+    return mask & 0x8888888888888888ull;
+}
+
+static inline FORCE_INLINE uint64_t compute_chunk_mask_neon(const char *ptr)
+{
+  uint8x16_t chunk = vld1q_u8((const unsigned char *)ptr);
+
+  // Trick: c < 32 || c == 34 can be factored as c ^ 2 < 33
+  // https://lemire.me/blog/2025/04/13/detect-control-characters-quotes-and-backslashes-efficiently-using-swar/
+  const uint8x16_t too_low_or_dbl_quote = vcltq_u8(veorq_u8(chunk, vdupq_n_u8(2)), vdupq_n_u8(33));
+
+  uint8x16_t has_backslash = vceqq_u8(chunk, vdupq_n_u8('\\'));
+  uint8x16_t needs_escape  = vorrq_u8(too_low_or_dbl_quote, has_backslash);
+  return neon_match_mask(needs_escape);
+}
+
+static inline FORCE_INLINE int string_scan_simd_neon(const char **ptr, const char *end, uint64_t *mask)
+{
+    while (*ptr + sizeof(uint8x16_t) <= end) {
+      uint64_t chunk_mask = compute_chunk_mask_neon(*ptr);
+      if (chunk_mask) {
+          *mask = chunk_mask;
+          return 1;
+      }
+      *ptr += sizeof(uint8x16_t);
+    }
+    return 0;
+}
+
+uint8x16x4_t load_uint8x16_4(const unsigned char *table) {
+  uint8x16x4_t tab;
+  tab.val[0] = vld1q_u8(table);
+  tab.val[1] = vld1q_u8(table+16);
+  tab.val[2] = vld1q_u8(table+32);
+  tab.val[3] = vld1q_u8(table+48);
+  return tab;
+}
+
+#endif /* ARM Neon Support.*/
+
+#if defined(__amd64__) || defined(__amd64) || defined(__x86_64__) || defined(__x86_64) || defined(_M_X64) || defined(_M_AMD64)
+
+#ifdef HAVE_X86INTRIN_H
+#include <x86intrin.h>
+
+#define HAVE_SIMD 1
+#define HAVE_SIMD_SSE2 1
+
+#ifdef HAVE_CPUID_H
+#define FIND_SIMD_IMPLEMENTATION_DEFINED 1
+
+#if defined(__clang__) || defined(__GNUC__)
+#define TARGET_SSE2 __attribute__((target("sse2")))
+#else
+#define TARGET_SSE2
+#endif
+
+#define _mm_cmpge_epu8(a, b) _mm_cmpeq_epi8(_mm_max_epu8(a, b), a)
+#define _mm_cmple_epu8(a, b) _mm_cmpge_epu8(b, a)
+#define _mm_cmpgt_epu8(a, b) _mm_xor_si128(_mm_cmple_epu8(a, b), _mm_set1_epi8(-1))
+#define _mm_cmplt_epu8(a, b) _mm_cmpgt_epu8(b, a)
+
+static inline TARGET_SSE2 FORCE_INLINE int compute_chunk_mask_sse2(const char *ptr)
+{
+    __m128i chunk         = _mm_loadu_si128((__m128i const*)ptr);
+    // Trick: c < 32 || c == 34 can be factored as c ^ 2 < 33
+    // https://lemire.me/blog/2025/04/13/detect-control-characters-quotes-and-backslashes-efficiently-using-swar/
+    __m128i too_low_or_dbl_quote = _mm_cmplt_epu8(_mm_xor_si128(chunk, _mm_set1_epi8(2)), _mm_set1_epi8(33));
+    __m128i has_backslash = _mm_cmpeq_epi8(chunk, _mm_set1_epi8('\\'));
+    __m128i needs_escape  = _mm_or_si128(too_low_or_dbl_quote, has_backslash);
+    return _mm_movemask_epi8(needs_escape);
+}
+
+static inline TARGET_SSE2 FORCE_INLINE int string_scan_simd_sse2(const char **ptr, const char *end, int *mask)
+{
+    while (*ptr + sizeof(__m128i) <= end) {
+      int chunk_mask = compute_chunk_mask_sse2(*ptr);
+      if (chunk_mask) {
+          *mask = chunk_mask;
+          return 1;
+      }
+      *ptr += sizeof(__m128i);
+    }
+
+    return 0;
+}
+
+#include <cpuid.h>
+#endif /* HAVE_CPUID_H */
+
+static inline SIMD_Implementation find_simd_implementation(void)
+{
+    // TODO Revisit. I think the SSE version now only uses SSE2 instructions.
+    if (__builtin_cpu_supports("sse2")) {
+        return SIMD_SSE2;
+    }
+
+    return SIMD_NONE;
+}
+
+#endif /* HAVE_X86INTRIN_H */
+#endif /* X86_64 Support */
+
+#endif /* JSON_ENABLE_SIMD */
+
+#ifndef FIND_SIMD_IMPLEMENTATION_DEFINED
+static inline SIMD_Implementation find_simd_implementation(void)
+{
+    return SIMD_NONE;
+}
+#endif

data/ext/json/ext/vendor/fpconv.c

@@ -0,0 +1,479 @@
+// Boost Software License - Version 1.0 - August 17th, 2003
+//
+// Permission is hereby granted, free of charge, to any person or organization
+// obtaining a copy of the software and accompanying documentation covered by
+// this license (the "Software") to use, reproduce, display, distribute,
+// execute, and transmit the Software, and to prepare derivative works of the
+// Software, and to permit third-parties to whom the Software is furnished to
+// do so, all subject to the following:
+//
+// The copyright notices in the Software and this entire statement, including
+// the above license grant, this restriction and the following disclaimer,
+// must be included in all copies of the Software, in whole or in part, and
+// all derivative works of the Software, unless such copies or derivative
+// works are solely in the form of machine-executable object code generated by
+// a source language processor.
+//
+// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+// FITNESS FOR A PARTICULAR PURPOSE, TITLE AND NON-INFRINGEMENT. IN NO EVENT
+// SHALL THE COPYRIGHT HOLDERS OR ANYONE DISTRIBUTING THE SOFTWARE BE LIABLE
+// FOR ANY DAMAGES OR OTHER LIABILITY, WHETHER IN CONTRACT, TORT OR OTHERWISE,
+// ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
+// DEALINGS IN THE SOFTWARE.
+
+// The contents of this file is extracted from https://github.com/night-shift/fpconv
+// It was slightly modified to append ".0" to plain floats, for use with the https://github.com/ruby/json package.
+
+#include <stdbool.h>
+#include <string.h>
+#include <stdint.h>
+
+#define npowers     87
+#define steppowers  8
+#define firstpower -348 /* 10 ^ -348 */
+
+#define expmax     -32
+#define expmin     -60
+
+typedef struct Fp {
+    uint64_t frac;
+    int exp;
+} Fp;
+
+static const Fp powers_ten[] = {
+    { 18054884314459144840U, -1220 }, { 13451937075301367670U, -1193 },
+    { 10022474136428063862U, -1166 }, { 14934650266808366570U, -1140 },
+    { 11127181549972568877U, -1113 }, { 16580792590934885855U, -1087 },
+    { 12353653155963782858U, -1060 }, { 18408377700990114895U, -1034 },
+    { 13715310171984221708U, -1007 }, { 10218702384817765436U, -980 },
+    { 15227053142812498563U, -954 },  { 11345038669416679861U, -927 },
+    { 16905424996341287883U, -901 },  { 12595523146049147757U, -874 },
+    { 9384396036005875287U,  -847 },  { 13983839803942852151U, -821 },
+    { 10418772551374772303U, -794 },  { 15525180923007089351U, -768 },
+    { 11567161174868858868U, -741 },  { 17236413322193710309U, -715 },
+    { 12842128665889583758U, -688 },  { 9568131466127621947U,  -661 },
+    { 14257626930069360058U, -635 },  { 10622759856335341974U, -608 },
+    { 15829145694278690180U, -582 },  { 11793632577567316726U, -555 },
+    { 17573882009934360870U, -529 },  { 13093562431584567480U, -502 },
+    { 9755464219737475723U,  -475 },  { 14536774485912137811U, -449 },
+    { 10830740992659433045U, -422 },  { 16139061738043178685U, -396 },
+    { 12024538023802026127U, -369 },  { 17917957937422433684U, -343 },
+    { 13349918974505688015U, -316 },  { 9946464728195732843U,  -289 },
+    { 14821387422376473014U, -263 },  { 11042794154864902060U, -236 },
+    { 16455045573212060422U, -210 },  { 12259964326927110867U, -183 },
+    { 18268770466636286478U, -157 },  { 13611294676837538539U, -130 },
+    { 10141204801825835212U, -103 },  { 15111572745182864684U, -77 },
+    { 11258999068426240000U, -50 },   { 16777216000000000000U, -24 },
+    { 12500000000000000000U,   3 },   { 9313225746154785156U,   30 },
+    { 13877787807814456755U,  56 },   { 10339757656912845936U,  83 },
+    { 15407439555097886824U, 109 },   { 11479437019748901445U, 136 },
+    { 17105694144590052135U, 162 },   { 12744735289059618216U, 189 },
+    { 9495567745759798747U,  216 },   { 14149498560666738074U, 242 },
+    { 10542197943230523224U, 269 },   { 15709099088952724970U, 295 },
+    { 11704190886730495818U, 322 },   { 17440603504673385349U, 348 },
+    { 12994262207056124023U, 375 },   { 9681479787123295682U,  402 },
+    { 14426529090290212157U, 428 },   { 10748601772107342003U, 455 },
+    { 16016664761464807395U, 481 },   { 11933345169920330789U, 508 },
+    { 17782069995880619868U, 534 },   { 13248674568444952270U, 561 },
+    { 9871031767461413346U,  588 },   { 14708983551653345445U, 614 },
+    { 10959046745042015199U, 641 },   { 16330252207878254650U, 667 },
+    { 12166986024289022870U, 694 },   { 18130221999122236476U, 720 },
+    { 13508068024458167312U, 747 },   { 10064294952495520794U, 774 },
+    { 14996968138956309548U, 800 },   { 11173611982879273257U, 827 },
+    { 16649979327439178909U, 853 },   { 12405201291620119593U, 880 },
+    { 9242595204427927429U,  907 },   { 13772540099066387757U, 933 },
+    { 10261342003245940623U, 960 },   { 15290591125556738113U, 986 },
+    { 11392378155556871081U, 1013 },  { 16975966327722178521U, 1039 },
+    { 12648080533535911531U, 1066 }
+};
+
+static Fp find_cachedpow10(int exp, int* k)
+{
+    const double one_log_ten = 0.30102999566398114;
+
+    int approx = (int)(-(exp + npowers) * one_log_ten);
+    int idx = (approx - firstpower) / steppowers;
+
+    while(1) {
+        int current = exp + powers_ten[idx].exp + 64;
+
+        if(current < expmin) {
+            idx++;
+            continue;
+        }
+
+        if(current > expmax) {
+            idx--;
+            continue;
+        }
+
+        *k = (firstpower + idx * steppowers);
+
+        return powers_ten[idx];
+    }
+}
+
+#define fracmask  0x000FFFFFFFFFFFFFU
+#define expmask   0x7FF0000000000000U
+#define hiddenbit 0x0010000000000000U
+#define signmask  0x8000000000000000U
+#define expbias   (1023 + 52)
+
+#define absv(n) ((n) < 0 ? -(n) : (n))
+#define minv(a, b) ((a) < (b) ? (a) : (b))
+
+static const uint64_t tens[] = {
+    10000000000000000000U, 1000000000000000000U, 100000000000000000U,
+    10000000000000000U, 1000000000000000U, 100000000000000U,
+    10000000000000U, 1000000000000U, 100000000000U,
+    10000000000U, 1000000000U, 100000000U,
+    10000000U, 1000000U, 100000U,
+    10000U, 1000U, 100U,
+    10U, 1U
+};
+
+static inline uint64_t get_dbits(double d)
+{
+    union {
+        double   dbl;
+        uint64_t i;
+    } dbl_bits = { d };
+
+    return dbl_bits.i;
+}
+
+static Fp build_fp(double d)
+{
+    uint64_t bits = get_dbits(d);
+
+    Fp fp;
+    fp.frac = bits & fracmask;
+    fp.exp = (bits & expmask) >> 52;
+
+    if(fp.exp) {
+        fp.frac += hiddenbit;
+        fp.exp -= expbias;
+
+    } else {
+        fp.exp = -expbias + 1;
+    }
+
+    return fp;
+}
+
+static void normalize(Fp* fp)
+{
+    while ((fp->frac & hiddenbit) == 0) {
+        fp->frac <<= 1;
+        fp->exp--;
+    }
+
+    int shift = 64 - 52 - 1;
+    fp->frac <<= shift;
+    fp->exp -= shift;
+}
+
+static void get_normalized_boundaries(Fp* fp, Fp* lower, Fp* upper)
+{
+    upper->frac = (fp->frac << 1) + 1;
+    upper->exp  = fp->exp - 1;
+
+    while ((upper->frac & (hiddenbit << 1)) == 0) {
+        upper->frac <<= 1;
+        upper->exp--;
+    }
+
+    int u_shift = 64 - 52 - 2;
+
+    upper->frac <<= u_shift;
+    upper->exp = upper->exp - u_shift;
+
+
+    int l_shift = fp->frac == hiddenbit ? 2 : 1;
+
+    lower->frac = (fp->frac << l_shift) - 1;
+    lower->exp = fp->exp - l_shift;
+
+
+    lower->frac <<= lower->exp - upper->exp;
+    lower->exp = upper->exp;
+}
+
+static Fp multiply(Fp* a, Fp* b)
+{
+    const uint64_t lomask = 0x00000000FFFFFFFF;
+
+    uint64_t ah_bl = (a->frac >> 32)    * (b->frac & lomask);
+    uint64_t al_bh = (a->frac & lomask) * (b->frac >> 32);
+    uint64_t al_bl = (a->frac & lomask) * (b->frac & lomask);
+    uint64_t ah_bh = (a->frac >> 32)    * (b->frac >> 32);
+
+    uint64_t tmp = (ah_bl & lomask) + (al_bh & lomask) + (al_bl >> 32);
+    /* round up */
+    tmp += 1U << 31;
+
+    Fp fp = {
+        ah_bh + (ah_bl >> 32) + (al_bh >> 32) + (tmp >> 32),
+        a->exp + b->exp + 64
+    };
+
+    return fp;
+}
+
+static void round_digit(char* digits, int ndigits, uint64_t delta, uint64_t rem, uint64_t kappa, uint64_t frac)
+{
+    while (rem < frac && delta - rem >= kappa &&
+           (rem + kappa < frac || frac - rem > rem + kappa - frac)) {
+
+        digits[ndigits - 1]--;
+        rem += kappa;
+    }
+}
+
+static int generate_digits(Fp* fp, Fp* upper, Fp* lower, char* digits, int* K)
+{
+    uint64_t wfrac = upper->frac - fp->frac;
+    uint64_t delta = upper->frac - lower->frac;
+
+    Fp one;
+    one.frac = 1ULL << -upper->exp;
+    one.exp  = upper->exp;
+
+    uint64_t part1 = upper->frac >> -one.exp;
+    uint64_t part2 = upper->frac & (one.frac - 1);
+
+    int idx = 0, kappa = 10;
+    const uint64_t* divp;
+    /* 1000000000 */
+    for(divp = tens + 10; kappa > 0; divp++) {
+
+        uint64_t div = *divp;
+        unsigned digit = (unsigned) (part1 / div);
+
+        if (digit || idx) {
+            digits[idx++] = digit + '0';
+        }
+
+        part1 -= digit * div;
+        kappa--;
+
+        uint64_t tmp = (part1 <<-one.exp) + part2;
+        if (tmp <= delta) {
+            *K += kappa;
+            round_digit(digits, idx, delta, tmp, div << -one.exp, wfrac);
+
+            return idx;
+        }
+    }
+
+    /* 10 */
+    const uint64_t* unit = tens + 18;
+
+    while(true) {
+        part2 *= 10;
+        delta *= 10;
+        kappa--;
+
+        unsigned digit = (unsigned) (part2 >> -one.exp);
+        if (digit || idx) {
+            digits[idx++] = digit + '0';
+        }
+
+        part2 &= one.frac - 1;
+        if (part2 < delta) {
+            *K += kappa;
+            round_digit(digits, idx, delta, part2, one.frac, wfrac * *unit);
+
+            return idx;
+        }
+
+        unit--;
+    }
+}
+
+static int grisu2(double d, char* digits, int* K)
+{
+    Fp w = build_fp(d);
+
+    Fp lower, upper;
+    get_normalized_boundaries(&w, &lower, &upper);
+
+    normalize(&w);
+
+    int k;
+    Fp cp = find_cachedpow10(upper.exp, &k);
+
+    w     = multiply(&w,     &cp);
+    upper = multiply(&upper, &cp);
+    lower = multiply(&lower, &cp);
+
+    lower.frac++;
+    upper.frac--;
+
+    *K = -k;
+
+    return generate_digits(&w, &upper, &lower, digits, K);
+}
+
+static int emit_digits(char* digits, int ndigits, char* dest, int K, bool neg)
+{
+    int exp = absv(K + ndigits - 1);
+
+    int max_trailing_zeros = 7;
+
+    if(neg) {
+        max_trailing_zeros -= 1;
+    }
+
+    /* write plain integer */
+    if(K >= 0 && (exp < (ndigits + max_trailing_zeros))) {
+
+        memcpy(dest, digits, ndigits);
+        memset(dest + ndigits, '0', K);
+
+        /* add a .0 to mark this as a float. */
+        dest[ndigits + K] = '.';
+        dest[ndigits + K + 1] = '0';
+
+        return ndigits + K + 2;
+    }
+
+    /* write decimal w/o scientific notation */
+    if(K < 0 && (K > -7 || exp < 10)) {
+        int offset = ndigits - absv(K);
+        /* fp < 1.0 -> write leading zero */
+        if(offset <= 0) {
+            offset = -offset;
+            dest[0] = '0';
+            dest[1] = '.';
+            memset(dest + 2, '0', offset);
+            memcpy(dest + offset + 2, digits, ndigits);
+
+            return ndigits + 2 + offset;
+
+        /* fp > 1.0 */
+        } else {
+            memcpy(dest, digits, offset);
+            dest[offset] = '.';
+            memcpy(dest + offset + 1, digits + offset, ndigits - offset);
+
+            return ndigits + 1;
+        }
+    }
+
+    /* write decimal w/ scientific notation */
+    ndigits = minv(ndigits, 18 - neg);
+
+    int idx = 0;
+    dest[idx++] = digits[0];
+
+    if(ndigits > 1) {
+        dest[idx++] = '.';
+        memcpy(dest + idx, digits + 1, ndigits - 1);
+        idx += ndigits - 1;
+    }
+
+    dest[idx++] = 'e';
+
+    char sign = K + ndigits - 1 < 0 ? '-' : '+';
+    dest[idx++] = sign;
+
+    int cent = 0;
+
+    if(exp > 99) {
+        cent = exp / 100;
+        dest[idx++] = cent + '0';
+        exp -= cent * 100;
+    }
+    if(exp > 9) {
+        int dec = exp / 10;
+        dest[idx++] = dec + '0';
+        exp -= dec * 10;
+
+    } else if(cent) {
+        dest[idx++] = '0';
+    }
+
+    dest[idx++] = exp % 10 + '0';
+
+    return idx;
+}
+
+static int filter_special(double fp, char* dest)
+{
+    if(fp == 0.0) {
+        dest[0] = '0';
+        dest[1] = '.';
+        dest[2] = '0';
+        return 3;
+    }
+
+    uint64_t bits = get_dbits(fp);
+
+    bool nan = (bits & expmask) == expmask;
+
+    if(!nan) {
+        return 0;
+    }
+
+    if(bits & fracmask) {
+        dest[0] = 'n'; dest[1] = 'a'; dest[2] = 'n';
+
+    } else {
+        dest[0] = 'i'; dest[1] = 'n'; dest[2] = 'f';
+    }
+
+    return 3;
+}
+
+/* Fast and accurate double to string conversion based on Florian Loitsch's
+ * Grisu-algorithm[1].
+ *
+ * Input:
+ * fp -> the double to convert, dest -> destination buffer.
+ * The generated string will never be longer than 28 characters.
+ * Make sure to pass a pointer to at least 28 bytes of memory.
+ * The emitted string will not be null terminated.
+ *
+ * Output:
+ * The number of written characters.
+ *
+ * Exemplary usage:
+ *
+ * void print(double d)
+ * {
+ *      char buf[28 + 1] // plus null terminator
+ *      int str_len = fpconv_dtoa(d, buf);
+ *
+ *      buf[str_len] = '\0';
+ *      printf("%s", buf);
+ * }
+ *
+ */
+static int fpconv_dtoa(double d, char dest[28])
+{
+    char digits[18];
+
+    int str_len = 0;
+    bool neg = false;
+
+    if(get_dbits(d) & signmask) {
+        dest[0] = '-';
+        str_len++;
+        neg = true;
+    }
+
+    int spec = filter_special(d, dest + str_len);
+
+    if(spec) {
+        return str_len + spec;
+    }
+
+    int K = 0;
+    int ndigits = grisu2(d, digits, &K);
+
+    str_len += emit_digits(digits, ndigits, dest + str_len, K, neg);
+
+    return str_len;
+}