1 // __ _____ _____ _____
2 // __| | __| | | | JSON for Modern C++
3 // | | |__ | | | | | | version 3.11.2
4 // |_____|_____|_____|_|___| https://github.com/nlohmann/json
5 //
6 // SPDX-FileCopyrightText: 2009 Florian Loitsch <https://florian.loitsch.com/>
7 // SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
8 // SPDX-License-Identifier: MIT
9
10 #pragma once
11
12 #include <array> // array
13 #include <cmath> // signbit, isfinite
14 #include <cstdint> // intN_t, uintN_t
15 #include <cstring> // memcpy, memmove
16 #include <limits> // numeric_limits
17 #include <type_traits> // conditional
18
19 #include <nlohmann/detail/macro_scope.hpp>
20
21 NLOHMANN_JSON_NAMESPACE_BEGIN
22 namespace detail
23 {
24
25 /*!
26 @brief implements the Grisu2 algorithm for binary to decimal floating-point
27 conversion.
28
29 This implementation is a slightly modified version of the reference
30 implementation which may be obtained from
31 http://florian.loitsch.com/publications (bench.tar.gz).
32
33 The code is distributed under the MIT license, Copyright (c) 2009 Florian Loitsch.
34
35 For a detailed description of the algorithm see:
36
37 [1] Loitsch, "Printing Floating-Point Numbers Quickly and Accurately with
38 Integers", Proceedings of the ACM SIGPLAN 2010 Conference on Programming
39 Language Design and Implementation, PLDI 2010
40 [2] Burger, Dybvig, "Printing Floating-Point Numbers Quickly and Accurately",
41 Proceedings of the ACM SIGPLAN 1996 Conference on Programming Language
42 Design and Implementation, PLDI 1996
43 */
44 namespace dtoa_impl
45 {
46
47 template<typename Target, typename Source>
reinterpret_bits(const Source source)48 Target reinterpret_bits(const Source source)
49 {
50 static_assert(sizeof(Target) == sizeof(Source), "size mismatch");
51
52 Target target;
53 std::memcpy(&target, &source, sizeof(Source));
54 return target;
55 }
56
57 struct diyfp // f * 2^e
58 {
59 static constexpr int kPrecision = 64; // = q
60
61 std::uint64_t f = 0;
62 int e = 0;
63
diyfpdetail::dtoa_impl::diyfp64 constexpr diyfp(std::uint64_t f_, int e_) noexcept : f(f_), e(e_) {}
65
66 /*!
67 @brief returns x - y
68 @pre x.e == y.e and x.f >= y.f
69 */
subdetail::dtoa_impl::diyfp70 static diyfp sub(const diyfp& x, const diyfp& y) noexcept
71 {
72 JSON_ASSERT(x.e == y.e);
73 JSON_ASSERT(x.f >= y.f);
74
75 return {x.f - y.f, x.e};
76 }
77
78 /*!
79 @brief returns x * y
80 @note The result is rounded. (Only the upper q bits are returned.)
81 */
muldetail::dtoa_impl::diyfp82 static diyfp mul(const diyfp& x, const diyfp& y) noexcept
83 {
84 static_assert(kPrecision == 64, "internal error");
85
86 // Computes:
87 // f = round((x.f * y.f) / 2^q)
88 // e = x.e + y.e + q
89
90 // Emulate the 64-bit * 64-bit multiplication:
91 //
92 // p = u * v
93 // = (u_lo + 2^32 u_hi) (v_lo + 2^32 v_hi)
94 // = (u_lo v_lo ) + 2^32 ((u_lo v_hi ) + (u_hi v_lo )) + 2^64 (u_hi v_hi )
95 // = (p0 ) + 2^32 ((p1 ) + (p2 )) + 2^64 (p3 )
96 // = (p0_lo + 2^32 p0_hi) + 2^32 ((p1_lo + 2^32 p1_hi) + (p2_lo + 2^32 p2_hi)) + 2^64 (p3 )
97 // = (p0_lo ) + 2^32 (p0_hi + p1_lo + p2_lo ) + 2^64 (p1_hi + p2_hi + p3)
98 // = (p0_lo ) + 2^32 (Q ) + 2^64 (H )
99 // = (p0_lo ) + 2^32 (Q_lo + 2^32 Q_hi ) + 2^64 (H )
100 //
101 // (Since Q might be larger than 2^32 - 1)
102 //
103 // = (p0_lo + 2^32 Q_lo) + 2^64 (Q_hi + H)
104 //
105 // (Q_hi + H does not overflow a 64-bit int)
106 //
107 // = p_lo + 2^64 p_hi
108
109 const std::uint64_t u_lo = x.f & 0xFFFFFFFFu;
110 const std::uint64_t u_hi = x.f >> 32u;
111 const std::uint64_t v_lo = y.f & 0xFFFFFFFFu;
112 const std::uint64_t v_hi = y.f >> 32u;
113
114 const std::uint64_t p0 = u_lo * v_lo;
115 const std::uint64_t p1 = u_lo * v_hi;
116 const std::uint64_t p2 = u_hi * v_lo;
117 const std::uint64_t p3 = u_hi * v_hi;
118
119 const std::uint64_t p0_hi = p0 >> 32u;
120 const std::uint64_t p1_lo = p1 & 0xFFFFFFFFu;
121 const std::uint64_t p1_hi = p1 >> 32u;
122 const std::uint64_t p2_lo = p2 & 0xFFFFFFFFu;
123 const std::uint64_t p2_hi = p2 >> 32u;
124
125 std::uint64_t Q = p0_hi + p1_lo + p2_lo;
126
127 // The full product might now be computed as
128 //
129 // p_hi = p3 + p2_hi + p1_hi + (Q >> 32)
130 // p_lo = p0_lo + (Q << 32)
131 //
132 // But in this particular case here, the full p_lo is not required.
133 // Effectively we only need to add the highest bit in p_lo to p_hi (and
134 // Q_hi + 1 does not overflow).
135
136 Q += std::uint64_t{1} << (64u - 32u - 1u); // round, ties up
137
138 const std::uint64_t h = p3 + p2_hi + p1_hi + (Q >> 32u);
139
140 return {h, x.e + y.e + 64};
141 }
142
143 /*!
144 @brief normalize x such that the significand is >= 2^(q-1)
145 @pre x.f != 0
146 */
normalizedetail::dtoa_impl::diyfp147 static diyfp normalize(diyfp x) noexcept
148 {
149 JSON_ASSERT(x.f != 0);
150
151 while ((x.f >> 63u) == 0)
152 {
153 x.f <<= 1u;
154 x.e--;
155 }
156
157 return x;
158 }
159
160 /*!
161 @brief normalize x such that the result has the exponent E
162 @pre e >= x.e and the upper e - x.e bits of x.f must be zero.
163 */
normalize_todetail::dtoa_impl::diyfp164 static diyfp normalize_to(const diyfp& x, const int target_exponent) noexcept
165 {
166 const int delta = x.e - target_exponent;
167
168 JSON_ASSERT(delta >= 0);
169 JSON_ASSERT(((x.f << delta) >> delta) == x.f);
170
171 return {x.f << delta, target_exponent};
172 }
173 };
174
175 struct boundaries
176 {
177 diyfp w;
178 diyfp minus;
179 diyfp plus;
180 };
181
182 /*!
183 Compute the (normalized) diyfp representing the input number 'value' and its
184 boundaries.
185
186 @pre value must be finite and positive
187 */
188 template<typename FloatType>
compute_boundaries(FloatType value)189 boundaries compute_boundaries(FloatType value)
190 {
191 JSON_ASSERT(std::isfinite(value));
192 JSON_ASSERT(value > 0);
193
194 // Convert the IEEE representation into a diyfp.
195 //
196 // If v is denormal:
197 // value = 0.F * 2^(1 - bias) = ( F) * 2^(1 - bias - (p-1))
198 // If v is normalized:
199 // value = 1.F * 2^(E - bias) = (2^(p-1) + F) * 2^(E - bias - (p-1))
200
201 static_assert(std::numeric_limits<FloatType>::is_iec559,
202 "internal error: dtoa_short requires an IEEE-754 floating-point implementation");
203
204 constexpr int kPrecision = std::numeric_limits<FloatType>::digits; // = p (includes the hidden bit)
205 constexpr int kBias = std::numeric_limits<FloatType>::max_exponent - 1 + (kPrecision - 1);
206 constexpr int kMinExp = 1 - kBias;
207 constexpr std::uint64_t kHiddenBit = std::uint64_t{1} << (kPrecision - 1); // = 2^(p-1)
208
209 using bits_type = typename std::conditional<kPrecision == 24, std::uint32_t, std::uint64_t >::type;
210
211 const auto bits = static_cast<std::uint64_t>(reinterpret_bits<bits_type>(value));
212 const std::uint64_t E = bits >> (kPrecision - 1);
213 const std::uint64_t F = bits & (kHiddenBit - 1);
214
215 const bool is_denormal = E == 0;
216 const diyfp v = is_denormal
217 ? diyfp(F, kMinExp)
218 : diyfp(F + kHiddenBit, static_cast<int>(E) - kBias);
219
220 // Compute the boundaries m- and m+ of the floating-point value
221 // v = f * 2^e.
222 //
223 // Determine v- and v+, the floating-point predecessor and successor if v,
224 // respectively.
225 //
226 // v- = v - 2^e if f != 2^(p-1) or e == e_min (A)
227 // = v - 2^(e-1) if f == 2^(p-1) and e > e_min (B)
228 //
229 // v+ = v + 2^e
230 //
231 // Let m- = (v- + v) / 2 and m+ = (v + v+) / 2. All real numbers _strictly_
232 // between m- and m+ round to v, regardless of how the input rounding
233 // algorithm breaks ties.
234 //
235 // ---+-------------+-------------+-------------+-------------+--- (A)
236 // v- m- v m+ v+
237 //
238 // -----------------+------+------+-------------+-------------+--- (B)
239 // v- m- v m+ v+
240
241 const bool lower_boundary_is_closer = F == 0 && E > 1;
242 const diyfp m_plus = diyfp(2 * v.f + 1, v.e - 1);
243 const diyfp m_minus = lower_boundary_is_closer
244 ? diyfp(4 * v.f - 1, v.e - 2) // (B)
245 : diyfp(2 * v.f - 1, v.e - 1); // (A)
246
247 // Determine the normalized w+ = m+.
248 const diyfp w_plus = diyfp::normalize(m_plus);
249
250 // Determine w- = m- such that e_(w-) = e_(w+).
251 const diyfp w_minus = diyfp::normalize_to(m_minus, w_plus.e);
252
253 return {diyfp::normalize(v), w_minus, w_plus};
254 }
255
256 // Given normalized diyfp w, Grisu needs to find a (normalized) cached
257 // power-of-ten c, such that the exponent of the product c * w = f * 2^e lies
258 // within a certain range [alpha, gamma] (Definition 3.2 from [1])
259 //
260 // alpha <= e = e_c + e_w + q <= gamma
261 //
262 // or
263 //
264 // f_c * f_w * 2^alpha <= f_c 2^(e_c) * f_w 2^(e_w) * 2^q
265 // <= f_c * f_w * 2^gamma
266 //
267 // Since c and w are normalized, i.e. 2^(q-1) <= f < 2^q, this implies
268 //
269 // 2^(q-1) * 2^(q-1) * 2^alpha <= c * w * 2^q < 2^q * 2^q * 2^gamma
270 //
271 // or
272 //
273 // 2^(q - 2 + alpha) <= c * w < 2^(q + gamma)
274 //
275 // The choice of (alpha,gamma) determines the size of the table and the form of
276 // the digit generation procedure. Using (alpha,gamma)=(-60,-32) works out well
277 // in practice:
278 //
279 // The idea is to cut the number c * w = f * 2^e into two parts, which can be
280 // processed independently: An integral part p1, and a fractional part p2:
281 //
282 // f * 2^e = ( (f div 2^-e) * 2^-e + (f mod 2^-e) ) * 2^e
283 // = (f div 2^-e) + (f mod 2^-e) * 2^e
284 // = p1 + p2 * 2^e
285 //
286 // The conversion of p1 into decimal form requires a series of divisions and
287 // modulos by (a power of) 10. These operations are faster for 32-bit than for
288 // 64-bit integers, so p1 should ideally fit into a 32-bit integer. This can be
289 // achieved by choosing
290 //
291 // -e >= 32 or e <= -32 := gamma
292 //
293 // In order to convert the fractional part
294 //
295 // p2 * 2^e = p2 / 2^-e = d[-1] / 10^1 + d[-2] / 10^2 + ...
296 //
297 // into decimal form, the fraction is repeatedly multiplied by 10 and the digits
298 // d[-i] are extracted in order:
299 //
300 // (10 * p2) div 2^-e = d[-1]
301 // (10 * p2) mod 2^-e = d[-2] / 10^1 + ...
302 //
303 // The multiplication by 10 must not overflow. It is sufficient to choose
304 //
305 // 10 * p2 < 16 * p2 = 2^4 * p2 <= 2^64.
306 //
307 // Since p2 = f mod 2^-e < 2^-e,
308 //
309 // -e <= 60 or e >= -60 := alpha
310
311 constexpr int kAlpha = -60;
312 constexpr int kGamma = -32;
313
314 struct cached_power // c = f * 2^e ~= 10^k
315 {
316 std::uint64_t f;
317 int e;
318 int k;
319 };
320
321 /*!
322 For a normalized diyfp w = f * 2^e, this function returns a (normalized) cached
323 power-of-ten c = f_c * 2^e_c, such that the exponent of the product w * c
324 satisfies (Definition 3.2 from [1])
325
326 alpha <= e_c + e + q <= gamma.
327 */
get_cached_power_for_binary_exponent(int e)328 inline cached_power get_cached_power_for_binary_exponent(int e)
329 {
330 // Now
331 //
332 // alpha <= e_c + e + q <= gamma (1)
333 // ==> f_c * 2^alpha <= c * 2^e * 2^q
334 //
335 // and since the c's are normalized, 2^(q-1) <= f_c,
336 //
337 // ==> 2^(q - 1 + alpha) <= c * 2^(e + q)
338 // ==> 2^(alpha - e - 1) <= c
339 //
340 // If c were an exact power of ten, i.e. c = 10^k, one may determine k as
341 //
342 // k = ceil( log_10( 2^(alpha - e - 1) ) )
343 // = ceil( (alpha - e - 1) * log_10(2) )
344 //
345 // From the paper:
346 // "In theory the result of the procedure could be wrong since c is rounded,
347 // and the computation itself is approximated [...]. In practice, however,
348 // this simple function is sufficient."
349 //
350 // For IEEE double precision floating-point numbers converted into
351 // normalized diyfp's w = f * 2^e, with q = 64,
352 //
353 // e >= -1022 (min IEEE exponent)
354 // -52 (p - 1)
355 // -52 (p - 1, possibly normalize denormal IEEE numbers)
356 // -11 (normalize the diyfp)
357 // = -1137
358 //
359 // and
360 //
361 // e <= +1023 (max IEEE exponent)
362 // -52 (p - 1)
363 // -11 (normalize the diyfp)
364 // = 960
365 //
366 // This binary exponent range [-1137,960] results in a decimal exponent
367 // range [-307,324]. One does not need to store a cached power for each
368 // k in this range. For each such k it suffices to find a cached power
369 // such that the exponent of the product lies in [alpha,gamma].
370 // This implies that the difference of the decimal exponents of adjacent
371 // table entries must be less than or equal to
372 //
373 // floor( (gamma - alpha) * log_10(2) ) = 8.
374 //
375 // (A smaller distance gamma-alpha would require a larger table.)
376
377 // NB:
378 // Actually this function returns c, such that -60 <= e_c + e + 64 <= -34.
379
380 constexpr int kCachedPowersMinDecExp = -300;
381 constexpr int kCachedPowersDecStep = 8;
382
383 static constexpr std::array<cached_power, 79> kCachedPowers =
384 {
385 {
386 { 0xAB70FE17C79AC6CA, -1060, -300 },
387 { 0xFF77B1FCBEBCDC4F, -1034, -292 },
388 { 0xBE5691EF416BD60C, -1007, -284 },
389 { 0x8DD01FAD907FFC3C, -980, -276 },
390 { 0xD3515C2831559A83, -954, -268 },
391 { 0x9D71AC8FADA6C9B5, -927, -260 },
392 { 0xEA9C227723EE8BCB, -901, -252 },
393 { 0xAECC49914078536D, -874, -244 },
394 { 0x823C12795DB6CE57, -847, -236 },
395 { 0xC21094364DFB5637, -821, -228 },
396 { 0x9096EA6F3848984F, -794, -220 },
397 { 0xD77485CB25823AC7, -768, -212 },
398 { 0xA086CFCD97BF97F4, -741, -204 },
399 { 0xEF340A98172AACE5, -715, -196 },
400 { 0xB23867FB2A35B28E, -688, -188 },
401 { 0x84C8D4DFD2C63F3B, -661, -180 },
402 { 0xC5DD44271AD3CDBA, -635, -172 },
403 { 0x936B9FCEBB25C996, -608, -164 },
404 { 0xDBAC6C247D62A584, -582, -156 },
405 { 0xA3AB66580D5FDAF6, -555, -148 },
406 { 0xF3E2F893DEC3F126, -529, -140 },
407 { 0xB5B5ADA8AAFF80B8, -502, -132 },
408 { 0x87625F056C7C4A8B, -475, -124 },
409 { 0xC9BCFF6034C13053, -449, -116 },
410 { 0x964E858C91BA2655, -422, -108 },
411 { 0xDFF9772470297EBD, -396, -100 },
412 { 0xA6DFBD9FB8E5B88F, -369, -92 },
413 { 0xF8A95FCF88747D94, -343, -84 },
414 { 0xB94470938FA89BCF, -316, -76 },
415 { 0x8A08F0F8BF0F156B, -289, -68 },
416 { 0xCDB02555653131B6, -263, -60 },
417 { 0x993FE2C6D07B7FAC, -236, -52 },
418 { 0xE45C10C42A2B3B06, -210, -44 },
419 { 0xAA242499697392D3, -183, -36 },
420 { 0xFD87B5F28300CA0E, -157, -28 },
421 { 0xBCE5086492111AEB, -130, -20 },
422 { 0x8CBCCC096F5088CC, -103, -12 },
423 { 0xD1B71758E219652C, -77, -4 },
424 { 0x9C40000000000000, -50, 4 },
425 { 0xE8D4A51000000000, -24, 12 },
426 { 0xAD78EBC5AC620000, 3, 20 },
427 { 0x813F3978F8940984, 30, 28 },
428 { 0xC097CE7BC90715B3, 56, 36 },
429 { 0x8F7E32CE7BEA5C70, 83, 44 },
430 { 0xD5D238A4ABE98068, 109, 52 },
431 { 0x9F4F2726179A2245, 136, 60 },
432 { 0xED63A231D4C4FB27, 162, 68 },
433 { 0xB0DE65388CC8ADA8, 189, 76 },
434 { 0x83C7088E1AAB65DB, 216, 84 },
435 { 0xC45D1DF942711D9A, 242, 92 },
436 { 0x924D692CA61BE758, 269, 100 },
437 { 0xDA01EE641A708DEA, 295, 108 },
438 { 0xA26DA3999AEF774A, 322, 116 },
439 { 0xF209787BB47D6B85, 348, 124 },
440 { 0xB454E4A179DD1877, 375, 132 },
441 { 0x865B86925B9BC5C2, 402, 140 },
442 { 0xC83553C5C8965D3D, 428, 148 },
443 { 0x952AB45CFA97A0B3, 455, 156 },
444 { 0xDE469FBD99A05FE3, 481, 164 },
445 { 0xA59BC234DB398C25, 508, 172 },
446 { 0xF6C69A72A3989F5C, 534, 180 },
447 { 0xB7DCBF5354E9BECE, 561, 188 },
448 { 0x88FCF317F22241E2, 588, 196 },
449 { 0xCC20CE9BD35C78A5, 614, 204 },
450 { 0x98165AF37B2153DF, 641, 212 },
451 { 0xE2A0B5DC971F303A, 667, 220 },
452 { 0xA8D9D1535CE3B396, 694, 228 },
453 { 0xFB9B7CD9A4A7443C, 720, 236 },
454 { 0xBB764C4CA7A44410, 747, 244 },
455 { 0x8BAB8EEFB6409C1A, 774, 252 },
456 { 0xD01FEF10A657842C, 800, 260 },
457 { 0x9B10A4E5E9913129, 827, 268 },
458 { 0xE7109BFBA19C0C9D, 853, 276 },
459 { 0xAC2820D9623BF429, 880, 284 },
460 { 0x80444B5E7AA7CF85, 907, 292 },
461 { 0xBF21E44003ACDD2D, 933, 300 },
462 { 0x8E679C2F5E44FF8F, 960, 308 },
463 { 0xD433179D9C8CB841, 986, 316 },
464 { 0x9E19DB92B4E31BA9, 1013, 324 },
465 }
466 };
467
468 // This computation gives exactly the same results for k as
469 // k = ceil((kAlpha - e - 1) * 0.30102999566398114)
470 // for |e| <= 1500, but doesn't require floating-point operations.
471 // NB: log_10(2) ~= 78913 / 2^18
472 JSON_ASSERT(e >= -1500);
473 JSON_ASSERT(e <= 1500);
474 const int f = kAlpha - e - 1;
475 const int k = (f * 78913) / (1 << 18) + static_cast<int>(f > 0);
476
477 const int index = (-kCachedPowersMinDecExp + k + (kCachedPowersDecStep - 1)) / kCachedPowersDecStep;
478 JSON_ASSERT(index >= 0);
479 JSON_ASSERT(static_cast<std::size_t>(index) < kCachedPowers.size());
480
481 const cached_power cached = kCachedPowers[static_cast<std::size_t>(index)];
482 JSON_ASSERT(kAlpha <= cached.e + e + 64);
483 JSON_ASSERT(kGamma >= cached.e + e + 64);
484
485 return cached;
486 }
487
488 /*!
489 For n != 0, returns k, such that pow10 := 10^(k-1) <= n < 10^k.
490 For n == 0, returns 1 and sets pow10 := 1.
491 */
find_largest_pow10(const std::uint32_t n,std::uint32_t & pow10)492 inline int find_largest_pow10(const std::uint32_t n, std::uint32_t& pow10)
493 {
494 // LCOV_EXCL_START
495 if (n >= 1000000000)
496 {
497 pow10 = 1000000000;
498 return 10;
499 }
500 // LCOV_EXCL_STOP
501 if (n >= 100000000)
502 {
503 pow10 = 100000000;
504 return 9;
505 }
506 if (n >= 10000000)
507 {
508 pow10 = 10000000;
509 return 8;
510 }
511 if (n >= 1000000)
512 {
513 pow10 = 1000000;
514 return 7;
515 }
516 if (n >= 100000)
517 {
518 pow10 = 100000;
519 return 6;
520 }
521 if (n >= 10000)
522 {
523 pow10 = 10000;
524 return 5;
525 }
526 if (n >= 1000)
527 {
528 pow10 = 1000;
529 return 4;
530 }
531 if (n >= 100)
532 {
533 pow10 = 100;
534 return 3;
535 }
536 if (n >= 10)
537 {
538 pow10 = 10;
539 return 2;
540 }
541
542 pow10 = 1;
543 return 1;
544 }
545
grisu2_round(char * buf,int len,std::uint64_t dist,std::uint64_t delta,std::uint64_t rest,std::uint64_t ten_k)546 inline void grisu2_round(char* buf, int len, std::uint64_t dist, std::uint64_t delta,
547 std::uint64_t rest, std::uint64_t ten_k)
548 {
549 JSON_ASSERT(len >= 1);
550 JSON_ASSERT(dist <= delta);
551 JSON_ASSERT(rest <= delta);
552 JSON_ASSERT(ten_k > 0);
553
554 // <--------------------------- delta ---->
555 // <---- dist --------->
556 // --------------[------------------+-------------------]--------------
557 // M- w M+
558 //
559 // ten_k
560 // <------>
561 // <---- rest ---->
562 // --------------[------------------+----+--------------]--------------
563 // w V
564 // = buf * 10^k
565 //
566 // ten_k represents a unit-in-the-last-place in the decimal representation
567 // stored in buf.
568 // Decrement buf by ten_k while this takes buf closer to w.
569
570 // The tests are written in this order to avoid overflow in unsigned
571 // integer arithmetic.
572
573 while (rest < dist
574 && delta - rest >= ten_k
575 && (rest + ten_k < dist || dist - rest > rest + ten_k - dist))
576 {
577 JSON_ASSERT(buf[len - 1] != '0');
578 buf[len - 1]--;
579 rest += ten_k;
580 }
581 }
582
583 /*!
584 Generates V = buffer * 10^decimal_exponent, such that M- <= V <= M+.
585 M- and M+ must be normalized and share the same exponent -60 <= e <= -32.
586 */
grisu2_digit_gen(char * buffer,int & length,int & decimal_exponent,diyfp M_minus,diyfp w,diyfp M_plus)587 inline void grisu2_digit_gen(char* buffer, int& length, int& decimal_exponent,
588 diyfp M_minus, diyfp w, diyfp M_plus)
589 {
590 static_assert(kAlpha >= -60, "internal error");
591 static_assert(kGamma <= -32, "internal error");
592
593 // Generates the digits (and the exponent) of a decimal floating-point
594 // number V = buffer * 10^decimal_exponent in the range [M-, M+]. The diyfp's
595 // w, M- and M+ share the same exponent e, which satisfies alpha <= e <= gamma.
596 //
597 // <--------------------------- delta ---->
598 // <---- dist --------->
599 // --------------[------------------+-------------------]--------------
600 // M- w M+
601 //
602 // Grisu2 generates the digits of M+ from left to right and stops as soon as
603 // V is in [M-,M+].
604
605 JSON_ASSERT(M_plus.e >= kAlpha);
606 JSON_ASSERT(M_plus.e <= kGamma);
607
608 std::uint64_t delta = diyfp::sub(M_plus, M_minus).f; // (significand of (M+ - M-), implicit exponent is e)
609 std::uint64_t dist = diyfp::sub(M_plus, w ).f; // (significand of (M+ - w ), implicit exponent is e)
610
611 // Split M+ = f * 2^e into two parts p1 and p2 (note: e < 0):
612 //
613 // M+ = f * 2^e
614 // = ((f div 2^-e) * 2^-e + (f mod 2^-e)) * 2^e
615 // = ((p1 ) * 2^-e + (p2 )) * 2^e
616 // = p1 + p2 * 2^e
617
618 const diyfp one(std::uint64_t{1} << -M_plus.e, M_plus.e);
619
620 auto p1 = static_cast<std::uint32_t>(M_plus.f >> -one.e); // p1 = f div 2^-e (Since -e >= 32, p1 fits into a 32-bit int.)
621 std::uint64_t p2 = M_plus.f & (one.f - 1); // p2 = f mod 2^-e
622
623 // 1)
624 //
625 // Generate the digits of the integral part p1 = d[n-1]...d[1]d[0]
626
627 JSON_ASSERT(p1 > 0);
628
629 std::uint32_t pow10{};
630 const int k = find_largest_pow10(p1, pow10);
631
632 // 10^(k-1) <= p1 < 10^k, pow10 = 10^(k-1)
633 //
634 // p1 = (p1 div 10^(k-1)) * 10^(k-1) + (p1 mod 10^(k-1))
635 // = (d[k-1] ) * 10^(k-1) + (p1 mod 10^(k-1))
636 //
637 // M+ = p1 + p2 * 2^e
638 // = d[k-1] * 10^(k-1) + (p1 mod 10^(k-1)) + p2 * 2^e
639 // = d[k-1] * 10^(k-1) + ((p1 mod 10^(k-1)) * 2^-e + p2) * 2^e
640 // = d[k-1] * 10^(k-1) + ( rest) * 2^e
641 //
642 // Now generate the digits d[n] of p1 from left to right (n = k-1,...,0)
643 //
644 // p1 = d[k-1]...d[n] * 10^n + d[n-1]...d[0]
645 //
646 // but stop as soon as
647 //
648 // rest * 2^e = (d[n-1]...d[0] * 2^-e + p2) * 2^e <= delta * 2^e
649
650 int n = k;
651 while (n > 0)
652 {
653 // Invariants:
654 // M+ = buffer * 10^n + (p1 + p2 * 2^e) (buffer = 0 for n = k)
655 // pow10 = 10^(n-1) <= p1 < 10^n
656 //
657 const std::uint32_t d = p1 / pow10; // d = p1 div 10^(n-1)
658 const std::uint32_t r = p1 % pow10; // r = p1 mod 10^(n-1)
659 //
660 // M+ = buffer * 10^n + (d * 10^(n-1) + r) + p2 * 2^e
661 // = (buffer * 10 + d) * 10^(n-1) + (r + p2 * 2^e)
662 //
663 JSON_ASSERT(d <= 9);
664 buffer[length++] = static_cast<char>('0' + d); // buffer := buffer * 10 + d
665 //
666 // M+ = buffer * 10^(n-1) + (r + p2 * 2^e)
667 //
668 p1 = r;
669 n--;
670 //
671 // M+ = buffer * 10^n + (p1 + p2 * 2^e)
672 // pow10 = 10^n
673 //
674
675 // Now check if enough digits have been generated.
676 // Compute
677 //
678 // p1 + p2 * 2^e = (p1 * 2^-e + p2) * 2^e = rest * 2^e
679 //
680 // Note:
681 // Since rest and delta share the same exponent e, it suffices to
682 // compare the significands.
683 const std::uint64_t rest = (std::uint64_t{p1} << -one.e) + p2;
684 if (rest <= delta)
685 {
686 // V = buffer * 10^n, with M- <= V <= M+.
687
688 decimal_exponent += n;
689
690 // We may now just stop. But instead look if the buffer could be
691 // decremented to bring V closer to w.
692 //
693 // pow10 = 10^n is now 1 ulp in the decimal representation V.
694 // The rounding procedure works with diyfp's with an implicit
695 // exponent of e.
696 //
697 // 10^n = (10^n * 2^-e) * 2^e = ulp * 2^e
698 //
699 const std::uint64_t ten_n = std::uint64_t{pow10} << -one.e;
700 grisu2_round(buffer, length, dist, delta, rest, ten_n);
701
702 return;
703 }
704
705 pow10 /= 10;
706 //
707 // pow10 = 10^(n-1) <= p1 < 10^n
708 // Invariants restored.
709 }
710
711 // 2)
712 //
713 // The digits of the integral part have been generated:
714 //
715 // M+ = d[k-1]...d[1]d[0] + p2 * 2^e
716 // = buffer + p2 * 2^e
717 //
718 // Now generate the digits of the fractional part p2 * 2^e.
719 //
720 // Note:
721 // No decimal point is generated: the exponent is adjusted instead.
722 //
723 // p2 actually represents the fraction
724 //
725 // p2 * 2^e
726 // = p2 / 2^-e
727 // = d[-1] / 10^1 + d[-2] / 10^2 + ...
728 //
729 // Now generate the digits d[-m] of p1 from left to right (m = 1,2,...)
730 //
731 // p2 * 2^e = d[-1]d[-2]...d[-m] * 10^-m
732 // + 10^-m * (d[-m-1] / 10^1 + d[-m-2] / 10^2 + ...)
733 //
734 // using
735 //
736 // 10^m * p2 = ((10^m * p2) div 2^-e) * 2^-e + ((10^m * p2) mod 2^-e)
737 // = ( d) * 2^-e + ( r)
738 //
739 // or
740 // 10^m * p2 * 2^e = d + r * 2^e
741 //
742 // i.e.
743 //
744 // M+ = buffer + p2 * 2^e
745 // = buffer + 10^-m * (d + r * 2^e)
746 // = (buffer * 10^m + d) * 10^-m + 10^-m * r * 2^e
747 //
748 // and stop as soon as 10^-m * r * 2^e <= delta * 2^e
749
750 JSON_ASSERT(p2 > delta);
751
752 int m = 0;
753 for (;;)
754 {
755 // Invariant:
756 // M+ = buffer * 10^-m + 10^-m * (d[-m-1] / 10 + d[-m-2] / 10^2 + ...) * 2^e
757 // = buffer * 10^-m + 10^-m * (p2 ) * 2^e
758 // = buffer * 10^-m + 10^-m * (1/10 * (10 * p2) ) * 2^e
759 // = buffer * 10^-m + 10^-m * (1/10 * ((10*p2 div 2^-e) * 2^-e + (10*p2 mod 2^-e)) * 2^e
760 //
761 JSON_ASSERT(p2 <= (std::numeric_limits<std::uint64_t>::max)() / 10);
762 p2 *= 10;
763 const std::uint64_t d = p2 >> -one.e; // d = (10 * p2) div 2^-e
764 const std::uint64_t r = p2 & (one.f - 1); // r = (10 * p2) mod 2^-e
765 //
766 // M+ = buffer * 10^-m + 10^-m * (1/10 * (d * 2^-e + r) * 2^e
767 // = buffer * 10^-m + 10^-m * (1/10 * (d + r * 2^e))
768 // = (buffer * 10 + d) * 10^(-m-1) + 10^(-m-1) * r * 2^e
769 //
770 JSON_ASSERT(d <= 9);
771 buffer[length++] = static_cast<char>('0' + d); // buffer := buffer * 10 + d
772 //
773 // M+ = buffer * 10^(-m-1) + 10^(-m-1) * r * 2^e
774 //
775 p2 = r;
776 m++;
777 //
778 // M+ = buffer * 10^-m + 10^-m * p2 * 2^e
779 // Invariant restored.
780
781 // Check if enough digits have been generated.
782 //
783 // 10^-m * p2 * 2^e <= delta * 2^e
784 // p2 * 2^e <= 10^m * delta * 2^e
785 // p2 <= 10^m * delta
786 delta *= 10;
787 dist *= 10;
788 if (p2 <= delta)
789 {
790 break;
791 }
792 }
793
794 // V = buffer * 10^-m, with M- <= V <= M+.
795
796 decimal_exponent -= m;
797
798 // 1 ulp in the decimal representation is now 10^-m.
799 // Since delta and dist are now scaled by 10^m, we need to do the
800 // same with ulp in order to keep the units in sync.
801 //
802 // 10^m * 10^-m = 1 = 2^-e * 2^e = ten_m * 2^e
803 //
804 const std::uint64_t ten_m = one.f;
805 grisu2_round(buffer, length, dist, delta, p2, ten_m);
806
807 // By construction this algorithm generates the shortest possible decimal
808 // number (Loitsch, Theorem 6.2) which rounds back to w.
809 // For an input number of precision p, at least
810 //
811 // N = 1 + ceil(p * log_10(2))
812 //
813 // decimal digits are sufficient to identify all binary floating-point
814 // numbers (Matula, "In-and-Out conversions").
815 // This implies that the algorithm does not produce more than N decimal
816 // digits.
817 //
818 // N = 17 for p = 53 (IEEE double precision)
819 // N = 9 for p = 24 (IEEE single precision)
820 }
821
822 /*!
823 v = buf * 10^decimal_exponent
824 len is the length of the buffer (number of decimal digits)
825 The buffer must be large enough, i.e. >= max_digits10.
826 */
827 JSON_HEDLEY_NON_NULL(1)
grisu2(char * buf,int & len,int & decimal_exponent,diyfp m_minus,diyfp v,diyfp m_plus)828 inline void grisu2(char* buf, int& len, int& decimal_exponent,
829 diyfp m_minus, diyfp v, diyfp m_plus)
830 {
831 JSON_ASSERT(m_plus.e == m_minus.e);
832 JSON_ASSERT(m_plus.e == v.e);
833
834 // --------(-----------------------+-----------------------)-------- (A)
835 // m- v m+
836 //
837 // --------------------(-----------+-----------------------)-------- (B)
838 // m- v m+
839 //
840 // First scale v (and m- and m+) such that the exponent is in the range
841 // [alpha, gamma].
842
843 const cached_power cached = get_cached_power_for_binary_exponent(m_plus.e);
844
845 const diyfp c_minus_k(cached.f, cached.e); // = c ~= 10^-k
846
847 // The exponent of the products is = v.e + c_minus_k.e + q and is in the range [alpha,gamma]
848 const diyfp w = diyfp::mul(v, c_minus_k);
849 const diyfp w_minus = diyfp::mul(m_minus, c_minus_k);
850 const diyfp w_plus = diyfp::mul(m_plus, c_minus_k);
851
852 // ----(---+---)---------------(---+---)---------------(---+---)----
853 // w- w w+
854 // = c*m- = c*v = c*m+
855 //
856 // diyfp::mul rounds its result and c_minus_k is approximated too. w, w- and
857 // w+ are now off by a small amount.
858 // In fact:
859 //
860 // w - v * 10^k < 1 ulp
861 //
862 // To account for this inaccuracy, add resp. subtract 1 ulp.
863 //
864 // --------+---[---------------(---+---)---------------]---+--------
865 // w- M- w M+ w+
866 //
867 // Now any number in [M-, M+] (bounds included) will round to w when input,
868 // regardless of how the input rounding algorithm breaks ties.
869 //
870 // And digit_gen generates the shortest possible such number in [M-, M+].
871 // Note that this does not mean that Grisu2 always generates the shortest
872 // possible number in the interval (m-, m+).
873 const diyfp M_minus(w_minus.f + 1, w_minus.e);
874 const diyfp M_plus (w_plus.f - 1, w_plus.e );
875
876 decimal_exponent = -cached.k; // = -(-k) = k
877
878 grisu2_digit_gen(buf, len, decimal_exponent, M_minus, w, M_plus);
879 }
880
881 /*!
882 v = buf * 10^decimal_exponent
883 len is the length of the buffer (number of decimal digits)
884 The buffer must be large enough, i.e. >= max_digits10.
885 */
886 template<typename FloatType>
887 JSON_HEDLEY_NON_NULL(1)
grisu2(char * buf,int & len,int & decimal_exponent,FloatType value)888 void grisu2(char* buf, int& len, int& decimal_exponent, FloatType value)
889 {
890 static_assert(diyfp::kPrecision >= std::numeric_limits<FloatType>::digits + 3,
891 "internal error: not enough precision");
892
893 JSON_ASSERT(std::isfinite(value));
894 JSON_ASSERT(value > 0);
895
896 // If the neighbors (and boundaries) of 'value' are always computed for double-precision
897 // numbers, all float's can be recovered using strtod (and strtof). However, the resulting
898 // decimal representations are not exactly "short".
899 //
900 // The documentation for 'std::to_chars' (https://en.cppreference.com/w/cpp/utility/to_chars)
901 // says "value is converted to a string as if by std::sprintf in the default ("C") locale"
902 // and since sprintf promotes floats to doubles, I think this is exactly what 'std::to_chars'
903 // does.
904 // On the other hand, the documentation for 'std::to_chars' requires that "parsing the
905 // representation using the corresponding std::from_chars function recovers value exactly". That
906 // indicates that single precision floating-point numbers should be recovered using
907 // 'std::strtof'.
908 //
909 // NB: If the neighbors are computed for single-precision numbers, there is a single float
910 // (7.0385307e-26f) which can't be recovered using strtod. The resulting double precision
911 // value is off by 1 ulp.
912 #if 0
913 const boundaries w = compute_boundaries(static_cast<double>(value));
914 #else
915 const boundaries w = compute_boundaries(value);
916 #endif
917
918 grisu2(buf, len, decimal_exponent, w.minus, w.w, w.plus);
919 }
920
921 /*!
922 @brief appends a decimal representation of e to buf
923 @return a pointer to the element following the exponent.
924 @pre -1000 < e < 1000
925 */
926 JSON_HEDLEY_NON_NULL(1)
927 JSON_HEDLEY_RETURNS_NON_NULL
append_exponent(char * buf,int e)928 inline char* append_exponent(char* buf, int e)
929 {
930 JSON_ASSERT(e > -1000);
931 JSON_ASSERT(e < 1000);
932
933 if (e < 0)
934 {
935 e = -e;
936 *buf++ = '-';
937 }
938 else
939 {
940 *buf++ = '+';
941 }
942
943 auto k = static_cast<std::uint32_t>(e);
944 if (k < 10)
945 {
946 // Always print at least two digits in the exponent.
947 // This is for compatibility with printf("%g").
948 *buf++ = '0';
949 *buf++ = static_cast<char>('0' + k);
950 }
951 else if (k < 100)
952 {
953 *buf++ = static_cast<char>('0' + k / 10);
954 k %= 10;
955 *buf++ = static_cast<char>('0' + k);
956 }
957 else
958 {
959 *buf++ = static_cast<char>('0' + k / 100);
960 k %= 100;
961 *buf++ = static_cast<char>('0' + k / 10);
962 k %= 10;
963 *buf++ = static_cast<char>('0' + k);
964 }
965
966 return buf;
967 }
968
969 /*!
970 @brief prettify v = buf * 10^decimal_exponent
971
972 If v is in the range [10^min_exp, 10^max_exp) it will be printed in fixed-point
973 notation. Otherwise it will be printed in exponential notation.
974
975 @pre min_exp < 0
976 @pre max_exp > 0
977 */
978 JSON_HEDLEY_NON_NULL(1)
979 JSON_HEDLEY_RETURNS_NON_NULL
format_buffer(char * buf,int len,int decimal_exponent,int min_exp,int max_exp)980 inline char* format_buffer(char* buf, int len, int decimal_exponent,
981 int min_exp, int max_exp)
982 {
983 JSON_ASSERT(min_exp < 0);
984 JSON_ASSERT(max_exp > 0);
985
986 const int k = len;
987 const int n = len + decimal_exponent;
988
989 // v = buf * 10^(n-k)
990 // k is the length of the buffer (number of decimal digits)
991 // n is the position of the decimal point relative to the start of the buffer.
992
993 if (k <= n && n <= max_exp)
994 {
995 // digits[000]
996 // len <= max_exp + 2
997
998 std::memset(buf + k, '0', static_cast<size_t>(n) - static_cast<size_t>(k));
999 // Make it look like a floating-point number (#362, #378)
1000 buf[n + 0] = '.';
1001 buf[n + 1] = '0';
1002 return buf + (static_cast<size_t>(n) + 2);
1003 }
1004
1005 if (0 < n && n <= max_exp)
1006 {
1007 // dig.its
1008 // len <= max_digits10 + 1
1009
1010 JSON_ASSERT(k > n);
1011
1012 std::memmove(buf + (static_cast<size_t>(n) + 1), buf + n, static_cast<size_t>(k) - static_cast<size_t>(n));
1013 buf[n] = '.';
1014 return buf + (static_cast<size_t>(k) + 1U);
1015 }
1016
1017 if (min_exp < n && n <= 0)
1018 {
1019 // 0.[000]digits
1020 // len <= 2 + (-min_exp - 1) + max_digits10
1021
1022 std::memmove(buf + (2 + static_cast<size_t>(-n)), buf, static_cast<size_t>(k));
1023 buf[0] = '0';
1024 buf[1] = '.';
1025 std::memset(buf + 2, '0', static_cast<size_t>(-n));
1026 return buf + (2U + static_cast<size_t>(-n) + static_cast<size_t>(k));
1027 }
1028
1029 if (k == 1)
1030 {
1031 // dE+123
1032 // len <= 1 + 5
1033
1034 buf += 1;
1035 }
1036 else
1037 {
1038 // d.igitsE+123
1039 // len <= max_digits10 + 1 + 5
1040
1041 std::memmove(buf + 2, buf + 1, static_cast<size_t>(k) - 1);
1042 buf[1] = '.';
1043 buf += 1 + static_cast<size_t>(k);
1044 }
1045
1046 *buf++ = 'e';
1047 return append_exponent(buf, n - 1);
1048 }
1049
1050 } // namespace dtoa_impl
1051
1052 /*!
1053 @brief generates a decimal representation of the floating-point number value in [first, last).
1054
1055 The format of the resulting decimal representation is similar to printf's %g
1056 format. Returns an iterator pointing past-the-end of the decimal representation.
1057
1058 @note The input number must be finite, i.e. NaN's and Inf's are not supported.
1059 @note The buffer must be large enough.
1060 @note The result is NOT null-terminated.
1061 */
1062 template<typename FloatType>
1063 JSON_HEDLEY_NON_NULL(1, 2)
1064 JSON_HEDLEY_RETURNS_NON_NULL
to_chars(char * first,const char * last,FloatType value)1065 char* to_chars(char* first, const char* last, FloatType value)
1066 {
1067 static_cast<void>(last); // maybe unused - fix warning
1068 JSON_ASSERT(std::isfinite(value));
1069
1070 // Use signbit(value) instead of (value < 0) since signbit works for -0.
1071 if (std::signbit(value))
1072 {
1073 value = -value;
1074 *first++ = '-';
1075 }
1076
1077 #ifdef __GNUC__
1078 #pragma GCC diagnostic push
1079 #pragma GCC diagnostic ignored "-Wfloat-equal"
1080 #endif
1081 if (value == 0) // +-0
1082 {
1083 *first++ = '0';
1084 // Make it look like a floating-point number (#362, #378)
1085 *first++ = '.';
1086 *first++ = '0';
1087 return first;
1088 }
1089 #ifdef __GNUC__
1090 #pragma GCC diagnostic pop
1091 #endif
1092
1093 JSON_ASSERT(last - first >= std::numeric_limits<FloatType>::max_digits10);
1094
1095 // Compute v = buffer * 10^decimal_exponent.
1096 // The decimal digits are stored in the buffer, which needs to be interpreted
1097 // as an unsigned decimal integer.
1098 // len is the length of the buffer, i.e. the number of decimal digits.
1099 int len = 0;
1100 int decimal_exponent = 0;
1101 dtoa_impl::grisu2(first, len, decimal_exponent, value);
1102
1103 JSON_ASSERT(len <= std::numeric_limits<FloatType>::max_digits10);
1104
1105 // Format the buffer like printf("%.*g", prec, value)
1106 constexpr int kMinExp = -4;
1107 // Use digits10 here to increase compatibility with version 2.
1108 constexpr int kMaxExp = std::numeric_limits<FloatType>::digits10;
1109
1110 JSON_ASSERT(last - first >= kMaxExp + 2);
1111 JSON_ASSERT(last - first >= 2 + (-kMinExp - 1) + std::numeric_limits<FloatType>::max_digits10);
1112 JSON_ASSERT(last - first >= std::numeric_limits<FloatType>::max_digits10 + 6);
1113
1114 return dtoa_impl::format_buffer(first, len, decimal_exponent, kMinExp, kMaxExp);
1115 }
1116
1117 } // namespace detail
1118 NLOHMANN_JSON_NAMESPACE_END
1119