1 | #include <cstring> |
2 | #include <cstdint> |
3 | #include <array> |
4 | #include <cmath> |
5 | |
6 | namespace simdjson { |
7 | namespace internal { |
8 | /*! |
9 | implements the Grisu2 algorithm for binary to decimal floating-point |
10 | conversion. |
11 | Adapted from JSON for Modern C++ |
12 | |
13 | This implementation is a slightly modified version of the reference |
14 | implementation which may be obtained from |
15 | http://florian.loitsch.com/publications (bench.tar.gz). |
16 | The code is distributed under the MIT license, Copyright (c) 2009 Florian |
17 | Loitsch. For a detailed description of the algorithm see: [1] Loitsch, "Printing |
18 | Floating-Point Numbers Quickly and Accurately with Integers", Proceedings of the |
19 | ACM SIGPLAN 2010 Conference on Programming Language Design and Implementation, |
20 | PLDI 2010 [2] Burger, Dybvig, "Printing Floating-Point Numbers Quickly and |
21 | Accurately", Proceedings of the ACM SIGPLAN 1996 Conference on Programming |
22 | Language Design and Implementation, PLDI 1996 |
23 | */ |
24 | namespace dtoa_impl { |
25 | |
26 | template <typename Target, typename Source> |
27 | Target reinterpret_bits(const Source source) { |
28 | static_assert(sizeof(Target) == sizeof(Source), "size mismatch" ); |
29 | |
30 | Target target; |
31 | std::memcpy(dest: &target, src: &source, n: sizeof(Source)); |
32 | return target; |
33 | } |
34 | |
35 | struct diyfp // f * 2^e |
36 | { |
37 | static constexpr int kPrecision = 64; // = q |
38 | |
39 | std::uint64_t f = 0; |
40 | int e = 0; |
41 | |
42 | constexpr diyfp(std::uint64_t f_, int e_) noexcept : f(f_), e(e_) {} |
43 | |
44 | /*! |
45 | @brief returns x - y |
46 | @pre x.e == y.e and x.f >= y.f |
47 | */ |
48 | static diyfp sub(const diyfp &x, const diyfp &y) noexcept { |
49 | |
50 | return {x.f - y.f, x.e}; |
51 | } |
52 | |
53 | /*! |
54 | @brief returns x * y |
55 | @note The result is rounded. (Only the upper q bits are returned.) |
56 | */ |
57 | static diyfp mul(const diyfp &x, const diyfp &y) noexcept { |
58 | static_assert(kPrecision == 64, "internal error" ); |
59 | |
60 | // Computes: |
61 | // f = round((x.f * y.f) / 2^q) |
62 | // e = x.e + y.e + q |
63 | |
64 | // Emulate the 64-bit * 64-bit multiplication: |
65 | // |
66 | // p = u * v |
67 | // = (u_lo + 2^32 u_hi) (v_lo + 2^32 v_hi) |
68 | // = (u_lo v_lo ) + 2^32 ((u_lo v_hi ) + (u_hi v_lo )) + |
69 | // 2^64 (u_hi v_hi ) = (p0 ) + 2^32 ((p1 ) + (p2 )) |
70 | // + 2^64 (p3 ) = (p0_lo + 2^32 p0_hi) + 2^32 ((p1_lo + |
71 | // 2^32 p1_hi) + (p2_lo + 2^32 p2_hi)) + 2^64 (p3 ) = |
72 | // (p0_lo ) + 2^32 (p0_hi + p1_lo + p2_lo ) + 2^64 (p1_hi + |
73 | // p2_hi + p3) = (p0_lo ) + 2^32 (Q ) + 2^64 (H ) = (p0_lo ) + |
74 | // 2^32 (Q_lo + 2^32 Q_hi ) + 2^64 (H ) |
75 | // |
76 | // (Since Q might be larger than 2^32 - 1) |
77 | // |
78 | // = (p0_lo + 2^32 Q_lo) + 2^64 (Q_hi + H) |
79 | // |
80 | // (Q_hi + H does not overflow a 64-bit int) |
81 | // |
82 | // = p_lo + 2^64 p_hi |
83 | |
84 | const std::uint64_t u_lo = x.f & 0xFFFFFFFFu; |
85 | const std::uint64_t u_hi = x.f >> 32u; |
86 | const std::uint64_t v_lo = y.f & 0xFFFFFFFFu; |
87 | const std::uint64_t v_hi = y.f >> 32u; |
88 | |
89 | const std::uint64_t p0 = u_lo * v_lo; |
90 | const std::uint64_t p1 = u_lo * v_hi; |
91 | const std::uint64_t p2 = u_hi * v_lo; |
92 | const std::uint64_t p3 = u_hi * v_hi; |
93 | |
94 | const std::uint64_t p0_hi = p0 >> 32u; |
95 | const std::uint64_t p1_lo = p1 & 0xFFFFFFFFu; |
96 | const std::uint64_t p1_hi = p1 >> 32u; |
97 | const std::uint64_t p2_lo = p2 & 0xFFFFFFFFu; |
98 | const std::uint64_t p2_hi = p2 >> 32u; |
99 | |
100 | std::uint64_t Q = p0_hi + p1_lo + p2_lo; |
101 | |
102 | // The full product might now be computed as |
103 | // |
104 | // p_hi = p3 + p2_hi + p1_hi + (Q >> 32) |
105 | // p_lo = p0_lo + (Q << 32) |
106 | // |
107 | // But in this particular case here, the full p_lo is not required. |
108 | // Effectively we only need to add the highest bit in p_lo to p_hi (and |
109 | // Q_hi + 1 does not overflow). |
110 | |
111 | Q += std::uint64_t{1} << (64u - 32u - 1u); // round, ties up |
112 | |
113 | const std::uint64_t h = p3 + p2_hi + p1_hi + (Q >> 32u); |
114 | |
115 | return {h, x.e + y.e + 64}; |
116 | } |
117 | |
118 | /*! |
119 | @brief normalize x such that the significand is >= 2^(q-1) |
120 | @pre x.f != 0 |
121 | */ |
122 | static diyfp normalize(diyfp x) noexcept { |
123 | |
124 | while ((x.f >> 63u) == 0) { |
125 | x.f <<= 1u; |
126 | x.e--; |
127 | } |
128 | |
129 | return x; |
130 | } |
131 | |
132 | /*! |
133 | @brief normalize x such that the result has the exponent E |
134 | @pre e >= x.e and the upper e - x.e bits of x.f must be zero. |
135 | */ |
136 | static diyfp normalize_to(const diyfp &x, |
137 | const int target_exponent) noexcept { |
138 | const int delta = x.e - target_exponent; |
139 | |
140 | return {x.f << delta, target_exponent}; |
141 | } |
142 | }; |
143 | |
144 | struct boundaries { |
145 | diyfp w; |
146 | diyfp minus; |
147 | diyfp plus; |
148 | }; |
149 | |
150 | /*! |
151 | Compute the (normalized) diyfp representing the input number 'value' and its |
152 | boundaries. |
153 | @pre value must be finite and positive |
154 | */ |
155 | template <typename FloatType> boundaries compute_boundaries(FloatType value) { |
156 | |
157 | // Convert the IEEE representation into a diyfp. |
158 | // |
159 | // If v is denormal: |
160 | // value = 0.F * 2^(1 - bias) = ( F) * 2^(1 - bias - (p-1)) |
161 | // If v is normalized: |
162 | // value = 1.F * 2^(E - bias) = (2^(p-1) + F) * 2^(E - bias - (p-1)) |
163 | |
164 | static_assert(std::numeric_limits<FloatType>::is_iec559, |
165 | "internal error: dtoa_short requires an IEEE-754 " |
166 | "floating-point implementation" ); |
167 | |
168 | constexpr int kPrecision = |
169 | std::numeric_limits<FloatType>::digits; // = p (includes the hidden bit) |
170 | constexpr int kBias = |
171 | std::numeric_limits<FloatType>::max_exponent - 1 + (kPrecision - 1); |
172 | constexpr int kMinExp = 1 - kBias; |
173 | constexpr std::uint64_t kHiddenBit = std::uint64_t{1} |
174 | << (kPrecision - 1); // = 2^(p-1) |
175 | |
176 | using bits_type = typename std::conditional<kPrecision == 24, std::uint32_t, |
177 | std::uint64_t>::type; |
178 | |
179 | const std::uint64_t bits = reinterpret_bits<bits_type>(value); |
180 | const std::uint64_t E = bits >> (kPrecision - 1); |
181 | const std::uint64_t F = bits & (kHiddenBit - 1); |
182 | |
183 | const bool is_denormal = E == 0; |
184 | const diyfp v = is_denormal |
185 | ? diyfp(F, kMinExp) |
186 | : diyfp(F + kHiddenBit, static_cast<int>(E) - kBias); |
187 | |
188 | // Compute the boundaries m- and m+ of the floating-point value |
189 | // v = f * 2^e. |
190 | // |
191 | // Determine v- and v+, the floating-point predecessor and successor if v, |
192 | // respectively. |
193 | // |
194 | // v- = v - 2^e if f != 2^(p-1) or e == e_min (A) |
195 | // = v - 2^(e-1) if f == 2^(p-1) and e > e_min (B) |
196 | // |
197 | // v+ = v + 2^e |
198 | // |
199 | // Let m- = (v- + v) / 2 and m+ = (v + v+) / 2. All real numbers _strictly_ |
200 | // between m- and m+ round to v, regardless of how the input rounding |
201 | // algorithm breaks ties. |
202 | // |
203 | // ---+-------------+-------------+-------------+-------------+--- (A) |
204 | // v- m- v m+ v+ |
205 | // |
206 | // -----------------+------+------+-------------+-------------+--- (B) |
207 | // v- m- v m+ v+ |
208 | |
209 | const bool lower_boundary_is_closer = F == 0 && E > 1; |
210 | const diyfp m_plus = diyfp(2 * v.f + 1, v.e - 1); |
211 | const diyfp m_minus = lower_boundary_is_closer |
212 | ? diyfp(4 * v.f - 1, v.e - 2) // (B) |
213 | : diyfp(2 * v.f - 1, v.e - 1); // (A) |
214 | |
215 | // Determine the normalized w+ = m+. |
216 | const diyfp w_plus = diyfp::normalize(x: m_plus); |
217 | |
218 | // Determine w- = m- such that e_(w-) = e_(w+). |
219 | const diyfp w_minus = diyfp::normalize_to(x: m_minus, target_exponent: w_plus.e); |
220 | |
221 | return {.w: diyfp::normalize(x: v), .minus: w_minus, .plus: w_plus}; |
222 | } |
223 | |
224 | // Given normalized diyfp w, Grisu needs to find a (normalized) cached |
225 | // power-of-ten c, such that the exponent of the product c * w = f * 2^e lies |
226 | // within a certain range [alpha, gamma] (Definition 3.2 from [1]) |
227 | // |
228 | // alpha <= e = e_c + e_w + q <= gamma |
229 | // |
230 | // or |
231 | // |
232 | // f_c * f_w * 2^alpha <= f_c 2^(e_c) * f_w 2^(e_w) * 2^q |
233 | // <= f_c * f_w * 2^gamma |
234 | // |
235 | // Since c and w are normalized, i.e. 2^(q-1) <= f < 2^q, this implies |
236 | // |
237 | // 2^(q-1) * 2^(q-1) * 2^alpha <= c * w * 2^q < 2^q * 2^q * 2^gamma |
238 | // |
239 | // or |
240 | // |
241 | // 2^(q - 2 + alpha) <= c * w < 2^(q + gamma) |
242 | // |
243 | // The choice of (alpha,gamma) determines the size of the table and the form of |
244 | // the digit generation procedure. Using (alpha,gamma)=(-60,-32) works out well |
245 | // in practice: |
246 | // |
247 | // The idea is to cut the number c * w = f * 2^e into two parts, which can be |
248 | // processed independently: An integral part p1, and a fractional part p2: |
249 | // |
250 | // f * 2^e = ( (f div 2^-e) * 2^-e + (f mod 2^-e) ) * 2^e |
251 | // = (f div 2^-e) + (f mod 2^-e) * 2^e |
252 | // = p1 + p2 * 2^e |
253 | // |
254 | // The conversion of p1 into decimal form requires a series of divisions and |
255 | // modulos by (a power of) 10. These operations are faster for 32-bit than for |
256 | // 64-bit integers, so p1 should ideally fit into a 32-bit integer. This can be |
257 | // achieved by choosing |
258 | // |
259 | // -e >= 32 or e <= -32 := gamma |
260 | // |
261 | // In order to convert the fractional part |
262 | // |
263 | // p2 * 2^e = p2 / 2^-e = d[-1] / 10^1 + d[-2] / 10^2 + ... |
264 | // |
265 | // into decimal form, the fraction is repeatedly multiplied by 10 and the digits |
266 | // d[-i] are extracted in order: |
267 | // |
268 | // (10 * p2) div 2^-e = d[-1] |
269 | // (10 * p2) mod 2^-e = d[-2] / 10^1 + ... |
270 | // |
271 | // The multiplication by 10 must not overflow. It is sufficient to choose |
272 | // |
273 | // 10 * p2 < 16 * p2 = 2^4 * p2 <= 2^64. |
274 | // |
275 | // Since p2 = f mod 2^-e < 2^-e, |
276 | // |
277 | // -e <= 60 or e >= -60 := alpha |
278 | |
279 | constexpr int kAlpha = -60; |
280 | constexpr int kGamma = -32; |
281 | |
282 | struct cached_power // c = f * 2^e ~= 10^k |
283 | { |
284 | std::uint64_t f; |
285 | int e; |
286 | int k; |
287 | }; |
288 | |
289 | /*! |
290 | For a normalized diyfp w = f * 2^e, this function returns a (normalized) cached |
291 | power-of-ten c = f_c * 2^e_c, such that the exponent of the product w * c |
292 | satisfies (Definition 3.2 from [1]) |
293 | alpha <= e_c + e + q <= gamma. |
294 | */ |
295 | inline cached_power get_cached_power_for_binary_exponent(int e) { |
296 | // Now |
297 | // |
298 | // alpha <= e_c + e + q <= gamma (1) |
299 | // ==> f_c * 2^alpha <= c * 2^e * 2^q |
300 | // |
301 | // and since the c's are normalized, 2^(q-1) <= f_c, |
302 | // |
303 | // ==> 2^(q - 1 + alpha) <= c * 2^(e + q) |
304 | // ==> 2^(alpha - e - 1) <= c |
305 | // |
306 | // If c were an exact power of ten, i.e. c = 10^k, one may determine k as |
307 | // |
308 | // k = ceil( log_10( 2^(alpha - e - 1) ) ) |
309 | // = ceil( (alpha - e - 1) * log_10(2) ) |
310 | // |
311 | // From the paper: |
312 | // "In theory the result of the procedure could be wrong since c is rounded, |
313 | // and the computation itself is approximated [...]. In practice, however, |
314 | // this simple function is sufficient." |
315 | // |
316 | // For IEEE double precision floating-point numbers converted into |
317 | // normalized diyfp's w = f * 2^e, with q = 64, |
318 | // |
319 | // e >= -1022 (min IEEE exponent) |
320 | // -52 (p - 1) |
321 | // -52 (p - 1, possibly normalize denormal IEEE numbers) |
322 | // -11 (normalize the diyfp) |
323 | // = -1137 |
324 | // |
325 | // and |
326 | // |
327 | // e <= +1023 (max IEEE exponent) |
328 | // -52 (p - 1) |
329 | // -11 (normalize the diyfp) |
330 | // = 960 |
331 | // |
332 | // This binary exponent range [-1137,960] results in a decimal exponent |
333 | // range [-307,324]. One does not need to store a cached power for each |
334 | // k in this range. For each such k it suffices to find a cached power |
335 | // such that the exponent of the product lies in [alpha,gamma]. |
336 | // This implies that the difference of the decimal exponents of adjacent |
337 | // table entries must be less than or equal to |
338 | // |
339 | // floor( (gamma - alpha) * log_10(2) ) = 8. |
340 | // |
341 | // (A smaller distance gamma-alpha would require a larger table.) |
342 | |
343 | // NB: |
344 | // Actually this function returns c, such that -60 <= e_c + e + 64 <= -34. |
345 | |
346 | constexpr int kCachedPowersMinDecExp = -300; |
347 | constexpr int kCachedPowersDecStep = 8; |
348 | |
349 | static constexpr std::array<cached_power, 79> kCachedPowers = {._M_elems: { |
350 | {.f: 0xAB70FE17C79AC6CA, .e: -1060, .k: -300}, {.f: 0xFF77B1FCBEBCDC4F, .e: -1034, .k: -292}, |
351 | {.f: 0xBE5691EF416BD60C, .e: -1007, .k: -284}, {.f: 0x8DD01FAD907FFC3C, .e: -980, .k: -276}, |
352 | {.f: 0xD3515C2831559A83, .e: -954, .k: -268}, {.f: 0x9D71AC8FADA6C9B5, .e: -927, .k: -260}, |
353 | {.f: 0xEA9C227723EE8BCB, .e: -901, .k: -252}, {.f: 0xAECC49914078536D, .e: -874, .k: -244}, |
354 | {.f: 0x823C12795DB6CE57, .e: -847, .k: -236}, {.f: 0xC21094364DFB5637, .e: -821, .k: -228}, |
355 | {.f: 0x9096EA6F3848984F, .e: -794, .k: -220}, {.f: 0xD77485CB25823AC7, .e: -768, .k: -212}, |
356 | {.f: 0xA086CFCD97BF97F4, .e: -741, .k: -204}, {.f: 0xEF340A98172AACE5, .e: -715, .k: -196}, |
357 | {.f: 0xB23867FB2A35B28E, .e: -688, .k: -188}, {.f: 0x84C8D4DFD2C63F3B, .e: -661, .k: -180}, |
358 | {.f: 0xC5DD44271AD3CDBA, .e: -635, .k: -172}, {.f: 0x936B9FCEBB25C996, .e: -608, .k: -164}, |
359 | {.f: 0xDBAC6C247D62A584, .e: -582, .k: -156}, {.f: 0xA3AB66580D5FDAF6, .e: -555, .k: -148}, |
360 | {.f: 0xF3E2F893DEC3F126, .e: -529, .k: -140}, {.f: 0xB5B5ADA8AAFF80B8, .e: -502, .k: -132}, |
361 | {.f: 0x87625F056C7C4A8B, .e: -475, .k: -124}, {.f: 0xC9BCFF6034C13053, .e: -449, .k: -116}, |
362 | {.f: 0x964E858C91BA2655, .e: -422, .k: -108}, {.f: 0xDFF9772470297EBD, .e: -396, .k: -100}, |
363 | {.f: 0xA6DFBD9FB8E5B88F, .e: -369, .k: -92}, {.f: 0xF8A95FCF88747D94, .e: -343, .k: -84}, |
364 | {.f: 0xB94470938FA89BCF, .e: -316, .k: -76}, {.f: 0x8A08F0F8BF0F156B, .e: -289, .k: -68}, |
365 | {.f: 0xCDB02555653131B6, .e: -263, .k: -60}, {.f: 0x993FE2C6D07B7FAC, .e: -236, .k: -52}, |
366 | {.f: 0xE45C10C42A2B3B06, .e: -210, .k: -44}, {.f: 0xAA242499697392D3, .e: -183, .k: -36}, |
367 | {.f: 0xFD87B5F28300CA0E, .e: -157, .k: -28}, {.f: 0xBCE5086492111AEB, .e: -130, .k: -20}, |
368 | {.f: 0x8CBCCC096F5088CC, .e: -103, .k: -12}, {.f: 0xD1B71758E219652C, .e: -77, .k: -4}, |
369 | {.f: 0x9C40000000000000, .e: -50, .k: 4}, {.f: 0xE8D4A51000000000, .e: -24, .k: 12}, |
370 | {.f: 0xAD78EBC5AC620000, .e: 3, .k: 20}, {.f: 0x813F3978F8940984, .e: 30, .k: 28}, |
371 | {.f: 0xC097CE7BC90715B3, .e: 56, .k: 36}, {.f: 0x8F7E32CE7BEA5C70, .e: 83, .k: 44}, |
372 | {.f: 0xD5D238A4ABE98068, .e: 109, .k: 52}, {.f: 0x9F4F2726179A2245, .e: 136, .k: 60}, |
373 | {.f: 0xED63A231D4C4FB27, .e: 162, .k: 68}, {.f: 0xB0DE65388CC8ADA8, .e: 189, .k: 76}, |
374 | {.f: 0x83C7088E1AAB65DB, .e: 216, .k: 84}, {.f: 0xC45D1DF942711D9A, .e: 242, .k: 92}, |
375 | {.f: 0x924D692CA61BE758, .e: 269, .k: 100}, {.f: 0xDA01EE641A708DEA, .e: 295, .k: 108}, |
376 | {.f: 0xA26DA3999AEF774A, .e: 322, .k: 116}, {.f: 0xF209787BB47D6B85, .e: 348, .k: 124}, |
377 | {.f: 0xB454E4A179DD1877, .e: 375, .k: 132}, {.f: 0x865B86925B9BC5C2, .e: 402, .k: 140}, |
378 | {.f: 0xC83553C5C8965D3D, .e: 428, .k: 148}, {.f: 0x952AB45CFA97A0B3, .e: 455, .k: 156}, |
379 | {.f: 0xDE469FBD99A05FE3, .e: 481, .k: 164}, {.f: 0xA59BC234DB398C25, .e: 508, .k: 172}, |
380 | {.f: 0xF6C69A72A3989F5C, .e: 534, .k: 180}, {.f: 0xB7DCBF5354E9BECE, .e: 561, .k: 188}, |
381 | {.f: 0x88FCF317F22241E2, .e: 588, .k: 196}, {.f: 0xCC20CE9BD35C78A5, .e: 614, .k: 204}, |
382 | {.f: 0x98165AF37B2153DF, .e: 641, .k: 212}, {.f: 0xE2A0B5DC971F303A, .e: 667, .k: 220}, |
383 | {.f: 0xA8D9D1535CE3B396, .e: 694, .k: 228}, {.f: 0xFB9B7CD9A4A7443C, .e: 720, .k: 236}, |
384 | {.f: 0xBB764C4CA7A44410, .e: 747, .k: 244}, {.f: 0x8BAB8EEFB6409C1A, .e: 774, .k: 252}, |
385 | {.f: 0xD01FEF10A657842C, .e: 800, .k: 260}, {.f: 0x9B10A4E5E9913129, .e: 827, .k: 268}, |
386 | {.f: 0xE7109BFBA19C0C9D, .e: 853, .k: 276}, {.f: 0xAC2820D9623BF429, .e: 880, .k: 284}, |
387 | {.f: 0x80444B5E7AA7CF85, .e: 907, .k: 292}, {.f: 0xBF21E44003ACDD2D, .e: 933, .k: 300}, |
388 | {.f: 0x8E679C2F5E44FF8F, .e: 960, .k: 308}, {.f: 0xD433179D9C8CB841, .e: 986, .k: 316}, |
389 | {.f: 0x9E19DB92B4E31BA9, .e: 1013, .k: 324}, |
390 | }}; |
391 | |
392 | // This computation gives exactly the same results for k as |
393 | // k = ceil((kAlpha - e - 1) * 0.30102999566398114) |
394 | // for |e| <= 1500, but doesn't require floating-point operations. |
395 | // NB: log_10(2) ~= 78913 / 2^18 |
396 | const int f = kAlpha - e - 1; |
397 | const int k = (f * 78913) / (1 << 18) + static_cast<int>(f > 0); |
398 | |
399 | const int index = (-kCachedPowersMinDecExp + k + (kCachedPowersDecStep - 1)) / |
400 | kCachedPowersDecStep; |
401 | |
402 | const cached_power cached = kCachedPowers[static_cast<std::size_t>(index)]; |
403 | |
404 | return cached; |
405 | } |
406 | |
407 | /*! |
408 | For n != 0, returns k, such that pow10 := 10^(k-1) <= n < 10^k. |
409 | For n == 0, returns 1 and sets pow10 := 1. |
410 | */ |
411 | inline int find_largest_pow10(const std::uint32_t n, std::uint32_t &pow10) { |
412 | // LCOV_EXCL_START |
413 | if (n >= 1000000000) { |
414 | pow10 = 1000000000; |
415 | return 10; |
416 | } |
417 | // LCOV_EXCL_STOP |
418 | else if (n >= 100000000) { |
419 | pow10 = 100000000; |
420 | return 9; |
421 | } else if (n >= 10000000) { |
422 | pow10 = 10000000; |
423 | return 8; |
424 | } else if (n >= 1000000) { |
425 | pow10 = 1000000; |
426 | return 7; |
427 | } else if (n >= 100000) { |
428 | pow10 = 100000; |
429 | return 6; |
430 | } else if (n >= 10000) { |
431 | pow10 = 10000; |
432 | return 5; |
433 | } else if (n >= 1000) { |
434 | pow10 = 1000; |
435 | return 4; |
436 | } else if (n >= 100) { |
437 | pow10 = 100; |
438 | return 3; |
439 | } else if (n >= 10) { |
440 | pow10 = 10; |
441 | return 2; |
442 | } else { |
443 | pow10 = 1; |
444 | return 1; |
445 | } |
446 | } |
447 | |
448 | inline void grisu2_round(char *buf, int len, std::uint64_t dist, |
449 | std::uint64_t delta, std::uint64_t rest, |
450 | std::uint64_t ten_k) { |
451 | |
452 | // <--------------------------- delta ----> |
453 | // <---- dist ---------> |
454 | // --------------[------------------+-------------------]-------------- |
455 | // M- w M+ |
456 | // |
457 | // ten_k |
458 | // <------> |
459 | // <---- rest ----> |
460 | // --------------[------------------+----+--------------]-------------- |
461 | // w V |
462 | // = buf * 10^k |
463 | // |
464 | // ten_k represents a unit-in-the-last-place in the decimal representation |
465 | // stored in buf. |
466 | // Decrement buf by ten_k while this takes buf closer to w. |
467 | |
468 | // The tests are written in this order to avoid overflow in unsigned |
469 | // integer arithmetic. |
470 | |
471 | while (rest < dist && delta - rest >= ten_k && |
472 | (rest + ten_k < dist || dist - rest > rest + ten_k - dist)) { |
473 | buf[len - 1]--; |
474 | rest += ten_k; |
475 | } |
476 | } |
477 | |
478 | /*! |
479 | Generates V = buffer * 10^decimal_exponent, such that M- <= V <= M+. |
480 | M- and M+ must be normalized and share the same exponent -60 <= e <= -32. |
481 | */ |
482 | inline void grisu2_digit_gen(char *buffer, int &length, int &decimal_exponent, |
483 | diyfp M_minus, diyfp w, diyfp M_plus) { |
484 | static_assert(kAlpha >= -60, "internal error" ); |
485 | static_assert(kGamma <= -32, "internal error" ); |
486 | |
487 | // Generates the digits (and the exponent) of a decimal floating-point |
488 | // number V = buffer * 10^decimal_exponent in the range [M-, M+]. The diyfp's |
489 | // w, M- and M+ share the same exponent e, which satisfies alpha <= e <= |
490 | // gamma. |
491 | // |
492 | // <--------------------------- delta ----> |
493 | // <---- dist ---------> |
494 | // --------------[------------------+-------------------]-------------- |
495 | // M- w M+ |
496 | // |
497 | // Grisu2 generates the digits of M+ from left to right and stops as soon as |
498 | // V is in [M-,M+]. |
499 | |
500 | std::uint64_t delta = |
501 | diyfp::sub(x: M_plus, y: M_minus) |
502 | .f; // (significand of (M+ - M-), implicit exponent is e) |
503 | std::uint64_t dist = |
504 | diyfp::sub(x: M_plus, y: w) |
505 | .f; // (significand of (M+ - w ), implicit exponent is e) |
506 | |
507 | // Split M+ = f * 2^e into two parts p1 and p2 (note: e < 0): |
508 | // |
509 | // M+ = f * 2^e |
510 | // = ((f div 2^-e) * 2^-e + (f mod 2^-e)) * 2^e |
511 | // = ((p1 ) * 2^-e + (p2 )) * 2^e |
512 | // = p1 + p2 * 2^e |
513 | |
514 | const diyfp one(std::uint64_t{1} << -M_plus.e, M_plus.e); |
515 | |
516 | auto p1 = static_cast<std::uint32_t>( |
517 | M_plus.f >> |
518 | -one.e); // p1 = f div 2^-e (Since -e >= 32, p1 fits into a 32-bit int.) |
519 | std::uint64_t p2 = M_plus.f & (one.f - 1); // p2 = f mod 2^-e |
520 | |
521 | // 1) |
522 | // |
523 | // Generate the digits of the integral part p1 = d[n-1]...d[1]d[0] |
524 | |
525 | std::uint32_t pow10; |
526 | const int k = find_largest_pow10(n: p1, pow10); |
527 | |
528 | // 10^(k-1) <= p1 < 10^k, pow10 = 10^(k-1) |
529 | // |
530 | // p1 = (p1 div 10^(k-1)) * 10^(k-1) + (p1 mod 10^(k-1)) |
531 | // = (d[k-1] ) * 10^(k-1) + (p1 mod 10^(k-1)) |
532 | // |
533 | // M+ = p1 + p2 * 2^e |
534 | // = d[k-1] * 10^(k-1) + (p1 mod 10^(k-1)) + p2 * 2^e |
535 | // = d[k-1] * 10^(k-1) + ((p1 mod 10^(k-1)) * 2^-e + p2) * 2^e |
536 | // = d[k-1] * 10^(k-1) + ( rest) * 2^e |
537 | // |
538 | // Now generate the digits d[n] of p1 from left to right (n = k-1,...,0) |
539 | // |
540 | // p1 = d[k-1]...d[n] * 10^n + d[n-1]...d[0] |
541 | // |
542 | // but stop as soon as |
543 | // |
544 | // rest * 2^e = (d[n-1]...d[0] * 2^-e + p2) * 2^e <= delta * 2^e |
545 | |
546 | int n = k; |
547 | while (n > 0) { |
548 | // Invariants: |
549 | // M+ = buffer * 10^n + (p1 + p2 * 2^e) (buffer = 0 for n = k) |
550 | // pow10 = 10^(n-1) <= p1 < 10^n |
551 | // |
552 | const std::uint32_t d = p1 / pow10; // d = p1 div 10^(n-1) |
553 | const std::uint32_t r = p1 % pow10; // r = p1 mod 10^(n-1) |
554 | // |
555 | // M+ = buffer * 10^n + (d * 10^(n-1) + r) + p2 * 2^e |
556 | // = (buffer * 10 + d) * 10^(n-1) + (r + p2 * 2^e) |
557 | // |
558 | buffer[length++] = static_cast<char>('0' + d); // buffer := buffer * 10 + d |
559 | // |
560 | // M+ = buffer * 10^(n-1) + (r + p2 * 2^e) |
561 | // |
562 | p1 = r; |
563 | n--; |
564 | // |
565 | // M+ = buffer * 10^n + (p1 + p2 * 2^e) |
566 | // pow10 = 10^n |
567 | // |
568 | |
569 | // Now check if enough digits have been generated. |
570 | // Compute |
571 | // |
572 | // p1 + p2 * 2^e = (p1 * 2^-e + p2) * 2^e = rest * 2^e |
573 | // |
574 | // Note: |
575 | // Since rest and delta share the same exponent e, it suffices to |
576 | // compare the significands. |
577 | const std::uint64_t rest = (std::uint64_t{p1} << -one.e) + p2; |
578 | if (rest <= delta) { |
579 | // V = buffer * 10^n, with M- <= V <= M+. |
580 | |
581 | decimal_exponent += n; |
582 | |
583 | // We may now just stop. But instead look if the buffer could be |
584 | // decremented to bring V closer to w. |
585 | // |
586 | // pow10 = 10^n is now 1 ulp in the decimal representation V. |
587 | // The rounding procedure works with diyfp's with an implicit |
588 | // exponent of e. |
589 | // |
590 | // 10^n = (10^n * 2^-e) * 2^e = ulp * 2^e |
591 | // |
592 | const std::uint64_t ten_n = std::uint64_t{pow10} << -one.e; |
593 | grisu2_round(buf: buffer, len: length, dist, delta, rest, ten_k: ten_n); |
594 | |
595 | return; |
596 | } |
597 | |
598 | pow10 /= 10; |
599 | // |
600 | // pow10 = 10^(n-1) <= p1 < 10^n |
601 | // Invariants restored. |
602 | } |
603 | |
604 | // 2) |
605 | // |
606 | // The digits of the integral part have been generated: |
607 | // |
608 | // M+ = d[k-1]...d[1]d[0] + p2 * 2^e |
609 | // = buffer + p2 * 2^e |
610 | // |
611 | // Now generate the digits of the fractional part p2 * 2^e. |
612 | // |
613 | // Note: |
614 | // No decimal point is generated: the exponent is adjusted instead. |
615 | // |
616 | // p2 actually represents the fraction |
617 | // |
618 | // p2 * 2^e |
619 | // = p2 / 2^-e |
620 | // = d[-1] / 10^1 + d[-2] / 10^2 + ... |
621 | // |
622 | // Now generate the digits d[-m] of p1 from left to right (m = 1,2,...) |
623 | // |
624 | // p2 * 2^e = d[-1]d[-2]...d[-m] * 10^-m |
625 | // + 10^-m * (d[-m-1] / 10^1 + d[-m-2] / 10^2 + ...) |
626 | // |
627 | // using |
628 | // |
629 | // 10^m * p2 = ((10^m * p2) div 2^-e) * 2^-e + ((10^m * p2) mod 2^-e) |
630 | // = ( d) * 2^-e + ( r) |
631 | // |
632 | // or |
633 | // 10^m * p2 * 2^e = d + r * 2^e |
634 | // |
635 | // i.e. |
636 | // |
637 | // M+ = buffer + p2 * 2^e |
638 | // = buffer + 10^-m * (d + r * 2^e) |
639 | // = (buffer * 10^m + d) * 10^-m + 10^-m * r * 2^e |
640 | // |
641 | // and stop as soon as 10^-m * r * 2^e <= delta * 2^e |
642 | |
643 | int m = 0; |
644 | for (;;) { |
645 | // Invariant: |
646 | // M+ = buffer * 10^-m + 10^-m * (d[-m-1] / 10 + d[-m-2] / 10^2 + ...) |
647 | // * 2^e |
648 | // = buffer * 10^-m + 10^-m * (p2 ) |
649 | // * 2^e = buffer * 10^-m + 10^-m * (1/10 * (10 * p2) ) * 2^e = |
650 | // buffer * 10^-m + 10^-m * (1/10 * ((10*p2 div 2^-e) * 2^-e + |
651 | // (10*p2 mod 2^-e)) * 2^e |
652 | // |
653 | p2 *= 10; |
654 | const std::uint64_t d = p2 >> -one.e; // d = (10 * p2) div 2^-e |
655 | const std::uint64_t r = p2 & (one.f - 1); // r = (10 * p2) mod 2^-e |
656 | // |
657 | // M+ = buffer * 10^-m + 10^-m * (1/10 * (d * 2^-e + r) * 2^e |
658 | // = buffer * 10^-m + 10^-m * (1/10 * (d + r * 2^e)) |
659 | // = (buffer * 10 + d) * 10^(-m-1) + 10^(-m-1) * r * 2^e |
660 | // |
661 | buffer[length++] = static_cast<char>('0' + d); // buffer := buffer * 10 + d |
662 | // |
663 | // M+ = buffer * 10^(-m-1) + 10^(-m-1) * r * 2^e |
664 | // |
665 | p2 = r; |
666 | m++; |
667 | // |
668 | // M+ = buffer * 10^-m + 10^-m * p2 * 2^e |
669 | // Invariant restored. |
670 | |
671 | // Check if enough digits have been generated. |
672 | // |
673 | // 10^-m * p2 * 2^e <= delta * 2^e |
674 | // p2 * 2^e <= 10^m * delta * 2^e |
675 | // p2 <= 10^m * delta |
676 | delta *= 10; |
677 | dist *= 10; |
678 | if (p2 <= delta) { |
679 | break; |
680 | } |
681 | } |
682 | |
683 | // V = buffer * 10^-m, with M- <= V <= M+. |
684 | |
685 | decimal_exponent -= m; |
686 | |
687 | // 1 ulp in the decimal representation is now 10^-m. |
688 | // Since delta and dist are now scaled by 10^m, we need to do the |
689 | // same with ulp in order to keep the units in sync. |
690 | // |
691 | // 10^m * 10^-m = 1 = 2^-e * 2^e = ten_m * 2^e |
692 | // |
693 | const std::uint64_t ten_m = one.f; |
694 | grisu2_round(buf: buffer, len: length, dist, delta, rest: p2, ten_k: ten_m); |
695 | |
696 | // By construction this algorithm generates the shortest possible decimal |
697 | // number (Loitsch, Theorem 6.2) which rounds back to w. |
698 | // For an input number of precision p, at least |
699 | // |
700 | // N = 1 + ceil(p * log_10(2)) |
701 | // |
702 | // decimal digits are sufficient to identify all binary floating-point |
703 | // numbers (Matula, "In-and-Out conversions"). |
704 | // This implies that the algorithm does not produce more than N decimal |
705 | // digits. |
706 | // |
707 | // N = 17 for p = 53 (IEEE double precision) |
708 | // N = 9 for p = 24 (IEEE single precision) |
709 | } |
710 | |
711 | /*! |
712 | v = buf * 10^decimal_exponent |
713 | len is the length of the buffer (number of decimal digits) |
714 | The buffer must be large enough, i.e. >= max_digits10. |
715 | */ |
716 | inline void grisu2(char *buf, int &len, int &decimal_exponent, diyfp m_minus, |
717 | diyfp v, diyfp m_plus) { |
718 | |
719 | // --------(-----------------------+-----------------------)-------- (A) |
720 | // m- v m+ |
721 | // |
722 | // --------------------(-----------+-----------------------)-------- (B) |
723 | // m- v m+ |
724 | // |
725 | // First scale v (and m- and m+) such that the exponent is in the range |
726 | // [alpha, gamma]. |
727 | |
728 | const cached_power cached = get_cached_power_for_binary_exponent(e: m_plus.e); |
729 | |
730 | const diyfp c_minus_k(cached.f, cached.e); // = c ~= 10^-k |
731 | |
732 | // The exponent of the products is = v.e + c_minus_k.e + q and is in the range |
733 | // [alpha,gamma] |
734 | const diyfp w = diyfp::mul(x: v, y: c_minus_k); |
735 | const diyfp w_minus = diyfp::mul(x: m_minus, y: c_minus_k); |
736 | const diyfp w_plus = diyfp::mul(x: m_plus, y: c_minus_k); |
737 | |
738 | // ----(---+---)---------------(---+---)---------------(---+---)---- |
739 | // w- w w+ |
740 | // = c*m- = c*v = c*m+ |
741 | // |
742 | // diyfp::mul rounds its result and c_minus_k is approximated too. w, w- and |
743 | // w+ are now off by a small amount. |
744 | // In fact: |
745 | // |
746 | // w - v * 10^k < 1 ulp |
747 | // |
748 | // To account for this inaccuracy, add resp. subtract 1 ulp. |
749 | // |
750 | // --------+---[---------------(---+---)---------------]---+-------- |
751 | // w- M- w M+ w+ |
752 | // |
753 | // Now any number in [M-, M+] (bounds included) will round to w when input, |
754 | // regardless of how the input rounding algorithm breaks ties. |
755 | // |
756 | // And digit_gen generates the shortest possible such number in [M-, M+]. |
757 | // Note that this does not mean that Grisu2 always generates the shortest |
758 | // possible number in the interval (m-, m+). |
759 | const diyfp M_minus(w_minus.f + 1, w_minus.e); |
760 | const diyfp M_plus(w_plus.f - 1, w_plus.e); |
761 | |
762 | decimal_exponent = -cached.k; // = -(-k) = k |
763 | |
764 | grisu2_digit_gen(buffer: buf, length&: len, decimal_exponent, M_minus, w, M_plus); |
765 | } |
766 | |
767 | /*! |
768 | v = buf * 10^decimal_exponent |
769 | len is the length of the buffer (number of decimal digits) |
770 | The buffer must be large enough, i.e. >= max_digits10. |
771 | */ |
772 | template <typename FloatType> |
773 | void grisu2(char *buf, int &len, int &decimal_exponent, FloatType value) { |
774 | static_assert(diyfp::kPrecision >= std::numeric_limits<FloatType>::digits + 3, |
775 | "internal error: not enough precision" ); |
776 | |
777 | // If the neighbors (and boundaries) of 'value' are always computed for |
778 | // double-precision numbers, all float's can be recovered using strtod (and |
779 | // strtof). However, the resulting decimal representations are not exactly |
780 | // "short". |
781 | // |
782 | // The documentation for 'std::to_chars' |
783 | // (https://en.cppreference.com/w/cpp/utility/to_chars) says "value is |
784 | // converted to a string as if by std::sprintf in the default ("C") locale" |
785 | // and since sprintf promotes float's to double's, I think this is exactly |
786 | // what 'std::to_chars' does. On the other hand, the documentation for |
787 | // 'std::to_chars' requires that "parsing the representation using the |
788 | // corresponding std::from_chars function recovers value exactly". That |
789 | // indicates that single precision floating-point numbers should be recovered |
790 | // using 'std::strtof'. |
791 | // |
792 | // NB: If the neighbors are computed for single-precision numbers, there is a |
793 | // single float |
794 | // (7.0385307e-26f) which can't be recovered using strtod. The resulting |
795 | // double precision value is off by 1 ulp. |
796 | #if 0 |
797 | const boundaries w = compute_boundaries(static_cast<double>(value)); |
798 | #else |
799 | const boundaries w = compute_boundaries(value); |
800 | #endif |
801 | |
802 | grisu2(buf, len, decimal_exponent, m_minus: w.minus, v: w.w, m_plus: w.plus); |
803 | } |
804 | |
805 | /*! |
806 | @brief appends a decimal representation of e to buf |
807 | @return a pointer to the element following the exponent. |
808 | @pre -1000 < e < 1000 |
809 | */ |
810 | inline char *append_exponent(char *buf, int e) { |
811 | |
812 | if (e < 0) { |
813 | e = -e; |
814 | *buf++ = '-'; |
815 | } else { |
816 | *buf++ = '+'; |
817 | } |
818 | |
819 | auto k = static_cast<std::uint32_t>(e); |
820 | if (k < 10) { |
821 | // Always print at least two digits in the exponent. |
822 | // This is for compatibility with printf("%g"). |
823 | *buf++ = '0'; |
824 | *buf++ = static_cast<char>('0' + k); |
825 | } else if (k < 100) { |
826 | *buf++ = static_cast<char>('0' + k / 10); |
827 | k %= 10; |
828 | *buf++ = static_cast<char>('0' + k); |
829 | } else { |
830 | *buf++ = static_cast<char>('0' + k / 100); |
831 | k %= 100; |
832 | *buf++ = static_cast<char>('0' + k / 10); |
833 | k %= 10; |
834 | *buf++ = static_cast<char>('0' + k); |
835 | } |
836 | |
837 | return buf; |
838 | } |
839 | |
840 | /*! |
841 | @brief prettify v = buf * 10^decimal_exponent |
842 | If v is in the range [10^min_exp, 10^max_exp) it will be printed in fixed-point |
843 | notation. Otherwise it will be printed in exponential notation. |
844 | @pre min_exp < 0 |
845 | @pre max_exp > 0 |
846 | */ |
847 | inline char *format_buffer(char *buf, int len, int decimal_exponent, |
848 | int min_exp, int max_exp) { |
849 | |
850 | const int k = len; |
851 | const int n = len + decimal_exponent; |
852 | |
853 | // v = buf * 10^(n-k) |
854 | // k is the length of the buffer (number of decimal digits) |
855 | // n is the position of the decimal point relative to the start of the buffer. |
856 | |
857 | if (k <= n && n <= max_exp) { |
858 | // digits[000] |
859 | // len <= max_exp + 2 |
860 | |
861 | std::memset(s: buf + k, c: '0', n: static_cast<size_t>(n) - static_cast<size_t>(k)); |
862 | // Make it look like a floating-point number (#362, #378) |
863 | buf[n + 0] = '.'; |
864 | buf[n + 1] = '0'; |
865 | return buf + (static_cast<size_t>(n)) + 2; |
866 | } |
867 | |
868 | if (0 < n && n <= max_exp) { |
869 | // dig.its |
870 | // len <= max_digits10 + 1 |
871 | std::memmove(dest: buf + (static_cast<size_t>(n) + 1), src: buf + n, |
872 | n: static_cast<size_t>(k) - static_cast<size_t>(n)); |
873 | buf[n] = '.'; |
874 | return buf + (static_cast<size_t>(k) + 1U); |
875 | } |
876 | |
877 | if (min_exp < n && n <= 0) { |
878 | // 0.[000]digits |
879 | // len <= 2 + (-min_exp - 1) + max_digits10 |
880 | |
881 | std::memmove(dest: buf + (2 + static_cast<size_t>(-n)), src: buf, |
882 | n: static_cast<size_t>(k)); |
883 | buf[0] = '0'; |
884 | buf[1] = '.'; |
885 | std::memset(s: buf + 2, c: '0', n: static_cast<size_t>(-n)); |
886 | return buf + (2U + static_cast<size_t>(-n) + static_cast<size_t>(k)); |
887 | } |
888 | |
889 | if (k == 1) { |
890 | // dE+123 |
891 | // len <= 1 + 5 |
892 | |
893 | buf += 1; |
894 | } else { |
895 | // d.igitsE+123 |
896 | // len <= max_digits10 + 1 + 5 |
897 | |
898 | std::memmove(dest: buf + 2, src: buf + 1, n: static_cast<size_t>(k) - 1); |
899 | buf[1] = '.'; |
900 | buf += 1 + static_cast<size_t>(k); |
901 | } |
902 | |
903 | *buf++ = 'e'; |
904 | return append_exponent(buf, e: n - 1); |
905 | } |
906 | |
907 | } // namespace dtoa_impl |
908 | |
909 | /*! |
910 | The format of the resulting decimal representation is similar to printf's %g |
911 | format. Returns an iterator pointing past-the-end of the decimal representation. |
912 | @note The input number must be finite, i.e. NaN's and Inf's are not supported. |
913 | @note The buffer must be large enough. |
914 | @note The result is NOT null-terminated. |
915 | */ |
916 | char *to_chars(char *first, const char *last, double value) { |
917 | static_cast<void>(last); // maybe unused - fix warning |
918 | bool negative = std::signbit(x: value); |
919 | if (negative) { |
920 | value = -value; |
921 | *first++ = '-'; |
922 | } |
923 | |
924 | if (value == 0) // +-0 |
925 | { |
926 | *first++ = '0'; |
927 | // Make it look like a floating-point number (#362, #378) |
928 | *first++ = '.'; |
929 | *first++ = '0'; |
930 | return first; |
931 | } |
932 | // Compute v = buffer * 10^decimal_exponent. |
933 | // The decimal digits are stored in the buffer, which needs to be interpreted |
934 | // as an unsigned decimal integer. |
935 | // len is the length of the buffer, i.e. the number of decimal digits. |
936 | int len = 0; |
937 | int decimal_exponent = 0; |
938 | dtoa_impl::grisu2(buf: first, len, decimal_exponent, value); |
939 | // Format the buffer like printf("%.*g", prec, value) |
940 | constexpr int kMinExp = -4; |
941 | constexpr int kMaxExp = std::numeric_limits<double>::digits10; |
942 | |
943 | return dtoa_impl::format_buffer(buf: first, len, decimal_exponent, min_exp: kMinExp, |
944 | max_exp: kMaxExp); |
945 | } |
946 | } // namespace internal |
947 | } // namespace simdjson |